From 7e5180ecffe21a2c18f54fa33f553c829f616e46 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 8 May 2026 21:15:09 +0900
Subject: [PATCH 01/48] =?UTF-8?q?=E2=9C=A8=20Create=20perturbation=5Fexpla?=
 =?UTF-8?q?iner.py?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 ..._explainers.py => eigenvalue_explainer.py} |  0
 explainability/perturbation_explainer.py      | 38 +++++++++++++++++++
 2 files changed, 38 insertions(+)
 rename explainability/{eigenvalue_explainers.py => eigenvalue_explainer.py} (100%)
 create mode 100644 explainability/perturbation_explainer.py

diff --git a/explainability/eigenvalue_explainers.py b/explainability/eigenvalue_explainer.py
similarity index 100%
rename from explainability/eigenvalue_explainers.py
rename to explainability/eigenvalue_explainer.py
diff --git a/explainability/perturbation_explainer.py b/explainability/perturbation_explainer.py
new file mode 100644
index 0000000..8a451e4
--- /dev/null
+++ b/explainability/perturbation_explainer.py
@@ -0,0 +1,38 @@
+from pytorch_grad_cam import AblationCAM, ScoreCAM, FEM
+from explainability.cam_explainer import CAMExplainer
+
+
+class AblationCAMExplainer(CAMExplainer):
+    """
+    Zero out activations and measure how the output drops
+    (this repository includes a fast batched implementation)
+    """
+    def __init__(self, batch_size: int = 32, ratio_channels_to_ablate: float = 1.0, **kwargs):
+        self.batch_size = batch_size
+        self.ratio_channels_to_ablate = ratio_channels_to_ablate
+        super().__init__(**kwargs)
+
+    def _build_cam(self):
+        return AblationCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn,
+                           batch_size=self.batch_size, ratio_channels_to_ablate=self.ratio_channels_to_ablate)
+
+
+class ScoreCAMExplainer(CAMExplainer):
+    """
+    Perbutate the image by the scaled activation
+    mesaure how the output drops 
+    """
+    def _build_cam(self):
+        return ScoreCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
+
+class FEMExplainer(CAMExplainer):
+    """
+    A gradient free method that binarizes activations
+    by an activation > mean + k * std rule
+    """
+    def __init__(self, k: int = 2, **kwargs):
+        self.k = k
+        super().__init__(**kwargs)
+
+    def _build_cam(self):
+        return FEM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn, k=self.k)
\ No newline at end of file

From 102ef5743c6a98d2063206725f7220dd88adf7c2 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 8 May 2026 21:18:27 +0900
Subject: [PATCH 02/48] =?UTF-8?q?=E2=9C=A8=20Create=20misc=5Fexplainer.py?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/misc_explainer.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)
 create mode 100644 explainability/misc_explainer.py

diff --git a/explainability/misc_explainer.py b/explainability/misc_explainer.py
new file mode 100644
index 0000000..db6afa0
--- /dev/null
+++ b/explainability/misc_explainer.py
@@ -0,0 +1,13 @@
+from pytorch_grad_cam import LayerCAM, RandomCAM
+from explainability.cam_explainer import CAMExplainer
+
+
+class LayerCAMExplainer(CAMExplainer):
+    """
+    Spatially weight the activations by positive gradients.
+    Works better especially in lower layers
+    """
+    def _build_cam(self):
+        return LayerCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
+
+

From 065a228b2947a6c282285894298e9155e44162b1 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 8 May 2026 21:25:39 +0900
Subject: [PATCH 03/48] =?UTF-8?q?=F0=9F=93=9D=20Add=20=5F=5Frepr=5F=5F=20f?=
 =?UTF-8?q?or=20all=20explainer=20classes?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/base_explainer.py         |  8 +++++--
 explainability/cam_explainer.py          | 12 ++++++++--
 explainability/eigenvalue_explainer.py   |  3 +++
 explainability/gradient_explainer.py     | 29 ++----------------------
 explainability/perturbation_explainer.py |  7 +++++-
 5 files changed, 27 insertions(+), 32 deletions(-)

diff --git a/explainability/base_explainer.py b/explainability/base_explainer.py
index 1a54186..951c6f1 100644
--- a/explainability/base_explainer.py
+++ b/explainability/base_explainer.py
@@ -26,8 +26,10 @@ def __init__(
         
         self.device = "cuda:0" if torch.cuda.is_available() else 'cpu'
         self.margin_ratio = margin_ratio
+        self.conf_thres = conf_thres
         self.category = category
         self.model_name = model_name
+        self.dataset = dataset
         self.branch_level = branch_level
         self.l_stage_idx = l_stage_idx
                 
@@ -109,5 +111,7 @@ def _get_transform(self, img_path: str):
             
     def __repr__(self):
         return (f"{self.__class__.__name__}("
-                f"model={self.model_name}, branch={self.branch}, "
-                f"device={self.device})")
\ No newline at end of file
+                f"model={self.model_name}, dataset={self.dataset}, "
+                f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
+                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, "
+                f"category={self.category},device={self.device})")
\ No newline at end of file
diff --git a/explainability/cam_explainer.py b/explainability/cam_explainer.py
index aa5fa96..6254366 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/cam_explainer.py
@@ -44,7 +44,7 @@ def display_cam_on_image(self, img_path):
             deprocess_image(tensor),  # (H,W,3), range(0~1), np.ndarray
             grayscale_cam, # (H,W), range(0~1), np.ndarray
             use_rgb=True, 
-            colormap = self.colormap
+            colormap = self.colormap,
             image_weight=self.image_weight)
         
         return heatmap # (H,W,3), range(0~255), np.uint8
@@ -53,5 +53,13 @@ def display_inverse_cam(self, img_path):
         grayscale_cam, _ = self.display_heatmap(img_path)
         
         return 1 - grayscale_cam # (H,W), range(0~1), np.float32
-        
+    
+    def __repr__(self):
+        return (f"{self.__class__.__name__}("
+                f"model={self.model_name}, dataset={self.dataset}, "
+                f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
+                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, "
+                f"aug_smooth={self.aug_smooth}, eigen_smooth={self.eigen_smooth}, "
+                f"colormap={self.colormap}, image_weight={self.image_weight}, "
+                f"category={self.category},device={self.device})")
         
\ No newline at end of file
diff --git a/explainability/eigenvalue_explainer.py b/explainability/eigenvalue_explainer.py
index 7f53174..5eb95c7 100644
--- a/explainability/eigenvalue_explainer.py
+++ b/explainability/eigenvalue_explainer.py
@@ -31,6 +31,9 @@ def _build_cam(self):
         return KPCA_CAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn,
                         kernel=self.kernel, gamma=self.gamma)
     
+    def __repr__(self):
+        return super().__repr__().rstrip(")") + f", kernel={self.kernel}, gamma={self.gamma})"
+
 class SegEigenCAMExplainer(CAMExplainer):
     """
     Like EigenCAM but with gradient weighting (absolute gradients ⊙ activations )
diff --git a/explainability/gradient_explainer.py b/explainability/gradient_explainer.py
index f50bff9..0282fd6 100644
--- a/explainability/gradient_explainer.py
+++ b/explainability/gradient_explainer.py
@@ -1,6 +1,6 @@
 from pytorch_grad_cam import (
     GradCAM, GradCAMPlusPlus, GradCAMElementWise,
-    XGradCAM, HiResCAM, FullGrad, ShapleyCAM, FinerCAM,
+    XGradCAM, HiResCAM
 )
 from explainability.cam_explainer import CAMExplainer
 
@@ -44,29 +44,4 @@ def _build_cam(self):
         return HiResCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
 
 
-class FullGradExplainer(CAMExplainer):
-    """
-    Computes the gradients of the biases from all over the network, and then sums them
-    """
-    def _build_cam(self):
-        return FullGrad(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
-
-
-class ShapleyCAMExplainer(CAMExplainer):
-    """
-    Weight the activations using the gradient and Hessian-vector product.
-    """
-    def _build_cam(self):
-        return ShapleyCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
-
-
-class FinerCAMExplainer(CAMExplainer):
-    """
-    Improves fine-grained classification by comparing similar classes, suppressing shared features and highlighting discriminative details.
-    """
-    def __init__(self, base_method=GradCAM, **kwargs):
-        self.base_method = base_method
-        super().__init__(**kwargs)
-
-    def _build_cam(self):
-        return FinerCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn, base_method=self.base_method)
\ No newline at end of file
+    
diff --git a/explainability/perturbation_explainer.py b/explainability/perturbation_explainer.py
index 8a451e4..28730ed 100644
--- a/explainability/perturbation_explainer.py
+++ b/explainability/perturbation_explainer.py
@@ -15,6 +15,8 @@ def __init__(self, batch_size: int = 32, ratio_channels_to_ablate: float = 1.0,
     def _build_cam(self):
         return AblationCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn,
                            batch_size=self.batch_size, ratio_channels_to_ablate=self.ratio_channels_to_ablate)
+    def __repr__(self):
+        return super().__repr__().rstrip(")") + f", ratio={self.ratio_channels_to_ablate})"
 
 
 class ScoreCAMExplainer(CAMExplainer):
@@ -35,4 +37,7 @@ def __init__(self, k: int = 2, **kwargs):
         super().__init__(**kwargs)
 
     def _build_cam(self):
-        return FEM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn, k=self.k)
\ No newline at end of file
+        return FEM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn, k=self.k)
+    
+    def __repr__(self):
+        return super().__repr__().rstrip(")") + f", k={self.k})"
\ No newline at end of file

From 857b308b1aa6ee5566492312177cff7496874502 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 16:03:55 +0900
Subject: [PATCH 04/48] =?UTF-8?q?=E2=9C=A8=20Add=20=5Fresolve=5Fblock=5Fid?=
 =?UTF-8?q?x?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/base_explainer.py | 42 +++++++++++++++++++++++---------
 explainability/cam_explainer.py  |  2 +-
 2 files changed, 32 insertions(+), 12 deletions(-)

diff --git a/explainability/base_explainer.py b/explainability/base_explainer.py
index 951c6f1..6054772 100644
--- a/explainability/base_explainer.py
+++ b/explainability/base_explainer.py
@@ -20,7 +20,8 @@ def __init__(
         conf_thres: float = 0.5,
         category: int = 1, # 0: real | 1: fake
         branch_level: str = "both", # low | high | both
-        l_stage_idx: int = -1 # gcvit low branch stage index (0~3)
+        l_stage_idx: int = -1, # gcvit low branch stage index (0~3)
+        block_idx: int = -1 # vit, gcvit block index 
     ):  
         self._validate_inputs(model_name, dataset, branch_level)
         
@@ -32,6 +33,7 @@ def __init__(
         self.dataset = dataset
         self.branch_level = branch_level
         self.l_stage_idx = l_stage_idx
+        self.block_idx = block_idx
                 
         # Parse Model Metadata
         self.model_variant = model_name.split("_")[-1] # b0, b5
@@ -60,27 +62,45 @@ def _validate_inputs(self, model_name: str, dataset: str, branch_level: str):
     def _get_reshape_fn(self):
         return reshape_transform_vit if self.transformer_type == "vit" else reshape_transform_gcvit    
     
-    def _resolve_target_layers(self) -> list[torch.nn.Module]:
+    def _resolve_target_layers(self) -> list:
         layers = []
 
         if self.transformer_type == "gcvit":
-            n_stages = len(self.model.l_gcvit.stages)
-            idx = self.l_stage_idx if self.l_stage_idx >= 0 else n_stages + self.l_stage_idx
+            n_stages  = len(self.model.l_gcvit.stages)
+            stage_idx = self.l_stage_idx if self.l_stage_idx >= 0 else n_stages + self.l_stage_idx
+
+            if not (0 <= stage_idx < n_stages):
+                raise ValueError(f"l_stage_idx={self.l_stage_idx} 범위 초과 (stage 수: {n_stages})")
 
-            if not (0 <= idx < n_stages):
-                raise IndexError(f"l_stage_idx {self.l_stage_idx} out of range (n={n_stages})")
             if self.branch_level in ("low", "both"):
-                layers.append(self.model.l_gcvit.stages[idx].blocks[-1].norm2)
+                n_blocks = len(self.model.l_gcvit.stages[stage_idx].blocks)
+                blk      = self._resolve_block_idx(self.block_idx, n_blocks, f"l_gcvit stage[{stage_idx}]")
+                layers.append(self.model.l_gcvit.stages[stage_idx].blocks[blk].norm2)
+
             if self.branch_level in ("high", "both"):
-                layers.append(self.model.h_gcvit.stages[0].blocks[-1].norm2)
+                n_blocks = len(self.model.h_gcvit.stages[0].blocks)
+                blk      = self._resolve_block_idx(self.block_idx, n_blocks, "h_gcvit stage[0]")
+                layers.append(self.model.h_gcvit.stages[0].blocks[blk].norm2)
 
         else:  # vit
             if self.branch_level in ("low", "both"):
-                layers.append(self.model.l_vit.blocks[-1].norm1)
+                n_blocks = len(self.model.l_vit.blocks)
+                blk      = self._resolve_block_idx(self.block_idx, n_blocks, "l_vit")
+                layers.append(self.model.l_vit.blocks[blk].norm1)
+
             if self.branch_level in ("high", "both"):
-                layers.append(self.model.h_vit.blocks[-1].norm1)
+                n_blocks = len(self.model.h_vit.blocks)
+                blk      = self._resolve_block_idx(self.block_idx, n_blocks, "h_vit")
+                layers.append(self.model.h_vit.blocks[blk].norm1)
 
         return layers
+
+    @staticmethod
+    def _resolve_block_idx(block_idx: int, n_blocks: int, name: str) -> int:
+        idx = block_idx if block_idx >= 0 else n_blocks + block_idx
+        if not (0 <= idx < n_blocks):
+            raise ValueError(f"block_idx={block_idx} 범위 초과 ({name} block 수: {n_blocks})")
+        return idx
     
     def _get_face_bbox(self, img: np.ndarray) -> List[float]:
         result = self.face_detector.detect_batch([img], [1.0])
@@ -113,5 +133,5 @@ def __repr__(self):
         return (f"{self.__class__.__name__}("
                 f"model={self.model_name}, dataset={self.dataset}, "
                 f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
-                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, "
+                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, block_idx={self.block_idx},"
                 f"category={self.category},device={self.device})")
\ No newline at end of file
diff --git a/explainability/cam_explainer.py b/explainability/cam_explainer.py
index 6254366..331ea74 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/cam_explainer.py
@@ -58,7 +58,7 @@ def __repr__(self):
         return (f"{self.__class__.__name__}("
                 f"model={self.model_name}, dataset={self.dataset}, "
                 f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
-                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, "
+                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, block_idx={self.block_idx},"
                 f"aug_smooth={self.aug_smooth}, eigen_smooth={self.eigen_smooth}, "
                 f"colormap={self.colormap}, image_weight={self.image_weight}, "
                 f"category={self.category},device={self.device})")

From c1dd9412ce20f0b37c9b29a9a03f0d384fa05196 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 20:04:18 +0900
Subject: [PATCH 05/48] =?UTF-8?q?=F0=9F=92=9A=20Add=20PR=20CI=20workflow?=
 =?UTF-8?q?=20with=20ruff=20lint=20and=20pytest?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .github/workflows/issue-auto-label.yml  | 14 ------------
 .github/workflows/issue-stale-clean.yml | 22 -------------------
 .github/workflows/pr-ci.yml             | 29 +++++++++++++++++++++++++
 requirements.txt                        |  5 ++++-
 4 files changed, 33 insertions(+), 37 deletions(-)
 delete mode 100644 .github/workflows/issue-auto-label.yml
 delete mode 100644 .github/workflows/issue-stale-clean.yml
 create mode 100644 .github/workflows/pr-ci.yml

diff --git a/.github/workflows/issue-auto-label.yml b/.github/workflows/issue-auto-label.yml
deleted file mode 100644
index ceec4a2..0000000
--- a/.github/workflows/issue-auto-label.yml
+++ /dev/null
@@ -1,14 +0,0 @@
-name: Auto Label Issues
-
-on:
-  issues:
-    types: [opened]
-
-jobs:
-  label:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Label Bug
-        uses: actions-ecosystem/action-add-labels@v1
-        with:
-          labels: bug
diff --git a/.github/workflows/issue-stale-clean.yml b/.github/workflows/issue-stale-clean.yml
deleted file mode 100644
index 7f90ca7..0000000
--- a/.github/workflows/issue-stale-clean.yml
+++ /dev/null
@@ -1,22 +0,0 @@
-name: Close stale issues
-
-on:
-  schedule:
-    - cron: "0 0 * * *"
-
-jobs:
-  stale:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/stale@v9
-        with:
-          days-before-stale: 14
-          days-before-close: 7
-          stale-issue-label: "stale"
-          stale-issue-message: >
-            ⏳ 14일 동안 활동이 없어 자동으로 stale 처리되었습니다.
-            추가 정보가 있다면 댓글로 다시 활성화해주세요!
-
-          close-issue-message: >
-            🔒 7일 동안 활동이 없어 Issue가 자동으로 닫혔습니다.
-            필요하면 언제든 재오픈해주세요 🙂
diff --git a/.github/workflows/pr-ci.yml b/.github/workflows/pr-ci.yml
new file mode 100644
index 0000000..4bbaf9f
--- /dev/null
+++ b/.github/workflows/pr-ci.yml
@@ -0,0 +1,29 @@
+name: PR CI
+
+on:
+  pull_request:
+    branches: [main, develop]
+
+jobs:
+  lint-and-test:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+          cache: "pip"
+
+      - name: Install dependencies
+        run: pip install -r requirements.txt
+
+      - name: Lint (ruff)
+        run: ruff check . --output-format=github
+
+      - name: Format check (ruff)
+        run: ruff format --check .
+
+      - name: Test (pytest)
+        run: pytest --tb=short -q
\ No newline at end of file
diff --git a/requirements.txt b/requirements.txt
index ce368d3..72e2059 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -40,4 +40,7 @@ seaborn>=0.13.2
 colorama>=0.4.6
 
 # --- [Distributed Task Queue] ---
-celery>=5.6.3
\ No newline at end of file
+celery>=5.6.3
+
+# --- [Linting & Formatting] ---
+ruff>=0.9.0
\ No newline at end of file

From c9f1041fc637f6d2423276be1c3b0e56058061dc Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 20:08:05 +0900
Subject: [PATCH 06/48] =?UTF-8?q?=F0=9F=92=9A=20Add=20PR=20auto-labeler=20?=
 =?UTF-8?q?with=20path-based=20label=20rules?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .github/labeler.yml            | 31 +++++++++++++++++++++++++++++++
 .github/workflows/pr-label.yml | 18 ++++++++++++++++++
 2 files changed, 49 insertions(+)
 create mode 100644 .github/labeler.yml
 create mode 100644 .github/workflows/pr-label.yml

diff --git a/.github/labeler.yml b/.github/labeler.yml
new file mode 100644
index 0000000..8f69fee
--- /dev/null
+++ b/.github/labeler.yml
@@ -0,0 +1,31 @@
+model:
+  - changed-files:
+    - any-glob-to-any-file: ["src/model/**", "models/**", "*.pt", "*.pth"]
+
+data:
+  - changed-files:
+    - any-glob-to-any-file: ["data/**", "datasets/**", "src/data/**"]
+
+training:
+  - changed-files:
+    - any-glob-to-any-file: ["train*.py", "src/train/**", "configs/**"]
+
+api:
+  - changed-files:
+    - any-glob-to-any-file: ["api/**", "routers/**", "app.py"]
+
+test:
+  - changed-files:
+    - any-glob-to-any-file: ["tests/**", "test_*.py"]
+
+docs:
+  - changed-files:
+    - any-glob-to-any-file: ["docs/**", "*.md", "*.rst"]
+
+ci:
+  - changed-files:
+    - any-glob-to-any-file: [".github/**"]
+
+dependencies:
+  - changed-files:
+    - any-glob-to-any-file: ["requirements*.txt", "pyproject.toml", "setup.py"]
\ No newline at end of file
diff --git a/.github/workflows/pr-label.yml b/.github/workflows/pr-label.yml
new file mode 100644
index 0000000..df3de40
--- /dev/null
+++ b/.github/workflows/pr-label.yml
@@ -0,0 +1,18 @@
+name: PR Auto Label
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+jobs:
+  label:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pull-requests: write
+
+    steps:
+      - uses: actions/labeler@v5
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
+          sync-labels: true  # PR 파일 변경 시 라벨 재동기화
\ No newline at end of file

From b84b267d078863269b70a4b77088513f56b9d722 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 20:26:44 +0900
Subject: [PATCH 07/48] =?UTF-8?q?=E2=9C=A8=20Add=20cam=5Fmult=5Fimage.py?=
 =?UTF-8?q?=20for=20testing?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/metrics/__init__.py       |  0
 explainability/metrics/cam_mult_image.py | 37 ++++++++++++++++++++++++
 2 files changed, 37 insertions(+)
 create mode 100644 explainability/metrics/__init__.py
 create mode 100644 explainability/metrics/cam_mult_image.py

diff --git a/explainability/metrics/__init__.py b/explainability/metrics/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/explainability/metrics/cam_mult_image.py b/explainability/metrics/cam_mult_image.py
new file mode 100644
index 0000000..c46751e
--- /dev/null
+++ b/explainability/metrics/cam_mult_image.py
@@ -0,0 +1,37 @@
+import torch
+import numpy as np
+from typing import List, Callable
+from explainability.metrics.perturbation_confidence import PerturbationConfidenceMetric
+
+
+def multiply_tensor_with_cam(input_tensor: torch.Tensor,
+                             cam: torch.Tensor):
+    """ Multiply an input tensor (after normalization)
+        with a pixel attribution map
+    """
+    return input_tensor * cam
+
+
+class CamMultImageConfidenceChange(PerturbationConfidenceMetric):
+    def __init__(self):
+        super(CamMultImageConfidenceChange,
+              self).__init__(multiply_tensor_with_cam)
+
+
+class DropInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(DropInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(DropInConfidence, self).__call__(*args, **kwargs)
+        scores = -scores
+        return np.maximum(scores, 0)
+
+
+class IncreaseInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(IncreaseInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(IncreaseInConfidence, self).__call__(*args, **kwargs)
+        return np.float32(scores > 0)
\ No newline at end of file

From cf97ec79ac1161ab28eee40d7887852237fdcea1 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 20:29:26 +0900
Subject: [PATCH 08/48] =?UTF-8?q?=E2=9C=A8=20Add=20perturbation=5Fconfiden?=
 =?UTF-8?q?ce.py=20for=20testing?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../metrics/perturbation_confidence.py        | 109 ++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 explainability/metrics/perturbation_confidence.py

diff --git a/explainability/metrics/perturbation_confidence.py b/explainability/metrics/perturbation_confidence.py
new file mode 100644
index 0000000..211ab8a
--- /dev/null
+++ b/explainability/metrics/perturbation_confidence.py
@@ -0,0 +1,109 @@
+import torch
+import numpy as np
+from typing import List, Callable
+
+import numpy as np
+import cv2
+
+
+class PerturbationConfidenceMetric:
+    def __init__(self, perturbation):
+        self.perturbation = perturbation
+
+    def __call__(self, input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module,
+                 return_visualization=False,
+                 return_diff=True):
+
+        if return_diff:
+            with torch.no_grad():
+                outputs = model(input_tensor)
+                scores = [target(output).cpu().numpy()
+                          for target, output in zip(targets, outputs)]
+                scores = np.float32(scores)
+
+        batch_size = input_tensor.size(0)
+        perturbated_tensors = []
+        for i in range(batch_size):
+            cam = cams[i]
+            tensor = self.perturbation(input_tensor[i, ...].cpu(),
+                                       torch.from_numpy(cam))
+            tensor = tensor.to(input_tensor.device)
+            perturbated_tensors.append(tensor.unsqueeze(0))
+        perturbated_tensors = torch.cat(perturbated_tensors)
+
+        with torch.no_grad():
+            outputs_after_imputation = model(perturbated_tensors)
+        scores_after_imputation = [
+            target(output).cpu().numpy() for target, output in zip(
+                targets, outputs_after_imputation)]
+        scores_after_imputation = np.float32(scores_after_imputation)
+
+        if return_diff:
+            result = scores_after_imputation - scores
+        else:
+            result = scores_after_imputation
+
+        if return_visualization:
+            return result, perturbated_tensors
+        else:
+            return result
+
+
+class RemoveMostRelevantFirst:
+    def __init__(self, percentile, imputer):
+        self.percentile = percentile
+        self.imputer = imputer
+
+    def __call__(self, input_tensor, mask):
+        imputer = self.imputer
+        if self.percentile != 'auto':
+            threshold = np.percentile(mask.cpu().numpy(), self.percentile)
+            binary_mask = np.float32(mask < threshold)
+        else:
+            _, binary_mask = cv2.threshold(
+                np.uint8(mask * 255), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
+
+        binary_mask = torch.from_numpy(binary_mask)
+        binary_mask = binary_mask.to(mask.device)
+        return imputer(input_tensor, binary_mask)
+
+
+class RemoveLeastRelevantFirst(RemoveMostRelevantFirst):
+    def __init__(self, percentile, imputer):
+        super(RemoveLeastRelevantFirst, self).__init__(percentile, imputer)
+
+    def __call__(self, input_tensor, mask):
+        return super(RemoveLeastRelevantFirst, self).__call__(
+            input_tensor, 1 - mask)
+
+
+class AveragerAcrossThresholds:
+    def __init__(
+        self,
+        imputer,
+        percentiles=[
+            10,
+            20,
+            30,
+            40,
+            50,
+            60,
+            70,
+            80,
+            90]):
+        self.imputer = imputer
+        self.percentiles = percentiles
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module):
+        scores = []
+        for percentile in self.percentiles:
+            imputer = self.imputer(percentile)
+            scores.append(imputer(input_tensor, cams, targets, model))
+        return np.mean(np.float32(scores), axis=0)
\ No newline at end of file

From 92c30077936c7c2f60091574d95e065fa752dbeb Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 9 May 2026 20:31:34 +0900
Subject: [PATCH 09/48] =?UTF-8?q?=E2=9C=A8=20Add=20road.py=20for=20testing?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/metrics/road.py | 154 +++++++++++++++++++++++++++++++++
 1 file changed, 154 insertions(+)
 create mode 100644 explainability/metrics/road.py

diff --git a/explainability/metrics/road.py b/explainability/metrics/road.py
new file mode 100644
index 0000000..905f84e
--- /dev/null
+++ b/explainability/metrics/road.py
@@ -0,0 +1,154 @@
+import torch
+import numpy as np
+from scipy.sparse import lil_matrix, csc_matrix
+from scipy.sparse.linalg import spsolve
+from typing import List, Callable
+from explainability.metrics.perturbation_confidence import (
+    PerturbationConfidenceMetric, AveragerAcrossThresholds,
+    RemoveMostRelevantFirst, RemoveLeastRelevantFirst
+)
+
+# The weights of the surrounding pixels
+neighbors_weights = [((1, 1), 1 / 12),
+                     ((0, 1), 1 / 6),
+                     ((-1, 1), 1 / 12),
+                     ((1, -1), 1 / 12),
+                     ((0, -1), 1 / 6),
+                     ((-1, -1), 1 / 12),
+                     ((1, 0), 1 / 6),
+                     ((-1, 0), 1 / 6)]
+
+
+class NoisyLinearImputer:
+    def __init__(self,
+                 noise: float = 0.01,
+                 weighting: List[float] = neighbors_weights):
+        """
+                Noisy linear imputation.
+                noise: magnitude of noise to add (absolute, set to 0 for no noise)
+                weighting: Weights of the neighboring pixels in the computation.
+                List of tuples of (offset, weight)
+        """
+        self.noise = noise
+        self.weighting = neighbors_weights
+
+    @staticmethod
+    def add_offset_to_indices(indices, offset, mask_shape):
+        """ Add the corresponding offset to the indices.
+    Return new indices plus a valid bit-vector. """
+        cord1 = indices % mask_shape[1]
+        cord0 = indices // mask_shape[1]
+        cord0 += offset[0]
+        cord1 += offset[1]
+        valid = ((cord0 < 0) | (cord1 < 0) |
+                 (cord0 >= mask_shape[0]) |
+                 (cord1 >= mask_shape[1]))
+        return ~valid, indices + offset[0] * mask_shape[1] + offset[1]
+
+    @staticmethod
+    def setup_sparse_system(mask, img, neighbors_weights):
+        """ Vectorized version to set up the equation system.
+                mask: (H, W)-tensor of missing pixels.
+                Image: (H, W, C)-tensor of all values.
+                Return (N,N)-System matrix, (N,C)-Right hand side for each of the C channels.
+        """
+        maskflt = mask.flatten()
+        imgflat = img.reshape((img.shape[0], -1))
+    # Indices that are imputed in the flattened mask:
+        indices = np.argwhere(maskflt == 0).flatten()
+        coords_to_vidx = np.zeros(len(maskflt), dtype=int)
+        coords_to_vidx[indices] = np.arange(len(indices))
+        numEquations = len(indices)
+    # System matrix:
+        A = lil_matrix((numEquations, numEquations))
+        b = np.zeros((numEquations, img.shape[0]))
+    # Sum of weights assigned:
+        sum_neighbors = np.ones(numEquations)
+        for n in neighbors_weights:
+            offset, weight = n[0], n[1]
+            # Take out outliers
+            valid, new_coords = NoisyLinearImputer.add_offset_to_indices(
+                indices, offset, mask.shape)
+            valid_coords = new_coords[valid]
+            valid_ids = np.argwhere(valid == 1).flatten()
+            # Add values to the right hand-side
+            has_values_coords = valid_coords[maskflt[valid_coords] > 0.5]
+            has_values_ids = valid_ids[maskflt[valid_coords] > 0.5]
+            b[has_values_ids, :] -= weight * imgflat[:, has_values_coords].T
+            # Add weights to the system (left hand side)
+# Find coordinates in the system.
+            has_no_values = valid_coords[maskflt[valid_coords] < 0.5]
+            variable_ids = coords_to_vidx[has_no_values]
+            has_no_values_ids = valid_ids[maskflt[valid_coords] < 0.5]
+            A[has_no_values_ids, variable_ids] = weight
+            # Reduce weight for invalid
+            sum_neighbors[np.argwhere(valid == 0).flatten()] = \
+                sum_neighbors[np.argwhere(valid == 0).flatten()] - weight
+
+        A[np.arange(numEquations), np.arange(numEquations)] = -sum_neighbors
+        return A, b
+
+    def __call__(self, img: torch.Tensor, mask: torch.Tensor):
+        """ Our linear inputation scheme. """
+        """
+		This is the function to do the linear infilling
+		img: original image (C,H,W)-tensor;
+		mask: mask; (H,W)-tensor
+
+		"""
+        imgflt = img.reshape(img.shape[0], -1)
+        maskflt = mask.reshape(-1)
+    # Indices that need to be imputed.
+        indices_linear = np.argwhere(maskflt == 0).flatten()
+        # Set up sparse equation system, solve system.
+        A, b = NoisyLinearImputer.setup_sparse_system(
+            mask.numpy(), img.numpy(), neighbors_weights)
+        res = torch.tensor(spsolve(csc_matrix(A), b), dtype=torch.float)
+
+        # Fill the values with the solution of the system.
+        img_infill = imgflt.clone()
+        img_infill[:, indices_linear] = res.t() + self.noise * \
+            torch.randn_like(res.t())
+
+        return img_infill.reshape_as(img)
+
+
+class ROADMostRelevantFirst(PerturbationConfidenceMetric):
+    def __init__(self, percentile=80):
+        super(ROADMostRelevantFirst, self).__init__(
+            RemoveMostRelevantFirst(percentile, NoisyLinearImputer()))
+
+
+class ROADLeastRelevantFirst(PerturbationConfidenceMetric):
+    def __init__(self, percentile=20):
+        super(ROADLeastRelevantFirst, self).__init__(
+            RemoveLeastRelevantFirst(percentile, NoisyLinearImputer()))
+
+
+class ROADMostRelevantFirstAverage(AveragerAcrossThresholds):
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        super(ROADMostRelevantFirstAverage, self).__init__(
+            ROADMostRelevantFirst, percentiles)
+
+
+class ROADLeastRelevantFirstAverage(AveragerAcrossThresholds):
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        super(ROADLeastRelevantFirstAverage, self).__init__(
+            ROADLeastRelevantFirst, percentiles)
+
+
+class ROADCombined:
+    def __init__(self, percentiles=[10, 20, 30, 40, 50, 60, 70, 80, 90]):
+        self.percentiles = percentiles
+        self.morf_averager = ROADMostRelevantFirstAverage(percentiles)
+        self.lerf_averager = ROADLeastRelevantFirstAverage(percentiles)
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 cams: np.ndarray,
+                 targets: List[Callable],
+                 model: torch.nn.Module):
+
+        scores_lerf = self.lerf_averager(input_tensor, cams, targets, model)
+        scores_morf = self.morf_averager(input_tensor, cams, targets, model)
+        return (scores_lerf - scores_morf) / 2
\ No newline at end of file

From f060223d0e33f9d95302c03d3381721950fcb3e9 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 09:34:58 +0900
Subject: [PATCH 10/48] =?UTF-8?q?=E2=9A=B0=EF=B8=8F=20Remove=20SegEigenCAM?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/eigenvalue_explainer.py | 14 +-------------
 1 file changed, 1 insertion(+), 13 deletions(-)

diff --git a/explainability/eigenvalue_explainer.py b/explainability/eigenvalue_explainer.py
index 5eb95c7..79f0927 100644
--- a/explainability/eigenvalue_explainer.py
+++ b/explainability/eigenvalue_explainer.py
@@ -1,4 +1,4 @@
-from pytorch_grad_cam import EigenCAM, EigenGradCAM, KPCA_CAM, SegEigenCAM
+from pytorch_grad_cam import EigenCAM, EigenGradCAM, KPCA_CAM
 from explainability.cam_explainer import CAMExplainer
 
 class EigenCAMExplainer(CAMExplainer):
@@ -33,15 +33,3 @@ def _build_cam(self):
     
     def __repr__(self):
         return super().__repr__().rstrip(")") + f", kernel={self.kernel}, gamma={self.gamma})"
-
-class SegEigenCAMExplainer(CAMExplainer):
-    """
-    Like EigenCAM but with gradient weighting (absolute gradients ⊙ activations )
-    before SVD and sign correction to fix SVD sign ambiguity
-    designed for semantic segmentation
-    """
-    def _build_cam(self):
-        return SegEigenCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn)
-    
-    
-

From 36b504964b1498b1aee4d7d50dc2f089ec971800 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 10:07:54 +0900
Subject: [PATCH 11/48] =?UTF-8?q?=E2=9A=B0=EF=B8=8F=20Remove=20AblationCAM?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/perturbation_explainer.py | 21 +--------------------
 1 file changed, 1 insertion(+), 20 deletions(-)

diff --git a/explainability/perturbation_explainer.py b/explainability/perturbation_explainer.py
index 28730ed..1753954 100644
--- a/explainability/perturbation_explainer.py
+++ b/explainability/perturbation_explainer.py
@@ -1,24 +1,5 @@
-from pytorch_grad_cam import AblationCAM, ScoreCAM, FEM
+from pytorch_grad_cam import ScoreCAM, FEM
 from explainability.cam_explainer import CAMExplainer
-
-
-class AblationCAMExplainer(CAMExplainer):
-    """
-    Zero out activations and measure how the output drops
-    (this repository includes a fast batched implementation)
-    """
-    def __init__(self, batch_size: int = 32, ratio_channels_to_ablate: float = 1.0, **kwargs):
-        self.batch_size = batch_size
-        self.ratio_channels_to_ablate = ratio_channels_to_ablate
-        super().__init__(**kwargs)
-
-    def _build_cam(self):
-        return AblationCAM(model=self.model, target_layers=self.target_layers, reshape_transform=self.reshape_fn,
-                           batch_size=self.batch_size, ratio_channels_to_ablate=self.ratio_channels_to_ablate)
-    def __repr__(self):
-        return super().__repr__().rstrip(")") + f", ratio={self.ratio_channels_to_ablate})"
-
-
 class ScoreCAMExplainer(CAMExplainer):
     """
     Perbutate the image by the scaled activation

From d24335674b700854cfbe850c87b77578615f7162 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 11:47:37 +0900
Subject: [PATCH 12/48] =?UTF-8?q?=E2=9C=A8=20Add=20remove=5Fpadding=5Fand?=
 =?UTF-8?q?=5Fresize?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/utils/image.py | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/explainability/utils/image.py b/explainability/utils/image.py
index 83b90fa..92a5ba4 100644
--- a/explainability/utils/image.py
+++ b/explainability/utils/image.py
@@ -44,4 +44,18 @@ def deprocess_image(tensor):
     img = (img * std) + mean
     img = np.clip(img, 0, 1) # range(0~1)
 
-    return np.float32(img)
\ No newline at end of file
+    return np.float32(img)
+
+def remove_padding_and_resize(cam, orig_shape, target_shape):
+    orig_h, orig_w = orig_shape[:2] # (H,W,3)
+    target_h, target_w = target_shape[2:] # (1,3,H,W)
+    
+    scale = max(target_h, target_w) / max(orig_h, orig_w)
+    new_h, new_w = int(orig_h * scale), int(orig_w * scale)
+    
+    pad_h = (target_h - new_h) // 2
+    pad_w = (target_w - new_w) // 2
+    
+    # Remove Padding
+    unpadded_cam = cam[pad_h:pad_h+new_h, pad_w:pad_w+new_w]
+    return cv2.resize(unpadded_cam, (orig_w, orig_h))
\ No newline at end of file

From 294c7153f2e87d5a08a7c2b2aaafb7e07e297ec1 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 11:53:36 +0900
Subject: [PATCH 13/48] =?UTF-8?q?=E2=99=BB=EF=B8=8F=20Refactor=20display?=
 =?UTF-8?q?=5Fheatmap=20with=20new=20arguements?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_explainer.py | 40 +++++++++++++++++----------------
 1 file changed, 21 insertions(+), 19 deletions(-)

diff --git a/explainability/cam_explainer.py b/explainability/cam_explainer.py
index 331ea74..837a7db 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/cam_explainer.py
@@ -3,27 +3,32 @@
 import numpy as np
 from abc import ABC, abstractmethod
 from explainability.utils.model_targets import BinaryClassifierOutputTarget
-from explainability.utils.image import show_cam_on_image, deprocess_image
+from explainability.utils.image import show_cam_on_image, deprocess_image, remove_padding_and_resize
 from .base_explainer import BaseExplainer
 class CAMExplainer(BaseExplainer, ABC):
     def __init__(self, 
                  aug_smooth: bool = False,
                  eigen_smooth: bool = False,
-                 colormap: int = cv2.COLORMAP_JET,
-                 image_weight: float = 0.5,
                  **kwargs):
         super().__init__(**kwargs)
         self.aug_smooth = aug_smooth # aug_smooth has the effect of beter centering the CAM around the objects
         self.eigen_smooth = eigen_smooth # # eigen_smooth has the effect of removing a lot of noise
-        self.colormap = colormap
-        self.image_weight = image_weight
         self.cam = self._build_cam()
     
     @abstractmethod
     def _build_cam(self):
         ...
 
-    def display_heatmap(self, img_path):
+    def _build_grayscale_cam(self, img_path: str):
+        """
+        Generate a grayscale CAM and its associated image tensors.
+
+        Returns:
+            tuple: (grayscale_cam, tensor, face)
+                - grayscale_cam (np.ndarray): CAM result in (H, W), range [0, 1].
+                - tensor (torch.Tensor): Preprocessed input in (1, 3, H, W).
+                - face (np.ndarray): Original image in (H, W, 3), uint8.
+        """
         face, tensor = self._get_transform(img_path)
         tensor = tensor.to(self.device) # (1,3,H,W)
         targets = [BinaryClassifierOutputTarget(self.category)]
@@ -35,25 +40,22 @@ def display_heatmap(self, img_path):
             eigen_smooth = self.eigen_smooth
         ) # (1,H,W)
         
-        return grayscale_cam[0], tensor 
+        return grayscale_cam[0], tensor, face 
+        
+    def display_heatmap(self, img_path: str, colormap: int = cv2.COLORMAP_JET, image_weight: float = 0.5,):
+        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)     
         
-    def display_cam_on_image(self, img_path):
-        grayscale_cam, tensor = self.display_heatmap(img_path)        
+        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)   
         
         heatmap = show_cam_on_image(
-            deprocess_image(tensor),  # (H,W,3), range(0~1), np.ndarray
-            grayscale_cam, # (H,W), range(0~1), np.ndarray
-            use_rgb=True, 
-            colormap = self.colormap,
-            image_weight=self.image_weight)
+            np.float32(face) / 255.0,  # (H,W,3), range(0~1), np.ndarray
+            cam_recovered, # (H,W), range(0~1), np.ndarray
+            use_rgb =True, 
+            colormap = colormap,
+            image_weight= image_weight)
         
         return heatmap # (H,W,3), range(0~255), np.uint8
     
-    def display_inverse_cam(self, img_path):
-        grayscale_cam, _ = self.display_heatmap(img_path)
-        
-        return 1 - grayscale_cam # (H,W), range(0~1), np.float32
-    
     def __repr__(self):
         return (f"{self.__class__.__name__}("
                 f"model={self.model_name}, dataset={self.dataset}, "

From fc436b8cb5a2a346510aa9c14558ef13b5d4a283 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 12:43:17 +0900
Subject: [PATCH 14/48] =?UTF-8?q?=E2=9C=A8=20Add=20display=5Fcontour=5Fon?=
 =?UTF-8?q?=5Fimage?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_explainer.py | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/explainability/cam_explainer.py b/explainability/cam_explainer.py
index 837a7db..4b597ed 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/cam_explainer.py
@@ -42,9 +42,10 @@ def _build_grayscale_cam(self, img_path: str):
         
         return grayscale_cam[0], tensor, face 
         
-    def display_heatmap(self, img_path: str, colormap: int = cv2.COLORMAP_JET, image_weight: float = 0.5,):
+    def display_heatmap_on_image(self, img_path: str, 
+                                 colormap: int = cv2.COLORMAP_JET, 
+                                 image_weight: float = 0.5,):
         grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)     
-        
         cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)   
         
         heatmap = show_cam_on_image(
@@ -56,6 +57,20 @@ def display_heatmap(self, img_path: str, colormap: int = cv2.COLORMAP_JET, image
         
         return heatmap # (H,W,3), range(0~255), np.uint8
     
+    def display_contour_on_image(self, img_path: str,
+                                 threshold: float = 0.5,
+                                 thickness: int = 2,
+                                 ):
+        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)
+        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
+        
+        mask = (cam_recovered > threshold).astype(np.uint8) * 255 
+        contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+        annotated_img = cv2.drawContours(face.copy(), contours, -1, (255, 0, 0), thickness)
+        
+        return annotated_img
+        
+    
     def __repr__(self):
         return (f"{self.__class__.__name__}("
                 f"model={self.model_name}, dataset={self.dataset}, "

From 3a6b9bbef0bdbe6af67d0be3736a6ad08651deba Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 15:00:14 +0900
Subject: [PATCH 15/48] =?UTF-8?q?=E2=9C=A8=20Add=20display=5Fbbox=5Fon=5Fi?=
 =?UTF-8?q?mage?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_explainer.py | 147 ++++++++++++++++++++++----------
 1 file changed, 103 insertions(+), 44 deletions(-)

diff --git a/explainability/cam_explainer.py b/explainability/cam_explainer.py
index 4b597ed..bec4ffd 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/cam_explainer.py
@@ -5,14 +5,41 @@
 from explainability.utils.model_targets import BinaryClassifierOutputTarget
 from explainability.utils.image import show_cam_on_image, deprocess_image, remove_padding_and_resize
 from .base_explainer import BaseExplainer
+
+
+def _draw_label(img: np.ndarray, text: str, x: int, y: int, prob_ratio: float = 1.0):
+    """Confidence label with a clean pill-style background.
+    
+    prob_ratio: 0~1, controls bg color (green→red)
+    """
+    font       = cv2.FONT_HERSHEY_SIMPLEX
+    font_scale = 0.52
+    thickness  = 1
+    pad        = 4
+
+    (tw, th), baseline = cv2.getTextSize(text, font, font_scale, thickness)
+
+    rx1, ry1 = x - pad,      y - th - pad
+    rx2, ry2 = x + tw + pad, y + baseline + pad
+
+    r = int(min(255, 2 * prob_ratio * 255))
+    g = int(min(255, 2 * (1 - prob_ratio) * 255))
+    bg_color = (r, g, 30)  # RGB: green → red
+
+    overlay = img.copy()
+    cv2.rectangle(overlay, (rx1, ry1), (rx2, ry2), bg_color, -1)
+    cv2.addWeighted(overlay, 0.75, img, 0.25, 0, img)
+    cv2.putText(img, text, (x, y), font, font_scale, (255, 255, 255), thickness, cv2.LINE_AA)
+
+
 class CAMExplainer(BaseExplainer, ABC):
     def __init__(self, 
                  aug_smooth: bool = False,
                  eigen_smooth: bool = False,
                  **kwargs):
         super().__init__(**kwargs)
-        self.aug_smooth = aug_smooth # aug_smooth has the effect of beter centering the CAM around the objects
-        self.eigen_smooth = eigen_smooth # # eigen_smooth has the effect of removing a lot of noise
+        self.aug_smooth = aug_smooth
+        self.eigen_smooth = eigen_smooth
         self.cam = self._build_cam()
     
     @abstractmethod
@@ -20,63 +47,95 @@ def _build_cam(self):
         ...
 
     def _build_grayscale_cam(self, img_path: str):
-        """
-        Generate a grayscale CAM and its associated image tensors.
-
-        Returns:
-            tuple: (grayscale_cam, tensor, face)
-                - grayscale_cam (np.ndarray): CAM result in (H, W), range [0, 1].
-                - tensor (torch.Tensor): Preprocessed input in (1, 3, H, W).
-                - face (np.ndarray): Original image in (H, W, 3), uint8.
-        """
         face, tensor = self._get_transform(img_path)
-        tensor = tensor.to(self.device) # (1,3,H,W)
+        tensor = tensor.to(self.device)
         targets = [BinaryClassifierOutputTarget(self.category)]
         
         grayscale_cam = self.cam(
-            input_tensor = tensor, # (1,3,H,W)
-            targets = targets,
-            aug_smooth = self.aug_smooth,
-            eigen_smooth = self.eigen_smooth
-        ) # (1,H,W)
-        
-        return grayscale_cam[0], tensor, face 
-        
-    def display_heatmap_on_image(self, img_path: str, 
+            input_tensor=tensor,
+            targets=targets,
+            aug_smooth=self.aug_smooth,
+            eigen_smooth=self.eigen_smooth,
+        )
+        return grayscale_cam[0], tensor, face
+
+    def _prepare_cam(self, img_path: str):
+        """Build grayscale CAM and recover it to original face resolution."""
+        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)
+        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
+        return cam_recovered, face
+    
+    def _get_binary_mask(self, cam: np.ndarray, threshold: float | str):
+        if threshold != 'auto':
+            t = np.percentile(cam, threshold * 100)
+            binary_mask = (cam > t).astype(np.uint8) * 255
+        else:
+            _, binary_mask = cv2.threshold(
+                np.uint8(cam * 255), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU
+            )
+        return binary_mask
+
+    def display_heatmap_on_image(self, 
+                                 img_path: str, 
                                  colormap: int = cv2.COLORMAP_JET, 
-                                 image_weight: float = 0.5,):
-        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)     
-        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)   
+                                 image_weight: float = 0.5):
+        cam_recovered, face = self._prepare_cam(img_path)
         
         heatmap = show_cam_on_image(
-            np.float32(face) / 255.0,  # (H,W,3), range(0~1), np.ndarray
-            cam_recovered, # (H,W), range(0~1), np.ndarray
-            use_rgb =True, 
-            colormap = colormap,
-            image_weight= image_weight)
-        
-        return heatmap # (H,W,3), range(0~255), np.uint8
+            np.float32(face) / 255.0,
+            cam_recovered,
+            use_rgb=True,
+            colormap=colormap,
+            image_weight=image_weight,
+        )
+        return heatmap  # (H,W,3), range(0~255), np.uint8
     
-    def display_contour_on_image(self, img_path: str,
-                                 threshold: float = 0.5,
-                                 thickness: int = 2,
-                                 ):
-        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)
-        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
+    def display_contour_on_image(self, 
+                                 img_path: str,
+                                 threshold: float | str = 'auto',
+                                 thickness: int = 2):
+        cam_recovered, face = self._prepare_cam(img_path)
         
-        mask = (cam_recovered > threshold).astype(np.uint8) * 255 
-        contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+        binary_mask = self._get_binary_mask(cam_recovered, threshold)
+            
+        contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
         annotated_img = cv2.drawContours(face.copy(), contours, -1, (255, 0, 0), thickness)
-        
         return annotated_img
-        
     
+    def display_bbox_on_image(self, 
+                              img_path: str,
+                              threshold: float | str = 'auto',
+                              thickness: int = 2):
+        cam_recovered, face = self._prepare_cam(img_path)
+        
+        binary_mask = self._get_binary_mask(cam_recovered, threshold)
+        
+        contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+        
+        annotated_img = face.copy()
+        for c in contours:
+            x, y, w, h = cv2.boundingRect(c)
+            roi_prob = float(np.max(cam_recovered[y:y + h, x:x + w]))
+
+            r = int(min(255, 2 * roi_prob * 255))
+            g = int(min(255, 2 * (1 - roi_prob) * 255))
+            box_color = (r, g, 30)  # RGB
+
+            cv2.rectangle(annotated_img, (x, y), (x + w, y + h), box_color, thickness)
+            _draw_label(
+                annotated_img,
+                text=f"{roi_prob * 100:.1f}%",
+                x=x,
+                y=max(y - 6, 14),
+                prob_ratio=roi_prob,
+            )
+
+        return annotated_img
+
     def __repr__(self):
         return (f"{self.__class__.__name__}("
                 f"model={self.model_name}, dataset={self.dataset}, "
                 f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
                 f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, block_idx={self.block_idx},"
                 f"aug_smooth={self.aug_smooth}, eigen_smooth={self.eigen_smooth}, "
-                f"colormap={self.colormap}, image_weight={self.image_weight}, "
-                f"category={self.category},device={self.device})")
-        
\ No newline at end of file
+                f"category={self.category},device={self.device})")
\ No newline at end of file

From 62561ddde6d60f4cdf3e63298c3d256edde86d3e Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Mon, 11 May 2026 19:11:13 +0900
Subject: [PATCH 16/48] =?UTF-8?q?=E2=9C=A8=20Add=20BinaryClassifierOutputS?=
 =?UTF-8?q?igmoidTarget?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/utils/model_targets.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/explainability/utils/model_targets.py b/explainability/utils/model_targets.py
index 62ba03a..5ec0870 100644
--- a/explainability/utils/model_targets.py
+++ b/explainability/utils/model_targets.py
@@ -1,3 +1,5 @@
+import torch
+
 class BinaryClassifierOutputTarget:
     def __init__(self, category):
         self.category = category
@@ -8,4 +10,15 @@ def __call__(self, model_output):
         else:
             sign = -1
         return model_output * sign
+    
+class BinaryClassifierOutputSigmoidTarget:
+    def __init__(self, category):
+        self.category = category
+
+    def __call__(self, model_output):
+        if self.category == 1:
+            sign = 1
+        else:
+            sign = -1
+        return torch.sigmoid(model_output) * sign
     
\ No newline at end of file

From 01cc4fe1294d8d67c6865396c788fa9f190bb5a7 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Tue, 12 May 2026 19:13:01 +0900
Subject: [PATCH 17/48] =?UTF-8?q?=E2=9C=A8=20Create=20CAMEvaluator?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_evaluator.py | 42 +++++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)
 create mode 100644 explainability/cam_evaluator.py

diff --git a/explainability/cam_evaluator.py b/explainability/cam_evaluator.py
new file mode 100644
index 0000000..d7f4017
--- /dev/null
+++ b/explainability/cam_evaluator.py
@@ -0,0 +1,42 @@
+import numpy as np
+from typing import List, Dict, Literal
+from explainability.utils.model_targets import BinaryClassifierOutputSigmoidTarget
+from explainability.cam_explainer import CAMExplainer
+from explainability.metrics.road import (
+    ROADLeastRelevantFirstAverage, ROADMostRelevantFirstAverage,
+    ROADCombined
+)
+
+METRIC_REGISTRY = {
+    "MORF": ROADMostRelevantFirstAverage,
+    "LERF": ROADLeastRelevantFirstAverage,
+    "Combined": ROADCombined,
+}
+
+class CAMEvaluator:
+    def __init__(self,
+                 cam_explainer: CAMExplainer,
+                 cam_metric: Literal["Combined", "MORF", "LERF"] = "Combined", # Combined, MORF, LERF
+                 percentiles: List[int] = [10,20,30,40,50,60,70,80,90],
+                 ):
+        self.cam_explainer = cam_explainer
+        self.cam_metric_name = cam_metric
+        self.percentiles = percentiles
+        self.cam_metric = METRIC_REGISTRY[cam_metric](percentiles)
+        
+    def evaluate(self, img_path: str) -> float:
+        grayscale_cam, tensor, _ = self.cam_explainer._build_grayscale_cam(img_path)
+        targets = [BinaryClassifierOutputSigmoidTarget(self.cam_explainer.category)]
+        
+        return self.cam_metric(
+            input_tensor = tensor, # (1,C,H,W)
+            cams = grayscale_cam[np.newaxis, ...], #(1,H,W)
+            targets = targets,
+            model = self.cam_explainer.model, 
+        )[0]
+    
+    def __repr__(self):
+        return (
+            f"CAMEvaluator(explainer={self.cam_explainer.__class__.__name__}, "
+            f"metric={self.cam_metric_name}, percentiles={self.percentiles})"
+        )
\ No newline at end of file

From d67a42638a97bbc3a447438f77b160586ef92d4e Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 13 May 2026 09:32:29 +0900
Subject: [PATCH 18/48] =?UTF-8?q?=F0=9F=8E=A8=20Add=20Reproductability=20f?=
 =?UTF-8?q?or=20same=20result?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_evaluator.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/explainability/cam_evaluator.py b/explainability/cam_evaluator.py
index d7f4017..a4aa46a 100644
--- a/explainability/cam_evaluator.py
+++ b/explainability/cam_evaluator.py
@@ -1,4 +1,5 @@
 import numpy as np
+import torch
 from typing import List, Dict, Literal
 from explainability.utils.model_targets import BinaryClassifierOutputSigmoidTarget
 from explainability.cam_explainer import CAMExplainer
@@ -18,13 +19,18 @@ def __init__(self,
                  cam_explainer: CAMExplainer,
                  cam_metric: Literal["Combined", "MORF", "LERF"] = "Combined", # Combined, MORF, LERF
                  percentiles: List[int] = [10,20,30,40,50,60,70,80,90],
+                 seed: int = 2026
                  ):
+        self.seed = seed
         self.cam_explainer = cam_explainer
         self.cam_metric_name = cam_metric
         self.percentiles = percentiles
         self.cam_metric = METRIC_REGISTRY[cam_metric](percentiles)
         
     def evaluate(self, img_path: str) -> float:
+        torch.manual_seed(self.seed)
+        np.random.seed(self.seed)
+        
         grayscale_cam, tensor, _ = self.cam_explainer._build_grayscale_cam(img_path)
         targets = [BinaryClassifierOutputSigmoidTarget(self.cam_explainer.category)]
         

From 9d0ce5b91ef7463defc45820d1d55e86bf06ac65 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 13 May 2026 09:47:01 +0900
Subject: [PATCH 19/48] =?UTF-8?q?=F0=9F=9A=9A=20Move=20base,cam,gradient,e?=
 =?UTF-8?q?igenvalue,perturbation,misc=20explainer.py=20to=20explainer=20f?=
 =?UTF-8?q?older?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/__init__.py                          | 13 +++++++++++++
 explainability/cam_evaluator.py                     |  2 +-
 explainability/{ => explainer}/base_explainer.py    |  0
 explainability/{ => explainer}/cam_explainer.py     |  2 +-
 .../{ => explainer}/eigenvalue_explainer.py         |  2 +-
 .../{ => explainer}/gradient_explainer.py           |  2 +-
 explainability/{ => explainer}/misc_explainer.py    |  2 +-
 .../{ => explainer}/perturbation_explainer.py       |  2 +-
 8 files changed, 19 insertions(+), 6 deletions(-)
 rename explainability/{ => explainer}/base_explainer.py (100%)
 rename explainability/{ => explainer}/cam_explainer.py (98%)
 rename explainability/{ => explainer}/eigenvalue_explainer.py (95%)
 rename explainability/{ => explainer}/gradient_explainer.py (96%)
 rename explainability/{ => explainer}/misc_explainer.py (84%)
 rename explainability/{ => explainer}/perturbation_explainer.py (92%)

diff --git a/explainability/__init__.py b/explainability/__init__.py
index e69de29..b547123 100644
--- a/explainability/__init__.py
+++ b/explainability/__init__.py
@@ -0,0 +1,13 @@
+from explainability.explainer.base_explainer import BaseExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
+from explainability.explainer.gradient_explainer import (
+    GradCAMExplainer, GradCAMPlusPlusExplainer, GradCAMElementWiseExplainer,
+    XGradCAMExplainer, HiResCAMExplainer
+)
+from explainability.explainer.eigenvalue_explainer import (
+    EigenCAMExplainer, EigenGradCAMExplainer, KPCACAMExplainer
+)
+from explainability.explainer.perturbation_explainer import (
+    ScoreCAMExplainer, FEMExplainer
+)
+from explainability.explainer.misc_explainer import LayerCAMExplainer
\ No newline at end of file
diff --git a/explainability/cam_evaluator.py b/explainability/cam_evaluator.py
index a4aa46a..c325811 100644
--- a/explainability/cam_evaluator.py
+++ b/explainability/cam_evaluator.py
@@ -2,7 +2,7 @@
 import torch
 from typing import List, Dict, Literal
 from explainability.utils.model_targets import BinaryClassifierOutputSigmoidTarget
-from explainability.cam_explainer import CAMExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
 from explainability.metrics.road import (
     ROADLeastRelevantFirstAverage, ROADMostRelevantFirstAverage,
     ROADCombined
diff --git a/explainability/base_explainer.py b/explainability/explainer/base_explainer.py
similarity index 100%
rename from explainability/base_explainer.py
rename to explainability/explainer/base_explainer.py
diff --git a/explainability/cam_explainer.py b/explainability/explainer/cam_explainer.py
similarity index 98%
rename from explainability/cam_explainer.py
rename to explainability/explainer/cam_explainer.py
index bec4ffd..5764ef3 100644
--- a/explainability/cam_explainer.py
+++ b/explainability/explainer/cam_explainer.py
@@ -4,7 +4,7 @@
 from abc import ABC, abstractmethod
 from explainability.utils.model_targets import BinaryClassifierOutputTarget
 from explainability.utils.image import show_cam_on_image, deprocess_image, remove_padding_and_resize
-from .base_explainer import BaseExplainer
+from explainability.explainer.base_explainer import BaseExplainer
 
 
 def _draw_label(img: np.ndarray, text: str, x: int, y: int, prob_ratio: float = 1.0):
diff --git a/explainability/eigenvalue_explainer.py b/explainability/explainer/eigenvalue_explainer.py
similarity index 95%
rename from explainability/eigenvalue_explainer.py
rename to explainability/explainer/eigenvalue_explainer.py
index 79f0927..9aa2633 100644
--- a/explainability/eigenvalue_explainer.py
+++ b/explainability/explainer/eigenvalue_explainer.py
@@ -1,5 +1,5 @@
 from pytorch_grad_cam import EigenCAM, EigenGradCAM, KPCA_CAM
-from explainability.cam_explainer import CAMExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
 
 class EigenCAMExplainer(CAMExplainer):
     """
diff --git a/explainability/gradient_explainer.py b/explainability/explainer/gradient_explainer.py
similarity index 96%
rename from explainability/gradient_explainer.py
rename to explainability/explainer/gradient_explainer.py
index 0282fd6..3767fc7 100644
--- a/explainability/gradient_explainer.py
+++ b/explainability/explainer/gradient_explainer.py
@@ -2,7 +2,7 @@
     GradCAM, GradCAMPlusPlus, GradCAMElementWise,
     XGradCAM, HiResCAM
 )
-from explainability.cam_explainer import CAMExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
 
 
 class GradCAMExplainer(CAMExplainer):
diff --git a/explainability/misc_explainer.py b/explainability/explainer/misc_explainer.py
similarity index 84%
rename from explainability/misc_explainer.py
rename to explainability/explainer/misc_explainer.py
index db6afa0..85cb38f 100644
--- a/explainability/misc_explainer.py
+++ b/explainability/explainer/misc_explainer.py
@@ -1,5 +1,5 @@
 from pytorch_grad_cam import LayerCAM, RandomCAM
-from explainability.cam_explainer import CAMExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
 
 
 class LayerCAMExplainer(CAMExplainer):
diff --git a/explainability/perturbation_explainer.py b/explainability/explainer/perturbation_explainer.py
similarity index 92%
rename from explainability/perturbation_explainer.py
rename to explainability/explainer/perturbation_explainer.py
index 1753954..cb460aa 100644
--- a/explainability/perturbation_explainer.py
+++ b/explainability/explainer/perturbation_explainer.py
@@ -1,5 +1,5 @@
 from pytorch_grad_cam import ScoreCAM, FEM
-from explainability.cam_explainer import CAMExplainer
+from explainability.explainer.cam_explainer import CAMExplainer
 class ScoreCAMExplainer(CAMExplainer):
     """
     Perbutate the image by the scaled activation

From 8c43e5e9f153978c48ff20fffd74f312d1a4c0f8 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 13 May 2026 15:34:47 +0900
Subject: [PATCH 20/48] =?UTF-8?q?=E2=9C=A8=20Create=20cam=20router=20for?=
 =?UTF-8?q?=20DeepFake=20XAI?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/main.py         | 51 ++++++++++++++++++++++++++-------------------
 App/routes/cam.py   |  3 +++
 App/utils/common.py | 12 +++++++++--
 3 files changed, 42 insertions(+), 24 deletions(-)
 create mode 100644 App/routes/cam.py

diff --git a/App/main.py b/App/main.py
index 530812c..95c3c97 100644
--- a/App/main.py
+++ b/App/main.py
@@ -1,44 +1,50 @@
 import os
 import sys
-from pathlib import Path
 import warnings
 import logging
+from pathlib import Path
 from dotenv import load_dotenv
 
 # ------ 상위 폴더 경로 설정 --------
+REQUIRED_PACKAGES = ["deepguard", "inference", "explainability", "preprocess"]
+
 current_file = Path(__file__).resolve()
-project_root = current_file.parent.parent  # App의 상위 폴더
+project_root = current_file.parent.parent 
 
-if str(project_root) not in sys.path:
-    sys.path.insert(0, str(project_root))
+for pkg in REQUIRED_PACKAGES:
+    pkg_path = str(project_root / pkg)
+    if pkg_path not in sys.path:
+        sys.path.insert(0, pkg_path)
 
+# ------ .env 파일 가져오기 ------
 load_dotenv()
 
 # -------- Huggingface_Hub 인증 ----------
 if os.getenv("HF_TOKEN"):
     os.environ["HF_TOKEN"] = os.getenv("HF_TOKEN")
+else:
+    print("[Warning] HF_TOKEN is not set")
 
 from fastapi import FastAPI
 from fastapi.staticfiles import StaticFiles
 from fastapi.exceptions import RequestValidationError
 from starlette.exceptions import HTTPException as StarletteHTTPException
 from fastapi.middleware.cors import CORSMiddleware
-from starlette.middleware.sessions import SessionMiddleware
-from routes import auth, home, inference, image, video
-from utils.common import lifespan
-from utils import exc_handler, middleware
+# from starlette.middleware.sessions import SessionMiddleware
+from routes import auth, home, inference, image, video, cam
+from utils import exc_handler, middleware, common
 
 
-# 가상 인스턴스 생성
-app = FastAPI(lifespan=lifespan)
+# 가상 FastAPI 인스턴스 생성
+app = FastAPI(lifespan=common.lifespan)
 
-# StaticFile 등록하기
+# StaticFile 등록 (이미지, 비디오 파일)
 app.mount("/static", StaticFiles(directory="static"), name="static")
 
-# Cross Origin Resource Sharing
+# CORS Setup for cross-origin requests
+# FrontEnd: React, BackEnd: FastAPI
 origins_str = os.getenv("CORS_ALLOWED_ORIGINS")
 allowed_origins = [origin.strip() for origin in origins_str.split(",")]
-
 app.add_middleware(CORSMiddleware,
                    allow_origins=allowed_origins,
                    allow_methods=["*"],
@@ -52,16 +58,17 @@
 # app.add_middleware(SessionMiddleware, secret_key=SECRET_KEY, max_age=3600)
 
 # 세션 미들웨어 등록 - Redis 이용
-app.add_middleware(middleware.RedisSessionMiddleware, max_age=7200)
+app.add_middleware(middleware.RedisSessionMiddleware, max_age=3600)
 
-# 라우터 등록
-app.include_router(auth.router)
-app.include_router(home.router)
-app.include_router(inference.router)
-app.include_router(image.router)
-app.include_router(video.router)
+# 라우터 등록 (View: Router, Controller: Service)
+app.include_router(auth.router) # 로그인, 로그아웃, 회원가입
+app.include_router(home.router) # 세션 유저 정보 가져오기
+app.include_router(inference.router) # 이미지, 비디오 비동기 추론 접수, 추론 값 가져오기 
+app.include_router(image.router) # 이미지 삭제 및 가져오기 
+app.include_router(video.router) # 비디오 삭제 및 가져오기
+app.include_router(cam.router) # 딥페이크 이미지, 비디오 위조 흔적 표시
 
-# Custom HTTPException Handler
+# 커스텀 예외 처리: HTTPException
 app.add_exception_handler(StarletteHTTPException, exc_handler.custom_http_exception_handler)
-# Custom RequestValidationError Handler
+# 커스텀 예외 처리: RequestValidationError
 app.add_exception_handler(RequestValidationError, exc_handler.validation_exception_handler)
\ No newline at end of file
diff --git a/App/routes/cam.py b/App/routes/cam.py
new file mode 100644
index 0000000..eb0a78c
--- /dev/null
+++ b/App/routes/cam.py
@@ -0,0 +1,3 @@
+from fastapi import APIRouter
+
+router = APIRouter(prefix="/cam", tags=["cam"])
\ No newline at end of file
diff --git a/App/utils/common.py b/App/utils/common.py
index c49d71c..a72d30c 100644
--- a/App/utils/common.py
+++ b/App/utils/common.py
@@ -2,12 +2,20 @@
 from db.database import engine
 from contextlib import asynccontextmanager
 
+# Uvicorn이 앱 시작/종료 시 asynccontextmanager를 호출함
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # FastAPI 인스턴스 기동시 필요한 작업 수행
+    
+    # ── Startup ──────────────────────────────────────────
+    # Event Loop 위에서 실행, DB 연결 풀 초기화
     print("Starting up...")
+    
+    # yield 이후 app이 실제 요청을 처리하기 시작함
     yield
     
-    # FastAPI 인스턴스 종료시 필요한 작업 수행
+    # ── Shutdown ─────────────────────────────────────────
+    # 새 요청은 차단되고, 진행 중인 요청 완료 후 여기로 진입
     print("Shutting down...")
+    
+    # SQLAlchemy 비동기 엔진의 커넥션 풀 전체 반환 및 종료
     await engine.dispose()
\ No newline at end of file

From ef7fb4c714ef3e91118892825bbd616cebf45ae4 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 13 May 2026 18:36:08 +0900
Subject: [PATCH 21/48] =?UTF-8?q?=F0=9F=94=A5=20Remove=20legacy=20file(set?=
 =?UTF-8?q?up.py)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 setup.py | 30 ------------------------------
 1 file changed, 30 deletions(-)
 delete mode 100644 setup.py

diff --git a/setup.py b/setup.py
deleted file mode 100644
index 16d3c3c..0000000
--- a/setup.py
+++ /dev/null
@@ -1,30 +0,0 @@
-import setuptools
-
-with open("README.md", mode='r', encoding='utf-8') as fh:
-    long_description = fh.read()
-
-setuptools.setup(
-    name = 'deepguard',
-    version = '0.2.1',
-    author = 'seoyunje',
-    author_email = "seoyunje2001@gmail.com",
-    description = "Multi-Scale Efficient Vision Transformer for Robust Deepfake Detection",
-    long_description = long_description,
-    long_description_content_type ='text/markdown',
-    url = 'https://github.com/HanMoonSub/DeepGuard',
-    project_urls = {
-        "Bug Tracker": "https://github.com/HanMoonSub/DeepGuard/issues",
-        "Source Code": "https://github.com/HanMoonSub/DeepGuard",
-    },
-    classifiers = [
-        'Programming Language :: Python :: 3',
-        'Programming Language :: Python :: 3.10',
-        'Programming Language :: Python :: 3.11',
-        'License :: OSI Approved :: MIT License',
-        'Operating System :: OS Independent', 
-        'Topic :: Scientific/Engineering :: Artificial Intelligence',  
-    ],
-    packages = setuptools.find_packages(exclude=['Attention', 'preprocess','inference']),
-    python_requires = '>=3.10',
-    install_requires = []
-)
\ No newline at end of file

From 91d4981e580f59a78a4b7907c2a71628bc2842d6 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 13 May 2026 18:47:49 +0900
Subject: [PATCH 22/48] =?UTF-8?q?=E2=9C=A8=20Create=20pyproject.toml?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .gitignore                    |  1 +
 App/celery_app.py             | 10 +------
 App/main.py                   | 15 -----------
 App/services/inference_svc.py |  2 +-
 App/services/video_svc.py     |  2 +-
 pyproject.toml                | 51 +++++++++++++++++++++++++++++++++++
 6 files changed, 55 insertions(+), 26 deletions(-)
 create mode 100644 pyproject.toml

diff --git a/.gitignore b/.gitignore
index da944b5..d0e86dc 100644
--- a/.gitignore
+++ b/.gitignore
@@ -6,6 +6,7 @@
 npm-debug.log*
 yarn-debug.log*
 yarn-error.log*
+*.egg-info
 
 # [FastAPI / Python - App 폴더 관련]
 **/__pycache__/
diff --git a/App/celery_app.py b/App/celery_app.py
index 027ba15..df5fd33 100644
--- a/App/celery_app.py
+++ b/App/celery_app.py
@@ -1,18 +1,10 @@
 import os
-import sys
-from pathlib import Path
 from celery import Celery
 from dotenv import load_dotenv
 
-current_file = Path(__file__).resolve()
-project_root = current_file.parent.parent
-
-if str(project_root) not in sys.path:
-    sys.path.insert(0, str(project_root))
-
+# ------ .env 파일 가져오기 ------
 load_dotenv()
 
-# 기본값 설정
 REDIS_HOST = os.getenv("REDIS_HOST", "localhost")
 REDIS_PORT = os.getenv("REDIS_PORT", 6379)
 
diff --git a/App/main.py b/App/main.py
index 95c3c97..d0cf85e 100644
--- a/App/main.py
+++ b/App/main.py
@@ -1,21 +1,6 @@
 import os
-import sys
-import warnings
-import logging
-from pathlib import Path
 from dotenv import load_dotenv
 
-# ------ 상위 폴더 경로 설정 --------
-REQUIRED_PACKAGES = ["deepguard", "inference", "explainability", "preprocess"]
-
-current_file = Path(__file__).resolve()
-project_root = current_file.parent.parent 
-
-for pkg in REQUIRED_PACKAGES:
-    pkg_path = str(project_root / pkg)
-    if pkg_path not in sys.path:
-        sys.path.insert(0, pkg_path)
-
 # ------ .env 파일 가져오기 ------
 load_dotenv()
 
diff --git a/App/services/inference_svc.py b/App/services/inference_svc.py
index 3ca92d6..02a8f60 100644
--- a/App/services/inference_svc.py
+++ b/App/services/inference_svc.py
@@ -207,7 +207,7 @@ async def run_inference():
                 if result["analysis"].get("frame_results"):
                     try:
                         async with celery_db_conn() as conn:
-                            await video_svc.save_video_frame_results(
+                            await video_svc.save_video_frame_result(
                                 conn, video_id, result["analysis"]["frame_results"]
                             )
                     except Exception as e:
diff --git a/App/services/video_svc.py b/App/services/video_svc.py
index cdf5ac4..4455e1d 100644
--- a/App/services/video_svc.py
+++ b/App/services/video_svc.py
@@ -347,7 +347,7 @@ async def save_video_meta_result(conn: Connection, video_id: int, analysis: dict
     
 # 비디오 상세 결과 값 저장하기 
 # 비디오 프레임 별 딥페이크 점수, 얼굴 신뢰도, 얼굴 비율, 얼굴 조도 반환
-async def save_video_frame_results(conn: Connection, video_id: int, frame_results: list):
+async def save_video_frame_result(conn: Connection, video_id: int, frame_results: list):
     try:
         query = """
             INSERT INTO video_frame_result
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 0000000..2d4ad22
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,51 @@
+[build-system]
+requires = ["setuptools>=68", "wheel"]
+build-backend = "setuptools.build_meta"
+
+# ─────────────────────────────────────────────
+# [project]
+# PyPI에 올라가는 패키지 메타데이터
+# pip install deepguard 했을 때 보이는 정보들
+# ─────────────────────────────────────────────
+[project]
+name = "deepguard"           # pip install deepguard 할 때 이 이름
+version = "0.2.1"            # pip install deepguard==0.2.1 으로 버전 고정 가능
+description = "Multi-Scale Efficient Vision Transformer for Robust Deepfake Detection"
+readme = "README.md"         # PyPI 페이지에 표시될 설명 (마크다운)
+requires-python = ">=3.10"   # 이 버전 미만이면 pip install 자체를 막음
+license = { text = "MIT" }
+authors = [
+    { name = "seoyunje", email = "seoyunje2001@gmail.com" }
+]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+
+dependencies = ["torch>=2.10.0", 
+                "torchmetrics>=1.6.1",
+                "timm>=1.0.12",
+                "ultralytics>=8.4.41",
+                "grad-cam>=1.5.5"
+                ]
+
+[project.urls]
+Homepage = "https://github.com/HanMoonSub/DeepGuard"
+"Bug Tracker" = "https://github.com/HanMoonSub/DeepGuard/issues"
+"Source Code" = "https://github.com/HanMoonSub/DeepGuard"
+
+
+[tool.setuptools.packages.find]
+
+include = [
+    "deepguard*",       # 모델, 레이어, 데이터셋 등 핵심 패키지
+    "inference*",       # ImagePredictor, VideoPredictor
+    "explainability*",  # CAM 기반 XAI
+    "preprocess*",      # 얼굴 검출, 프레임 추출 등
+]
+
+exclude = ["App*", "client*"]
\ No newline at end of file

From 27066ad1f566da0756de6346e7f91e4ae146a701 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 14 May 2026 10:58:27 +0900
Subject: [PATCH 23/48] =?UTF-8?q?=F0=9F=9A=9A=20Rename=20Cam=20Router=20->?=
 =?UTF-8?q?=20Explain=20Router?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/main.py           | 4 ++--
 App/routes/cam.py     | 3 ---
 App/routes/explain.py | 5 +++++
 3 files changed, 7 insertions(+), 5 deletions(-)
 delete mode 100644 App/routes/cam.py
 create mode 100644 App/routes/explain.py

diff --git a/App/main.py b/App/main.py
index d0cf85e..14fafc2 100644
--- a/App/main.py
+++ b/App/main.py
@@ -16,7 +16,7 @@
 from starlette.exceptions import HTTPException as StarletteHTTPException
 from fastapi.middleware.cors import CORSMiddleware
 # from starlette.middleware.sessions import SessionMiddleware
-from routes import auth, home, inference, image, video, cam
+from routes import auth, home, inference, image, video, explain
 from utils import exc_handler, middleware, common
 
 
@@ -51,7 +51,7 @@
 app.include_router(inference.router) # 이미지, 비디오 비동기 추론 접수, 추론 값 가져오기 
 app.include_router(image.router) # 이미지 삭제 및 가져오기 
 app.include_router(video.router) # 비디오 삭제 및 가져오기
-app.include_router(cam.router) # 딥페이크 이미지, 비디오 위조 흔적 표시
+app.include_router(explain.router) # 딥페이크 이미지, 비디오 위조 흔적 표시
 
 # 커스텀 예외 처리: HTTPException
 app.add_exception_handler(StarletteHTTPException, exc_handler.custom_http_exception_handler)
diff --git a/App/routes/cam.py b/App/routes/cam.py
deleted file mode 100644
index eb0a78c..0000000
--- a/App/routes/cam.py
+++ /dev/null
@@ -1,3 +0,0 @@
-from fastapi import APIRouter
-
-router = APIRouter(prefix="/cam", tags=["cam"])
\ No newline at end of file
diff --git a/App/routes/explain.py b/App/routes/explain.py
new file mode 100644
index 0000000..f012bc9
--- /dev/null
+++ b/App/routes/explain.py
@@ -0,0 +1,5 @@
+from fastapi import APIRouter, status 
+from sqlalchemy import Connection
+from db.database import context_get_conn
+
+router = APIRouter(prefix="/explain", tags=["explain"])

From 573ab4b90369a3a429e7751098e8871ecdafe52c Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 14 May 2026 11:12:30 +0900
Subject: [PATCH 24/48] =?UTF-8?q?=E2=9C=A8=20Create=20explain=5Fimage?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

diff --git a/App/routes/explain.py b/App/routes/explain.py
index f012bc9..7aba577 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -1,5 +1,22 @@
-from fastapi import APIRouter, status 
+import io
+from fastapi import APIRouter, status, Depends
+from fastapi.responses import StreamingResponse
 from sqlalchemy import Connection
 from db.database import context_get_conn
+from schemas.explain_schema import ExplainRequest
+from services import session_svc, explain_svc
 
 router = APIRouter(prefix="/explain", tags=["explain"])
+
+@router.post("/image/{image_id}", status_code=status.HTTP_200_OK, 
+             response_class=StreamingResponse , summary="딥페이크 이미지 위조 흔적 시각화(CAM)")
+async def explain_image(
+    image_id: int,
+    explain_req: ExplainRequest,
+    conn: Connection = Depends(context_get_conn),
+    session_user = Depends(session_svc.get_session_user_opt), # 비회원 허용
+):
+  img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
+  return StreamingResponse(io.BytesIO(img_bytes), media_type="image/png")
+    
+    
\ No newline at end of file

From 00a8416bd73f7f6a6d2be560a6df57e2eaa59aea Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 14 May 2026 11:39:54 +0900
Subject: [PATCH 25/48] =?UTF-8?q?=F0=9F=8E=A8=20Add=20response=5Fclass=20&?=
 =?UTF-8?q?=20response=5Fmodel=20in=20inference=20routes?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/inference.py | 24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/App/routes/inference.py b/App/routes/inference.py
index ba856af..2afff51 100644
--- a/App/routes/inference.py
+++ b/App/routes/inference.py
@@ -1,17 +1,19 @@
 from fastapi import (
-    APIRouter, Depends, status, Form,
-    File, UploadFile
-)
+    APIRouter, UploadFile, status,
+    Depends, Form, File)
+from fastapi.responses import JSONResponse
 from fastapi.exceptions import HTTPException
 from services import image_svc, session_svc, inference_svc, video_svc
 from sqlalchemy import Connection
 from db.database import context_get_conn
-from schemas.video_schema import VideoDetailResponse
+from schemas.video_schema import VideoDetailResponse, VideoDetailData
+from schemas.image_schema import ImageData_indi
 
 
 router = APIRouter(prefix="/inference", tags=["inference"])
 
-@router.post("/image", status_code=status.HTTP_202_ACCEPTED, summary="딥페이크 비동기 이미지 추론 접수") 
+@router.post("/image", status_code=status.HTTP_202_ACCEPTED,
+             response_class=JSONResponse, summary="딥페이크 비동기 이미지 추론 접수") 
 async def predict_image(
                         imagefile: UploadFile = File(...), # 사용자가 업로드한 이미지 객체
                         version_type: str = Form(...), # deepguard1, deepguard2
@@ -43,7 +45,8 @@ async def predict_image(
         "status": "success",
     }
 
-@router.get("/image/{image_id}", status_code=status.HTTP_200_OK, summary="딥페이크 비동기 이미지 추론 결과값 가져오기")
+@router.get("/image/{image_id}", status_code=status.HTTP_200_OK,
+            response_model=ImageData_indi, summary="딥페이크 비동기 이미지 추론 결과값 가져오기")
 async def get_image_result(
                             image_id: int,
                             conn: Connection = Depends(context_get_conn),
@@ -62,7 +65,8 @@ async def get_image_result(
         )
     return image_data
 
-@router.post("/video", status_code=status.HTTP_202_ACCEPTED, summary="딥페이크 비동기 비디오 추론 접수")
+@router.post("/video", status_code=status.HTTP_202_ACCEPTED, 
+             response_class=JSONResponse, summary="딥페이크 비동기 비디오 추론 접수")
 async def predict_video(
                         videofile: UploadFile = File(...), # 사용자가 업로드한 비디오 객체
                         version_type: str = Form(...), # deepguard1, deepguard2
@@ -94,7 +98,8 @@ async def predict_video(
         "status": "success",
     }
 
-@router.get("/video/{video_id}", status_code=status.HTTP_200_OK, summary="딥페이크 비디오 비디오 추론 결과값 가져오기")
+@router.get("/video/{video_id}", status_code=status.HTTP_200_OK, 
+            response_model=VideoDetailData, summary="딥페이크 비디오 비디오 추론 결과값 가져오기")
 async def get_video_result(
                            video_id: int,
                            conn: Connection = Depends(context_get_conn),
@@ -117,7 +122,8 @@ async def get_video_result(
     
     return video_data
 
-@router.get("/video/{video_id}/detail", status_code=status.HTTP_200_OK, summary="딥페이크 상세 분석 결과값 가져오기")
+@router.get("/video/{video_id}/detail", status_code=status.HTTP_200_OK, 
+            response_model=VideoDetailResponse ,summary="딥페이크 상세 분석 결과값 가져오기")
 async def get_video_detail(
                            video_id: int,
                            conn: Connection = Depends(context_get_conn),

From 302aca4be6aa255db1df8d1622f3d57ed314554c Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 14:38:08 +0900
Subject: [PATCH 26/48] =?UTF-8?q?=F0=9F=93=9D=20Add=20python-dotenv=20&=20?=
 =?UTF-8?q?cryptography?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 requirements.txt | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/requirements.txt b/requirements.txt
index 72e2059..a7bc9ab 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -5,6 +5,7 @@ python-multipart>=0.0.22
 pydantic[email]>=2.12.5
 aiofiles>=25.1.0
 itsdangerous>=2.2.0
+python-dotenv>=0.9.9
 
 # --- [Database & ORM] ---
 sqlalchemy>=2.0.48
@@ -16,6 +17,7 @@ redis>=7.4.0
 # --- [Security & Auth] ---
 bcrypt>=4.0.1
 passlib[bcrypt]>=1.7.4
+cryptography>=42.0.0
 
 # --- [Data Analysis & Math] ---
 numpy>=1.23.0,<3.0.0

From c8e5566645ff7b13428618fa509ac1a0b5d3cc6a Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 14:45:11 +0900
Subject: [PATCH 27/48] =?UTF-8?q?=E2=9A=B0=EF=B8=8F=20Remove=20debugging?=
 =?UTF-8?q?=20print=20in=20upload=5Fvideo,=20delete=5Fvideo?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/video_svc.py | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/App/services/video_svc.py b/App/services/video_svc.py
index 4455e1d..fd69f20 100644
--- a/App/services/video_svc.py
+++ b/App/services/video_svc.py
@@ -52,8 +52,6 @@ async def upload_video(user_email: str | None, videofile: UploadFile) -> str:
                 detail="비디오 파일을 저장하는 중 오류가 발생했습니다."
             )
 
-        print(f"Upload Succeeded: {upload_video_loc}")
-
         # 6. DB 저장용 경로 반환
         return upload_video_loc[1:].replace("\\", "/")
     
@@ -73,7 +71,6 @@ async def delete_video(video_loc: str):
 
         if os.path.exists(file_path):
             os.remove(file_path)
-            print(f"Video file removed: {file_path}")
         else:
             print(f"Video file not found: {file_path}")
 

From b430278d61d254a4583df3a426ea90e1345fc9ee Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 14:55:43 +0900
Subject: [PATCH 28/48] =?UTF-8?q?=E2=9C=A8=20Add=20validation=20and=20desc?=
 =?UTF-8?q?ription=20to=20image,=20video=20inference=20params?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/inference.py | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/App/routes/inference.py b/App/routes/inference.py
index 2afff51..b538ebb 100644
--- a/App/routes/inference.py
+++ b/App/routes/inference.py
@@ -8,7 +8,7 @@
 from db.database import context_get_conn
 from schemas.video_schema import VideoDetailResponse, VideoDetailData
 from schemas.image_schema import ImageData_indi
-
+from typing import Literal
 
 router = APIRouter(prefix="/inference", tags=["inference"])
 
@@ -16,9 +16,9 @@
              response_class=JSONResponse, summary="딥페이크 비동기 이미지 추론 접수") 
 async def predict_image(
                         imagefile: UploadFile = File(...), # 사용자가 업로드한 이미지 객체
-                        version_type: str = Form(...), # deepguard1, deepguard2
-                        model_type: str = Form(...), # fast model, pro moel 
-                        domain_type: str = Form(...), # model 학습시 사용한 dataset 종류
+                        version_type: Literal["v1","v2"] = Form("v2", description="모델 엔진 버전"),
+                        model_type: Literal["fast","pro"] = Form("fast", description="추론 모드 (fast: 속도 우선, pro: 정확도 우선)"), 
+                        domain_type: Literal["서양인","동양인"] = Form("서양인", description="학습 데이터셋 도메인"),
                         conn: Connection = Depends(context_get_conn), # 이미지 File 저장
                         session_user = Depends(session_svc.get_session_user_opt) # Signed Cookie 없을 시 None 반환
                         ):
@@ -69,9 +69,9 @@ async def get_image_result(
              response_class=JSONResponse, summary="딥페이크 비동기 비디오 추론 접수")
 async def predict_video(
                         videofile: UploadFile = File(...), # 사용자가 업로드한 비디오 객체
-                        version_type: str = Form(...), # deepguard1, deepguard2
-                        model_type: str = Form(...), # fast model, pro moel 
-                        domain_type: str = Form(...), # model 학습시 사용한 dataset 종류
+                        version_type: Literal["v1","v2"] = Form("v2", description="모델 엔진 버전"),
+                        model_type: Literal["fast","pro"] = Form("fast", description="추론 모드 (fast: 속도 우선, pro: 정확도 우선)"), 
+                        domain_type: Literal["서양인","동양인"] = Form("서양인", description="학습 데이터셋 도메인"),
                         conn: Connection = Depends(context_get_conn), # 비디오 File 저장
                         session_user = Depends(session_svc.get_session_user_opt) # Signed Cookie 없을 시 None 반환
                         ):

From 13bb4c6ce6c87f3f3ec73bb3e6212e8440fe52a2 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 15:13:00 +0900
Subject: [PATCH 29/48] =?UTF-8?q?=E2=9A=B0=EF=B8=8F=20Remove=20Debugging?=
 =?UTF-8?q?=20print=20upload=5Fimage,=20delete=5Fimage?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/image_svc.py | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/App/services/image_svc.py b/App/services/image_svc.py
index 9a92d68..e094f6f 100644
--- a/App/services/image_svc.py
+++ b/App/services/image_svc.py
@@ -53,8 +53,6 @@ async def upload_image(user_email: str | None, imagefile: UploadFile) -> str:
                 detail="이미지 파일을 저장하는 중 오류가 발생했습니다."
             )
 
-        print(f"Upload Succeeded: {upload_image_loc}")
-
         # 6. DB 저장용 경로 반환
         return upload_image_loc[1:].replace("\\", "/")
     
@@ -75,7 +73,6 @@ async def delete_image(image_loc: str):
 
         if os.path.exists(file_path):
             os.remove(file_path)
-            print(f"File removed: {file_path}")
         else:
             print(f"File not found: {file_path}")
 

From da1ab758db21c86f6b924b855dc8f2f75b5acece Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 16:17:57 +0900
Subject: [PATCH 30/48] =?UTF-8?q?=E2=9C=A8=20Create=20ExplainRequest?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py         |  7 +++--
 App/schemas/explain_schema.py | 58 +++++++++++++++++++++++++++++++++++
 2 files changed, 62 insertions(+), 3 deletions(-)
 create mode 100644 App/schemas/explain_schema.py

diff --git a/App/routes/explain.py b/App/routes/explain.py
index 7aba577..3d35db9 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -4,7 +4,7 @@
 from sqlalchemy import Connection
 from db.database import context_get_conn
 from schemas.explain_schema import ExplainRequest
-from services import session_svc, explain_svc
+from services import session_svc # explain_svc
 
 router = APIRouter(prefix="/explain", tags=["explain"])
 
@@ -16,7 +16,8 @@ async def explain_image(
     conn: Connection = Depends(context_get_conn),
     session_user = Depends(session_svc.get_session_user_opt), # 비회원 허용
 ):
-  img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
-  return StreamingResponse(io.BytesIO(img_bytes), media_type="image/png")
+    return None
+#   img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
+#   return StreamingResponse(io.BytesIO(img_bytes), media_type="image/png")
     
     
\ No newline at end of file
diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
new file mode 100644
index 0000000..8ad5823
--- /dev/null
+++ b/App/schemas/explain_schema.py
@@ -0,0 +1,58 @@
+from typing import Literal
+from pydantic import BaseModel, Field, model_validator
+
+# ── Branch별 허용 기법 ────────────────────────────────────
+#
+# [LOW branch] 고해상도 feature map → 국소 위조 흔적 포착에 특화
+#   - hirescam   : activation * gradient element-wise 곱, 업샘플링 없이 원본 해상도 유지
+#                  → 픽셀 단위 경계가 가장 선명 (aug_smooth: O, eigen_smooth: X)
+#   - gradcamelementwise : 각 위치별 gradient 부호 고려 후 ReLU → 노이즈 억제, 국소 영역 정밀 (aug_smooth: O, eigen_smooth: O)
+#   - layercam   : positive gradient로 activation 공간 가중합산
+#                  → 얕은 레이어에서도 안정적, 텍스처/경계 검출에 강함 (aug_smooth: O, eigen_smooth: O)
+#
+# [HIGH branch] 저해상도 semantic feature map → 전체 구조 파악에 특화
+#   - eigengradcam : activation * gradient 의 첫 번째 주성분
+#                    → 클래스 판별력 유지하면서 GradCAM보다 부드럽고 EigenCAM보다 정확 (aug_smooth: O, eigen_smooth: X)
+#   - gradcamplusplus   : 2차 gradient 활용 → 다중 위조 영역 동시 포착, 작은 영역에 강함 (aug_smooth: O, eigen_smooth: X)
+#   - xgradcam     : gradient를 normalized activation으로 스케일링
+#                    → attribution 합이 logit 변화량과 수학적으로 일치, 충실도 높음 (aug_smooth: O, eigen_smooth: O)
+
+_LOW_ALLOWED  = {"hirescam", "gradcamelementwise", "layercam"}
+_HIGH_ALLOWED = {"eigengradcam", "gradcamplusplus", "xgradcam"}
+
+class ExplainRequest(BaseModel):
+    branch_level: Literal["low","high"] = Field("high", 
+                                                description="브랜치 레벨. low: 국소 위조 흔적 포착, high: 전역적 위조 흔적 포착")
+    explainer_type: str = Field("eigengradcam",
+                                description = (
+                                    "선택 가능한 XAI 기법. "
+                                    "low: [hirescam, gradcamelementwise, layercam], "
+                                    "high: [eigengradcam, gradcamplusplus, xgradcam]"
+                                ))
+    display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", 
+                                                                description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
+    category: Literal[0, 1] = Field(1, 
+                                    description="판단 클래스 인덱스 (0: Real / 1: Fake)")
+    overlay_ratio: float = Field(
+        0.5, ge=0.0, le=1.0,
+        description = "Heatmap 투명도 (0: 히트맵만 강조, 1: 원본 이미지 위주)"
+    )
+    aug_smooth: bool = Field(
+        False,
+        description = "TTA(Test Time Augmentation) 적용 여부. 히트맵을 더 객체 중심적으로 정렬"
+    )
+    eigen_smooth: bool = Field(
+        False,
+        description = "PCA 기반 노이즈 제거. 지배적인 패턴만 남김"
+    )
+    
+    @model_validator(mode="after")
+    def validate_explainer_for_branch(self) -> "ExplainRequest":
+        allowed = {"low": _LOW_ALLOWED, "high": _HIGH_ALLOWED}
+        if self.explainer_type not in allowed[self.branch_level]:
+            raise ValueError(
+                f"branch_level='{self.branch_level}'에서 허용된 기법: "
+                f"{sorted(allowed[self.branch_level])} "
+                f"(입력값: '{self.explainer_type}')"
+            )
+        return self
\ No newline at end of file

From 086af0025600b310ef4d266a506cad15eef3c459 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 16:30:03 +0900
Subject: [PATCH 31/48] =?UTF-8?q?=E2=99=BB=EF=B8=8F=20Refactor=20BinaryCla?=
 =?UTF-8?q?ssifierOutputSigmoidTarget=20with=20removing=20category?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/cam_evaluator.py       |  2 +-
 explainability/utils/model_targets.py | 10 +---------
 2 files changed, 2 insertions(+), 10 deletions(-)

diff --git a/explainability/cam_evaluator.py b/explainability/cam_evaluator.py
index c325811..d215def 100644
--- a/explainability/cam_evaluator.py
+++ b/explainability/cam_evaluator.py
@@ -32,7 +32,7 @@ def evaluate(self, img_path: str) -> float:
         np.random.seed(self.seed)
         
         grayscale_cam, tensor, _ = self.cam_explainer._build_grayscale_cam(img_path)
-        targets = [BinaryClassifierOutputSigmoidTarget(self.cam_explainer.category)]
+        targets = [BinaryClassifierOutputSigmoidTarget()]
         
         return self.cam_metric(
             input_tensor = tensor, # (1,C,H,W)
diff --git a/explainability/utils/model_targets.py b/explainability/utils/model_targets.py
index 5ec0870..ba63a85 100644
--- a/explainability/utils/model_targets.py
+++ b/explainability/utils/model_targets.py
@@ -12,13 +12,5 @@ def __call__(self, model_output):
         return model_output * sign
     
 class BinaryClassifierOutputSigmoidTarget:
-    def __init__(self, category):
-        self.category = category
-
     def __call__(self, model_output):
-        if self.category == 1:
-            sign = 1
-        else:
-            sign = -1
-        return torch.sigmoid(model_output) * sign
-    
\ No newline at end of file
+        return torch.sigmoid(model_output)
\ No newline at end of file

From b6bfd16659fa790b52100a18c4d5373ed7134a97 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Fri, 15 May 2026 18:14:16 +0900
Subject: [PATCH 32/48] =?UTF-8?q?=E2=9C=A8=20Create=20PNGStreamingResponse?=
 =?UTF-8?q?(media=5Ftype=3D"image/png")?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/utils/responses.py | 4 ++++
 1 file changed, 4 insertions(+)
 create mode 100644 App/utils/responses.py

diff --git a/App/utils/responses.py b/App/utils/responses.py
new file mode 100644
index 0000000..4647454
--- /dev/null
+++ b/App/utils/responses.py
@@ -0,0 +1,4 @@
+from fastapi.responses import StreamingResponse
+
+class PNGStreamingResponse(StreamingResponse):
+    media_type = "image/png"
\ No newline at end of file

From dd81bba64050582ff677f36f11c219e52448751a Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 16 May 2026 13:38:18 +0900
Subject: [PATCH 33/48] =?UTF-8?q?=E2=9C=A8=20Create=20explain=5Fsvc.py?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py         | 13 +++---
 App/routes/inference.py       |  9 +++-
 App/schemas/explain_schema.py | 31 ++++----------
 App/services/explain_svc.py   | 79 +++++++++++++++++++++++++++++++++++
 App/services/inference_svc.py | 78 +++++++++++++++++-----------------
 5 files changed, 140 insertions(+), 70 deletions(-)
 create mode 100644 App/services/explain_svc.py

diff --git a/App/routes/explain.py b/App/routes/explain.py
index 3d35db9..ea1b1e9 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -1,23 +1,22 @@
 import io
 from fastapi import APIRouter, status, Depends
-from fastapi.responses import StreamingResponse
+from utils.responses import PNGStreamingResponse
 from sqlalchemy import Connection
 from db.database import context_get_conn
 from schemas.explain_schema import ExplainRequest
-from services import session_svc # explain_svc
+from services import session_svc, explain_svc
 
 router = APIRouter(prefix="/explain", tags=["explain"])
 
 @router.post("/image/{image_id}", status_code=status.HTTP_200_OK, 
-             response_class=StreamingResponse , summary="딥페이크 이미지 위조 흔적 시각화(CAM)")
+             response_class=PNGStreamingResponse , summary="딥페이크 이미지 위조 흔적 시각화(CAM)")
 async def explain_image(
     image_id: int,
     explain_req: ExplainRequest,
     conn: Connection = Depends(context_get_conn),
-    session_user = Depends(session_svc.get_session_user_opt), # 비회원 허용
+    session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수 
 ):
-    return None
-#   img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
-#   return StreamingResponse(io.BytesIO(img_bytes), media_type="image/png")
+      img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
+      return PNGStreamingResponse(io.BytesIO(img_bytes))
     
     
\ No newline at end of file
diff --git a/App/routes/inference.py b/App/routes/inference.py
index b538ebb..9d04a32 100644
--- a/App/routes/inference.py
+++ b/App/routes/inference.py
@@ -127,8 +127,15 @@ async def get_video_result(
 async def get_video_detail(
                            video_id: int,
                            conn: Connection = Depends(context_get_conn),
-                           session_user = Depends(session_svc.get_session_user_prt) # 로그인 유저만 가능
+                           session_user = Depends(session_svc.get_session_user_prt) # 로그인 필수
                            ):
+    video_data = await video_svc.get_video_result(conn, video_id)
+    if video_data.status != "SUCCESS":
+        raise HTTPException(
+            status_code = status.HTTP_400_BAD_REQUEST,
+            detail = f"비디오 상세 분석은 추론이 성공한 비디오만 가능합니다"
+        )
+    
     meta = await video_svc.get_video_meta_result(conn, video_id)
     frames = await video_svc.get_video_frame_results(conn, video_id)
 
diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
index 8ad5823..58c8b86 100644
--- a/App/schemas/explain_schema.py
+++ b/App/schemas/explain_schema.py
@@ -21,30 +21,13 @@
 _HIGH_ALLOWED = {"eigengradcam", "gradcamplusplus", "xgradcam"}
 
 class ExplainRequest(BaseModel):
-    branch_level: Literal["low","high"] = Field("high", 
-                                                description="브랜치 레벨. low: 국소 위조 흔적 포착, high: 전역적 위조 흔적 포착")
-    explainer_type: str = Field("eigengradcam",
-                                description = (
-                                    "선택 가능한 XAI 기법. "
-                                    "low: [hirescam, gradcamelementwise, layercam], "
-                                    "high: [eigengradcam, gradcamplusplus, xgradcam]"
-                                ))
-    display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", 
-                                                                description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
-    category: Literal[0, 1] = Field(1, 
-                                    description="판단 클래스 인덱스 (0: Real / 1: Fake)")
-    overlay_ratio: float = Field(
-        0.5, ge=0.0, le=1.0,
-        description = "Heatmap 투명도 (0: 히트맵만 강조, 1: 원본 이미지 위주)"
-    )
-    aug_smooth: bool = Field(
-        False,
-        description = "TTA(Test Time Augmentation) 적용 여부. 히트맵을 더 객체 중심적으로 정렬"
-    )
-    eigen_smooth: bool = Field(
-        False,
-        description = "PCA 기반 노이즈 제거. 지배적인 패턴만 남김"
-    )
+    branch_level: Literal["low","high"] = Field("high", description="브랜치 레벨\nlow: 국소 위조 흔적 포착\nhigh: 전역적 위조 흔적 포착")
+    explainer_type: str = Field("eigengradcam", description = ("선택 가능한 XAI 기법. low: [hirescam, gradcamelementwise, layercam], ""high: [eigengradcam, gradcamplusplus, xgradcam]"))
+    display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
+    category: Literal[0, 1] = Field(1, description="판단 클래스 인덱스 (0: Real / 1: Fake)")
+    overlay_ratio: float = Field(0.5, ge=0.0, le=1.0, description = "Heatmap 투명도 (0: 히트맵만 강조, 1: 원본 이미지 위주)")
+    aug_smooth: bool = Field(False, description = "TTA(Test Time Augmentation) 적용 여부. 히트맵을 더 객체 중심적으로 정렬")
+    eigen_smooth: bool = Field(False, description = "PCA 기반 노이즈 제거. 지배적인 패턴만 남김")
     
     @model_validator(mode="after")
     def validate_explainer_for_branch(self) -> "ExplainRequest":
diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
new file mode 100644
index 0000000..9ad6f7d
--- /dev/null
+++ b/App/services/explain_svc.py
@@ -0,0 +1,79 @@
+import cv2
+import asyncio
+from functools import partial
+from fastapi import status
+from fastapi.exceptions import HTTPException
+from sqlalchemy import Connection
+from schemas.explain_schema import ExplainRequest
+from services import image_svc, inference_svc
+from explainability import (
+    HiResCAMExplainer, GradCAMElementWiseExplainer, LayerCAMExplainer,
+    EigenGradCAMExplainer, GradCAMPlusPlusExplainer, XGradCAMExplainer,
+)
+
+EXPLAINER_REGISTRY = {
+    # LOW branch: 국소 위조 흔적 포착 
+    "hirescam":   HiResCAMExplainer,
+    "gradcamelementwise": GradCAMElementWiseExplainer,
+    "layercam":   LayerCAMExplainer,
+    # HIGH branch: 전역 위조 흔적 포착
+    "eigengradcam": EigenGradCAMExplainer,
+    "gradcamplusplus":   GradCAMPlusPlusExplainer,
+    "xgradcam":     XGradCAMExplainer,
+}
+
+_explainer_cache: dict = {}
+
+def _get_or_create_explainer(model_name: str, dataset: str, explain_req: ExplainRequest):
+    cache_key = (model_name, dataset,
+                 explain_req.explainer_type, explain_req.branch_level,
+                 explain_req.category, explain_req.aug_smooth, explain_req.eigen_smooth)
+    
+    if cache_key not in _explainer_cache:
+        _explainer_cache[cache_key] = EXPLAINER_REGISTRY[explain_req.explainer_type](
+            model_name = model_name,
+            dataset = dataset,
+            branch_level = explain_req.branch_level,
+            category = explain_req.category,
+            aug_smooth = explain_req.aug_smooth,
+            eigen_smooth = explain_req.eigen_smooth,
+        )
+    return _explainer_cache[cache_key]
+        
+def _run_visualization(explainer, image_path: str, explain_req: ExplainRequest) -> bytes:
+    if explain_req.display_type == "heatmap":
+        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio)
+    elif explain_req.display_type == "contour":
+        img = explainer.display_contour_on_image(image_path)
+    else:  
+        img = explainer.display_bbox_on_image(image_path)
+ 
+    img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
+    _, buf = cv2.imencode(".png", img_bgr)
+    return buf.tobytes()
+
+async def explain_image(conn: Connection, image_id: int, explain_req: ExplainRequest):
+    result = await image_svc.get_image_result(conn, image_id)
+    
+    if result.status != "SUCCESS":
+        raise HTTPException(
+            status_code = status.HTTP_400_BAD_REQUEST,
+            detail = "이미지 위조 흔적 분석은 추론이 성공한 이미지만 가능합니다"
+        )
+    
+    model_name, dataset = inference_svc.MODEL_CONFIG[result.version_type][result.model_type][result.domain_type]
+    
+    explainer = _get_or_create_explainer(model_name, dataset, explain_req)
+    image_path = "." + result.image_loc
+    
+    try:
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(
+            None, 
+            partial(_run_visualization, explainer, image_path, explain_req)
+        )
+    except Exception as e:
+        raise HTTPException(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            detail="알수없는 이유로 문제가 발생하였습니다."
+            )
diff --git a/App/services/inference_svc.py b/App/services/inference_svc.py
index 02a8f60..1a1d440 100644
--- a/App/services/inference_svc.py
+++ b/App/services/inference_svc.py
@@ -1,17 +1,15 @@
 import asyncio
-from inference.image_predictor_prt import ImagePredictor
-from inference.video_predictor_prt import VideoPredictor
-from inference.utils import PredictorError
+from typing import Literal
 from fastapi import status, Depends
 from fastapi.exceptions import HTTPException
 from sqlalchemy import text, Connection
 from sqlalchemy.exc import SQLAlchemyError
 from services import image_svc, video_svc
-from db.database import context_get_conn, celery_db_conn, engine
+from db.database import celery_db_conn
 from celery_app import celery_app
-
-image_cache = {}
-video_cache = {}
+from inference.image_predictor_prt import ImagePredictor
+from inference.video_predictor_prt import VideoPredictor
+from inference.utils import PredictorError
 
 # 사용자 모델 설정 변수명 
 MODEL_CONFIG = {
@@ -31,29 +29,46 @@
     }
 }
 
-# 사용자 이미지 딥페이크 여부 판단 로직
-def predict_image(image_loc: str, version_type: str, model_type: str, domain_type: str):
-    
-    # image_loc: static 폴더 내 저장된 사용자 이미지 경로
-    # version_type: v1, v2
-    # model_type: fast, pro
-    # domain_type: 서양인, 동양인
-    
-    model_name, dataset = MODEL_CONFIG[version_type][model_type][domain_type]
+_image_cache: dict = {}
+_video_cache: dict = {}
 
-    # 캐시 확인 및 모델 초기화
+ # 캐시 확인 및 모델 초기화
+def _get_or_create_predictor(
+    model_name: str,
+    dataset: str,
+    predictor_mode: Literal["image", "video"],
+) -> ImagePredictor | VideoPredictor:
     cache_key = (model_name, dataset)
-    if cache_key not in image_cache:
-        image_cache[cache_key] = ImagePredictor(
+ 
+    if predictor_mode == "image":
+        if cache_key not in _image_cache:
+            _image_cache[cache_key] = ImagePredictor(
+                margin_ratio=0.2,
+                conf_thres=0.5,
+                model_name=model_name,
+                dataset=dataset,
+            )
+        return _image_cache[cache_key]
+ 
+    if cache_key not in _video_cache:
+        _video_cache[cache_key] = VideoPredictor(
             margin_ratio=0.2,
             conf_thres=0.5,
             model_name=model_name,
-            dataset=dataset
+            dataset=dataset,
         )
-    predictor = image_cache[cache_key]
+    return _video_cache[cache_key]
+
+# 사용자 이미지 딥페이크 여부 판단 로직
+def predict_image(image_loc: str, version_type: str, model_type: str, domain_type: str):
+    
+    model_name, dataset = MODEL_CONFIG[version_type][model_type][domain_type]
+
+    predictor = _get_or_create_predictor(model_name, dataset, "image")
+    image_path = "." + image_loc
     
     try:
-        analysis = predictor.predict_img("." + image_loc)
+        analysis = predictor.predict_img(image_path)
         
         print(f"딥페이크 확률 값: {analysis['prob']}, 얼굴 신뢰도: {analysis['face_conf']}, 얼굴 비율: {analysis['face_ratio']}, 얼굴 밝기: {analysis['face_brightness']}")
     
@@ -130,26 +145,13 @@ async def run_inference():
 # 사용자 비디오 딥페이크 여부 판단 로직
 def predict_video(video_loc: str, version_type: str, model_type: str, domain_type: str):
     
-    # video_loc: static 폴더 내 저장된 사용자 비디오 경로
-    # version_type: v1, v2
-    # model_type: fast, pro
-    # domain_type: 서양인, 동양인
-    
     model_name, dataset = MODEL_CONFIG[version_type][model_type][domain_type]
 
-    # 캐시 확인 및 모델 초기화
-    cache_key = (model_name, dataset)
-    if cache_key not in video_cache:
-        video_cache[cache_key] = VideoPredictor(
-            margin_ratio=0.2,
-            conf_thres=0.5,
-            model_name=model_name,
-            dataset=dataset
-        )
-    predictor = video_cache[cache_key]
+    predictor = _get_or_create_predictor(model_name, dataset, "video")
+    video_path = "." + video_loc
     
     try:
-        analysis = predictor.predict_video("." + video_loc)
+        analysis = predictor.predict_video(video_path)
         
         print(f"딥페이크 확률 값: {analysis['prob']}, 얼굴 신뢰도: {analysis['face_conf']}, 얼굴 비율: {analysis['face_ratio']}, 얼굴 밝기: {analysis['face_brightness']}")
     

From bc55e95038f696b5048eaa2301ca7b440f0a77a8 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 16 May 2026 14:04:54 +0900
Subject: [PATCH 34/48] =?UTF-8?q?=F0=9F=8E=A8=20Add=20threshold=20argument?=
 =?UTF-8?q?=20in=20ExplainRequest?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py                     | 15 ++++++++++++---
 App/schemas/explain_schema.py             |  1 +
 App/services/explain_svc.py               | 16 +++++-----------
 explainability/explainer/cam_explainer.py |  7 ++++++-
 4 files changed, 24 insertions(+), 15 deletions(-)

diff --git a/App/routes/explain.py b/App/routes/explain.py
index ea1b1e9..d15a71d 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -1,10 +1,11 @@
 import io
 from fastapi import APIRouter, status, Depends
+from fastapi.exceptions import HTTPException
 from utils.responses import PNGStreamingResponse
 from sqlalchemy import Connection
 from db.database import context_get_conn
 from schemas.explain_schema import ExplainRequest
-from services import session_svc, explain_svc
+from services import session_svc, explain_svc, image_svc
 
 router = APIRouter(prefix="/explain", tags=["explain"])
 
@@ -16,7 +17,15 @@ async def explain_image(
     conn: Connection = Depends(context_get_conn),
     session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수 
 ):
-      img_bytes = await explain_svc.explain_image(conn, image_id, explain_req)
-      return PNGStreamingResponse(io.BytesIO(img_bytes))
+    result = await image_svc.get_image_result(conn, image_id)
+    
+    if result.status != "SUCCESS":
+        raise HTTPException(
+            status_code = status.HTTP_400_BAD_REQUEST,
+            detail = "이미지 위조 흔적 분석은 추론이 성공한 이미지만 가능합니다"
+        )
+    
+    img_bytes = await explain_svc.explain_image(conn, result, explain_req)
+    return PNGStreamingResponse(io.BytesIO(img_bytes))
     
     
\ No newline at end of file
diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
index 58c8b86..b730a51 100644
--- a/App/schemas/explain_schema.py
+++ b/App/schemas/explain_schema.py
@@ -26,6 +26,7 @@ class ExplainRequest(BaseModel):
     display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
     category: Literal[0, 1] = Field(1, description="판단 클래스 인덱스 (0: Real / 1: Fake)")
     overlay_ratio: float = Field(0.5, ge=0.0, le=1.0, description = "Heatmap 투명도 (0: 히트맵만 강조, 1: 원본 이미지 위주)")
+    threshold: float = Field(0.5, ge=0.5, le=1.0, description="contour/bbox 이진화 임계값 (0.0~1.0)")
     aug_smooth: bool = Field(False, description = "TTA(Test Time Augmentation) 적용 여부. 히트맵을 더 객체 중심적으로 정렬")
     eigen_smooth: bool = Field(False, description = "PCA 기반 노이즈 제거. 지배적인 패턴만 남김")
     
diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
index 9ad6f7d..6bd0e95 100644
--- a/App/services/explain_svc.py
+++ b/App/services/explain_svc.py
@@ -5,6 +5,7 @@
 from fastapi.exceptions import HTTPException
 from sqlalchemy import Connection
 from schemas.explain_schema import ExplainRequest
+from schemas.image_schema import ImageData_indi
 from services import image_svc, inference_svc
 from explainability import (
     HiResCAMExplainer, GradCAMElementWiseExplainer, LayerCAMExplainer,
@@ -42,24 +43,17 @@ def _get_or_create_explainer(model_name: str, dataset: str, explain_req: Explain
         
 def _run_visualization(explainer, image_path: str, explain_req: ExplainRequest) -> bytes:
     if explain_req.display_type == "heatmap":
-        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio)
+        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio, threshold=explain_req.threshold)
     elif explain_req.display_type == "contour":
-        img = explainer.display_contour_on_image(image_path)
+        img = explainer.display_contour_on_image(image_path, threshold=explain_req.threshold)
     else:  
-        img = explainer.display_bbox_on_image(image_path)
+        img = explainer.display_bbox_on_image(image_path, threshold=explain_req.threshold)
  
     img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
     _, buf = cv2.imencode(".png", img_bgr)
     return buf.tobytes()
 
-async def explain_image(conn: Connection, image_id: int, explain_req: ExplainRequest):
-    result = await image_svc.get_image_result(conn, image_id)
-    
-    if result.status != "SUCCESS":
-        raise HTTPException(
-            status_code = status.HTTP_400_BAD_REQUEST,
-            detail = "이미지 위조 흔적 분석은 추론이 성공한 이미지만 가능합니다"
-        )
+async def explain_image(conn: Connection, result: ImageData_indi, explain_req: ExplainRequest):
     
     model_name, dataset = inference_svc.MODEL_CONFIG[result.version_type][result.model_type][result.domain_type]
     
diff --git a/explainability/explainer/cam_explainer.py b/explainability/explainer/cam_explainer.py
index 5764ef3..36ea75b 100644
--- a/explainability/explainer/cam_explainer.py
+++ b/explainability/explainer/cam_explainer.py
@@ -78,9 +78,14 @@ def _get_binary_mask(self, cam: np.ndarray, threshold: float | str):
     def display_heatmap_on_image(self, 
                                  img_path: str, 
                                  colormap: int = cv2.COLORMAP_JET, 
-                                 image_weight: float = 0.5):
+                                 image_weight: float = 0.5,
+                                 threshold: float | str = "auto"
+                                 ):
         cam_recovered, face = self._prepare_cam(img_path)
         
+        binary_mask = self._get_binary_mask(cam_recovered, threshold)
+        cam_recovered = np.where(binary_mask > 0, cam_recovered, 0.0)
+                
         heatmap = show_cam_on_image(
             np.float32(face) / 255.0,
             cam_recovered,

From 1b12a809ea98cc1d45c40853d56317d07c907d06 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 16 May 2026 16:16:47 +0900
Subject: [PATCH 35/48] =?UTF-8?q?=E2=9C=A8=20Add=20DeepFake=20AI=20Explain?=
 =?UTF-8?q?ability?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../ms_eff_vit_b0_high_eigengradcam.JPG       | Bin 0 -> 9858 bytes
 .../ms_eff_vit_b0_high_gradcamplusplus.JPG    | Bin 0 -> 8904 bytes
 .../ms_eff_vit_b0_high_xgradcam.JPG           | Bin 0 -> 10401 bytes
 .../ms_eff_vit_b0_low_gradcamelementwise.JPG  | Bin 0 -> 12208 bytes
 .../ms_eff_vit_b0_low_hirescam.JPG            | Bin 0 -> 11089 bytes
 .../ms_eff_vit_b0_low_layercam.JPG            | Bin 0 -> 11790 bytes
 .../ms_eff_vit_b5_high_eigengradcam.JPG       | Bin 0 -> 10367 bytes
 .../ms_eff_vit_b5_high_gradcamplusplus.JPG    | Bin 0 -> 11970 bytes
 .../ms_eff_vit_b5_high_xgradcam.JPG           | Bin 0 -> 11378 bytes
 .../ms_eff_vit_b5_low_gradcamelementwise.JPG  | Bin 0 -> 11845 bytes
 .../ms_eff_vit_b5_low_hirescam.JPG            | Bin 0 -> 14089 bytes
 .../ms_eff_vit_b5_low_layercam.JPG            | Bin 0 -> 13648 bytes
 ...54\235\270_\352\260\200\354\247\234_1.JPG" | Bin 0 -> 14630 bytes
 README.md                                     |  49 ++++++++++++++++--
 14 files changed, 45 insertions(+), 4 deletions(-)
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_high_eigengradcam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_high_gradcamplusplus.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_high_xgradcam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_low_gradcamelementwise.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_low_hirescam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b0_low_layercam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_high_eigengradcam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_high_gradcamplusplus.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_high_xgradcam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_low_gradcamelementwise.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_low_hirescam.JPG
 create mode 100644 Images/deepfake-xai/ms_eff_vit_b5_low_layercam.JPG
 create mode 100644 "Images/deepfake-xai/\354\204\234\354\226\221\354\235\270_\352\260\200\354\247\234_1.JPG"

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_high_eigengradcam.JPG b/Images/deepfake-xai/ms_eff_vit_b0_high_eigengradcam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..cee217c01a14a182b8ab2dccdd7b20cfa6270b89
GIT binary patch
literal 9858
zcmeHrcUV(dv~TDTER+N>^dgXi7Dy0~DkO#sAS9thER+BVy{O2b3|N3j?*v3TBq4N^
zV!;B45Tyh}R7Mf0>Ih=N0yn7Rlz0EQ@4frI@BML4a+1CF*=O&)eyi=ZHYPTn0wi!4
zI}Cu2j}H*d`v5kc@lT?oDSiNeg988t002Y*f_!@b{JfSB0KkV!qyPkYZ9d-DsrbU?
zKL7{-@ZVXK2Z-?cw(w{R-VKmn<o#?u{htfq5=oBW!61|&@C`>053iC8!JqBstjuq1
z0TUAdAS>&y_FtItWo^RvKbZcdOa}S}MyC2GQ-qN+0%eLcF*Ps&{AYatfD0fJKn83I
zVB?wKj)17BFjGCfkVsuG-%xL|u1{#NUbI)39zs`N4`5~)9p>c|NRCqWCi_!D%r*b$
zx~{2A@io_ULpkU>gxQb-DA<??G9kv1=o1s@W8$l6X|WY(7Ht|G92QKD@=}ft4ho4h
zjW$>PHoGaWz1gj&YGD@P>t~9$#rzn-qs&!*?B)3J<GRNUbweZk^$;c|CVKh?dIknM
zyb(H)CqkmUqIE(d)qYH1OOEu3poB$HLPL}{C-m|TJsM@Mdh{s8*VNC;59RIai_-D-
zGW69!Abh=aynGE2I=+U6zPnKfV*@{gx9X4OeSLne9(FV$=-b-9K6>OJaxgh0Dw2l=
zLQhri`=<Z9*5pC-bJ714ejbm!F{U;VWUnZ)E$_D26sV3qN(W&?{8PxfCMNo3djA?{
zrnd>k-!}T+yYv4NS{7!$KBhjKdk>BHzC3{({(mR!Ad1DN>-f$ckNDo;P0Q0cF)=YS
zy}z&cze4b>ho^G9_kJoU?@i_Z2iO14i~MEe<Pe@^ALp6y#v_0YKtxDLSV&MrSV&k@
zL_`!QDbB+f2;MHSMN$?F*(nQ_-65x_sUoMaOJRqss;=rTEht<Y4pG)K)Posl!r-vY
zN%%xWMS)^KkhnMqCNC=w``^Acx&e}+eD-{@0(`0fen~z7NxqHS03}{172<{H%_#lj
z;^P+(6cQE@6$6U%2DC{4`1u3`_yq-ogamo<neQa;IY3ZSNJ<4^EiCQqC88P*GKkMC
z6II*WbOVeZexq*a9g!dgl-VJ>6S7M~Qws`1qKtMMo0!_zqA~lhb~pmj#nsK-gXH5&
z_VW**P$Q#`9y=a=BIfj&v*!|%l4)7lIk|cH1%>qTipr|$npy_4xuv!3(&hG!&YQPx
z_w@GN>AyQN${ibjF!6Bm+4S?77qfHo3yaGuZ&%;F|L}3`(<Wbh0D<pp{p9Rld`a^7
z;ujPY5ER+ui;w>}?-r006jDJ5OIbUMc!f)=8pMl&_GXqf-4IhV#J>T1M+^gH)RE72
zEpM{+jkEt6V+sF9oc(0%FTN%LTLk!c!V{1LSOWT&4}VD`J@^`Te*;hfX@>jv_^iWq
z0@rU!Zvf_d>r&im6f(7NPKH*Xcb;^zyNd(I?MQk+ZFxE>%o{b^k_}bGX1tz@C%L{w
z%K=WHGD5anJEo?^ymdFC_>`4=(#lbsWn8L$by1h8EgTdY0vmKvKN8QeA**!F92@A`
zUWJt`x<+{>aw^9VY@C*HdnL@E4=nK8b`cbOEzROsZTq1sLjnl8jic@!f+5{Tf*W*V
z36CNcs{5?ag_D2bdrS;gen~C56<z|9k!~OfA2KN6CWM>{fn^P7p<oHOG}%I`^U7&^
zM&0$_F)KJWnZ>}XTW+HH`jEZ78q?5`{V#fLmI-LcS}uV^mh_57xfupTNx{=P7<b?H
zc##Gd;jZGP-TWum(}TX}<tX7eEfmwx+ZheU=7Gwya-4+o6D?;bBiEPSD}CMKyzlRX
zXLRat@!)=N+?^NfxQV@SUuDO#Y2!cOKBKt-7*6#-aWClrg&d@2xKu+zU(c3dY2`_%
z!T@CN>f@OQW~oy0HDE8<EiO)!TiVfeg~#-gKMS~)ZFxNQOSDh20QT`GVtTB|fD+E?
z@aMg@?IF~zCjuX$ndrep^--?0)k}N7V}CeI?k7dCamiNhsa_n*yQ=m3BvdOA<>>`9
z5g%l4z)BkI_UJ9e6G|6|pJQ#>zbs4~6UOx67}~p9dPt@z3fnAAWu`^u_sv<o^e@uT
zR36gtJ(PR0Fz(EH>n-Rjf_LU#Yovb(+C*LFk#Ix0uA^f1OG+`LI>P{*S+SGp7I!@;
z+pRh&KAw12F}m8CBpXH*?^hhRTA+{6%o=NcpH#n8DfyO<beoe+$ya^Nl^(Cl;6t6Q
zxHDw;qNI?0KW0p}u?OT#>&nuAT2+8ryNBOA-2l*Pa2tSziMUk+TJPV8=M3uo9r0XJ
z348;PIY8O~G@jW2d_u6a-0(3FRbmfztL+s+i?c?c<Qhk=$)01Zd+ObhhfJH%ty7hc
zC6=8-(9dFqU_cE6e@0mh_lWOXd~K=iGIt)<Nn}2J8W8~N$*^J&$Do?-1RY{MDkxlb
zZZe@RXrOjiXEFWe)dx>jR8cNj(=X>Vg^N4=#`@)l_JDSEbzDe(;xd@M0f@Qe+V_=i
zU6Xa{j|HL(V%#pWpK)fS2IfH70QePuoy`7>%pA$Rm2)p`-HSVxbW}_rb__8pXKsn3
zn?sXk@Jd?A5eMjnp&lox)-?t`G-vPb)tc}Lx_C~}SAUX6sZ+#-MM<i2voXv0bjog|
zXOYA{S=XJ6dI?!abEvQVYITUR7c*PcSwHc0s6q7XWiKyvtF%Df70IrX2S#RufOE!d
zf_|26a8-{913}X8GfKfvr?@12$G>6||2yCD-}l7@;`a5F%?6;122YplG$=eh-LwG!
z=e2DB4!`~8M9N6(Fr{VpCChgk0L%3hYJX4EDkjMV^b!vw-NiI*%Q@SHAMtMVyUF~(
zaEZj8C%{HzVKeyGeW1Pqtw~d^Rh~&HMJBgrYgAL%l{VL}ON~T%m(Rhd2t`K+;bid0
z&idcy1a&;m+oi^&1>$zObn{n>-wWt)h3@lB9s*`Og1&MyqB*pJ?-CEVC@?JUmaejH
zM3?I95C>hmXAX^xU*J790=mI^tV0Y(YTSvyoLzfBp4u315qaksy(%2E+X)UT;X>j|
z;iUHIvgmnm%19;Apf8ubQ%%rCyttNWpO?bujezYDJ9Ynf>j#u};vBPiYYRiMt<r;o
zj509#S{a3$3=%9lihMaff8#s`3Z*6^WA|(T)G+xxm@;@UdDF}`X^X$;7q>%FX1zYL
z*7t4oU%x*0&0)B+5A)RfdEEx!!Iy994C{<h9r8|bcI0o$tg{v0FKz`|yfe1l2^>Kc
zKi@4p(SdxFoOF;$m5Xi~IEs|9<2N*#TcoZcd{>OCZz!3df&mIp_ts=JtBX-C1s|tN
zaU~Y5sov)iv|=KH+s)1Q3h3Ec9zPKVYOfT|`>intKWqi4Ds{$s)559=JpgEJ0k@Lt
zsHL9V#35V|aH(OgxTtz$Blp;$iozx`lyH+pxv1m_{n^%Jj(Ui@-L=7Afv2c>*%ra?
zU0$`<7rczb3h8L#JjXI9ACIqNPU@{+SJ?mz^?*tgcxHR?mp}C^!h1I+5@AM+n{FJB
zTU!6-Oi9;ZyfJ+qyfJGVfK?w?ckt+Y2UV>aKCF<j>y{_VJvL@j?hM~BmdDCn^&ny%
zOlrj+RR!F7WlUCD8$rVAx2CEl6KJ_qFvB$)g)M{)dY4+eN_bc0C68BurML-|We;nB
zWMhKtVp#R`1-iXZc~(A34X9Pg(O3Azp^5`gLH1GcCu->l7!!_LB*EnAn8!9)5}A@t
zy4@9I=z68bq)x{w!@<+u_wdlY>9NM&f2H-jS$@*Ba3IBmb9A3i)`1ibU>NODSoLg_
zXET`GpUSX*+Xmpya8=xc8Ou+y<sAsQ3TVMMD+*NlVMT-dZDuY!jXB4&wjYLe<-z`k
zm|+kQA7<!qKc?*0XbFi#dQVZ=gB@yuK!zN-nc7fru*EI7e2ThNN^erngv~d?deqlM
zjUDmJl!p;(Bk?LdYI`RYsdrbI;Uy<n*vI%LY-5|&g43>F=tbL}FWjsU0X^kz$#8af
z4CRTA0~qnkMka=@5^Q4V_?sE8O|y~lEjWbzUPvnQ5NtQ7zeC^yGopn0`~#EVkD^H3
z%_?a-U)vyjsd5gLU8@3gz+><lZ-!o6%e{89@ry;ba9{X=bDe5U11e|mK5PrxP*46B
zQ_JaEUW8rQjIfSxK72{}boftylfCJ0&QD)llSie0+5lYK0ANsL*@^n6#$eWT`rmsA
z2~Il#qfo<fQS%8MxjN%RZdd)Vg#*$Dj(<84I;Dxb+djENpdsc)OB8LG@nq!EOD{PC
zv_Gj<+mI7t)kTQM<K>O3<cDrDJqJ`%Fj^>dvIe$KhvORT;8w<5@<-hwr-Sw0yD`dy
z`s$ncX6vo%(ZkHd_xMj^Cy+NSmI#d<X><jW@N3x6^`uoVPnF9hCi?Z1>j9l-KQ)1n
zD77J}r@NI$zf8^qq3p!uSjVPbVuh&?y#lV3U(PjHL00AOf8afQ1}g@cpKG6B#rbhI
zyId%Zu-3~(*8OG59TjsvSJe)y`^-{bgrbtwLbK_2FA_Qm-tpYOVZ<K)@iNc1E{|>_
zB55<F9`)M?d)stYGj2uhenzjypBPvPT=V&jlIZSX@c}fydgs@Xfy%y?7fa}b=g%|!
zJWqQ}G9D^Y4+Quo;A^VTb@fL;K-&@}lk*m1UG(lZ%eUR%vxfqsTj&SQqctdoq=FX0
zt>0G+7vkQnZC^j9RsAt)1JJX-is!NZQsqKLD9?1eHcfZoo9P;@KrY2P#ZA}$Y29f%
z^yp8%XdX0}#F3=rs^J%|A8e72d0->;r7gX7Z|utlCc|EN(7c`OiTs&5mh3up!uf1z
z{2AgFfVrV-0Z~d<3uR-*Y1?wS%tWIo?6U-YwxL1(E)i1s3u<oM>T@;Y8cS`ra8o9k
z?z98tRhDJM_mOGq+5Vsfe5caEI$2y9EOxTMhC6zebs5$HLW<EZt*G#~Dd!|0++|9k
zlaC>fE)#lekeEumOfPkMmFcv_yY;z!yhjTW3AD3tVX49r7W82lf6x7Gff}<*5yg<M
zC~;{0DawAj(@~afGnD+;k-8mu`Zwt4E!AUY<5Agm@;8o2XOq;#3~gTB7Y@hUG+-~p
zxI_nYMlwWDxB!jI8Fj3cudtMR*Y3MY*G@hjr+dmM1m4%yh_3&Q&>Ps+>=%7w$&Rju
z)em_*rhqpepIdM%!|y{We43RqXNvy{LaJSczdBIf6jk*I%5H?a7mqx~;f@+zM)>6?
z#?1tbYyiI7*uH;Io&~AHIfI9#FHGr=Csut6I?OXZX@S4oPEMCy+Bg2q7hAh_#;8@7
zbdRiQky`xQrG2McHPQ$7!k82M+u&kwh4AAtcaF0GV|c9;&tvRgx0cI)(4@a$dpIYe
z>Z6=Xm-gq)C|H$EbEs$1_KKHUl3bgU-=;9k%~^Eg#n5S2)2e*#VeyXmaQ|5-f1~cI
zt5jYe`0=$1$|m;uU#Vg*SAd;Q9|W5Y(LZ?H!lz<mu`@xdV`pzQvgqeZ^091*8;ZGA
zyW$V!DrKts&>k8p5=z!ND!tF{vqL=|-r9P}oj=?mK)(1@O{Ib4X}T^@<7iLX+CWbq
zisi+pC9n-`3irFO+R!62R-28)Qv2G@23U3UxC+oOhjLy0lA5Jy>|+;ISWjciS)SEs
zm4XP;Yl8dXhar06jn7fzYksW$&>P(>+g4vs_bSdcu96X4@u13(*Yi2^z;fTH(0sNH
z-2QPehhz{AT_EAh^Y3j%*HU`@VQwaD<FbMX<M%8^h<xl<%3JY_+2AW?wBtVX`MQ^}
z1s$wm(%D0~db#mV52;4bGmq%(V^tN#OKw%N!mQoM=e3pi#+14<jqH#;L@`+frq@Bq
zecG2LMk!qHRs{0jr;Wp#0q6d|%=FZ*zAw8b`@Y8Q-^>nM&XSgq?aO&Be+C>TKlxof
zD<E}xRXrr9Nn+bpePbk8Gq<ifjUE(R@XD=B#m7E;g?2Z1ZMJn<q~Cr65apJg_R;y(
ziw%JMRq>|lxaLHS+TIhaaK*s>fcKfl5XCO>{DbXBIc`a-tVa6z*4O727;uB)NrX4a
zV&4&J%B3wAPl|q`E##hq%EIeqecg_lMLn-<Gh6J4+Es@%vGIR}yA-?1UaJEoct8E*
zqLET|3o0^o6&9W3a%jnn-cA`gmZf#xX%D{DnDr|qeT3Zt2i-lweY2fj9(omj^LG?H
z=a7K71NjMV8(MI*{hTPT=Au@GZGEHFUWl#(6;BrgNq=supjosXmn?d(7!8I!oUVu1
z*fnJ&y@94B)S%?g=s4VMDvD|+RbFX3*cD4-K)yU=FBA}ma$^X!0yH>VXB+hDUYb_v
z;P9WZ483`F*7^qp=S4QdLe9s!iy!UZHGCHRhJSNJB<b2_Iy3rLI-`l2%1a6lDwLmb
znJCJn-y^tsXSHORLI?0`lefd&Dy`TnDcf!~XC0A^oNBu`+Yg@h8E;y<Xp=W!7kdtR
zgN4|QA{f_HJbjRLpKR!8Vgf#<Sr<LS${a|0S7c=Wz-_t&(lUNHrci?cFWg~WmZ|P(
zngA2ZH(zx4Sj9YU+s>1YSe?GOZQN6h=Gi^bW=b&=S+T$}gx$xZB2!&ZAF_fbW6pKp
zxejHX+%a1ff9^z$gWtg5iSUJ~()2y7(HcAoJbhoa3Oaf$3wb)c2SHCd7c*Mv3FO|9
zu|LYlvKm7tgOV-AXERWdju~7#{qP%AdM*(`eGso8Vz7?w3LV%4%>Xu-m6sd5b+YdR
zsk}rRh9H}Ctg}e<w(Yq2>PR9ahCq2`j~XvY+;SO6!jXLG7K;BkT`K;Q44>YOA(&j0
zQToV6{lwmoG^H)FF6|Kdvor!(-Es3=;_3YaT9N?1mXXy_ckmqQHbs?X?5dgaz}hED
z7xZ+t<5$kE`g*k}Z&6fBdGGiJpzBom75kXj1;usG2QAreQSwLVN6ffcM{AdO7NDl;
z{)pb1O3bUP!Eq{FR2S_CEEjc@!6m4Bz4xRAbk-YI8*}g>BVZUCn=Q4C+B~>Bf{K}C
z4urh=K-DH?`hUiZTc~FU{9)T5xV+Q*h{s8Dc;WRo@5Z=YBbdWwEE%JGi0`il&!40&
z^p9p`sj`j7PM=4OFIznwF<;So!BWLmYac?ayz74DY<{7qS0CGOs`xqHY}g{C3X0Fy
zPW7q1`50QMEtg3QMjqQ=*pDh#nJ8cxD2Znh)^eSE38(3KLJ9BW+VrncQn<<zB~&p-
z$_zJ@Ls97W4X-EF>+!R0o3KRU4?68rz@Qj~HPR03;XAiEJ1aRL(a|@5{OxQ9y{2fn
zOr#wCeu}8K$IRdJH<`W}_Y67Sz5yusp3G}%j$HoYB5R>PzW=M`_X^gPrb8*GKsY%6
zNf?=c7wy5yT%fgEP2}zl<72()OHN)OV*7BTtrkIET^}0M5T`ro=2x1_DWhxYUfB+l
zX!{+)=SXiV-&rIeWGm=}6lR+$tv~;6KfYuu2HHKO>f~SyNr6NoQ@yd!MHmXDO3x;m
z0Wob-Utd$YxwVf8O-vi_=2Q`!Mst$w;Ax6;$sX?Fdh4Sa=b%$TQYQmRy0~&@e+v!$
zs7F~7_vYJqZ`?Ddq(wJGuVu-ACC#`EBc(o0x8)dB46#acs4_o=jS0u@2iMs@0v`aa
ztT6{aB1P&8<Xg07Q@AQEc!2|IxwHUKVR=ES-ak@?>{~VKunX`X`4qT6>B-K)o@L^1
z(S06s>xaK*lZOIWhqUj1(dhjB>yLtzB`OB~%(i@v4!s!$E7^+7CJH?_?)fD+X6RyO
zqu5N$_z{g_IR!W@6YuKRI{Wl>;c!o_@CLyB>L=Qzl&pHExlgV_T`#C39T*`;(FVn9
z!cyZ9#aShN#tzD2>dDqOur~Hk(TnWFSAV#6x+UYKYIHr6(6~z2ExHO0TKN{Y9lS%h
zO{q(P63bOsY^~eR-||{5^V*VkW;mw5BPCAa%c9YB^n4s??n{UD3&ujQ+A2~GEw`U}
zRzA{x{mEoo+IaTqq3PUU@_iVE4yidj;LUilR5T}f?it3xs?j*IW83gJ<66ne;K;Uj
zPZDZ479o8_1;kd99Rx@70}764EGkD+xMjL=C41HcDMz*#TQa=q$NYVt^3#zWTRk!#
zft6>*+CSC?WOdJ8pixVp4;g%^eYoUD{8f=I{;@iN(mR{dbr2f;%my7XP1G1D9EYJB
zIoNLpW1uE1+(U5LdEdfPNC0xtyQAva?`x2#hEHD;Hvl#u=V$lVAjpPKU%pXp8fsDg
zh8LE32V~mlfxl{3kdfukFKK^?%cQ|NRtuYtoD!=ZrdJ9qT%G4=A}jH_j}(OtFkNdh
zU`xOkg^%nbjJ98;lJ3uG$bNbL4xA#l)%8eDpO)9h2ZhO%%%1uh#@YILj>QEBA3M6i
zPHPe6Rpe=QH1^ZJ`r3mX)n*>@!~r}&3t~`t*~bBcpg&mQzGohnIkQrp$y5-Wcbn~T
zmqyRSA?`2IU(*IVx&G<U9aFDQ!f|1_b`iJlvE4_~k!cB^-dny4e`Pt<<6m_3DqQfF
zh$q&J@&kz>)(yvUQ8@Y9wZ*Z69&ZgD>OGQYD^Kgjj0FT&nWHHOx}73PhQh}C;;lUu
zVLsipg-y<p`XUOHM@${=$g0+3H+4UbQ)hM~qr_8ysF@w<=3*?Lt8TR)`NFu``Jj=B
zoKEjYYM0!y2d)q0xATMm@}+fk^+hMlm_G6xH9B$~lghIZukSSnd0tfbR@?ekTjbRo
zY_W4@B+Kbmp5K0{vRS&i(-Rndd41nsQGmAU^QKINONg{aS()Y)(haD-c%4VZ`mDKP
zS2E*?GMH)VrhaeoSBb_Or4}$(PwqBR)SVBe<Gt*h&rc`!!xyPHFQd+G={3o{y2kcW
zUNm((Ppj$Q!!BGkhFu%E2!FBz#HCT|B2K%4TRijkyw`9hz(T-~m{H#=X^J9)Q@^d&
z$leVZwsowd)Km03Q`M|j5;&uiEe{-jdnd5MfIPj^Q#><rqP-z>I~t$bSt6_^vE=ww
z2ol^rUtX*|?MQ<r_>{D60QO9EZVQZ0(X+Pqi66=7|HG+YQz2kiQ~j7g8{Cebnig=W
z^05QAHv<@G1ncVOgv!d|TxC2A>C8Uldv$sUoQY`ahkBCvFVJw~y()B~w=CXyh&!>>
zSP9M{5j^t#IL$eXxx?7mvZ?1kwVn`b6gSQIvwXf;F_+RebHBb!3XFR%@hzM8$vj}8
z{m&zh@`DPcv(n)OmB{B<>Lo~*+`c`(Em@m)jaL`>WbG;rpKQs=tn<Fz)(R`}l4W_<
zT*wdME;s{ciqr0(%dBP#RmoH*C~mf+WR5h!$;6E<+3oJ$t>|`|9+WBIGG)Nmhns6a
zyQZem!B4~@)vO-(DDE}}#Zh}ZCd4I2&EX31U}+QK#x!C2#|xaA0++aL=bkU(aqaiU
zRzYJ$Q%!;|Va`!w4@EJ+VOL|aiNloRY~KaDxzFQH9hz5Fic8*kqpl5!#2ju-C-qB7
zPZjS>;$_;i7bnrkflh|9MjfU)Ln<C)T*F+(-r{eyHK?lrJKX2ER-28vE*;c2W+FOj
zyfe0pE|sYjrl`@p7}#n_BKEGc9gSq8*l?O8VGM&CRMGIM_}-(65X)s3isie{n~8XB
ztd8y<V2aXrs$L#5Us!0Nt;ECn@DHqg<4Nn<p?^E{G!4?WD+CUm_pHTf?+ov+%_QeJ
zWHRQi9%^l(ziY?KTxP$lXmMUi7%Q#}72SEs)Hq2@bLR6{-w_MvZm?Oiv<s=D%CKG^
z65!Z+W~6MkDWap(wwutjmlg&VPD;&3iTAnEMt24X``dwCqFK(_D8Wo8yMlgN2*Qm7
zJ3EbhsVXs112k*F!@}*od$+pVa!@jLy{&-x=RNK&({vU-Emv9V5X>Z+ltUz(oUH8k
zi|wP=$8blF);1lCx(7=_GlWoR(@y8Rw?JnIwW_sQ!k}FEHkc}O3r?$)oAT&Q&Y9?2
z`xGw^ZU2(gN<1ZPx$F^eomXEISx+Is6Qw&rv~{i4j?U%rb`=q&$=y8v@Cd^)GL_mJ
z|JnpKy3{X!N@EPzFc`W^bLq;sZ`uJi<$l=%*?`?Kw~e~X(6+Dl6X>oeaY+B2DFhaS
zk|uEUxgpe~J5`5TpIcRW7%0N5opo?9oD3ap3_RFTiatj2vLCyUb5@H#8~w@|d1x%U
zo^sAqS2aEJJ}xYAlD!rsIxh~s9nEm29?W?`0-#A5Tb#--(lQ{_o!1r^cWV7Uw}oq9
ze6LbX>F714jBI_oe6|(zX89A28RUs!7WNcHtDG&R0vhS48MYUtP1C8awHz#GKQQjj
zT=@oIV(<2^vaR@!PH{ctpQPJ1%_#e{{M$Ob3Z)T)m-}V$%GG|&6XS_V>Ad29&DT>7
zyaa|m=;Qvo%@}QX3zcj<>Hw#!BS*0{+j4TYc!7Wftm%*ePZ$WwbbPvS2@Ek_qQ&<t
z@|WgZ3b#N~cA?tLC57M~ZUv?}Uj{=#!{fg-a$1EL;f~!^VNFt1mLbsTa|`yO4rtXG
zVs8bt(#~#rjVY$Wdga{Cu26%bcalZThG}iE6s~;*n3*}CrK(SP>Q~XnA#4*0mGK(z
zDnPc^Kx{_p?7cuEZ7S7}y~ryt%jxFf>r3mK922ham{ZG3<uUgrZoa3yJ$z5OZ`{Vi
F{{Z5~PAdQa

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_high_gradcamplusplus.JPG b/Images/deepfake-xai/ms_eff_vit_b0_high_gradcamplusplus.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..21c3a61a76ba5632410967bd308e1cf34cddf4ff
GIT binary patch
literal 8904
zcmeHqcU)81w{PeWDIy8dL<}|bKtc(iw8Ri<5)v?g(n}~pC<0?c1tcI4I#HBP3LsSh
zK?Mnd7^DQ0q5~+X*cEhe{OOA`^JebnzWc|0@7>S)yg%+fXXmWF&pK<Z?|1EeR?d${
zKb8SIv1k`GKuAakkSaI;KURbnP^nSj0D!wYKpy}ANCHHKYyiRnNeloG!e&PSL<O>t
z;Pkpw_45h<0sz9J8)^Uv!JM>UykF1(Y8!%|pLhSe0f>oVQ~?cQg(83O2omshxlr_P
zxub~whb#g|002cr|H%JvDpd57#{cH@pT%TmYG#2nMId1o8ZZPBVTwdp0{){p0DuTc
z1cU*8Cg8`4=<dj*q<Ex>NnE0FQ0UR%FyoM;u_mcO@g^{1Qxkygfz<e*km#@^jo`3|
zs5m>_cQ<b7YD9(F>G~ktP2J<2!Xl$E>C`Ynx`%g2dUS|&sP2LNGC<o@WNK`DY*<o|
zMrv$KTp}{n4)TX}q(J^TZ35YEOAQT2;+@fdX$VGkkiTL%cI=q(F>~Xi)Cd!pwY9a0
zshNqHnUO%lC^0QADJaz_E>ZI@1I}TIA=Ie&q^P5D8b1vN1s_dLvV$ZiM};E8gTfKP
zp`i$);2`r*BN!|+$S5e(9A*@1ZXRlhfLWP^!-64yxepEb+k1R6HRccBp&=$=F=4S`
zaY=~+Hee<YlRtz0uktCN=x?Y0A^if30v)6iH7qD8%vsR)|4gWnDZ&V5;r&08Yz&86
z*_!-2JzJBXWc-^z|7#=vPqDS%HZ%ko@-z0M)IZ%5!V>=PvK<q(|L3dYPjrI8pMt=w
zV9iNMN!gnGTaW)M1%J#4mYm?ZzZa<BNsa#p?f+*d|7`NGIKj?7CfLG176DEG2{AEo
zF;NL|F>y%=2}z**4gtqNu-r~*c_pxlvJzNnx2n4CKGi+idv+^9j3L^3Py<5)6%7+}
z6MZvXeFOcUMua3JC4o{v(2gA-eKjRD{r~Ou<2FEEQpin6QA7v=5SAAbkr(<g4A?6u
zrDB2-{j*B{)r5paM8(7<B&C2m1Pa|d0m4EeBEq5~Vq&6#`Yd!pa33HlFSctR%t2fM
z7bF2m0GVabYa}%t+Xun;srOpu!PGNSK*ilk$|~AAx_VH3IKslx${Ok9j6xs8xL^t1
zM3Rs1A#zA)Sa?Ka6eTe!IptVtTKd^@nOWKAb7)1yC8cHM6_t$Iy84F3CMJvB(b?5~
z?Rrmd-_V_5?#SKIv3t`q{MosOkLDLvR-ZmwTYvuI<;I7PpEkEXfB9|u>rY%l0Fggo
z{f+G3;gT2N5*8H|5taCfOGx;bpo_?hitU4m?{dIN1SKdy%rYcFj`W)LK`Bjh{CjXP
zbqc7c1z*wr@Dth}$o_M{&io%C`y1FlxE27?B0_@XiO2&E0M_)Hf{O-lQdsdw;mf%U
zi6ERM;_w+e&U7tJlkOufpVr&}DbmSsiAx2uC_#9UK*orrT2Ry!W`5Lm&KO;WlzOXi
zZcCVB!Av;MIWStZk`WGR=Elg|+cJ0^h3~pN_ibMhnZD5aWJ@FOSaq*crgA;(rfbM$
zeF>#^_|a*XhqMAfa<$}3sircK2GdWNU<F(;&yIa*B;amjBVJdFpnSXe^Dn2<y~Cpp
z;xOSE7Z)<ik9FhxvcA7@$%^}B5^s|7g*k4@erZc@GqmDtjF~-4W{acqhDBVpmJgmC
zWG$bM``r9E@|?!L#><VdiYtM))w^{J)TdvTq9`b<li7?Xh9j6_P*-*OWa`&BCa*@X
zzLr^Lq*q_<rKm{LYg9CaN;cFM18{`lN+v(N`5@QTNI`eB>n?TV41Z$lfleeHa+z{J
zMqIzzs!8~Myi{{#3Y_$`gV8CU@@$h`=y#qEZ@~|2Fc(VBaG$~>Z^70%cDEmld%NX5
z*P3~#lkuo+rqlbBc`|&BT&l@4#pDF*RIL^2Z^nP{G}TkK#qY@*$~vpO$-#Hw&hi{S
zGjA|L+IKXKe>sJ1GSuz^PDL~2R9UzBf=VeWd-ASyxI7}9!#7c;4wzu0iWeHWn$``;
zF~|>_tdQd0yW{Vu-D#CE^?hwE&afXobU&IGeptjdGl%o3Mwks=t(3khZ3W)(xVvg}
zlNaTr?(eNKaAx4<)`mf3B|O>>bKT;V#ybygSY&5g?!CBKP9BL>34@{a;KW8uFb#$k
z189SKYEN5#6|;n|I0fst1z(oc*KQIy<iZ#}8TptN9yNF@d9^KO9^$NegP(CmC0oBA
zx_<fuMciY$GdT{bSc~Ji)n+K%<@q`gXJ{oB>0ZQKeH1nu?Ec*Mv@xY>qLoEgUO&?u
zGtN91cdmH4?cH(5rEULNDR<udTZhfY3vlLkr@JgSt=oo`qiOYZHZd)6;}mjQ@ra2q
z1ODy#TBD+#(Ng_K&*y}O%@3Pev@ZDTy*E6N37QQW?JAUx(rxxIvh3Y}*9GhACl;#U
zBUV<`DY~x3d}1cwVv9-Qmgf;NFQ(%Xx&m@mdk%Z*)RZk51bQdj{%{?fa5B&`x%%SH
zJ3FQIZWfqcO83sMw;<6kQVVCJ=z0|?v1<W|+@o@p4$ZC=DXx9p&04<##(0#<2U=Lg
za?n$*Qj5F@3}pifZgvc&`QtFMiJ%o%nw$4{!oH!h3{;3_WG=?fPVWsyERc6lgNx;0
zav2rIV*AeT&YQU|<d){xD(0+I%n+x?`f0X6`!FRL<3#(9Hm^~fxxR>ZBBP0K9BNTP
zpz9Y#HBvaR$b0+0=yb2<?0(Ax>o-}${W%~=o9_sKrNZ;147B08ynLYchcx!+DN<}m
zS@$e{!Np^0KI2}xehzx+a}8*bj-xgc{D=>*ZPzwzyld=O%eeS<sq6lm^%g_Tgm#8w
z<V4ZKNpJM70gf*dQj^;I(!M#obPFu{;RBr;IllAFp)XoY2+NFY)+!Y5z)H#)!i1!U
zP-jI?-MAr|KGz;$NS}1Me<`VPHPiZnXSZ?-Jy#n4rk-a^8B+k^Gf;rUVH{1bVV?#l
zxa^WxkY*?xm`VmT<#*HB>TFOb-_kNDwV)XH*u(KPnfAokY>5HX$~e>;*PWu|Bu{si
z4{0IW?{0uhYti5y!n4eO>5H<SzQWAubt<tkQG2lZu5AO%+LGkzl-;1<5}ds+78(_)
zbfz2xB+vk2wgfR-FGUfY9fo2F%Gq}^i|Nn3fk?r&=sJ1lgxeOr9%Jvam>>MolABR#
z)$H<H$mQ8)k3*VuId>LB-ne-AVzoMf$k_`$t><4Sjzakh{<DQhn%k7U+ID^QRs#qV
z1r)Oa+Cx`+>thOos~PcU_iP+e(_img3)1A575+AsX1(4OoEOU`jg^6G>vmvfA>OME
z<V;FvJrA#g?NoY6oFrEK#tIv*qPX#+J?p*RM-5fTSmA~E8?neNAruI|=S*2q-lPZ<
zR9OcHDBwEr-OeLt&=^Wrc8kaJHFl9DxmFpU_>pm^?|58GWx>7ZE4N}4$j(FPC$wYs
z(au*Gop9M$9z}kvwc<5(JfVJ4M!XnYQ}KGroePp2)?|3;*+ow1hx#8Ba<K8yDYkpH
z|MRBCLK9+wSNQ5;Mpp(g*qYL^zB^`iG166y&(MloFiG^G>_~c4+sM<2mLDh7u(#J6
z(1WY=upZu_!?SpE60xTtn2zLk169fq7Wz*!tq4L*A`HL@BTIN!mT@FcTjvB+IvI3&
z(iiBEzz;7}n_jK(k3}ze*L%A`i>Xpbl?Lc5pCdJFv*!sMOctU1En|@H({l8lnSp{A
zVIrM~QW_8OxpVekV9ue;@JOv4B;jr;0lY*^I26XU84&J*g(RT%(?kQ+qgtCeJ^9eB
zD4?yjlD7+kt8DIHV90!R$gIY;m@zxh9=W!({|&AO5F*~YXn{E+|4GR2TKuYMG;$gd
zuKda`p6UA=3$@RIYeGhYJyzFd)*PHHP`%o|+d4JKBHi?AqqoMg9`;SA!fsI`F~*TP
z?Mk+c0r3`*Q{ccKfG`UizFZKJniuuR|84|$Q)VjrIpIX({`&?GX5X#`3n8@9yYC;(
z)+u(v+pJZnZ@#nW&E4aBQ$%H+wv5py8?T=(hBvwq9kdUea6lE9UD!I3top%TQM4or
z%bp45xTec{!nJn8&pm30?(*!Py%!MOsH2g1DDf5H-ckgZqGi2je50-U_o3<&aZ7Ro
z{-N%6W517F9c;j+rC71oG%r0mkpf9&E)r;jLH5f!);pfX360t{IY_ST+Tf>FRLI7g
zZpxc071edKGRdAwjX88y%@4p`re=?TcopRz0B7~vq}J%j&N9a-7lyf?vTW@|hMxTS
zRM2p_3?prgg79Q?R8F~F*UV`-AGAnYZ$;0yzYBmgy`8W#0KQfDe0HlfF(s#2!B8d^
z@~yNoXiv1>(CQh<iUh9`|Bq&r86%&<z)8xvT`}rsYKyiAEM{E&$4SO>6O;L>5_NsL
z;nEdfqLnynf5&|f11Xh5VcS$nd`DUahKt(`--|z4lZd&_f(aFibBgw_TQ{EfwTM62
zB^>^^-64<E-Ty49_2kSYYNDUT@-uI9=V;WvVEsvU%F}glEq%Q+d^ECgG$o%hN@G8U
zDzlWatMPR$p#hJ>JY>iNV@|Z8GW>n7WR9~13ZBu&hXh>~vj}{F8Du-YQb-F`n>#<`
zHWT2sm>1l#_*<@$Y{DU(`KI$7(sXYnvoaSDYma{3)USRUj$Xh^-x0}iJCdziMGp4U
zbtHXPZ>UBXo}NaazCo>VywF#;vI4|a;4`8-&RlA>zNzi8{H>|cucYqK7q3o7`vd26
z4X^Jis&a3cDpsV~;Oz6jvEu~D2i#BvxeBcx#_Css4}r8fW@&izlFjU5MIMXBzGift
zwqDdevzyAT7^3ydAw0ma))|s~CtgMoc8Q>_#z<Ewq(*`AZj38Hb{br|f=PgB#k(JK
zw7pER$<?N(s?9b!9W-f*;z(!Is}Z)0Ow6S;)JQkx`!;zx^YSTlx_6~*lNBRTx#Lc8
zfG9Evws7)&=xQBo=ll+WSm{FvtJ<eHGxy?#kr&_JMpGyDOr3pFq$Fiao21$0awBvi
z3+Xy@;5y-THgo-|Uve`~DurmkzsKNn34QMTtU!NqduIDf-;{-4a%-pM$`Oy0KK8JS
z7FRsP$l7%ZUNfTfx2Ho_+pl1sDLSgo_4W5h<Wu@Pc3Z*bvOKf{%&{>0Bdxy>Wt?0;
zt!Ws`MyJbJ+XDjAiS3lV#dG;VR0{RlVRe{kOk#jt%au~e$37ulz`8Y7%Fb??vbx~v
zfLhkCP=mXO8r^&PH)LjU{7P^a|L$Sy7Q^70eA59jcR#M$^b!_JhLM$WP{w-IXh$Jh
zDwu42VgsjfmmEup+FiY_MyMV1R|!wV9E09_Xgf9IUQh^#qQ%cDrx=YkNzbA0cPp<q
z>*}!4X`BT0C}GsO!<6BKIObDMA%CeAFWx!EbN=i!-QB~5GvQDPXnsugm#DE&I|HA{
z`UU7G38*t!2s^pj`i6T@7n{BBQKN!*i(fva_xIWG$8D52+)*X4YBi<Ux?ZPh*x6_V
zsxph+<Oo;Mj>%UnP@XrKzGS^5^8=4@)<v{@m*4A$cHYJLyobuT20E-sCo!KOJ19^|
zu%-S;z~myayqI)V6+uR{o>$v~Ywxjc!sY96leKH^>N0UC@yZIEeQlvyTZlbXE*)#`
zhhV{C+SRgC^j|wGJiziuoCFtc5Il)c*3jf_C+6q_S}RF{H=9|trsUVZg{V{bXx3R;
zsTQO&oH>1Ao6~NI9nHty01X)EL|(AtSI?r#8eCBOUifmA8V%M<uIcTp4<ttU312aM
zMDS(=jMN(%s8fbT!`a~t=e`Y{6VHDO&D^dLf)t^+rdo$2(kp#c9z<}F%J1*MTV2sv
zP|3y|AvG1p-`G9PTw8)N>w>o9$4O#NZU!-;h%pm^AVpZ!$OS|F5}MHc-WqQy*K*Ib
zG9rx7rqhH9LUfCrtZq*i9_Ij=azO#8f`a`dDn2?qJo~ICNyFRq9iG|ztTkE@(W+=(
zw7fTwOzOy48MSWp^PY)QPtb}<`AS+pk6nBdSKGU$&}e4d#)YdCVFQ(INzyNKbuO@{
z5eH?fQa+09L_WVE4V=z)?lTn4^h$q@xND12dZ(3yzVu`?XD_5tymheV^sfUK4GLD2
zIIeYmryjT*fPDu~eE2Qt3&J<Opr)dLM!E=YhDRUWeyO{DV5sm#+4JvYm+gZ0%(P6|
zAApP_BxaDSu><?wQ(MN1D-PEKPyO;!kLkYhXnS7kR>cnh@W=wl=DW}qB~xy{&>%f=
zi?y1xp%k*vZeqJPO?xo1Rd>e~+;?zL*W!Tj{XGvfMAjekf0--!(53HM3d3e%I9+kM
zr6Pct!*3*>ASMnb8wlmC=LGj%dHW3bwmqZgA<hyws~ks_@#>rjcj!awboCD~#WZxY
z@M13O;cO`6k{Z8F*9g}u&NzZlMM#Qj5z{VU-5tL%rfrHwhI=VJa^u7mHU4wj#u=*(
zSc_Yef?Lbf$Q?Ky=jJ$taR*7SUB~q`Z)+|T=#gHJf%(t)8<I_uSro6@(yU6=J>#Ps
zvbVhLeVy_#|A-WGp5r*)(X{&k*@&YXQhTWznF=4byHO(5VXgIyczz`xya=MyS>Fcp
z`~uu#7Hg<CAK%hnZ0FCt>F>37u}Mc9Z#Ca>*s{3X?0cX3TlH?shUW1N4@%&Ao$w{G
z)aOmR!cIb+OBE9^Dy<`9klTeb!&#iHrmTkGW=}@^S8x4U{cZ!X?C-)AbFSuL*`W7n
ztf}{UYYGPsoLU)u^C=;&F5-j97Yf_CIvy5~LUPty2HWn8-u1g%^f41j6Zz-)DgTOI
z8Vuep?__ShNK2)uU1ivStVZ-L!TZ#=R39=EcA9e0?$~usJ1fv8*|d8fe6IeS4Sb{y
zo)_&~IhK#++TNtFPNl~e^;J)}6t5T6$uLgDRNm)S<5i$V5-p$-FLSy3>Rp!u*$#|y
zBfm~iz%F2X2~myVzJA)I5{;Z$ElFygygOibBpNsoIW0TVL0`BmC84pLpT!wkEb*Ri
zh1T^+zwF}t+U;-+eV_F(zi#d-W36lcTkeZRUD4%?UlIJ`BhdU>=Ioc0h+CP=H5s%5
zMn+RT>ISD&{hZ4&V0@wM7W$CXGAVhu9PsVjbbI&@z~!?8@F%rjTBGmx*!g}>S}vDH
z7NfF5YEIxlxPRSH$(Zh~?R+k(=GA^H89sFP;17V5u;TJ+eJ=S|`VB~LMeVOT7OTw-
zCDm1xHNzg3mZ3!xE7vg%p44n*cE=+6_x?)$r;i-DS6+8x#&*AxgU`rX6KGGA7@M1U
z`B5u1Mw1=${`wc#dk-vGv(3@(Z_ayjk_TJ#jt|?mv*I9r<J2ZqewfRffTT}SpR{q2
zRPVw1i#aswhI+O5RCxDm_4tw1+G8xWS-<StEwv$*L02`@rnPzzxHX<3H8B8o5L6$f
z8Q#G2&fK^iGnZvlz%#Nat7rw`P_Rq4;l0dHu8jFgh}%rOxeG(6MlwYlDz?uSFTGwu
z%U{3HzD%TXqIriX&nh3g(fy|TZnzx(U9bLGSi*zr;Jci}0G(U|mq22lX6#;%TXX$t
z6T9mh<}Y8{MXJheO~5HOjZ$`|UEO7){>)-e9v^?OblLkAE!7K!3wNViY#hhGUhQIl
z^;4UF&!yHc<CUqpIpymj2;H}KVnJhuZJXlNmLCgN;L<N#GOUTpjjEGp0l??HZ+F{4
zEZb@ItQH90{n&t!C(Uu|qRQxFQkRA!hvx%5&N6bf7QCuq+K#*oTyxKHe$Qj&{Q+5&
zl<&m5j5;aZ15QtvZ^2j7Z$-}J06KN=TYvijzz&+wKDU`CLk-?`?N0>DMc|68n_sbP
zfbClSDif7oU{??CwCyX}r&Y|uWP`zah}n<}9k20$v;33DIjpsK%N=8ka_uk!7_|k8
zKp2St^=b<zd}%NjhC%h1EElviCHSlX@%V8pdgLui{7|br<3Xp*ldfNUv?dlpwSqW)
z&izuFx8I%q#<b&wckm6d*VR-$ai)pY%M%;a_@W}@wn{DG6Y1ndw$$yc<yFs>qwh%P
z0V><QL*8Ly=1=Mld00B<G77^P6$rDux#OdYRg%33-IBfCcK8My&3|yh8%f2S408Qf
zj?msD;(Pn*n_|16m<O?nV&CQ6kjGu7P)E!z@jbR%f_YCN#g42gilvnru<Z>ooHA<`
zk>XS2<RvA$L7$*#?5EsNyLNn}ijzh!IZ4ctY^C}T>8D>v311gcYSD_1cibg+)1W-f
zCFHWpRae8a{a0$9h7)(2IT%)@v6~Y;Ll|)jUnNKGW<veL1pg%5g1g(#hu7zfCg7&e
zo+Sh#YGP<@jh!v^56FFu^`Wb@@W^V}XwW&rOPRQ^go=9j*QiDWbAh<_;mO&rUrCLX
zfy({!SrRdG_##A}Z`rjGge+SB5v$2ysL26Gdv?(V+6YL*xvHZnw?Ss>y_JuynXBq8
zk<KBPJ~YmQLpOD4G>ilLfx3Z8Nez)qSsS5d5Alc}b`)#}->N$7DZluI<~iLPLU%LU
z8GYGT&F00CIsMkAV~V|@AI?n2=>CSWC62xeq&pdoWxXuprrldAqy(iV__`W4epj4I
z;PT?qp+yy;Us_f6gm3KKqFysElCy64#V9qvG#}UDH2M5Y^qFhoE!y|H4+hwJhrZsl
z%<Ew;1ikW}fApns*GSKdO`X}DdorZLw=dCi@U6VLc6xuP&-^QFP8mnXSP%1lwy*3=
zF{zh8N_K~pTvF5&dB|8-`ZhbTDg~^ly**-i3Ofk7nQN5EY+Cy5ThcbWU8sQzAhaLu
zJ6Q2(2wJ)iGz2}5l_2ZY!(COE=kh#Do7wH6zkA5IDmTi2ce&6}noe?{;Bt5tiahsX
z5OpCaqZn6{b|p0M`9pBZ!i3+Ps`s0g#c0Xc-P+WxlKP{P%5zBc;JQDR^2tJcb{WVx
zB=wa@LX1PJ5>C%GX`3>rBx4^-nvQTehz)(jb6`Wlcs@nuK0tn4zn4z<7?_y4-GVMa
z9L6xUCiEQYIk8~JDMpGA?NO6(BHjHFg@_HLy!_fvUolRE6o1Ki=Hn!lzarwelHE<!
zyE#jJaxn9(GuN~Is-^ttOwV+c1GhICa!TOoRro=C+Wy&#(JebGdJ7W1fBZmhd$RE0
zi?&x{o~80tl}7b$lN|Lt*v_*1LA1J$?<w2)om<wK4Hs|WPZA}1S&d82YWvFZwO+B8
zF0&3?Ib{$cVuO}z0^9RDt}yDjCkB+5C{)sZY-j8>{TUWG>b4KrTB(8uZ*ieHS~Wt%
zL8Y3&ysdheITkrZv*G8jRWxY?(^?y8B;z%!;NIS1d1HrnexH%RTF%nkFlmz93c+pT
zv@UX_++O5~v+~8}g#mUH-Q!y|=fMgp+|dPCTYjP!ok3@1buP`z%X@f0-SxoQ3ar&D
zIBhIcQM}>~tr!HVMun4u3Lq^p^-!9qS0vja>uTu(N?ud<ZZJ()HfnV*<@(xs$wB+q
zKE59akeSbf*~lG0j;9DnuW7YjO0&P`Tx>qc%Oint7I%a3uCl=4_8#bPFvnk3?}A=p
zq8~mCN`7b_`aAz#>BDtq0r*!DS|rW`yfcnsADuLDXkQ{q8ytC2a;SpTz|#hx^6)Bj
zwmBZwg<qwKCuerz+UpRiDxXv%8jDm$DxyXhlsK}JyD9Ay#R)ZHI;jUYGRpz+bKS_P
z^=9@_PZ{TvSz-X2IFuOlX@ilDs5%jTmxO6Ftp<J%P))$o-Gwux=g|u#LLLBokGwwN
zGoPkM0c5iKIdqa=3wky|yF|q}js21Pl5D82i}M>HD^2JeUQ!{qL-P1;)xdeo+9gr(
pD%yo~{?mq7#gOq}Tm0(fC>h^MRbUXgWV2C)F3kmXd71y1{~w2U7ZLyf

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_high_xgradcam.JPG b/Images/deepfake-xai/ms_eff_vit_b0_high_xgradcam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..2362109c09b168cc2f1bb88d169a75f735339a98
GIT binary patch
literal 10401
zcmeHsXIPV6w`S;wbcxcXm(WXS0wNfCCkYU+KmbYTy@=8+G4vt?q=z0bNN5Vut5Jba
z6{HzN!~%$7fr;<?o$ou>ne*e!%(>>8AG4F4tiAVo)>`+y>$RWb`QsG;x236tDS(cS
z4iHWI0FGDb7a`Fo9{|9{1^@y80IUE;x^n<}n#2SE&{^I_0T^jA9qnsh{NT?&fDHhk
zAAK(jV4>Bq)5?al3=n=#`}x!Q{}TWf<{e6-fhm*WFCO`5e3fJ}{!?zr&iYGcP*no}
zva|ms|BF+)>_0UAe@_23nABC(we(cMdTLr?YG6IEx}Lf^;6JMa0APSHfH&Yz2OO_5
zp79M257tvr2@F&A@Cx$uRz?N|s6=}NtEeffssQxQM+bW#{k+4)JiV`>0u5w8J$@=H
zhVn9ybq3q0+60?;`=ZQaL%p41ZK23mKcudg?0G{DHvMS5=z!n=?{E*X=m7t~FuiC4
ziNCDt(d0kXDiVhJp<X_Eju%b;p+PGdNc<y~sHiCAC=KPH&}%Acy1Ke5s_H80>Pj>X
zrLdU5aF1xEz%a>w7`W&ih73gohogc5#r_!d@C=FwH;{;kKzZr;c=&)ly}ZCmo*o)r
zN@{9e9!efw8fr>j8X8{OU^N|eA2m;jf4KKT{?mJKM5zB?zP*qt-u~VJ-hts^G&a;!
zBvk$m`oGI3jiP@#{ZHwqVWjEknS^?KgnM74Wy3!ms-y~5QqzL|TPG{4s%h%0{0BXK
zl|N+syFmYYBmXb4WvK6k)I<J>Jt*{V_fFm+|7X4Jk23tTbo`BuR`^?>nWc@n6DLmS
ztNeS9|1$-D)zF3<?YVyrP}-AX{}<B#%TE3^$-M(<Gdqelg^ynWOaLrQOw3G-EX+*I
ztSl_7Y&<7v9J2wrxY>F5fr0}3K>jmAqO#&bBGMvf_$8Djq~+ul6cq%;R5VmT>arjO
z&>tgoEUc_-C)oH-p5y}w^9zIim(y`CfQOaNp6)aQodkfMhmL`V?syPzmbR5L(YEM6
zyYxR29X$gh6Eh3z3AU3o1rj%ao{oWmo{@oxiIKKH)7_x80~mRjc*WH&FrRYpV37#n
zQ%}q)XO%Q2_5mFyKTBzNh9;e0JAH;<Ku}snR!$zI3D(lq(bY4#2r)G?x3F}A!r;y>
zt_Y-;x6d_S6gn(CA~GsECN?=G_15h>cQM&Hxq12b3kq=+l~wran%cVhmPf6mw)T$B
zuKt0+q2ZCyvGJ*C%FOKC{KDeu+WN*@>gKzx_g{AR_P-u{`+oT24=y?Y!{4y}iR?e%
z;-TT9XJlkxWch=Ojy{T(8F(0(#MPL2FF3Gxgq)I4Ph{mY&MGJNosiUU{0#IAon$*L
zrMW8o<qx#Kkp0(yCH-$9`zNq};aUW+GtkjGkAVkp9<Z?`KVLNzTfwFVF)+0e19a-c
zM#Y6fA!fXBOQ2n`lP`e9g(%*>;sgcwB2!#UnyaxlHm)1><&uv7-hQL|Jz~epAs%WL
ziM7I=U2p{F`q4e$Imnz)<f}3VF7mF5(vH11ng9w(C*4!Teh5Q68v030*J<4JPJ1}B
z3vDwj;;J&&A_}vmwdhfH?Mfv1YlGcBMG&56>rb7s>*X7m&3u{LA5$X=YHSx0Dle-T
z(%b<;nqTQB+x!fSrQ+{ArOMSPBu~{6#zi}8SUkq}>o_I`s?jc9;_q{#j5;`V2>hOA
zx(GW@nN|(b=CjY&ewNNY5}=snZMz9lbUPE`iNF1dW-Qu<r>{T~<3D^c^z8b*t|Y!h
zJ_me&irK}^W_NgGT`qPVlai4~%~t{|#4bSieCl}na~&Nq^nDOMCm@!MwIV-9kedys
zSxX6ndjg;>lbX>EWL`@<OXJtj8`~w;lM)bJu#X{YeZur$Rd{2eDbDUu0w|iQ=x4Yr
zaF*#iu(?Jt=!S5-0w7A&pm*CrR}M0^FAQ5s?^+i%ms)=dyse3m(xu|NyRMB2vrQ<v
zHoUF#zX4`ll&&t0XRWnvT#|#3Z8NE)&X-0Eg27G*p2dyuI!OL?=n~UpCiI7sH7J(g
zN2;i$7H!Cqghxxlk>J+rH+Nw93)hvxmHbm85D5$Et;Ly#SugHjO1scrIzJ~6RPHC-
z^3^)OLCTuX$djj%I=(f`hWZEPZa7+bO1(A0x3xuYY=oH1JI>#8D^~oRGdFQW8hdX&
zJsL9A^@h9eshw*AU^vF$vtOR1f!~{63(KT-6Y3YX838!!C?fmZG2qd2UfaS3?vXvu
zy7vd{#KVTmRko&S-$(ORw_Z+p*80w!BP5Adv>KA?)jd72j@0a?jaza<N=g7RKF3L;
z^w?ZSpIrMs-m<kSC(K=&bKu;<Mzef8bLEoh$?!%Yuz9Z6e%&dy&=7~FcO$Ymn}u3|
zkftgHQZ%r$a*<!`=f$eSY)-nf;qb`9Dl5sVZ!BK%(JEnXEe6pAzZ|m<pl_NFMdr~;
zJw94+vQe^<?X_wJkA_=pqu)@&qwRv2cHSACdG^yfb)r9Xpt~pPa)evK@QZ1ogRpZC
zeB^qz8vX8VT^k=}B|H21m|NdhN)^ubQ}$WPBJ0kR3_G5W1X(7$pPE&gt%}ThWX<)d
zW3vYDl^C{JnDSi80MX4Ba|6{dHouxLEn7Z&3_w{faaiJ7?|$1ZZJ|hEQy$xF|NcUW
zPYZ-FI){8pY^R$OX~-tAwVv&1{TN`|l0TGK-_WdC7o8n7<fNs@Wn!u5#ggDBH@bb#
zd!uVHS|3&rWZzD9?J-GAk~0^I(j~+SA>O^a`9kFD<KYm}9B^>i#Q;Cizn-VlR5EVi
zaaDK7g%UphRp-fI|MSY^-=C0|`1LKz9qRmFD4c>!Ktn#8+~T`3R20f(4J-7_Ui55A
zHP;ct_*Yp8kZ<UF;kxt-FAh$>n&^17*shkz;b9>J3CS7?aD7??kRe%bW{WZ@%g+xn
z!tZ1iRPudrz?5?g@x7O0C?3*rT+0Ef^9pyPDEF*f%ooK#7cGR(g0$iy7OU2px4MEJ
z-2G@OA#&#Do8M@PYh{Z6U1PZq3ZsL_N?Z3#=W92FHR}}$;fvC*-)KJeWb_I&hzHG+
zrU(HyeMJ3!wA+nmv~G5*lwO8e+e$R0EyReX>I<gh+q27K&`<7*I}>2qOc^8(!N_yv
zrbELy>GE|F!zXp?Z9+9tHz@hDVeZ!KrIGD!A`Xxoh1U;=kqz7^-th~WxM?C0qklEb
z?eyocsm^-KVW6gZ`Lrh~WlW6@>XqZ7*8u*xb0r`6vtj$AU)fxsMb3w;(!Ag+>D9zG
zXX5=;sAq#+6mi|TDCe!I`LwCrB?B`D#7~{d3E=i?s`>a&w)7dFNyPIIsVjFrK-_oV
z?Y@`Fy|36NuW+t1a!TR)_lsF{rE8a0W|C{ljs%CM18wr=Gui7_Z(VmrM0T+l3%Gdk
zEcYPK-2TyRqjg;r&sIF)B<RNX!m{~5ObG8cPI6DgJ=ZJ2!X`~d7sD_(jh03re{BK#
zR7fXpO~|8xpeWH-cV@m1U(XmO2L94x(Rzd|F-cD%Ew$~}Kf1KT=+jmSC!I|!&Qr}Q
zdAAYNdT&xl_qO;A*7JC6tJhB-KXBy=7y$On#cLyy1U&OsMpW5_GD0ptP#mv_f0FS!
zCNEeCjhO;S^-^<yH-Fk(zi43+la=)?Ve4x>@Z(e+TNX8$Hk^ir7XGXS0))|1e2y47
zhEj<fzE;QlHDk8PsYVFp*&e-TMsw$W<Bx9q@{R0uszAD3UZR1)_U$vzx2zUEhEKh9
za(^W_=PYD8v%(2zTo+eDw8K|#nX>99yB^t*r+R!!l`?8>5#5u9hwnD!1GmZQk-O`?
z+px-ZR1Zp*5Uyp3T+gfC*l%z!dzxt1-g>asG!)QIvhtmTWSO7@iyuzFa~aBjf+pL}
zO67<3!r7)oIOVBtw}N#?J1NFxhplt$YyH8){rcFi&$STcBVW$%H_AIjyswqK&Eo)Z
zABj{ycP_%yqIzn-qKDuceZc2nDOaW_&DT+pKzGS7BQ!AlGvQmsVQl1%J2~ikbbQ5&
zqt8b$D%HU@^^KJXe-!Gzv<SK~82zEYUX}Vp&;J|B-}nV4aieYt{fH2*Amm15OgC?w
zcv3uzO{#m=$=sTT#Xs|W4D_^tyJfdy!j@YDD-SuDka}y@KW8V21Ijm%U8Fqa@MD0|
zE^Lm4`JqZc+JkAx<6z<B9XCN$4zDMB81l@<9#~#gt^Mt<hR4dedCoxM8kHyMRcJv)
zBINpJwi2HY2CzxyyG_t>%#;E+C6qa(hCT91TLSIB9uC-)jLpa_sfvY84WBY3J-@HU
zJ#P7MgJ%(ce^5h<?k<0NjSe|Fx|sd-VeQv<vC(mM60K?CTn}WQf%diZITd*4F2}1y
zxm9aCKS2(ZdFp@W_r+3t$i#`e84Kr-_wK!SPk%*U?G{_nBi?olDB-=C2l}4$>L}Xc
z5JO%io*(xmI()ohZczXB34dT-evjDSO1WG5Mngue86HP_p%9gNPp_Q`&mNLcAo>kj
z-a~Zs^R((peAMeX!(IKBGg<dEgtO77B(20_PWT|RX&chcYA|tNc;aT`N}U7w{dvub
zB3tIJXCrKg0K?1^5GB4Q)u!Zur(d?u{~S1UKJT_HoE6u(_Tok6{6Q0YpZ2ZD-CyL|
zU%r#MFjL`~1qjT89Jep9;}i>=hw`$u_LoIaNseHuygA7cwlbT0cXvkjG|7=OOlI|_
zlgm`?9!L18YUr%+_Fb;oR_wlNab8n6LchnhNNVt!=cd50pGc8n7<}-tW^2mT`~D>{
z{PiiJe#;Z$XxpH=-Hx$6=);tk95QmvE-NB9`Br;op0i*bHm+qnUSds-ud}y`4DagV
zQ9eQKTxdC2`0zFkFLc^J2eyoqj99BlJb})h(j+j)G%VG3w*)xgh*LRQciw{4p#oIA
z=_U@6zg{;-f@AmK&@WD&S?ZKZyzy<YGRh@N#Fnr325RMpy;?uUdWxDkM5Zq?4Ka^8
zi<|wt`Rk_DF~I#8P}>-Mfsnx&KeF-k>x&<l_2FZ{<I$;{Z<C#3_0K?*JohG>xIxzs
z!w4Q=?E+|a94URN?^(rGUwt+ntsVm1utn_&blrjye^nC$R|j$~X%a)+YwiX~wzoxB
zUiHol=-p)K@k6NPxz?K1O~6~QZsHF+enO{JrRp;qO8T2r8_9IW2ohTGC)&Y23G*&)
zV`*tYfG;7$t`bPRCxnrouSihMkSQ1eO1tbjx`pb2t?ywp-WDicW5EqSST?qr*)H%(
z_va3Q)!0-5leNr~#PflqYg}Wv=WPiZ3~mo)Mv3+a(I@O|2-I3RZpVZFlLfA#FW?D|
ziGq#+XGv40i^X(mv1d5uG(Vhsz#cWZ`HAa!GbBWalmBMK$Zy<A7+JO`;>5b9i*EWy
zZKugpiw@4OeM1Cy)8H`p+Lu$O2c{lU(g)2s<TBwR)%B0)$i7v5hs-&U*WFJwdwf6@
zW9j>AhqVI8XO%bLZ>)L}6mK{DP#Tb_#Sju!Zi!Y3G&(8EAKX6%_;h0+V%)u9pN5bK
z;eGsB8Fz1#OjMd!bxg8>pXjD5Mk8q^D`ZW$jkm{CZhyu~^5Nb~PV*{l?!@eYy(YmM
zs3}QFBHMYyVCO7Jp<TfrA@}&~8hFm+gcJAgnVK-Rf;)rzR)ios(rcLq44e%m@t}dn
zU!c}+-Vx<D8%junX@;xMU9xX%iB3^@I!;}lQEbUIG_ck6lvYZ6m1g{LBWbT)M5{O+
zv`KQp=8VbeHq3@Ss-D|+{jteE6l<Syl>kKM2XvDSPTh<9u}1QfH8}=kT-XOT;NIVz
z5W`nfv9q@nBlrqw-T8t(U$?FOvYbA_xFNrh8tw#{$W6Z<YIK;;6u6Ql!$k<cV3&I{
zFFES$j|)HMbRS(jkh5Z2g!IF`lfnrRXO$SURM+aeyB>SL-&QhOXHK?HLt=v~ZsAm8
z>0p~Vo;fa2xI)JI5_Fd0QvQ+DItH&?!==wgC9jXkaN6j)=MRJKbT)l^$UNpiHap6Y
z9~Ty^hFM;p=CSO#K+T-iCY+fVo_5oklfFK84*fBN?Dr}g?ONf}Z|yRZFkI)B4!Pp8
z=c)ni<wXgvK>IadA4BP?m{8S=db8e33v9utZ%xL<aK*h%N&T4$SIXjLrQaCCo8E^m
zu{@Rte#(;+K&2FJeFv-?e%P+^tmcz!Z5`<>_fQmCEQK%X?QZOK=VWK)x}BY~sYlE`
zOLBgl=uLMEs&vm<vBbd$!KHn6)te#Fy>aw;(S1CWw<I!~-}z;KF>1(E;bQQf*2fp7
znBTUnL$=e3$&BpITvh1wXOfxJt~=7NJL{EBr^sO~XD_!_HH9F)uw!cKhuLMyY94lp
z9|KHBrbrhW)A&8i30U}~u6IXnZfhPfhfh#$2STqp;2m94v589u9>&yBJgX5!y;+xF
zobSiGpeSuFM0=z_W&aEhW2%{mLW$*wC8g7=jO`_Jom*Tj-8MXza<{@!!@=UTjqQ`Q
zQY;Q4dHToFvunU-Hc0Wp7LM2xDd!nsRu91D<F#ajba7iz)yIOEWQOI^S{kV<{KN2e
z{aYCoYbCEP>sxk$EwaW#x6z-p;w@h#L!8^+hym8BcJr=?!5*D!y>8T7vF@Ncd#1hP
zRmOerV?XA%?DE83;aw6wh&xRUL&jBcSM$SjB_-<X(-sKb=PbIq?pW}0+%89AXMrGN
zw!KI&)@{B~a>%LME_Zftw&hi(Oj@OE^(uBYV$lbOYAAc8`>NJua73hmd3gr^MEtfu
z(_{sEfh5V{IwV8dl*9Lq@8(w#m5EF5G=@Ce^7Z?~wzn?ztkAhS2Y;*2F$=UCFB5z@
zHcSmHtR|)uI)VJo*-she7Mx-3nn1Oga;xwiwD4;EZauuaL0MPSjHxWpWogDoo{mg0
z)@FOmkqHxql@+*qmOp|B0eQ3t(5C_>xXF-Kn5(@-Up0vjvxq?Id?i?WHn<JE4R6pT
z2l0hYgmhF5=}dXHs<As*(GGVa+;om~_9HUrORw#}f$zOE6?UpLa$d&8RB<8=XI^NA
zJ9&07={u@07fc7D=?)ttecVl^sxPA(1IUPbIY!yC5Nb6~;114w82uU~@80B?n;&yN
zviv&WI-J^<o;!R;loEC_w@)wjhtpJVEPO2Wh8z#wVyS9q;a#sYB&&@b{)V3{v0fMj
zZHXxm<U|bYOanKwUk?#`f%jq=paILast()bWYuMiY#OkgA*5j@^f8T36;zyXUmoc|
zaT%a4y;_~>kAXnWf(RCDn0xi(sv&g0dC^^=l#!(d=bK{L_VAfF($SZVS?laEd1mKp
zlI2L8i?C9x<7Q6F2O!$ma<ob*R6fH?z~%~H2l=|QuC>gxMPQ!5lxY~|U1Wx(oucSW
zTz`vd{z%a0nSvej9mGvUdJ@*rUlG&K(^eV%8MF#Km4PVALxNU38=95i&8+prbiZiH
z6Ys3`oNDPYTUI9NcJ>t*4h>b>Sq%_IZ>;B(@ACqINt`8xPzpxda=g|ErsdB$u8H?G
zWEps9hEL~LH3wE^St~f+dV%npJ7)&_D0L0j&O@zB=dckjXk74}N;_=G0(+auf+OW)
z;r=s%=Q_#+n_4KCl_`-d6@;2Ozo525C`1yI<dkTaZOs7PD3T?J&l|s<xuZ9u9*XvL
zPqtv4TpCh+s)?~;^TRsPM+BoCpeNVxx8fc$PIcgvJTozuC{r_0vi34K8*nkyPK*yj
z6h}Af7gV<{e}$uLSStixg*lo&fqJ)tT_OE*HX8Y5w`m6plI9#DR@m`1LSEiiO0EI}
zc&h=?*7Rxf@y!&>lM;ghs8tLSE!kT4r1r)iN_}x&^H6jJj=m4OJQU1$^+QuKB&&oH
zwC%oYX6@ysZ%?odv8I1`?V_uv78tFqTQ3*WAZOmVP~rVGj)ED4T|^hUX*Q})nWKh1
zV9D2~PFa+v<J^;B(2eOT#=0T<d{;>p2F+Wjly$I_33JDFGSU&Va;H){Y}-xmCPCws
z&k7LD%g|^6gh-APmOH;vT(A^v!IaHAeVQLmYrk&N!%~FC+ktt$bD_nftiWpSRDo^F
z-b=SaopT%oG?wcmuR0Cp7M#i_8&fl%;!R)J08(aRa8)wpIk*mj1yWJ~-NiM5X=bI?
z=sR2Uj5wruSl67gX-Wyzl0k*pR21Cy!JstlSyU1)3|Vblci7;vT&k2B9R|TfD^gP!
zoKotD7FiR#mc<FEgA;GF`9dt6=<x{wh~WzD8VNYH2op?A(rwZe$?`0tHmacSqG64N
z^|B6jEI>hAuqONc2sz!_z8VJe^7-mR^CgImU{<iN09|XsSL{s~B0%$8T(a*0foB%)
zsAeNx7`){c@)&=`il1`hCA!xzcSU}lnKjE9z=o{-0CFm|L{bvnu}&vF@i+4r1aC34
zPg^-zROc9<;=>^YHOz-ecZ<BTik%kTH~n%-{xJs?7*qGBW)}!)zMm0pv~)1{tp@m3
zYT7f)yw_&KZ>r(<Y{E78F0sq^r8?}{$QJw3A9`ToY}H(QP@4O>4Q$VSXW1#V2?;sn
z4Y9(lWuLn`La-#UQOaf&w=O13_thp2=|i!Zg7)afmb@*f48<G!S;M^2FH&xACKk78
zTc;6OEwoTXQlFe6z4n`_dlQr(H{mSxEmWID^M%^02EkShf?l!MuSO0Kx_2KuYAR6r
z7Ky{D>HCM|SabFTP5x+;AEs)(n;~D!(Twp=HeGnRp$6vQeSWuCFfjvspMCEd2d*pK
zjbLRgvu&zjgs-F57(w`{*0guS3zIOjG8!-a47m`ejclMXVHZQgE;*P9SNYb}g~2@I
zD2D`gmtT`M_-b<pQ@OHhs(_lS(5q7)2@Ba$Vm#Kb$U?QuDdLe+Y>v^@$z*w;2w5ns
zw{8zchyh;x;vQxep$^gfjAjf(+xz7`@XZyB2?U9udZ+lYo!c;c3eQwiDw$c#Op|>m
zd>|O)+eUYbzS*D_<Y(x+sHIcw=#77rbfK8e_UJ{@qqNm~b=7IkA~p`b(>8{zMZw#S
z$@HAXvt|cPzGr$o))H~e!6{eS9mU`D0xdOqS_w#<Ixk|sa&B1w*bt;OWa5KNtnn#-
zEq`+AVUaMk_L4;+{PLCD*@%I+X%WV<N*0-Cq^8&Xq0n;TC3%<iz_;JMsDw%{xT8g~
zv&4NcT;}cD8mMoWBb$*DpM%kJXTFX~O`k6IOi!!>PQ&OVM$MutVMbl)OaLc8Y#3-f
z)2=h*yVeO3X+02oHzUz4Nl;bL2!Qw)CirxU-bN@g*VHxqxQVfq>OL17jg9(2>6&S!
z@%~cm!h=R&upz0R*oucaruW?7=+4cJDJ<d87fieY7DMl-O9@YVwY%y~=i{A>CX1A_
zrB|THfzD5EXrnsDj!)LD&mf;yM9|T#&~#GDq5exR=1ljBHnkDQ4^UxfV=qNrOHP`Y
zK3^-c^{l@fX}Op^|F|x-4*MO*U@?bVR#VOWpy#Rsws?*<^QT0UBTTh)WfjuO9M`f*
z#DtX)#{n5(sAj_eG_m;<FDMwD`lJvT;)cn}p0pwJ(LGzsuBy(u#Oy3YTjjD7KWgQM
zqPT=-mh}3XsEqO>WuIKfD$UDh9n>F;;C<!BdV$R>x8Hg@$bNL9-TSY)oz6c!i~MGO
zGCpqCZ*a&*vkoQVx;0rY&!3Ou82t6(XYM!Q(tc9K^a?S9Tg71hD>%BH0FfVU#x-La
zu2>kH6D5Oq?m{;DYL%UL;Esb)UQMRU8=)Z7dTqsO1M4z}MYx`{?HAv-v&J<nwJVe_
z<L4-Pxt<)xQ1{fOEL^=zU-QG1+_kE%65-ZJEyhZq#DvRsl+#Hf)S_n7VH19Baf@RP
zuik{d+r70M-|cmyYVzQgCMT6xv|JuzQ7-`ptMkTA@rX%HiK13Q%wH?{-tk6pf{L`s
za36hK{Qzz&NvB@*y{|-q{CrbNy3a6=4Om<l>ILJKmfN<lGWYu2GmYSBgSXK2HDub&
z<MJ_@bQP7(j1G=4EXcO?y*kU~yq;U9ksPM9Yh3jkpy0G~0oZF`0<vP{v{X?u-do;f
znOoz-{#Kw{cryBjbNLo7PR|>cij5hhe-_hs(3ob(y4h`hkZ^C0C%t3G&2en3xrnUB
zW$AF!|LenC<-^++^0GPBZ(z%ozS&QN+z-mlTMz|g>o?Xu)fJJZ1=)hBhF=v(QfAcp
z;Fp~yOX_iE{tPBMvWje3`^SK<b9(Q?E)5gtN-b`bd9xx*FSzGsIk(?`flGbt8%LGU
zsvYN-NU$Tg_|~5`R+!Wsf`7xmMkNSlInNf1l#^{6Y%T(Qz~u?V5jap)o(FwZ4~jLC
z&)%<l3_VqvPtedN8YB0hjFgsAT!lo}Rxy|faRUs|VcRm(;Kra3IeaNj^z^j%SE;H~
z{DpY9*cS^7$n+BCrfx-jT}9r#PLmz>D`<IH5^CYy-qO{v-DEd-taNwTMA`^tW{jU{
zF_~eJm`g^|cat7;W3F<Lw=v=Af|(R+KLK0nm(;XwUe&fen-Z6uk?GQHyTT)+^<>5`
ztEyu_)9phxsh`6=PQXu|vSII3L)`DA7Ei+t@gWXtEe>Wi{1v@7<+Y(`)K`>{e~QeV
z7i4o3@(Qj9A+Y?RhuAp#NeJ7Z%n`~Z>PKp`%l)m8<2CH}NHJb;|F_*qL4wq^R@f;M
zgkn;q8s<PbSfHO1eSyP$H&rfU@j-7Bk@BwflX4ZbioUic(HL_RWE9{gC^C4kW;`MK
z5wk#skAUawUfxhh7P(t@!7Qha076G~_AkHt7WrZ1T0!#noZ=%9Kb<dqP2V$c9lB5F
z#*V%mMa<=|ZQkj^&94tm+0S3FvgRP`3m)X?FNmW~RA7i49<;L_&f1<)r=rTk0he`0
zu9DAQgW&U``{HnG{OZ}+r|-Dpj{ynIM|`uN5896bM%Rx4znJ1bH>-`6VBUW8Cf4pD
zZV%uVjsXQ)<$o(quad)#5*BYC(LenT-2+B{Q}TVY%b9c3b#_DE>*sLumY^s>v}9Z6
zek}Hij$af@*#Z6r7hEx<`f~gV?F~|D+}r4hSL9c9JFDmm7cNzbZ{^i4ayHXkH@x}9
z692jO81N9B@q_VA{E7xm+Y<X&6zk`7<ozv|Nk{OBCI|8;(fQ4H_G7@R)$ioS_?Muc
z#{jz0Um%M2d^@6aPvuFa<G=(gt8Dx1K3mT*p!eD_fGuBy3(oQUe(d?3rDH&U5tcy2
z5p0hEygw(j8g_32$JHa6-p^T3V1@yP??RSO)1KJmx=q3!14bqOwysd4e9t!O(o?6A
zVOvhiXG6Pn)u2~~BD3e;@!X)X^5TeZ`18j{Gz|AX{bHuE=e|PC2s<-zP$o<8Kg+p&
zWuN8gKSBxmC5fA)v1xD&_(Pz*@Z?8t|1u>%T3p(M_?HsLfWP3#o89_a`h)SxG2ky2
z6mDx?F9fyzg4cegww$N_o71ygemd*s|7O~QEK0Qi&RJI_?l)?zF8Q6`{mZK}4bHZZ
z1%>ncOXaTk7k|)w+0%R@eLw!?*1)*nQo*vTOvlOEQJmekjJ;;q(J<xwX8-olulDcG
zz>w8|fa*Bku0*B`t3P(;<N}qWE@9Q&n`W;mrK1iBHJ?kmOEe8=cYJhGw04irHFsES
zxEyI<EzbA%q}qtOS76Ife%8^>gSy9n%+obAwnF1m&GnU<*N=>1?;p`y906N>9v8Zd
zRsv~*fVGW=#z=`~53oVA$Cdiz0<HYV{$J%L-)}tF{(=jH%GDt5_?IXDIMO}ekBd`S
p<N85|j{nXg``wXxsdsqjFJsMj;@>gU(Kr+?U4<xnzven#_#eh!h86$-

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_low_gradcamelementwise.JPG b/Images/deepfake-xai/ms_eff_vit_b0_low_gradcamelementwise.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..97608a4da7138f39eb5ab9e16930000769023fae
GIT binary patch
literal 12208
zcmeHtcUY6pmv4}cAff~T5u{6yPLPg*v`7(B=qeCeC<392A|j}fUN!VyLIR-|3n)RV
zv=BPdk*YKm3*6wh{qElVWAE-h_ql)UByTb^?_|!Lb3Sv*yg3{_oB^DK=o#t(XlQ5v
zLDU!EaQ4`^ZjhTZ0AOqk00961CIB6c8sHeUMh^hcKpwdP=%{rX>ic>|=FvBR1pqkK
zvC0Qvr1r5=+nUr1z_&{MIJ)~k1pr4o`BKF|A5Z&Rj+|6^WyRC|U9U<^_+6(}kpln{
z6aT6IBU75hBQgGuO#cCsysW&Esw`MlPU*ZHSXDtzRRIk6_xb<;H~<ZB0vsveaF*_j
ztDm2@s*H>mTG}3Y+rdfN@wTT-kiEBzoV2VAKwT@y+uqUL$?v>_lZ%^|hWMxE*W%~h
zkQ(9^U}IThZyhIBH-ivgCuoQX%rV5>Q3WZkrOC#k9;6!N>Fw#{XMaA()58m`8l)lg
zdv;Z7{is_;NK@Sx>8xsgRqxLc)Ru<OpKA#W43rL3kiPBfA|t1wq9P+JFC#B6MI9l9
z4)*f150dgi3;#L6RVTEguba1@+ikD&M-$pR-1hg=5c2nTL#jI4JA)mNNU)THy#i87
zP7Y}=Wsg*llR_#eATNRCl;xe}9EAQPk97Q-y0^ct$8Xw5M;Rv%Cr>9YKQvVuaxy|P
ze=PdHizZb>e-r(8#ZTptI!0B;*U8?`=_>WoJW{BXELciT3HFzgrB!4vsmuJEaq2Qh
z!uXek{`ao@KO`+pb)=)J<I&o0`~E>5>g4lZ)3%43=8@_6gF9;Dj|z2KD(0A(nbl?f
zmE!*?g5Nz<$WedyH$bVsJpcco{QpekpH1%MMcvtf)Gd5C0nh<3($h20(=js8GcYkS
zGO=(Rr^=WG$bOQQgB!@p!wuv<bCzFR@a#F!b7#1Pq=iH;NJw6k<UKE=AOn&Y2T6jC
zCZS<uVq#%t;XHnv6U4{O2l~G*hb;gOCK^*3Zdw{4z%dRQS`M1SHh=)tmC{pP^pThT
zv(Ox)rK4wHWMXDHP90Et5^#)$mi8DOEj>LQ)j!kRrQQe7anPR<l)J)k+RUC&$cIxt
zJRz4!Sew`iG#~gZqTt~BfSHBs3^xz2sF?T#36LUK>5{UFs?JqiJ$(a12owfKSXf$F
zJ0hK&U0mHzXg~jez@Xreh=-9;j~+k4Bqk-Nq^3Q4j?K$2C@d;2DJ`q2uBol7Z)j|K
z^R|uD{;s35YjB7>Oc@y+8=sw<UszmP{<yNbwY{^u_vP!i{qIM7(Ew=wVC!$r{taIo
zRKAYU(b3W|9`QwUERcH9a?sHW$}ya}V#a9ib6Q9~oQYFAA(z<7EUaMu8R+0Uz``Y>
zI4in!#M*Dp{$q?i_)l^6H)H?sH4b2<rJ*X1mII&#So3863CH&Ry4QOM$Y*n0EBP?~
zD0VMECG<5fY9_0GDD0zi?Pd+R8LL7Z!aHK8GY#N(tVxS0xz{+%71m06fe_2Yw|wx7
z2c}@QJ#qlz+YJC80{E&%BtB+rnNlI$0!z7MZf(1xz}pA3>cS436J{+Xp|OPxcPC!D
zJDjcDO0pVR_Wq9D`v~<*CivP5iX%2SrF?j3bUh|)YL=Ijfr-%x=q!|L?HN|tAmG!4
zU6A;dFLZHqwGio?aPRg?uUeHQT}$*gnD<-7E~Cb^X=3tIgV8jL$3JmAnsYdf&U-1R
zuGDXSM%^mOUh~$hsh4!%H$^?EMl58B+??`7;ZBPg&fOR86;y|QnaAr*1r^9oMbL<g
zXN))%H$8g?Zi%|k`pC@si|RJ%es9OB#PFl~!j5`nAHRYP<z2Nzz440kCc%V|+d%E7
zl&a;ehc<Pej=f+}(YU%fmO%u9B=yAF37w1aY(CXn3O>{Ig0Er^0pF^zQ3r1wHg!8H
zavqprip5O0stWdz4hws@@w;y-7zyrDn?SMc53#|x<J_}zsKHYUsl`Nd^(|s!psFHm
z*6`CGwc2m-7u-bm1&UBbxTdaiEqi6jLsj}<RTGYCG5sBOU}2%OwCSwvX6BuxO6cT>
zeQ{g3&WWg|l0$%8H^xmjA4fa{n1kLrRN)D2?wF`G1jq7H3qj@9qB$;MMn6;9Apcgv
zD(F&z52ew{|I@P>?33CmeYz1Rbx_<mUnShBallyUJH>9NRHxTTseEJfz~9=YK0Wq9
zQ+_9#m%mwSmz0v_pw@ytl}rVd=&qcu*waIR;D*-0E5}1Xe<W~W9-OGbBgr!&634sR
zu<~W>^{#@Qt;)6*%OZjCCjwjX8<V-%)ft{gpUZcYQuzghqly}0)c9UM9_~_NCZXd?
zS(#==T*}|@E7<0>pVHP~3y}>h<Pr9uH8~+hO8IOcyUc)_zVX~<aR%-&1G~CFfrIMG
zLu3oNoAa(O#)*}a$nf*QrB^5JI5+aVjCovO&@yuxVZI<Y>&`#Xn-P7C9@Q@VKr7?t
z@Oir{jp)JLl&DY7Fl8=@@Y@$=0?LZ?#ZQ>HHPjF^&Ju%*V)VOvFMs9Hx==!PF0iaN
z2%yfWOgoJy-Od_zwmF##;rl*h=6PxKj)G{Up1s}JM|VTK59RqHLCvgk-MJT#a?f!e
zf&5~iaNj-8-CG}MwZ3i2X;|MtHXnVY^W5GW|DMgSKCu6(U2!h3ufTz;gByHt=t+|M
zEpwlT3x!exBq(Da^f}L&7`hRu=!h&fO=k6qxsKw#VrB+MAYH(q@T~|8KVdnq2BCN>
zGtf&NfoSv|`uVg2m!!Yg&Lk{8SUt%lt~l<t>bxhhG8W~L^?a#G`Nieb!c*GEDC%--
z!#37K8LdkdsI@}#u3kR>lJ@*&wUlA=Ae}d3bOGi2i@_PG>LCdBsN2kS+Xd>Faq)}p
zVW@?2WU1a+Hj=c2+XbUsIS9SfvHQ{=UvKkQQep9o3JXD9=@=?3*4V%A#~lK+IV~6@
z=k(AD1g!<X;jpzYw|}vYSlvD-{!B3OX>R`;6Km&hP96YL-9;)tzfNcsnHD!`cg7u*
zjPUpVpomkNUb=nJHm&vfQDvyH)^w_tDcx!7tBr^Z)(?gBV0i>~et#=*ex%R=FJYBc
z-4xURwI+c<Un>%Pnr$Z<G5BhMJ{fMf!Mbjx#|A4cu8TTZ@pGR3YZ|w~2>%|fsYpyW
zlz%#?&j`c8T{&|Ae?Kp>J}`(4$-iU#$WK_q0B2mxtcJdBX;%Q@oCj9`=L8^YDU{Uy
zokdVW=0J*A)|RgdXB)RN{thl#B8%R>?+&L?j{)r{1t^r*IHB!}B@f07TZolStp$AP
z{u+t4j(TN+;G=LpSX~ygo%v2NSS9v<2qMy+OnDXxyA!6`Yp0z3LcuIkJ*fqF#SwaN
z%Ia+HFI11FsC`??!zMJtig`N2JBUuTs_2Md8bJYpce<9RLh`+a`ZyT|8mSD_R&K>h
z=t-@*X!^YzDvp|rcza4hENJe$mX2w4bDO38cvY>qTLDebaxH4Gx{25eDYP;f)U6M9
z5olg2F?1bBWXhws<#_n&x)OhAM@%=ycJq9;$$V2jt&@&`b_f-K9B%1ela)EsBg=$$
zyu$+#m#{yGyIbSa>wkRbf>2{~1NQ98?lqfn<N;NBx}IF<_o8Xx30E}OE_vFf&fd@n
zM-hPo(y(7c+|%O;3A$>W4&iBvxA#wshK=-ueLtC(rYEqa@ei7Uzg0Zp>B)m?f`?!K
z2_L4m%ONA6psga@A>d}!A>gBPQ^#xaFUj7%+v{$l%lWtQkYBOrwwS78?(u}`w&kxP
z77{lyyh>U%+ZzPJzuzxrVD4BmE{c@?#<x^OSuRCiGi^+?s%u7`NEx;oFG$$d^`%q>
zD~sdZbxN7^j$2%vy+ltae*UBw6!+D+yvY|GjqqPA(=Nayqae`K^0gnoT4si!(73ja
zT$6HS5-2~R*SfyRZ*gMrT#VD*Tqq80khipXxr|B2mY8I!GqGsudytP)Tv@!5@wGhP
zDcJ9xLkr?pRve!ZPa|s3#)*+dIS=X^=+w^a8m%Gz_KlT(mO`BMRP*O`ZT937DT+SN
zMu9UJ;>B%9a|p|Lra{Trb6BB@vEqB1A7|Y;C<%V|xfd+-gyxbuflos+S;QE>Kw`;T
zg+;1}<s|(i$fEDFn`2C-kb15JZrVjW__1@*ag2q41lF$w(5^pkta^)&it~EqVh`@T
zbrUhVR2$P{2@VoAcz`_|S;w=Y+ey44Y>}YDhlWm1Jgo0BlX^9sVm%kYvF=FnGcU`-
zI)fz|6$c_hbV}zFOb5hzij<TK($l~YdCyNNY4?peT+|1KA?e&F8jI!<&h|!f?-k<b
z&_t!djl#Sdk=p!<pQabX)|Rw>F`1cA9gYgsyWEtcYFQgIOY(|daxt0+btwz^RdthU
zN-U1woBvEu-T(5f4@xyaoncg$V_fraJYwUTZ0<%$%Z_{B7-GpT_D)+xicn)T<<nzP
zDRDC8JaE8Fl)FN>zXfNmqkD~pqk}tokiIw^vt(TC7ozkka-?Cp;mQPFNBE*d8kjSQ
z$w6cfn7(8f^g-^;FvR4xc`gu$w7r&wwbk=hpb$D~<6kW04B>}e%*N}ABF*&iZuCEj
zwCWPFPdqFCQMfT)yD<;D&gqbIB?;lUX8R6i<|Wu`X@-dsL;y+^7rCO%=P)dlEVg&d
z!>?f&?HlqkF>;qJFKd*#T`nX^-e(EoY&B<*=3>DZ!HRN2q_61)GhFF)nr^e?LlneB
z534i+4dVy;;s<&RdCo~{E*$(aed%R)e%u)S)n!?8_yuyD_)dA%@5@ziVm8VfY@wGF
zklncqtZ;M|vFrP0;iRmR;V+Uy=z5S^*!WOK{2ZGOVs37|<EgoY?v%iRPJq_lmcYTo
z3;Ft5TWUw%|M71#AN&nJa9VDUCYuKpwzn^H5anJ{$6YG`S^V^R!H;|hSlj-sAP=p9
zM}zew2w5oDk`(fUFZI>y7XqIUs+j35x-dC$V!5mYY>fXEyY&Ut>w|Q*S!DcnG`8<W
zeElRiGIrs#>2ha%txa77i_-9MKhgFO3i){L1tpAFWPsy2RKK%hs!G{Ml#^msJMoFS
zGhP=S4{s>rDL3=dl!HdV$>bJtsv*|;mg0;n_pC|L&^w3`U0y0iS`AQGkYj%QI3`H3
z)g%W{tGA)U8#(@fX%U5!D4qO_%D^KYJ4qa@^rA-^cDO0K8=1qsDAU^P73<4O&v7{h
zb_RVw@>t*EVu*#nEIRSscoQu8MFDNB<Ur1N>Z5FH5TwVS+`K5lPp{2J?~`C$mb4V8
z0O{)JALyfkTk8*S8x2(dYPSCy+&e^@JQ3$<La1BmAz)<T^>-dZ-Ei2zyTQiNe71QF
zqeH;Z@6IN=Hi1YJCo^d7@eH{AhU7Nko%kGQ3u<YtZ0o`V4?%lw);>CQuFmW8(`~iS
z0<EX!Jc(KEB9K0DvD}6i@^XOBuHlCEn4Vp8@Om0LR?<q3t9sp0(AV7%Hh<qB--1|k
z#T4wIiXSYFX<W(~Z4|Jsi8xl)QogaxN=Zie%Eyw2Rjwg5gYcs`^VsA#T2jng%=Ms>
zWLvxEC__9Wi+9lji8lp|gG=exuq-;&dTuq&1&%w0)A1)=a{~D!&}+uMoNdA|X^;a*
zJSRQn!^`W>hQEHY8p;4$>v`nM&5T^avA=N7PtPUYLPbPaaM}xvJT#db+a<gopykq7
zOZc38=TbIm-6L-L`$&G&w>Z_;Z@Mfb>Kw0!BM6NdWHr=ksL^{ZW*2>dE%krgh}Kc_
z2)uv9D8UW*fsE)iTsUKN?ZvMz(hdPSV6sk$9%Y}-k09Je7srWm3exiU-7ktK<5Ats
z89u$9kKYcq4MwB<6f#+2LwBd1ySJr%+)^4;u26r~lo%lV4Q(BgKN&n4pQvC}-HuSt
z+AZ!%j1D>15Te<|SEwB2gPj`M*p}CIny}fAW+GtQ1*E0#N}eX3);lHx^znL7#Imyi
zHz4VYzR@mt33Hw$7K7Z)1Ml`WY(c7X)Ul+7VXIEdNT{21HnDVa@QWHw!<?=~8Q4E|
z<E_vGrV2p;_l>6V@zt-C#OO1^$|S$`z(IX51ZAB+D$Qb$EiEb9s@J+<8&i)V(9pcS
z2hPh9G;D~!=MR?<!=pj)6b+sRKK|X=^XJxpVD3J>)w`&16TV|b4a<yYoI~WWiA{>3
zJwp%g!-2`MHHQF2-p*8;9bNI~aP=<ruFJOUR}xE~Ym+_w@?SF3chA{RlAm7KBi!?J
z;5Om8U*7$iv;8?oqG_#l1X5*my{AhR8&ee5%rYagML(JCW4V-LckOyg-TgNeW3^4f
z<Ko0AL;n+VwD4;=IbI!0#EZLYozGM*8h*tE)>%~AY^o9MKp03L8f*EmAW!n=Sh=Si
z+))fadV@zkkyUoD?B4l(&(?EmnY<mu+0YFUekQ^vlPVXZ?>n~Q5$(H|?~D-Kl-3qp
z#+wDrpb_cn9{v_gWBw`{+M{nJyoODhOix}eiPbMeU#4uM3=RFo*`edm#z2EZK>T9d
zuTcF%KrV}fV#Fcfbhp4?VQxBR>WoGY+g-yg%_2r<Fx4KqeW_zx+!Q$kL`v7qs>4g=
zKFWwv6Qy3ex>wQK`sXAYWu~SiiR6r#G$93<^}L`CFH6=4TmP*Ki|m(!2!pF(;Kw&2
z@V>m|E*tu5w$Q57M}>5gQ2Sd24J5xxG2wD4d39L=Msy#fQ+oCUD1}s!b<dIHTD1i)
zpySo+w*T&xMsv$b!P+7ulqnkoKxt^+BF4%Y5BrO<!2C%8VtI+1!9meibIsd5hHP34
zylIMU!1;Vt4~)_ERfay^3R$%w0}%Pm&Z1=dGO;=-U)S5}N|zsv(<4$TQ?0I(v?ks$
z?BH}^d97{|vRFc)C(_Dr)p)VSz`%nNnK^6A0De|&FkfhR5*f-{HsC3Cmw`7hCN@$!
z&=H8#)O8wvV1U0G9WRN|&v`auEa5<};Q$7{hokfm*t_|7C%YR7@x@6b%`fI9-p0t5
zUU0kRXDw?AN3%jHn7gj?5Kzs(yXjzvJNfONV9YOfy@3Zh9ZN-4*I#Sgt0iH2oF0V|
z_4{_<ZI?b-5)^-23wo+_LIahALiqI7vq<V4aHXEE9&BK{(eC5Z#HAL#VrR+cKr#!Q
zywG?ZzrNIGI^fS#K+CN@SO2NU+?uAwSF-b`WJxz`o?5Jl`@s0_H56c%uz#tS1D?4`
zfTpK*3O5IuTK9~k`PUn5*`SYkvFDcY)~Ml42OF30Vs`>xctJm>uk(Du?AZyeGgX?R
zw&U0dolL-wyvT=FRW|H6^ZU+l1%3M-x+ON^6&S^{Mndm}K|Kdq`UuZgT)&93pMbcX
z@k9SB-ppMO0o#24*tg(k_yP1}>lSSo=~`LX$ij#3JY{m5>uv0gh&^I9=jzxYAp0mG
zH+nFxXInDY470%yM~n+<xr6}walnL<wm3njrHbj>1_pqra?fh{%&|4wq?p*zvAoTU
z`?BCFcJ6W<O|7OEn=8&am$@w?8uP%?5Xbx;WS)c7&3g!0EoZC%n%KCBHSFX3PcqZI
zJGW|;^H6#=dc+*!UAw2B68*L-|Ao<p!rL|VwU<F(an_WFic5$q2?)VqE}F<0a@A_W
zv~Jx!Q&d+bOU8)T$L;A;#W3IM;J6E8ZzwSYjMv=sPiT?l!{>lo@NnG4>%p!nUkJnW
zzySZjH`S-E1=KWM2)3n^2e@Apd{l&zkrNMX5XnaM*W2MDf7G7NI5oXo2#L<?$cb~a
zrI4BySPF#CR!QoSeI<Kb3KEntWjFP=Pj6VaYMQ|+l0a^(OQh$GzDzUlJ9e^lIj8xY
zn><zNrLO=>`JeL{Az2jcE-e<}$y;07Df?zA{(0P<l*{sR?<ECmic+tD5&;?tc2y?n
zp^fc`viEh&Uzu)tjKoE^b~Sx9m1H*?Of6ev6EF@QOAR_OweB7XRbYYiF)a@^b+g@|
zOe+X5D%SHGcL(1`T5z3_A=ytyueHrdP<}9`z02tZzQf{boSW`2J)a#P;|5l%8!jf@
zr@y|Vv~a$qqNI3JZ^!=On@${2vpu@tv?VqV5Z&0Zf+Z(fsqEu4XX&0sn4qJ-Xt=hL
ztg2=7pL%Ry*8RR7=b@XyEaRm@Q$-Mo9o_wNwRvQ_u}*8g)wX(klNUSChI77iDG69p
z*t6-71|r`6-#x2ztkzucKQTHrjLBEoM4~7ME|@*hy+2%Qqf4Rlhho(GcxH8{QmWn5
zN{ss%I^yDa3g19G*yKLP#hYxiinhpD6g6RFFB{!}CUlfPwnwbe2VfT)<Pzz99`4I0
z#46|_%B~pm$KHgKqw>@R&rT=Evt6u0BBNpUPf&|)#igqn*NFTnPQhS3u?6^OtaX7Q
zmstqx0I`KJi1M+xbp>lnkxX6k7!UG#0?)p7E=scGWT|z`26@Q*j6}C47BZ7`jX#xU
z#ZIF3CWO8!cG$#olZAmV#1)zW&-E~Kzpketk*FtU1<@^e{zjd*5heR3&pryrH9BS1
zW0jM?xRjHnF<0(_aVz$gw0gJDov18U*9ko{Yqb^!2~(ce%Jk&LP!hUr*m+~uG5#u=
zDKWtw@85YD;@{-#^(wX$*P?KupYsGC7N0?9Vx|rH)G7Fy9$c)h%@6q#<UCoesQd}T
zv#PRpS+;5Azq5qrwxjlG2lIR^8ynTC;dChpF@DXkVSc=|#SDEpnI~i%+Zbug+|TM`
z&Z5g_KH}k++p2=%G`mtJC+#M%efFP7o>%MVBWlJqca(A2a#I7Np5P^kc{>ZngPYDr
z4l96b#B=e)9Sfju<JPP4u$3>=pf!QbKXvU#?ry#!=KBlj;7!vt)c|2~(FcAuG#(Ie
z9y9gw=2ViE2y)Br+oc}`>TEZxz)McH;|HI2S~jkGyyM2J<T#6lf5#M_|3X0ZC#H<;
z_G1(KkuD5`Hy^-ee7r=?`@Boe1K>O#(#`NSUWt45)+c);#XJ%S0*m<psJvdNWCUc<
z>eOC7!~HjD-gyh7MzB*E61_&-^wznG`=yS=)AeADM$t(4Tf?ug>_S7@6c00JVOu^H
zcP%iloV~$JtTDh+kD<u={f^StQv1fvV=0}k&6o>pBtcYjQb`onSXjA)|GL_MZ@Gar
zZqhK^$nDo^CaIMqa)WfI@e+(fRLSrX*fR2-?TDRlV)dC#jM30DsAQ9jPH)!SfT5mS
zyxm{PE64BT)%g4&pl$ozU#F$5m~RPGLF<Kj;=)FDspgtcMf0O<0jafRi~6wl>k>7e
zDi43{Z@WM5&UQSQ4+h`5-Cq%;6)=o>7^Q+AYhTB-HQ&5g8{o93^0C0D!ngWyqVn;?
zpt2+pUqL_huLNjYmSA&byl4{_V=3+i-)rplxuK+_@ut;r3DWAFZ<f7xda)j|Zoo>E
zp+1;P^1bU`Ubmb9YS%nVQtm7RL2nWvbyf_eFRg2o$YHhBh`aC;HtO_oNoA=(fz=m<
zsej3Ou#4rnxbvWxp`_iYn!XQelB8w=N;K)8l+divy<u>+f3s#vRHVM|94PY(wxq4%
zx;y*9>9frOiON$^YVhWp9GbbbiXGQ~U<T#A^36v-mr}6ho819#MNa(sg){RM3Ni_q
zYiy`xRS~8XX4?kkt6dQvZvukpb3ek%KD;pvW=|5NEk^apb9zZi8=Gvd&Nx}JyykX}
z=328I7D^r)VRIn3O-v=``x^~8SIv9Y%BD?MNt05;LLTV#3WkL|DaZx1J2nglPb`)m
z_b=Pdpd?JDYDb_c(V_a7LYAMyXlhFAql_Y!gA`01&9Z9AvCJAeiq46N&qX(=X>qF7
z-p(IsF*Pildh{#w-BDO583Alkcpf7M->SoB?;oTeB^DH$s(tvo4{iD}wFtVM;7bS5
z;NkH$<EE?-L}Qk~#B5G58n^DFKGnYup(qtTdM{}*%W|StTlbVdxQ>>6Jug-OwF+NW
zIs4<nq6j3tx}(Ie*~e1_He&$=V}RnA9;H&|i?t&+Qa)YC)`@H@{V0+_z6HxP7UW*J
zVN|d2tWtNsZ2azcIK3sLHND6(8diqlRQycFG|HlE$+w2Wb8_396FdhB&ELLM=1q`g
z%2QQ~{!lqsUJxg*S1|t)O0M~qBs6v`4)xkm`Q+XP{I)c$Qvbc$q4mL}B=yI$`(@Rq
zB+bt*3G?7OZ{Mz8vU|4zksHs-$DXqo;K$_BU6pd+i=-WVwsl_?)LODY2n^Wl-9W;e
zNV0VxmlgQ49V=OZSI=%dGX}>n={$v*ps!)w5kv)#P+LCo`n;$4j-1J<rDnJ4s9a67
zq1mJ22<(&s#z2P>Zs5vLlDL^bJ#6m&>#*6}b@@QNPT&v45Lgi(pCoUd4vVhU`u_8G
zDo2^fIT=^z<7+GM&A8Tn<*w`@;9akOrRk5d;M&_xk0M{hDUAKFHO{`OQ*x4+WIH*6
z<5BTIyf0eJ^|0V5;o&l#x?EvH3kLCZL(|xpcDm#h3(3#JC*X;W!snK%GzX0}q<JjS
zx>k^D(8+vm(nk-L00sC=Qb+`Xrb5~sI{fgFlu7L|a+iB%rrGE--XQQv4=Bm-YJQ<I
zFYe;B^%hK=!m42b7EQM&nw$!!G)8UF@&i#^6r@)>R_a_v;M_^?`4y~b{nZZ&2J&~;
zqVAeo%wsKM$jfIjn2(s`l!yC{IJNLQ{OiJA-(g!*Uhm~>V`@(00DP2kc15GWV?p>}
zK_H>{oDuxhX_t-eX7`h9Y^t)eQSKRyKdz4Xo6EtBNt2#mV<lPQ7hPbYeP=Tr8%Xek
zmy%p;5Z}wU+^>ksL{r#UvsLnW<`R5Nfq<)e645PE1_yiA&f~Q&SV69G9~kjMUCe8+
zh3pE&j>Y|^5hO>e3UA3Id36%beNzliOG2UL76yz9aUw7C!#V=q{%H_7v6p^6D*y9n
z<j4!1R<NSkYOkQ4j~qR-Y$_qsnc{urH9zp74gq_IfTLYRX_@cp+^`*6uz5Q~BxysG
z>+iJ%`bg+84mPU&Ku(Sd)$QGxrIYFdlln^4%Y9mUL-w66K_2+JunK}|os`#gd;qbg
z$<eT5S?_HUeNh^dBXGTf34P*me6OUEn(mEWmodJ%iuizq6Ow%;agZbt9dm}VuAk(t
z;kNgCL}kTa->e_<daIOP(5L@CEU}Ri-hl_&_S;XY6qu{5oQJE+tXLyd{bUYmSTfgj
zkmQR?ll}xONUAyMlH{GD-ukEwvHdhxyt2BSSHpfJn#27N;KD7<IB&`?%W`iYkIY6#
z!Ei=vTC?XKh_f%TzfZ}jiBo$l*2tso)acEu6$49ZGnUv7`i_lsYUY&nSIzWW8gYZQ
zlt0C}Rzf(km#vsfy2V3Ydh|&j-z`>+U#wfUQ&otVgxz;6G&p0AJF8qI3v3&!EOVP7
z3r12cV!CX@KAX!_XXF~<#w~?jryJK&2~FP0t7MRRK@!H#3*D=2s+dc(D^t?X6c2EF
zE7MD<Xx|4u;cEcjW8T3KplaCjL#vz~m}l4JH;*0yP5iwtKYrLg_fN9o?NEQ!fYfmF
z8vz{l{m4e@@vhDaH2joIOV2*LN>%S~_gD7lm$AuhkrL=WzaM<dw0|+wH2dOr3rWnH
zy6Hg)cHyf<HT&$l{vfT{(02~;__HOG5(c|McGb#0Hm&`w{eFai;q4igQJx7(W2TAT
z`;Yq1Px1>47FZPcmDntPy5T;#iy)DFOrhAkD)LsRxtYIp0Rv8QsgUjRsbw%H<>`e5
z*wbE<tX<A%wV2brbQj99KR<<hI4E07J^^xymnoDVQeOC+5rMafc&m28x+z=3)b=_O
zdU-$Eau*p#0FB#}Un5xuK;a9w=!I;A0|4;A5On2U*(NI3U%|ufAo<NOPbiH2l7@TG
z<u}mEmBI<^`;@P8rK0f}WIE^O()M?3hLcX58#T{b@$0_+e8mh!Q69=mIP#Z-sI-eA
z0eN!10jC2jl+Gs!;U$qUt~DNj4^)<^MWPq>u|X1}VAMZg`qT8U9c4Vg{?n9tYV+4|
zMG9u>gkG@m^d*Yc+V`V}a**O%x(9o=#DB()zxx$tO!Z~hi3T<2ha+MEAF7LDLq%Xj
z^semBYh0e4B%dc|A+ykzAxcD-FHZx^R%~BprFt%jMg>;{(wKuQKYOnGK+8&YS~s7w
zy?Ts{^PwXvxG^Y9T3E3TkQkM@%Y0R*P3RnB_g>fdCfPRen6*)o-+PS@8aQ{e5<)0t
z!)qP$QRgrWcd!L{4L*Ej)&kH+OFkBC=?x_WDLcK5VjxPm>hkCcVyRL4*>UWAw8Q3v
z+9y$W;Y`Pxoy*C?>CdhaJselyMo#W$UPV3J>d_CW<!jjkr&?m@7EP_08($^HyhQ0H
zhfsKL%$MShhcv8RCiMDfsttUJ3+cRsGrpAfJW%QC&%NSb`O}rJCa!%M3^6~R?i}*K
zYRBrqN?b^a$i?8Cs}oNYA7OnlOfC3Rg6r_Ka2WD77XLoDjQuN)fK>){K4Ntx31ObK
zb>&pHfXBF-`J}IkK$%M)&*ci0q(3o1+}gjvRAe4oQ^(`QGs<sC$&hK(Gq!(l2NClk
zw&YUju_WVF4UT?Smjm(fuwQq7A3A}g=vzm2*TeeAH@|b|z38!Mi<Ac|Z`Dy_`QKKu
zUOM7ZGRy_oGu68o^Z~K1Iq^H#n0Q}w(nQW`X*75^-{ks7@4@oQZ}k%E{lA*$_M_OG
zZ(KBpHlziL9+dJyv@^0%JuvtB9Z`ofgZa{}Be`eaBlO@7#H5$XKYPXsvw{P(3#4ow
z-^Hra?l@xW#nVpiGlF#|F1inv+HIVRL$gQ5DAtgXRf-w4U|m7Qk82oc<BA?rvXq)9
zD0(itj;P8(`wm-rQ723CGG13C)USZGJ@H~dS(QYuhvAtUpn}Rf(DOYDyw}&&h95l|
z&q0#p*ex`ArN2L>H3uzWc|c02Hw0aMJ_mv>ckVh3r_}~x{^_yDL*wP3S(F}2181jx
zH}pgb=Q1<mc}8?|L@&&-*RnwiY$4R_#_w-_N4c%~ehABGQ|--a_ClmpU1#uIjWpr4
zFGKLgjhB)aY_}yPnlSCuzafcUGtfn!6Lsb;_+;KhMxqUzHl*pXw)J>F3*JyS6t`m`
zS~YYU_HEt$-qH`H9G6-*H_aQ+p0*#RsB+z$SACo-1Iq_D=HC7Ka-mF>F7oHiqI&%L
zZ?#2xHtuR%x+9tVLN~PTKyesNzDeD0`E0jWOat$FKHld1T8!f7jy2_6O}nHeo49R3
zJ0xxFMiR1YU9`E+psblP#!Fl<i(j<|A_>rgu2^OTf3c;t=%E7SiU93+x^OkoMFmU?
zM#FGDecCZVrZKkoC7g2UGqOH*DJ1iW<-@&Vacdt=j7kYHPfEWZmcr_S$uO5<(v?dC
zi%Br$=;|54WVlwFpo&8R_+#?ZCs2&_U+~V4Va6x(v2Wz>_4OK52H@Gxn?Kf73pW!Y
zSwE4Oy9g^uPN%sJ>ryipBB!G+d&yn5WltVX1ikE=0{cup;ogZwfw1;FFmZe!R2e0x
z&-truPo<<jpg6(~rOz7wbMUl*G0^oXn?3ePYH<Mxa9O^9gACiAQcR#<Su?UXOVCco
zR55Z+jpfe1-sjz|`2On=HMZt7o9)}@t@*2Ekrw?eejxz+B6~~uCEX#QJ5b=;R{64K
z%596=-!eahjkINd=O|+xj#D76YlAvs#`DA#zbxg0`9{EIyyUm08wR>+DMX$E&N&H2
zYZg(+Gu}w9LQae*=Vy9tb}$GFP;X-O2kn@SOVAj+?NxlcVbKNiN@mJmm`lD2>a-0g
z+UP8x<F_ni#$@P2gz{#zalH_(0!xx|2`3FL3tqWvxDI$c8q*LzJdhxQtcTz)A`#8)
zAZB}prd+8L!Ecs4YjKm^kt@^n!u|61<HJ<bbZf8A$L<Y3e$cLrmvl&B!poqSef&XX
z8I^<Yg#1`(TYs*BHhb~uSt%byP6QFqb}<3<z=6oNi`poPN{Iz-gVPzU2(7~~;Z5Or
zN<bc<3Llj1ymXLv7zth&HnOfowIi9cvxfknNeb17P<`}A2h5I7_9@Zq!?FJb(nwXx

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_low_hirescam.JPG b/Images/deepfake-xai/ms_eff_vit_b0_low_hirescam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..9a282e711c8720fedf577cbc790bf63ec47e8ae0
GIT binary patch
literal 11089
zcmeHsXIN8Pw{B>n6cHr|NRbW+H4vl(0U<O2A&>x}2!tSnPN=a0TS}xiB~ql5Ku~(M
z0Md)p(A{(eRFtBCg#&x<Z-3u8_xw2b-t*k&{x~yP$yjsEHOCz9m}8By4kr)i0cVW$
zjr0L@baa5wqYvP4fqohu>fr_en3@7W007`5fRRoEKz~Fr0RVKy(H;QCBbx5$Yo3*T
z{10FO0O-5d`2oyFeQZbPmyZqr|N7C-@zwun03;q8a0CNW62l)javkB7lf?KpU6qpj
z2h9M1001c||D^wkDP77jjQ@k_KQ|LZ0ivR=09991xu^(Lhp4J6T>|`PeE<LwfCpd!
z#|k)HU_9?080e=iFYkkwb9VK0!OCHLz2!rl{p1zp6yyOgtx!K_j3+kmq6_whhmWS@
zX7fwQiyp3;lGadD1yes=th<LHApmPhFtfrCJTYpnl3JJ9Szw{+q27Mp*g)rtq269T
zc=b?Ckw0cvKcbJj<wY*T0$knH5qkQ6jW{~f6!~i{At51hAxd(-0XO6o)zsAF6(I5u
zi0sh_S$vpJpmV6K4_@@I3G}ddOn`@9pog!|#p4N`U3`NAHARAgJY3b?oZX-<uC7p7
z7iT3`Sw%%xXIW=gB}G|RB_-EOP(@XUo1%-zU+i5me{=T>3h?@aw<|^->xK2k`UK*S
z&`^{Yk^ghi|6M$fAo`o>e+vJRj7MYCbpx=@fmpqx<MOdWWfh>ZiYiwBP_i5Zq5zZs
z5946+$6)-+LjQYL{$E1tGRzgDjyYbtZ@{1IEwTRpciQ&yxP1KR_*1&0lRt^0X^-~Y
zsZ*z5^8d>5e}&+Wo}(>ybl=|_^ytou{}0;#&rJSl<XE30%MLj*;lmk#E`XVd=>!ua
z^9iOCCz+W~vT(8<Vax*LIK#%t1LWi70rH%`ASfw(K|oyKJdcQ+h`5wASO&~@QC>+N
z1d#-RLC2HOF`qoia*BnEm6Z#`&%+P;-)@Jm0Guc3uF>%@(1`%(Iq4WU=?-55gpOP(
z(~*ll_R@bfI(i01rW4F3PqDBb4X8Z>pr>PCpl4)YVq!e<&vdtst^*i3na&C;>YO-d
z;mj=J&jkrj&O0f3rLq->82&7#<Ps2ZiiP_;4=<m%grt--NExbfNmWf<R}ZdlU}$7)
zX@x{t+t{Kpu2{Dl?jAUNU{G*KXc!^#xBF4iF|ovy)U@=BhmSHz`2~eV#U-U=N>z1D
zZC!msV^drE>yFN@?w;O}(Xnyb#N^cU!s62M%IdrKYwKUOws&^Fe%sqWIF^eJ!0@M7
ze@pfs<l;P%i=L5@fsy%GE;{;<ql1Bykx5wb#918+W@rC%B9QQtTvw9wDqBy9Dj_}t
zT>^$#xW$wg#J?Pi_J?HuHNhhOk0kqBuz$)m4PaxSJ5nA4CqN4@x7Tq95WWBOXQ<?_
zxRjP(&l#4^|486|xjUbG05M+jl7+3xw2MA@diNBwy1!u?!Xkc34>8`{C(6ogv2_Rl
zl{5LOHfTT!#Ki+I0$omqTv8r0`C9hbpi87OwK88XsZ@81I93<e6W>K&3Kj|4FqV6i
zB~Sk>K-uv9*%4tuJjueJ{iVOn>R0P%mv_qDPMI^`q7$iD350$jsm@G98^_%cvaV=R
zuAvVye__)n2nqqa$Uh=^MhU=el%8BqH@lPWkeT&kPC6XOTLPpfO@1`Qc*J>(GmeD0
zL?kVlIz3;{df5M*t~7pM6R+LZF@`wBl>>@tMuZKh_PzXh9SxVgI_JUqk@6tGGHSI#
zZ+I{(s5ok9SPp09stH%ttuvGUrm!E}(Z$G8_`z+M`{$rGS#hM&d(*~5myZ^;;Cj6%
z5dNs=EMl~2rvONCL!Cj42;r<y`1_2cqTn6z`{JfwjSc~g2P4py(njnns%F;26}9|>
z<Y73F$!2b^{a?^;r1CGy-S#Z{_ni(P?K5#(>^wr@@}fCS&mjZfoo&9OKozF$I60Q3
zv2nvL7;P__5^Li=%{VN<0adzM>nDA0@llsV%f3>h@J#H6r2#UW%&=7C)LU7^6@FgR
z)W$A_fU>&mHSlV&X3;7<;GIMvtABR;>S78C`AcS;w_>}^N_eV?#8MeJdy(Jjnwlm^
zoR^GLGLcb|@wk$+>!ymg{F$p9je!xJ<1Sf{hCThB<Ajn_AkqmP-1RZL=rU+cm@0Z=
zI=Z2h#DwXT5{ToDWWKLl%5n%`_u`2N9*~&)PJGN2*Y=*dzh;jpFLf8)MEla)06=1#
z-v2ID3<=DYQ5q>2^SRnusWIA8Mee5foWi+2H(8E}4Iqqd8Pj@s0yixg_bIj?UJT>>
z*U+yz;n|lz%8Tq1h;_XN!2Y$<w2=P9EVm{%?@7@jpUl(?#GL87;`_Wnc3`>0;-|Q2
zQ(MwnU<$^Acc-x7;>^n-nz<N%WTbWE!)mbuzReQrX^EV7GGwWmVO=f@cvzQh<Co}P
z`roe~e7^RRSKROG_M_{QyA@0SYAy3i`pI=q(Q)9E-@-yfdZ<Q)4+m2<n-K}@(7ezW
zpMR7x(<;h~`uFRV?^E-Hadf-IZgH!P;-YyXCfgmF2Hz-R`uwTkW{`8O<-Ky^J+XfG
ztbCOk_v5X5J|d(v*XUqD&P=F8VpUXlYW`TS8fAU|z<k2ub)$@wKs?W^yd&rW6@4K{
zOJTs-(UWmmPidz-<otOEx7hr_5V7*<%|)+eg8}S|m(nh9{W{4I<wV!3fj>bJM?0`y
z+-FyYFxJ<jMtM2)ZlsV~Qb3ATNe4PK3pH%Y*Knd4RbgQa2=T``_g0fe+`5|q5mDtu
zpS$|;whA+b1xYwx;hy4Xj1??|UmhKC6Owz$0(n+V)X5rJz!fjzO<-~%wQeI@xrKj7
zqhaz~GP!c2R+(pG-#AD<3~M0JR;qpP-3DtRB4r!lMn%1?bx>U98!`9ft93cnT)|XJ
zrDy%fFM>1Yeton=&d#4+t>x3&K@^Mve~(H?xuI$8DJsfkSvdfF?c;UJS?G)P^dVp(
z;_mw%aLTAKBe$k$&>`S_=Pd`%Lx7|7A)r94!p=mf`G@<fAK;;P<sVHlLdT?;i`B{t
zXf?jMUalc9RFG#}Mt2Kz-$z_dJD;TQ*>{7$RjszsVaNpKE{2n(fH*kM#+CU1oaDpY
zj7K(A#f6rs+x#p=?-+S8+O`#bqTO8aXJy`m7Dx+879qd0ceE}ifl;F-aC6s{GQ%qr
zfCap_GSaueOY4#rx<z?q<`sRdo(&l_qR-2aNGR<@5QLj$`-O|y$4X2c=D;^UAP5|T
z1@0`lSuPawG*l@tr&H@I;q=&o(~ETS3f|zM?QAn|q?xrBQxElIpxD;-?`vq4jBJ5u
z%`#KxHtX9U#g&$WtZ|U-6o21ZnNV+^$#?ptaJ%}XRn0jXIo!TBuLmsZ{=$8>H|320
z6MXq2wDuWalp)*|88`5v_FPvhm{udmkwNKIVK9ZDA)ijQqJa8S{nn8%mgfOex5ioT
zzi+EGStY4MxY~Hl@Fw(6o~8EJw(gYiQuFauH?@m>ACD5pK(JIBN{5ZYI@oUD5CE6`
zm}OlOMe?i15L+BZA;v$tHsH*uB=Bz=DtF{L_R75PZggnvxTg1dL@!?qI??p<!S7ot
z$EqbBt2TKQpV7f16m>aT6G_0-0(;*L(e6T@s{0~12z{<I!q<&Qc?GQE6*lLdwGODt
zQ>D1(!TmRI+{)yH$x#`c<|4F*!U29GZEiNrs<#<{O~nP03lb#?A^<M~QNQ_x6XAs_
z8sfKj)lt83mSb2{OCOHbnq`yuI=B(#rHfAzsO>x<fGqRW)G9fe0Lx}F+(Vu|KS4S1
zQ=H{D>oWb+k_$>XzL|-oa@q<h24?9BP+*Us5isa()X!59-;(vy&1lROQXM=`xjzbI
zK%JA$oGiQS+h<d0GW>J2tH`WA{)g_?Q)5+Pz+R1-xK4c@4%2vH&Gr?tVfE}Xh98Ln
zs4*0K%3MWQqy?)K3ECCQDz5EbnBXkXlL2M-`dC`?k9%;S-814eVhnLb_%@fiRR&T=
z%R;=zxg7VQ;YnpZWamP>>-!X@^l&Ie-zJx^7TdeS<(h?hWSSwPHAM*~Mwz^HY^%Xu
zM8CMrPOu>aM?S<8%bzdDk))0cE%5R6^LLlMYYzcWxefutDc15u25FM~sSa8loBJ33
zp}EMy3g_!ZFpkfoTWX4@$2zhVZMG7BzeKC|*&A@mu}R%PyA*vA^Z0Gt-o)I_7af8t
z!%OHhg_Yz(Zbo#uuOVh$bC~TKLp2&#YrP+`u%C$K{wN!l!Xy++tr@j~k6zxrYPxSN
z7D8=R-Kun`Su5?(@Cl%0jZx2Wdtz8!18NLt6>}UAmE8>_CjJo6^m{sfqNGngNMH;*
zAaIp)hdlPhG>Z9_fA}Vfd${|1Ta^_K8W`W0-xP0a&vrtj#I9VvA96Qa+4f~H*S5Q2
zJ3sC7{m>7(X*{{OJCpQ}2H$*3@IW)0O$m<8a<=l6gN%zu<v^N@<fS^fMohSc)a=4-
zQ+MJhyQx)k66NYcm&$71g=d%3Msf35XW;f5IzZazom^dUHQsf!gzEf`@(o&16F^n7
zdo}N=dZE48!tI2Z<}#w0bi%B8X93>!R9x^5Wf13rt1027=vCFsnPRSUl4zNY<qkTP
z)&l*m1|>$n9^bn;VHj#*Uz{>iOSfA=jnb|t<;R~>T}DO*qi4zA=Ro51$xYHf+;84U
z&>ZNh4Qij(mo3U0kVEbAuyeh2FTZ-}#j)|p{$YF=nqb0vVf&o%<}dhP`hC6!U#|UR
z-ylY;%zuwgG|m!B^8kN^kgrW&>7E=bbcm0fA-0$FY4itl=4V*&Ess#45-INk%}YKk
zIg2OBg_V=BvhpV~@nNfq(>T>ud`CdG;ZXhlBbKQ;xJ=$2F!KiItP)x@+*Ct@B%W8Y
z+I8~PMNF;X!?n$Q6v|^gtRkA2I>lU>Xz8QzXp-KBgKBH!()BwqDt?688M=C(nCnb2
zoFv*z-s9nWeg^$+>;Tza;Tm^3Kyy7qS}T^3eBI6S8QK|WBwuwW=m%xG0ox57J=J-2
z&J9+Tu=Hv%sZ<jV7{eP?L2iC6OEtcP)&H1$DV^)pL6Bg&LXOhlg4L-334cd>e>vH$
z!XQ=2MoKw{vE@j$b5`fuf`+nA<W_xTvP0U2e*TM2$ODt8_fHx2D79=Wla-rBw3hYi
zcaM`tfmkQlm=DOrQm~qynNlrS>#<(c1vEO-i`XsC*2bFRdIsDa6O+cpu%a6JOO^~j
zDW9mK)>+IIV11pRg{O9b6J=?)d&7Ve@X{U?|1oP*QSUr$Rmww)aMqvFmDibH`LG-L
zh1k6unyBV3VlBb<n5SXM@kHsp{FZ{2nZ3rNSQA_SGqmJa)NX4IQ%ElNAz<07?k_7&
zw8vJFy|j$q75tD&XSCV|tb58Ggoy=E?x2UqsJxKjEzL58&RbPoRFll1WFxMJua-$8
zRlj8{V+7(&3++0TQhvV)qqU<AxM`MewyBzljROI|>vxlg&`FTV2&Y9&zHnn%c6Q7|
z7$WR_Nvdu2x?C40OL-4o5;lY!*m)5XB9v_(u_f_KaeqSet6ki;LU7P|lP_j%s~Qa}
z`tRGyPw;Fw5}<aaJU2wBlCk0@%mZMKN1tG=Syr(D!*NH^hO|dTERQSg^`)zz&mLf^
zg|!Loe8Am1$;ECf3AmVBA!NoZL|WmFS7FLz_3CQsZR<JxgYXR9?<1LhiaDUXTb6wD
zEN8fX%<_dkBG6<zOOpb+fSLG3J)PgwC@bRK8*%T|pOsF3lKY^gK0o7-alOHmRrA9;
za1O4~@STSA(|ODI_H9a)qpry3_U($>UKK?`ifSd<Dk}zzqt6=SQBGYetK_!;5wUSk
zkKzcYlvV2Qw`g&1$A${AYeLZdE*Ee)vq-~FwnIRP%58Wss+S6hf}wUMlnj^2#f3>T
z)z2_(<NHyvxoe2~=BlNiLhJA5J;f<6nX!xZr5mV;iV#hP9c?FlexQZEpzQ7*J4#=B
ziJv72wQp~s4*|(lGrw9a4*|M_?S1Bf4m{!)o2iG$4~ST7ZKo%P<uwV!xlgo07P!Uu
zP_nJ&;0p&9?z1h36)jE&MwzX9*SE?qsD55-x)9Vxd=X}salpG(Hj4nYREA{<rh8F!
zMaC|Ow4lCp5*RFNl1tx%&U@!q`POnsK1ZB`=4?q#NCM5g5p<JxkbSc_grX|UlCium
zl~l#PR(n!xxpZl|thS^4XDC!(2C~vn`u6f2$Fy)ImuWHfG#{`J&E+1RwCXk21)$AU
zpDobD+2*_bux9$qEH=Iz<?=FQx4cYF{K}k?10V1tM4L*s_-%3a(TYlTvS~qJ=R&yA
z5R#k)yJ<};GcH0ju6HV?Klv0dGJ{SpO66Uxu8^zD!P#Y8ggVWcAqSV<7O_`>XA+??
z64i#<X6v`M&c6OiftXS5_<;4CCERsn6jfzZsaG9*PR{UKt8-*ETpuBKn;^VyMWQhH
zq4{yqG^}A+Y^cCZYtr;OWP0}gTsm}O(qVs?T<9JjCY*qp>%@VN94FjKdRU;dC{yXw
zn<AYO-RzOe4%+X&kJG@IWc!Yqo6yr*&mABfX{aPMHxE1e&3(H~mimixM}x6;3}5ee
zE!k<@_4Ay`vlu*8-CL1{&AWH!OMZG3PvizX><zwVX0LAYOVG2Uo%E-FW-IWo_<eC+
zvNQM}i9OJNq_Ol6z#@@WCnVGxM)|g?rpipp-mY1qVwEA8v1%PA=?UG<4{(wF)HGEK
z+u+I?-)-VBs(fhnD#NEyK^Rqy11R;q;febo=>*8%zLSgvYw0@BUKtn@y*0%ht&qY>
z;k6@iqosG~HtnpN`2D9&+>caIqbj+$VD6Xb8@sCZ9-xRu`ZT4Lkj5TpO<cHtggUJ_
zv+!}^3stIgUVH))=jmYru7!gmVpH!H|4h+#UaeT`#+vg4fO;&!*iPI>YN`J036R(T
zLWtiEY)o&*+jDvM^f1baN2JP|B!;%xs+k)77UXemx!CNbxJW|4qR5~hm^AK$4+L62
z52gzB2Y|9WQecmX<05Fag~C{`>cq#u`}pFpa*NJcYwyT&{v&yh>{R<iBGIC}dMkl*
zkZYRA^B}M;a>OOHRXVD}(+;8|f#EGn<Y{rVCY4R*=)ReN=h!yzbKur@JelL^pTSTa
zFLLPC65F1_0damH{s$Hyu?rW%ud4o>Dl=<<qRrsqrBm@-n2%4w-_{S`(b=IGdF7kF
zLXL~cfkO*O*augbx->hy?_D1B%te6~qRl-Wumg2QuR1-)+BMoAR<10%%%(yl4*`kM
z2g9bYO0`Y6f6rUf?5miYi*1pMGnhkw*>Tn**zg2gy;t-7mIF>}?T^5^qijM`XhE;Z
zSLd}Qr_LRZnQBv5fD}oc{M?}DYyuAEG_?Jrq1b26MnK+W>;~~!%OY76+3?0#uh160
zSYp-PVQ^-V`nrW^U5RRBo%V6$e&Pu*0Nc6Wduq%F-6?paW)`~EK9U?bug>!Omi0(#
z&voNlZ7UGeD}q&@cMU3`b<MC>WlRl-fptY%>NSXD=n!pQd?q(y_oDms9yB2G_C4Gi
zh*2}LhF+~@`B^7CW_e>TlWbq0!)vnV;DL4qMG~)WBFU_}ISvX2(<KT036a3c75SnI
z+V7bp+PaH?FPKGghFz$b=|^qFOO=n+o%r7PK1bK9PJ4=Ni+eIizc)^oDt@w<?ci^~
zHHpy#>-zMGXiXLC_(ZZNb(UngOnq5%%;uBJvb(`#L^cw`TEKztvL&8&a?dQW{-X@w
zh(EnM_u~lU<hF@lE$0pa`bPGe<yUMkJ0~&mpl|KWCqL8zHLPTPD^skS2W630`unDf
z@2`zTKJds^Ub%g>c+M7Y!Si*eL?)*=D~#V-)Bu;>{qX2nq6w-eBntr_%ko*gV6u<h
zkQN_sh(h-~7DcJ-d$dyWO}EC%E*5Ht+^KOP@mN+HFHb63^mM=)SdkLBDRhNhE@~G@
zqH0jiTO$R|4TaoLQB@1$0b+!NINZ25c-x90y~CsgY7i`q2q>YNz8+xnHx+rzf3H-n
zgqzaH7i(xLs#@x(W42<GR0|hfz)3tFDeF^QO#!=rV<8}l96DBf*5k8W7<KKDmEZ{`
z37kGp=|-Kr?wg#I2ev9b^_A39T(#A7vnW09D(&`4>g}Zu%U?S^N80;H{N)Sfq!jnA
z{Ip$V>zp8x7O&Q_-*G0v_(vuoUi1WXjT)58YX-v{XY8l`$k;87)#7MVcRw%wNKhXr
zg=oUBG>BICBQ0W}mkSR8FIe}+k(AdPmoT?9z5#psyDH1LQ=>Jmq5GO6)(vf@PiLE=
zh8@@BzRpFQzB$29?lB-u3k;P$t{qH(w-(<xQJk7Sa0MRFBVgeg^JBbWc9AR(bQAHn
zRkG^)((MNd!WCRV{ziDrg|<Bjm4Dm?HHU(s>~zYhY%ZnED+7k(#@_FzfTD6DpSV6Z
z;8udXjklN%0R~sD(eg*2&HEA^BS7?7-nUm;4WQE5h86BO*Gh(>v+EGAeH|zY82=T!
zk&G<#3WSCWbbJ_@XCO}08>#%-Dt=`j5zOyK4Vwbyyp$$0rroP?EN~kzSY@+h9g2b;
z?3NaKNtY<$EevtA<*LRN%B&D*94iF1v$$VDa$hO!AjFJLj<N@dMwyr_)^PUOoW0xR
zsY<KCv6n13Ji-PK;s$$oZSx~rwb0<C*|EKz#61&P{2}1-pNQKT0y^*1(r!6FxclJa
zwV$fT)#1i$K4s1%!fJirAt37cX@1?GGfKM(Ox;VhC<`>Y+;HUHb3F1aOx6>9`8zcK
z=W7oSw#T-%2203V9!_7*WIz0LuIsJF%5|wm?{9$9+7z;{^WJGvt!IK(;BDzgB_Uje
z{*E$lTsU-|ARirAv7$tIojCBvS~$}=a(kGRM%Hg<<9qJSN4clUHB;m8+cyDlg^g?8
z&2`r@M-=Epha679<JAP`SagmDn{yM9t|ZdXRiDQFy_;jf!nEXdU(680PFc-PhjNrF
zydZn3?DJZ8*)|)gxgofOj#Ng+A3R)q$6VS0@E&x9U3sVU8E^_Gp-P@Et6VLLxirFE
z?3P5*md0TOrImDGsAS>F^P5%+dz7L|?Ube`gJ@C)t6NGw$2DeX(+#fsP%6XS{5S$i
zr);SH$JIAO6K=dy)j3s7aEbLhHHf1OZS`A6dm{IrVOm#(pxRsRUKQ}uy&mCfJk#O-
zP{UKL?RXd%7l&4;{U+ZMWgzoTY$OV(JD|*KxM}zHAUv;yFjN5>^wb(u&bOR^p0$Q|
ztb4vwE+Cco$}OPAzSvtD+!BMaSyB9nF;+y+>}(6@#}c+k=44ua^J6@tKqJ&hAZf*n
zUBlnS2sWYjg@0_HkVGy5Ct8{TSzioW-FTjI8PzZ_;4~JNM@Y6`!fgiwFW}DA<FExK
z-w~&&R|R&&E1OBnZ@)V3mu9S*FQmsn3%zd}3~jnL)Y)Nb!+jblLeuYg0B}6w`Z?}P
zDx1dl#M(U=No!9(>?|u47d=&+k|?Eqi#Pq+5c7S!N9SF$Qv{RAaR22EzGe5}Y~ggn
z`PIDOa%IxOX$Wq|QbT^ge3+?xWc18x;fcZ%zz&5ZkbbGyMFzCqAV!RkH~`o!6@SQv
z<CGk5urT0SAs13Dr74P9#&y&V9KO!b)G_7|@I)xD{4*ZSNiTC+UllqPJ$JxkMUY3t
z=BHAD{a&6wuw6eSP`Q7;0GTh){&b$!{3fpJ7Z`l)ui9_$$6rag=yDOWtaA!C>C8_F
zMm_3VGyeF!^MK^DKH)}xzksJ`K$*mASnb`8x~C<3!!BwtIE%1YV=z}?Ozgr+sF@ma
z6j{$jBFe1!>op206`&WrtS*j*mr9}bDZC=grDZJAGjto;{bBhGs$AW#m{)-nwAOe;
zW8GvxO)pS5%IAqy?$nTw@xV4oa3>nsMaUm63~b<#D7Ja>AOZr0`esjvejX5ROiHla
zf|gWgKXqD7#9Nfasv=8&%X>_GQqi?3Eo5QsT_9ej=0JxPd81nLp-ZF$`D&FkS^mkb
zGq1yl^dMN8)oOaNQIxMt<D(_%=2c=*kQ_YFGWHDutLEA_ifk;d>_adRKg4+@@EU(;
z<bTxR7`tr?+AA|3Kw`ioZCbzf1v7%nvwSb&+vs1K%p=VT&DAX5MZtUVTz5J18{WRV
z0==Wew%0rQ56>x^vWnTM*jQ5Rilb_6v>uoLUfW;>#d7>dYWsfi@Ne(gvh~!t3GTcC
z*%kas**)!vi<Pb+CbmV>=}R}4&e>KC*Pfx5%Pw`q0NJel_c4;6$xAD@nQ`TwW&=uI
zAlx^e#rs~CQ+Nx7j?kz0v4q#_i&=aW-wwkRgj7?>kX7G><fIa=4I`IH`sAG5+bYSl
zX6L8})?RH$^iAHzC!GRe<V~(qIs)s^3fry}Gl?(h<GlfEeUduJmkWZkGiBa}n~1W-
zCnGBwcR9-RPidi_k+?j1I?Uu06iL^W`99jzWt4V|T7!FTutZSkmZ*x}!|KSWBrCIa
zU>~B8qER)TUmi0Vj+|Tjv;<`!^|tz7+BS77=I+oH%h|1Q!ARTOdYX`KY(48SqV-%i
ze&CJS?!?I~e0G-(UFP1}?AXZK2ergp44T{G3Pr(Sf>r}-NF?kcdGu9MV1$ND>Gybc
zw&dliBxZv49(1J7n=@`gGlZpGjdhfYw<^qU%=gxBD$5I=HVJoNYAN_&*Gy+mZFy`6
zZ1(lbe?|83)0b-RORqMjtZ}}RAiAsNH^#^3#<gEotohr2Gy5z`2M}DEH{)`*RFAUG
zKP&NXVkoUWomptbU|gGF$Fke5@~BxM_0vo<-v<?<J)(m%yG^QRK#!zM6pIkqqOYxp
zWMMM(b*Zb{_kB)u+OSW(;h-!houYSgv4}zWW6QDC97S{@>9A3T1LTzR9EmdYk<3FD
zW{xDj^e%;*;a?`uEd7mEy{W=#I&++2Ijc)^r>?6l)v`*JBb#O{i{`tOizFq3S>YI)
zs|h<{Z+9PlZ8h{+zY4vzX=!$D={e?S`N+?rcsjgOKL;Uxb&2YBYI3Y^l{4kERJDtQ
zjLq;Bw$VzfSl6@uRz<M!#sNfXs)K?WNHKFjPPI7v8Uu0NaLEfecN!T(UK$fKVcw6X
z>JXiQH8^7pC0>2V9AD3!Dt*84O#H<R7X!1*Qi6rG`k?<xrSe*-M!uegA(b`k63Pu1
zDlzHjnR!(GGX=6CqUeb!gQ-grcNY}@8878pKl$PM@4`>!k&Fv_iVD>{aK$WnblNkR
zyhMfV`_8rcUWMAiB8;JS0pr;@RhzN!oZ4m&mkf_>w&#w-lingTd~pY3XH`}!Y)wAI
zE2d~%BDf5hiu_z}?R{tB5}W34#y~trx&xN*ikSk*ODW&G!~k`rDrTe%$Hzho9c}mZ
z!Dp@7`3bb@jG1~#!E}_NXFk<%kcLwH%|67<PWnV#@>?nl$!d@=9<HmitNih8+CXtm
z841Hr9k78981S)F^RYO&PoqWVlaETnSLWDm(QcQ2*S7w_?!Lb+G8h0JEm2kd-o&38
zw@*CVCYnasDOOl4a(layNQmD%(K4mH{IPQMhPUzg6J*6QYvxo79l9LTf)c`deiq4a
zu`It?SOBMkn4RX~;moqqwsy^*|1c<r6%XVoPA^%EkfK-MG<w?{O)Kzz*`C{3<ZO`)
z-+30)Fs%f)z3NB9zIs>T9>+iPOik*O*f}6<*qmVim>zGbaXsEL%n?$vr1k#KxQTm&
zwz&TS4tmk>6uI}f(BS6chFxqJtLuZry$iV;_FD$*H8)+%*#j)H$$BGanjjwn4XOU(
ztvx$b=4AShaZb0T`Ionj-czV$bKXn9ohYd+=IdyFT&UNU#KxCI^id-#>kV7gviRYG
z$g-?NPnIA?w|n5GF|W)racQ=|9w``<RTnLDZ_G^*{@SfZFkOs0%3%6syxmrGk>j)b
zSKc+kXf>FRl5Tpx%lPP#2q(YPe<^jX+q#|*RsR~iG7fGCkIOalY$_L((vVHSAunVU
z6Ie3gB5_vttkZ&HM=~Xilc<nu&8BDi`x1&vMpfr>(GmUw*W4E{46uaSb*G6JelF?r
zJ+;jAqMIFVH=r&D^XH!NAB>xO7Pq8kKhz~=@jGc0$x3*D#mqS^k`W_Lmh|kkmGmzC
z{?paL?-J1y@rs^_Iu178GiqNr^qeof)j~`8w%n<)<{cxvd+pX0b5Uq#{pG=h3Bk^$
zH?fj}qV-459{3?(=xxJS+?I5GcjA>4`xsBYm>iD}coXIQ_z^yjs9aNJ`!Y7=NGfd9
ze}F;oInEp>#y?g-WD(@x`eNZ$hz11D4p(K>N-$jnE88lMp0Bl=mge=KGwHf^pCr-E
z9|9D%w=|^0WjPFynyO_Ubj-v-nKFVvB`Kh~>1y}f=UT?=C61O4Wttu};Xv;xpKH9M
zgyN;p&jTZIKXESr<zSsF-lB%hogZR|_q$G%I9ohhtWH>R3#?2bAVXqksZXHiIu#wx
zq{rQ0&15?X8%bcY&)_r<3#!WO0bz8CiqGhf&-mMQTC44QiSsCLG4Yh0)3v(QzF}wl
x^8IS%#AYJZNAkx_p`u{tuQvoG^9K8y-JH}J;FtZse>)hqPaQGYNa#A8`ZwaC<`n<{

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b0_low_layercam.JPG b/Images/deepfake-xai/ms_eff_vit_b0_low_layercam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..983d8f16b19d8dfc9b9a5d01afa37195b813eb7e
GIT binary patch
literal 11790
zcmeHtcUY6lw{EB^f`}%F6s4DhA|>=9NS78u0)(PeA)yA4A{JCYBGPLpN(n85-USuu
zT{==M^p12X$_;z(-`?k*d;U20-t*k&{&6Pxl3DYuS!=!Tn#nx*W=_UVz5y=kYU*eL
zC@3fZe&iS6<U8fKhM%(|0HCiA00RJkvjA!e1pp<vMgss)=!QB2sL6E-^4lmi?erT!
z2LMpEFS7&AklX0VvNHJsurHH8PVfGA0U*2_uw)u&qNx7h5lH4MJ&OA8dRa{LA9X4z
z82}(A=AZgMoKnP`()iz;{xg^)Zb(SW-+;=COLK}t<s~KMC8YuXQ5yh&0C)i$0H+B!
z`A&Vs$=lmQUQEo*OVq~R-PS?W&fQha&&EScT=a$*KvBid!^RHd;LU04aL?IINnqpc
zdjU>odnEx=sQwLok6R8-&e{Q32Y7&iiCqB3PR?FHMVWz4(NEsb)x*`n+lJH6)y2(A
z-cO1959{*e`f0Nmx3VJE-cjCIUGpyuvZTcQS1i81zM{U8qVCvxV&Zaga$+|m#3Up{
z$QmME{%+njej;vOJbxKbckr^qI(v9KySs6o8nm%>_wiQZ_VIDHmv^*rgxcEMLq%+D
zB<)4S#qDiGZ0se)MeHRd?PZ|ivJ#Hsw%mWYx3~M-yN3_f<qzNXc47`L4z3Pv-d<!j
z#KpM9{tWuR$|sqkzn%Vv^pi1?b>weh9c;WE)XA6fX+lMAKt;r*P5zZ+Q7M@liempx
zPf_fYjDHj8e{JOdDYldq?d|04PGfh+{^=g>;Q4=-Z5L<d)1Qt%(UFBeD`d0eF?a6V
zIYqI5>+yf3;ExvakR$i|dw`OAa{fOk|35qVXOcU(k!Q9qc?zFQ0B!-!(9qD*P@kct
zp*?%%%vn07^JI?cK#UjZnOH!qS6M(TSJ<uzaItalaa>{H7Ukv>6oLptSUJTc#lR8*
zU<mls2*sJRXX(z-0neWYg4tQv!T;Ok<O6`|EQJXLh>C(6K*>Zw#YAz^47g5SN@>VT
z^yw=7SD~P!qNbrebM_qFd9p&~MF1rQ6%{2l6%7qFd3~mMNWKrCW}>;oC9X<)8D?{a
z+Y=}e6rFLFN3FCGWZeIQSJD>y^c)@Y6_%^4eEb4}LSQMVw2Z8r{4I42O)YI5UAPGX
zX=-MUva@$^yyxVM_VV_*@9XCu@a%bTNa%|&Tuf|Sd_rPUGCngan~;;6mtRm;UQt<9
zT~k~4v8lPGwXMCQb6{|2cw}_!^Z57anc2Ddg~g@i&8_X7-M#&v2ZyJ)C;(J{!ulK8
zzr)2u#zjd@O+|g?6c+`hFZrTkqNd>zr@f>KJ7eQ{nOh?0EKn^vqqOlHkEHPrkS(^K
zj+s~LJKyFhv_Fvj=YT!^KSK65uzzrk1L&zJ$jPH(0;m9X9xAk<(xyUR^D+mOrh=ta
zcEH_{(~f2P?{un4CsOZq1)cyFeUiU#&pS^jX}~H6JBBBCo_sEH@t$H=pRPcDTI&36
zK$5G&tx3L1Y?Xg{ll_)!jAUo^2>=9ZnX%GoyD@$O_{wksXfc&3n0)yl@ZOsiqlzKb
zHt6ErJOe)mA(CNwb8l9@fX<RvWlC#6xF@rjtIodW31FJY>>egReLJ|aWMYHu1hDC&
zYe8@d7!jU%kZS4xW-^(;OE=c`>UY@{hfebs%q%?ZQf@y193xPXZLAf!b5)gXN3Tx(
zf+%0~T35R4o*jM~LE}}^RJwcw9b1OC9Ok>He;H!jC)>smd0J%NVvWZhf3P^lys8@k
z*+W=*5SaY=HeM?88i;jEf#>S|fuF5w@_hLxfG4N$5U<-%PmW5)Ul+)bB}=^4vxMG9
zmGTw|8}QE1)$e1AhLKQ}gYV8KfWT7)J!O^QGupx@fQ^t-Y_!@l;el=Vg5zKPE$n$p
zsR|q_$HENo-+tzQ(PbQYeL3{E*}p_nNX2gU8QLOzq{RI7@=&|CrJErc*l#)vB_{wM
z3o-&y#c-}YZ=ru%#tDFR`z70ZejEI9;jUnsrkQt<-atIUU`I-}f?zIp%)h)`O`kk*
z?NR74`kc?(nxSZR#m8ljTgwSU8LvDzH0-}%V6NcESZf2ibZO%Ttv8#{ye?fq{0X2f
zfRHsUE8KU~{`DAvB(qNB0gUX~Ek>Tw49%0Aywuf$sCJEz*~A?t^}lo;Tg+vVMOP{@
zUuU_~acw?i>>=p=&jOcpiTBpg7LUJ*oOTmn`0=?^yGL4wWnho)Fe`o#e@|5<veIng
z98L0J*5geCRiOIt;KsxlnHMtc)5Me-r0v<lw{J_H^QN*E#4WWzcnYMZXG$j+UmVsh
zxjLKx9(^VAk^koBvx3fV(1zFF{s<4`%H5ttTa3!#rLD}+ycE#fnQd&)mND@&eb1>{
z%7|(;_t^PP^C)pZR%IjhH?()cay1kL9R1*%Q__ze+jKYY_?%Mi_ueQnTly>RH2)^{
zo9lU@FGN&;?!Upc2OInL2qE4{ET$~|63%)DHl=U4#-VjA%whWb;p$&m&r`WrX{0jq
zrSJF8$K*arLn&ooG^~k`4f~Yo^Y}-}uKU4$*$Lp?b#nZfwUQd1N30KVa~iK1WYz(s
zM#?tD7>W*1@b`?Pg2nbOm4SDU70X)GyUq#mne#+%vTwY1N4Hl+k4EXc{d`KFHdg4D
z^<ma@#OOZZ3P<>lFnP_VdfiVh!@6YQK}N}Et-<uO+m@9p{I#~eAMp__v63p(!!J~A
z0qq4<%T`a=@ng5$MLFk7UhxjkcBam7T;*l<DDBR6B)Pm`AVxrBy7~1)(c_{mldW}N
ztfn-(+f>KL;AKgeIQA>#B3KSegBZ^wLc7FgaNGK~MW4yt@5l?sRi1NvloNE#iDKSz
zCp(|{D|hN$<+u=2M;-$AZEPrb0VMup;6PZO^$SSOJK<v-Fg{sr*<vQPK4`bFHD>9~
z{QcvV`%X#4P@T2g>CT;`*KbQS<G3r-$Gs|=&t2cn-^kE_&qd=&H2JT}Na{aykJqQG
z22;`xSqdcf5*Ll(iV{yRXg$K92Nrsau9NQ88PbO;;u;+1ChzBJVpDwO(R`gQ9B*PU
z(W88qtOC=B_tzaG!$*Ld=U7T-@u^z3WLKHt@&((wMNd1Icx4%SY$sH=zR~s7)Z$xz
zX}4RBW@%2oSja2tU7U>5L&!NgN_ANT*HoIWYbc7Y#}~xJROtBwwRQMo@yRo7rFkm}
znQ>OxsPe12LY}vA?Y6Mn8doUWLPFOuneb%$Yfo!W0QBTJS3?@>J5<jae}_@@ZBcKd
z;ebe<zY{3p_5CL)lBwh9L*+C4p$0rA+%~LHYu}eoZ2UF5h=3|S_y($aXDHNZ62p?0
zSKGcqgnw?GQ}7EHKDUCpj$vfc<N;31hN*%d#fJ};*e#jNT><esoCW{hhR)VLrHG#$
z)_mt|AWxT3FTS5|!%`@=hUN&vW{&5q`xC_{6ufC)+rMabgMKofXo}D#&F43wBLd<N
z=oxv7sZIJh+k2drnYd%iL3n$!Eu9t{V{si_0S%$;*16eNo;tiy&ADE5jYH7g>hNI2
zQK)~R<s^F+J8KJ@*-DWQ#`~MZ@SU~xGn*ow2IcR}Q-_o-zkZSW*oO6)xc_Y}-_)oo
z-jCDtJWFDOPIByeDUg0Ga7d9z5366;us0QA7&`K-!4H<#Kd##yH3+wK>%f-QWnR_l
zmCH9CmztN)=1{i?nkTX?8Esl<sHIv)3vDO3Gezj%OX`In$Cu?sFUU!C+@&>)^Rz^O
z<=)13*_)O}&KKZ+73iGxVX;AdqCqz1`N-C{TYxGsI$Cj^FhQr->jGVN19j(p4qC8v
zftlwEWxT9>GLYjBF~xNIa^Re(#M|2liCq4QqN;^wCX8^@$WK@rWGi|)5m;ec`%1k9
zxGY2=6XG>X_a*#sw*F3`3VrdR>}lThn#SMRDL!7q94<RN{fS}EeQ^R99l3__ZGLun
zgc<TRWuIpedH$z11@!*-M*CcCfn19qGHM7Ox2L<{5{D{>GHi_Z0zMF!g`2pE7Ok*Q
z5Q(83(c<&YzCufCgYijT-Gd85Av{d;6+Iu<`aqh~Fdk2Jiw7A4aTEK`BdkJ;Ai5*J
zLQ6p5+OUM_Od0I$BccadOsuLxK25fV^)RI5;&*@9LbjBZl!T8wG*E&wp+4997(aw-
zm13cTsDGQW%`!|>$A)Ef$W0!!HRv81CMY9{DqV-}10yM`ZOn#E2B5HeJiqWQ%4_Ed
z;@9uj^_7c$i93Jy3}sCbDy&O1M)jy-Ci{}Sso|dI((OG@jkZo~M^=d&jnzvn9(Ls6
zP04{mVa$>#Ygt6otmc`$2aJi8QcG<dkj$+hw>T-f2k?034)fUd=pK>Wgh@~2C(j;d
z9Y+PaqLEpo>3LxKB?|~l_mMHlbT7KGpc?m9XTu5>ZYk;@n7TW38~8TZeJR$x(yD$$
z<fA)$1?jC&&X&}t>%A=dHY>qXI;{g8c{R7B`}u1UpWCwB+j2Gwn-r%BV_3@-rfY(^
z{d(uGUHwo^)!@>aZPjh4yKt;<<4!`IKZm+{M9uuW2XYxgw$r2VSVk+@Ug<*vmF8^!
z-5~tvsQp(gM!@v}6qE&;^}ait&&iG-C<$hnTTwz5YH;0Z1H$55rg!9XukhJB8N3Lc
zD=0jY69C3cogMYsHuswB!2XO}-!I5FjYkRBd6bn)v97*9>jj^f-CaS5*l<A%K1*YV
z>wHo1UEaEN5A@!qen<NW+f?5*vOgRsY<^WcO}JH8yWAs)&GkY_P5KHqgAtZ+ZjfLR
zTN;=PCvNbD&KD9Vjg#IqpiBf7F<-5KRlqkj1#8T1J~2KW(M1Y?TnfBXK8*sd$s7wl
zbMpea$ve;bWvp##e?(@Pm@oYyax!Vcd|y0-HeaUnk|xDfJEQUA#3oR{w3x6x(k#~D
zlk#qbd7G4vI>(zM_><`vo_AA><e>;roQTEkEVhE#2bMh%_>g&X7Acc*t<o1zkcJsn
zx7~Tl5+#`V#^E7_VUlv?P^XH=qxfDwZ--!?OQJT#3R3x9)@L{8E;tW8jQ)sTdHHO&
z0D#P?N~%>s^-$Ik6(&CI-54LdDNC3ncD&g%5p`2`jqk&xy|74ho<Q#qnSZPqfk_2(
zL~A|@4IcHdwIT2}lA=$yC=?`t&o#EJ&nXSFVL*ixydipCz%aW4Eh2GItifQt0pW*&
zE80RD07X_I8u_S?I6;C3Gft59Nl9iZgLL@jO*#YgmE6lP1*2PBG5J0F`lzT)d_Kej
z-A~_B&lWX8-G*8CG_u>@?O6M^<A&!3>cf2Um@_=tbx?^&z&SA2>=kLL@qVVV@a+7e
zQ*0aOzr@qtJUDMJ<DSfb?2$+ErVjb%7jK#k&~-L-WG3Swlwpd}mBV$vr)$45Xe2!|
zRVxeQjS*YN9jtB<N{y@{0+p`7$16Mowx++0+?=ki*9>wj!a5_Rc=3Du8!Yd6h-Z#@
zC6+y2;`=V}1w_nm^n7;~>Gdl#Xhwf-i^ca-Q5JkI8e&JhtB~}IM5QZxB1WGfxuL>C
zl`K7|NF0&DQbP_XsB6I;TM8bEYhlS9k{T=r&jjVtIpyM1JwK-5oe0^j!I>#|ihP!3
zitIOVmeqc|uA1`;jEM-GHEgj|9z;+PfhCtPSSPNJ1WqllI(A+W*mtabum=&Cu~KTA
zD#F0;q-+wy`a~kCo4!h?GkBl7#sv-f-dL@7b+;(^`aXX{!S#<w#cnl8einfhm~C!K
z3|iR4R#}`B(r+-KaaYo?$>CAE_()FKk=pV=qVbh<%<F!i>*Y6)il#a~dA!JUbQBiS
z(t|y$ByKmEoB*g4rNx$U*?er<E;fE;CXw^}^@oSY?LTj|c(eY_c26RBUH*855MWFY
zG>(jce$2^VmR2nHvS3DjdxDV|k(RgsAymH*<{UYL<$t9n2oh}3NHvZDbg?5eb(COb
z{47)O7PC>-aL->OJqTCl&?zE3(k3mCmu|4y<gNs!Z+ep|W7Zy-R*j2$xDLrl)2?(4
zi|=h5Mnr%Wx6MBJC45osq91B94Kb{1+Aa#zNI{A+-ZGIEYHUD1_f<`tQ$njku~JUS
zzCH@v1;Vv&I&d*XDE?UA2IRKjwvrG{mukbJge}p(9<DdW;Z$N-b6x72hGR<Qo#}#Y
zjOVT4&;>+i2Ua>#H+ZCWDG<iG!?T^o)_gR2Rsw_ec2IiZrFr}eV!o9@Jzth+QFH&h
z-<##R*O%xoTSYhJt>5cSDMg!1R`x7Aml;}PvMvcn&*^)i23KpSz8;RiW8B)qI1HYD
zE~R3buNt;&_J9a9MIR(ADn0)A)=%o2t7OFRltCsz|4#m_EfMV2Jv({PKL_zk^J_=c
zFrCjkQHP-#9bFk!4Iq*xiHW}CaAeZHq2yH`OJS0DjblYOt|a^#CI{>JR6AKZJ36?4
z{EzKkT&EmrVLesxiHJ|)mVZ@uE`DZ?rH1puZi_YXJy>DpJOnQv@U#gxrTQEcpCh1#
zabk=z<<bsnl}9>s_6aB53PSbPo5&rM;x4@{tyYU;75#1Rf`5I)%+*MOxcES=_m*rr
zXx|QlF`8I|y6{5dhujxy+6%m#HlyxWsjx;^X==qimjf>1{TWfk_jDDah4pN4k3aWa
z^As4wzp`GQI*P{Q4@B(_f?AAmhnlf26}qYH?w3&rlU*&HoMnk8RHL`t@AFz4LfcWj
z=RhP^@!ev%)`>K*mRdPwz);GH^3+S#n#?lC_Cy~Zogg*REsqHr&Px`r8dj7Sj38GV
z^V~&s5xgQEUgLFwgZf`AuA1a!u*JK%i&>^6H!hiP(5x0G^4E>6T~w2yz=w}h4HT~c
z;gyk>{GTV@x-hCQ5|T8doKr=r=ra4^OJz1*ZfWt9rx$Wf<k9pOBR&DAqMyBr`7In$
zm?MPO!0M`qJ>V|F^^Vr_$f*d8drkpmlQegCOMN`pqP$FgLqJ!d#;wq7F=kjDUDIfy
z%*~@wSla9iu5t7(-!n$k7GYDe%7Q~B_|8RsO#j^|HrP(F#E<5rV66t<z5VNpTIs`s
zNcJCL+sRJ@_ci$|4x)Zx*q~YY@0J)GN3Ap*fIc=eUHV*J8q&Dtv5$<ncs|=20bJ3G
z?posc>!9+zd{mo;GIE#KXPZlQr^xNW!)ofIk*TT1>PV^vhud7yanU37hjH@_OFYGz
zuNdSN1$&$$b+a6$TNx}h=gZV5U{>gIA;y7X@LaxPly$0A_RB?M36`1ZvTSoVd!92S
z@61oKEoxP2=agOU;#Wym8*Ke*Y7Ft45<4dV#tx6UCo1t)FZH(WfoG^qb#kmN%%eP%
zmPS|HJq6OA(;u3qK*JvJ^v$|Y*S3E_9dsLAps_Q|FetRt+WG}4?^e5eK|H-uX{x7g
zV$~>@2ekOTA=7$uH29UqqGtl~^WB<8P%%R}*Q~KHo8vam?h>kBAelv2j+9#v*j9Qw
z{qR+#r8LA?wz1$wTNng6C5hX*wt!~k-YcA#oxo3RZJH<c&y{W%$X7^SuwYI*Z-<N3
z8^j~W&*)A*9b?00k?sr>jtuWtorAsu`*4#kH?|tMSr*&TJ`l3)j1KoMU=nE?r>nDh
zh<V>IIB=-czh(^Kw_p@`66;12NYM4m*Qv=+{XSsX-5jeb|0-7Ef~Tc*?)N2;{V@2V
zr*)58RjIiRuhrPBsZ!o}i-Pnj>t2aOOCz5YcE||cz}zQ3Kwnk9HcPj{Euq6=hsRrS
z^Kg!d|3nN`noR>ttL2NBtUj;WEWNYn!k&$tn=h;M+b_OWS!IeIaorISm@0zi4~m3I
zF_m$>My=S^wd7wdx}#^v2S5*9x>$1tw=0P4lzzx4(Eydnn@ALG9u9<0&Zp=W#u46V
zn475y5;Dsuvzc$^eIVY>{dNz;AA7XwTNJiBq&k>ts*nECi@Zm@c`4_>z+lE8Y;>Wp
zcBf`^eP_*LJl0=5$eeiNt}koPRO{-lM}+ZyvZ_(*jRz*1RnBEh^@Ff?=q-GC>7ve+
zEGT#jUE$h;df^zwYsk_#ALFFh{{-wbKx(?MURh8qOSH5MOOiloUjPp2B|<*f;g-fb
z4AEh$p#<oQ*PaicV{_e-2@6o}@%FA>%o}Otp`Ehx<0M<hAh;&4bfG-1zBH;bv+BeB
z935szf)9JOV*#H!J8qz2IzO)oafB#bH+KpjdDv}^?uW8*x>c;x2qIz$78e@*@|kZa
zz&kI*Pq9YfD<V)8c1oRF*1A!zg`+KZ=hk-0E;3#%N`6D*iAe9T$Q*2t_pJAE1eSO_
zd#g#!=v{a&YPtegUqG>9O7rAQI@qBSv1C@Lp0HqA8fqI|J)N{^gVD>OQ3cbpw!*!9
z9qX&a7Y@)hY_CuaY$&2e4~E7oJ>W9fmE9yMn=+v;=F@cSkRA+&0(U>pOuzwNHene)
z;8*YtPMC-v)4w;MLo!UJDJcl=&5M!Ab1e_dO&&%(K=Jsuk!&mMw<>c<(!8_7XR{})
z>tC<KmGs{EX5&(Lr+Q@R5!GBrCiFE+)^?-tk%kT~XaH5uJxxzju0W(qHd-BhOk2hk
zn_ihgpC(s00oDE_EUwQ3WeMEReTEm6xyT*vn5F@VwMlAG6`w9t095KaQz)90j=fZY
zT2{Gvn$I2&W5-o*vI3_*;2l?S;ia@sL`Vv_f^ydtBu|u#?q$A+7T~&2F36YnmcEpD
z9$wG5R)>)n;hZX}0nwv@)h4ljdMPwjwV$J21B~!hGI+T)>cA6`XNzMa@!c1RRJM4D
z7ZS|Yt}uq(y!aFI3*lY($XJP({G<QZuy;yXuM)AjkE7H`)}!F}*qov&KVM;c4vjy_
z{h~#J!jr-au!oQ(=eG1z)YvE$Z$9rGo%jf$^?JwJrQNrFxlHQ^+XK~W{73qE?&#U(
zk4$5GTwJv++K~}Ett>hD^YwcXaXn}4mrOHLqI>L{$Dmm0LOfJ&$lMdoyZf3wUKWkK
z`X+{UL-!6y$Noc3MQiL;zV2C}0$i(Rjv?1XkM#I=rXcs$imVIHXdwk2Raf@5x4@EC
zM1zoNb3j1bTsa`Xhk&WEqK=Nz@M#fR?C&xl#+Hj}z7~VBcU4Cli7dK>;zDMOwKdH)
zjec}aq_MZhXmT1i`jj3NZWd_CecZ`iP#k4{GrihbT2mf!6*!VOP(B?MoFCXdG$|a0
zM-}sC8}+Riz1`HO$LMN94ZcT`3@mW@GW<n?$qX;&A7O5>zz`y%_{wxh8D17#CJpR-
z0GG6a*o8Hy;aq{u@;Jda_R6fr+*gajxh|Ei;T9@57oXVH(cWj;DHf1#>MY_G6^IXM
z){HB-lb@EaM!Gs>+0e}+%OPojPwVd;Z;2Hfco#{d3CoG)IXmYo8a@wYWF$`g%8);F
ztYhGTpfBzbqR&}t7-BAJ_jrvqbx8N!PaS0c<(ETY9G?^P%-S+OE#hZse|V(@EZmmZ
z3hqP-4@-FG#}|MU;f>jdj_Ur@p}T{@{6-_;Dzurudn95c!-u25CbC61$#8G`4l9<v
zZ`slHeUhp3ga`M^!_~5*AkL1iHB;JmwdC9|8G__(;nrsc^AEEbMbAL*-(d`!>8k#Y
zv@?lZ4>$Sk4+$R(EVjPuZ9lbcpYw{N?D$<JPg{^_+CFR1bj^y6uBn!BKPVKKvH2yw
z6?_Sk79RJ$L9+*MgSL6(q-Ga3o$*Lt1<K|_dUH=8Ly?LDttng3#Xo-kc+_6l!{nO2
zbgYjk#v3bMOUTZW`e5voof(wi{jfM3xf&e_`B7xiw9grBrO>Tdkor*<l(if<to!J=
z*8*v&{iqU8o=_Gkm#&lS*a}U{SyST%m4u}$qCF?Lv%_DdVWg(+HQZ)sS2X2At*u6C
zMn8~Bn#Yr(4M8&|Z+-6(Xtm=dhJ;24v?Wf97MH0EXtQVux}K4eC2drOv>7o7gUWO7
zg9A!H5|NNnr+PLJcY}j;K49_V8&?O4Z#blA>ko92Tnjd4<@v)wn97T}>%P!T4YO7C
zHxFlwh5MQIgFW8=h|-%~rzKR6=vbCB-BK$ZPp}?l52<b6%|*1&XX`)kY<gz&DY=+y
z1u-^sKCeyYmcbX6P4f!vE$3|GJi^*drrANu!IW{L?i|K1Nq<bS*SU)i9^>Loz`2Z?
zFf~a#K^(LDKq&qmef($kEe^a;Euja-J&)7tzuB42_97iabc&3DB>xyZt{j%MWwN&R
z5fWCvq=<(ZAl$G4yS1~p`Oj>Mqp;%18Qgq#GIUqg%30LLMr^hHK9g?vE!=~e7R_7{
z%b!V&fdT~mWz5ZqIFk?C2vP(xSJ5;Ub}e4Q0m@x@lVnU~8NS9(o4tJl3~>J3!@E{;
zu~#p-koZ!@v3DBcQlz0FtWL6Y71#iVGnQTSFdKe8NdI!H5*0HUnF5TSWJxl$l7-i-
zcYYnK<_K|a(yF0*q&pFc)VO&83?sHagb?FJpMTM0co#1{%@;{om=l^_Pl-`;)wjri
zAyCIV42MHqvH6*`Vq?a<bhPNJ!cjZy10sXzAw?f$!tuSCxm{Qva0{A85mEEAMW-A%
zP=nr@0ClM5-ow3+7W@h$eVo=Fe(ftNY0BqWF`_gXoHS}BTje>N%%*RAiw|+$5IK%%
z7>F~jZ4|N`d5WrF<1V|Kc!uPp#qVtPGQkTkL^RBy?7x{A6n8A^X$5~hiJJ>l7^zEE
z85(7Fz2>5?0&g)cCoX!Tj-O95;ajOks%jn&7><+?#}mTDc#7#sIusq*Mz`SQH|3zu
zBih%UKL<`Cs;w(EdrF4=BQ(6WTZJZfgff9HtZDkDHqUNiAw2<Uc~uNb=)8oSS4%*1
zOZX24HIgCi6VMFO(zaaG#X?SZL^qBT?OUonD{50~SbJPh)wR0$cB^Q~l}BJ61}6~S
z`}_j`%6@I!|AEx7;zL_xv9zzhUw&YU+0#}KJkHx<3ZQ`9Z}UohW*BGZgVtb7s2^^%
z^v<()G3F|hkp)pT$ZPTw6->-RI%I6KoEdJsqfy(wi~JE3o@;+uKUi_Z`&DgizGE!Z
zycYUZ2V2og7gNhg<tkn`;3k{hi+M-V?v&QzZs=etH6;<cY|9I7-v{ruwz_M6z{RoV
zo`0HI_)I&~vR|JT2PppD9%eae$o+eT<3ph#gQjvf1BN$cgE#y)Xl4PL>Iu7@(HN4o
zhCq6bAG+2+BMQtmtyIE9*wdD43%kyXbkjcr?Ml_6;fQM^5?KgM+<Udi&%6_<tS_=e
z9}-J37M^ajnx1BZsKRnkxMPOAk5p;Sb<i3g2IT$2kiVktlL$>2BgzQU-FzzssdOlx
z5uUt+xWhT}1OzOlMSas<Z}U})3N(4C3>M-6`+=_WkmNL4J_lv79}v&+V>oZ)@CvCW
zxcSa1Fd`6%nxcw8!;}?1U6Fb|^zLzk<FBkQDSJ^0pAbEW)WgB>NuR=?`%2D_#?Ay#
z&ZpeoKoH^(aDV*Gqeb4R4*aX751!gH=TREyTm2vhTN;G(x#^s=x?u_bq8jyY`CNwl
zv95Bo&fQ5xEoU&PG0{xG!4{<unj8JLZj|wcipF#l`6Yu=BH14RBTu2WE&L-OBGo!&
z%rke@gb;n;n@tVxLgke3^6y!{UkJ@FbPcmFG$4NrWp$iG^$`lI`7M9=7bo^+b%oVe
zIrTyE4MYnamCdH>BTM_x9Sv0$2;nTiqxhm}i28xQ3aYK%ym%|=<KalH$>tsE#5#{@
z!TvPFs&#|jRk7K0cQAdrCw`<{qcT$BxYwDx1lM8!5F3?ha~Z9K1801I0XrTgA?-S|
z+xC5EG|oAnXTE-^rwbWyBRG@-N9UMMBMM1f)X=e^&F^BSL6^-FZ-XcuZwrCWVIi7k
z<V_Y~IB(cJn=jc{^jD1HcnC5CN|Ey+N{icJ8;IfjZ<qBi%q}<Y=WC*6qlRl{dL{Lr
zL}x(Vq*|C-Y0MBU+Rf)%G{h(iXK*CX4<zngiG)_~o^94=ZUIAO*qupJ((~+GqT+aO
zQ?#?0Q7*7>c&SHI+-Zi(1|Lo`#9`>RUvh3}$b~RS`2hLyu86qmb>9FV-5i;69#@tC
z@8wWfD?xG)>}V`0QrRpFJV{_u0V3U#8ei;Zmgg-CDO@W+Hy=U|Cpx-3tt6L&Gp^kf
zHY8wKO+(jnqV40#iZhBjxftTO)I#n;gf+lEsJJ?Icw?v6{L9H~uG$)R3A@UNV6C{!
zHA_0a8vk<(RO3X2Q4!S0zD?<}QIjUVOdVTli!L`=^)f0cVxCYKP`4k7hSLqsd>dlr
z?!?_6N8_VR@La)Y0RwPj&W$$-#!hVkBw^*@%r3m0M)W{Zcw*=$gzqy@$)_U|5gLVz
ztF}Dp@e-|?;BVD!Xur_$mqFBM^dv^D4VTOudsj(Amxi09NSE>?-e6&Ucs$o<phzKF
zBz*99$xy+SuKm0_8^-R%nDa>SSIP4uuiDaSICB=HGK&ZX{+4s-rhJm7Q+^FoNnRT&
z6=$P+x#|siNdjEbK{|hcNGwJ#*u^z}rp=Do%;vxl6C@Mm@6rG0MaaFF<<xe{3+n;B
zEgi4Wa=GL55df}#6&tH-0cFD<_?x)w0foiKGx$@F2N|@WJ(pO91`WCK8jyq=10*gM
z0*a1|-N?&`q+JoBP3t72os-S}iTaThG>gb3scYAx4Wzdpq}QyVN3d!20F1Ztk>JV_
zW4nPQ+;1yIQLt59ri8Pl-9?ip@C12JOC$Q=5ODmbo;x;(oJb#MDmh5<g;addJJgVq
z{3gP&S|As-@;p95H`+nH11dU*6mhNA2_g15y)n&Y8N^jYidGzOgp+b&ighT1rwi&3
zy{Z2GnHLkqLBO^8e%yj+vjJEVUXs^|i(+8_)?^ls->xy+XZ(3n>ixmbdh#aq=m{V!
z=XK~f>;#bbvhQe3<tJXSn0n1P(NpESwX@3hrvC}xVeSdwvcAfNir5o?*p>JD<BvC`
zU^<{4(eEb!x&`eM!0VS^|LAG>)09=~_}H4f$xnX#*4(dgV?NQ7vB{PF1kly9|LfTB
zH-B89%)puB`Y&XaX7V`zr_7-3uA1*ZHgv9UZ9e$R)orrbZJkrIe+Bd=aBE%g=R*SH
zf!}G&gXEYA*8l@;$Mx2azheIIMA0WY(-tJvcmfbvC-2i&l5IOb{!TNfve5AL*WUYM
zeW39Le_zI*s@{DE8@^;m#iy8*DF^zFtKa^`Lnh9qPjvb^z*+0J#g{(~4?O<(e&GE=
z!@@r{uKd2pA@=*h6dA#(>wy!%M;gKjpzp!^qqWCm_OjWx*2RuF@pn%E3Hy)t$p>dP
zr9wXhkdJ`;EyDCHx7MYORR$veMlfjD)Wb`rsfFz6H91<iYLDoQVN)9CY1_shZ8wgg
z{T(NO$d?;OKTlPV(JrAvz8u$^9$u5~oV0#TJWZ3aO4AzOVW0m``BT3zD>vhJ3$HCq
z^DhTV_fL>o{!Buq%J~W;IpM?;z^N(nBh|^R*-4|zzd?>+Cjb>x=xM6BWB+Fn2f60#
zD-wDPCS4@QG(tY0^7!%z;6={sVAiW-TclgRe+!%dm^6U3gh}cXK(JxP<GwHd(#pgM
L&L?`PoP7Qt@Asix

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_high_eigengradcam.JPG b/Images/deepfake-xai/ms_eff_vit_b5_high_eigengradcam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..186a1bcc836446a0fb8828b52da01e57ac69377f
GIT binary patch
literal 10367
zcmeHrcUV(Rw{Pf35s?Hi(xj8nNu&r0NRSQ*2_=A10+P^4kZJ*?7!jmHKokfip(xUu
ziV~0#X`zcEAPNXdvzLqS_ucn9&$;K1bMHOReeNG;vL}1Znzh%O-*3(A*?a$<{=EP=
zWMyt?4q#(r14Qlr0KZ?e&myAmegJ^2EdT-l0C)kMY$pKh`w|xbz-E<<2XOAoZ2N~p
z(d|EH03QIr{%lPGz_YIt*l(ZQuK<a){l7oG|6KqWqHp*<4P3W4{^Ak1&sXs+&cEdc
znHhh{9GZFnKxXE@<bQF>midRq|K{{xlSx}s`?!H7+(7HNj27Gge%wG?2k;;1000bt
z2=E2`Nx<*doMQeYQmBE3Mley`3m4+;tL_sLq!Hy6s-dN>sR1xN6&32`6W~jd@%BBB
z4>nTz(D6t~29GmRa)#S#+J>I?^~YPpg!?+i*r9!50(|swN~cZ=@)<@ML<NNg`I5Y3
zqJjd0i3U+dvVVCu*q8sPHDpg3hU5GUP-f=;wAgPM$^LUJk&%(=kvi%j;pa89^!4>M
zG_^Ifwbk}5)QHi+B(Es7V4~bV9hmtNeZujfBz#D)%pZqd-XRerBiV=uJkG$+%Mb32
z!@<?Oy>xJDT3R?SH7}fwmKsh+2d4|y($n_S@|OK)c%0AQ(L*D`1OJMR^U?4P^bPV2
zCK319(9)3A_{X6Ct77g`^mowzkp6v)`!)ur!+pI-zGnON<e!A9X~NaCj-&r)lGQbJ
zbqzKCot>e^A2R;UK>uq;{-0v&q#@47z~|4{L&E<N-qAPg|1R5s_>+Hb9sfYL-}r~H
z?{<IA9XN2nP~+c5{9h^fOS3=a_OJVUg6>}_^Z%gy{~Y9Bi`+MOe`QDRFX7*FfYSgT
zE-r2^P9AP9ZeAW9UOr*|eUAA+LWcx|#X;aB;vjLcqf$!8j!G&>iiyjr%PJ_VKvkh&
z84Vo`h_(_03i;!Njfa<)?*Jc=pC1U35SM`bZ<pVX0m8g&_G}^?Y_b4$VKxq7w%<Jf
z>HV#gYk!OWvrGT8u(5M+a&hzU9^m8OH)uKpU}xjtVCUrE;^N%jpV=<%_X9YExeg!G
zGT|0Mdhy7H0kz{ZN_gc=>$^dy(U0;v-r)%c_(a9TkAM{vm6TN=F!*s@J$-}IW(e~$
z7M50yXpEDyi!0U#=j(UgA5S2XA}&NmMaNvdmY9^Ba-E!+m7SBDmtR0BEu)rKR94aG
z4;mVqnp;}i+Mhh_Vf6Mr>mL{!XH85_P0!4}UR+v!v-0-c>e|-!&hFl)&tJZN`-6)O
z!0``Qe<S;MxP<p{v2${AaPs`Y#l{}FUpa(1xsGXZA2vbqc!i0`YRB^eO*2aByAR0e
zpgw}U!$<i<<zcTCw*El-3)z1TSi=7ivcG}-3)d__fP-y6c^txkQ-C-A<qSpW<RMhN
zF>vn)wa*j%g0Oq&%H0emLg2ElRVN*=@*taRvEHfGy1K^mIen+@;lPaa(K*TzjGt<2
z2ob*3C)>{N)@T0QHmdd0tl>1&Ip_Or{yv~w7CJaWdCC2tEh?u@y$}-hZCtQ0k*j_2
zq1C6{%JC6GkIt*kwy&B-0osx0*Rk4eEqaUe0gRrf{cDfLVQiLiP9Uku?JCnH_e{uM
zKs-{sf8Fi2v#cVu<y03{C7vA7F0F^keUjx~^G@wDi;_g_rUiYu_7ie*zTab-JNMiV
zr?@1)Fw}2AP24w1$_D7{+L}wSq=LB<Pf69ux#G5aH296NckMVVCUBQ-`kMEHmAvmV
z>24D(@-9=C$v_AWbi8J``eX8-==hS7?lqM$3wl}5NpaoKi%=cMjM^2xN-}kb0uxzT
z$pe{Begt_;Tyzm;kQ9C70jRMh7-=0VtEvEuq%yT#8t;yw$lmvn2;jk9u4BMPlqUea
zRY8D({0!MpgJ%7Sz^PmHR5ivK_iexHVvXN>8(I$w<or_Lsl7l6`r)iqU3T$INgbQ5
zHE5__Q&Y?7y-8g(!`3z5=_$Q<d$hO`wDOY6jl9s$3Ue3DsdhIXP=CI>w+XGifs8NZ
z78C6AIeSuxO|&jm4pn;q8w<`dPUPosX@QXY3|Wh-MSF_{u-GZtP8D->xnJy;z%yFK
zJ<{!vQ)-2GYdKqn2@R6PVGKul<ab9IH?0lF(*#Vr)B6h&;SgkHJkxLOMY8+wb<O+=
z>)@Lc0k$PBeR>trs=on;N^d&bIUCOIcF~5`3?$1JFKufSN@%V@^SVAZRuO|wr|ueh
zI#c^4<S)LVZ{VoDp_g(tQA3ksdMEFlC4ZJm&a(C~qWyfYTJ+Ik-L&Z9n|pSYF(p)H
zZw+rZr;}*O>o0CAjcj(&j$>U|ObODzPr&w$8p0=iUPRYM*hwzWdx40uuROFI^pxXS
z)lv)T`;+#=C)8c7ucTZIIQ$zx+7UH^ng(+@`3_@o+MQarI?=haDQQcCd6P9JqD$nV
zi5lzyQOsWLBJO#<;9y$}I2h7gNL-m!-sHl5o22E}YniLyA)CjvtPyGe1z@8Cj}aMZ
zNkLj(7g1iMIz6t+-m+u*s(^dK2=3!L#!ev}F2@5iAPF((`di6{ex;nfT18{Q!*C9@
zt`})qBk1;sAI*8ku1a3}vBbh~^}`{22eQ#Q1@D}QHKWxf&*T!Ib`shPQ_$CajDq3M
z!`q-y#_`!LtuMj8QN<VeihHZC#)?hCvA4xF2&GEM&xIha-l_~0!15INJ~Qr2#g^Fg
zZ0xhwHua#a$VI&+d6k-@9W$E`Xc-tpYWh&lgG5-9&Pt*0`aR)}t1acY;YCa5hma+g
zU(3G%iM}uJiZ)yB8E;mM;tcE;q5X@|S3~z=o-#HyB_<6i{F{2*70n-iBy-a0dTHUs
z*(LUQj<q*D$cu)%aE)0M;(5T~IV;Vj3xvhFRyn|8r67cSX{B%2M`iSWUhe3NgjW$;
zPn|d)eAUjc5cm2v`D2?`PSJ_aXg!B@bR#m!j1!<~w7#x9V0Uust(W^;*mwDjU`6*@
zbZmGf_np9P^U!dF+pPSFYR8fKvo9sU$><xJx1y^(^Iy#thI7_~M|Jeg2FLE;5(7NR
z-57s$ZM~J2yCK>jpH*xHu^)i-Xk3GuSqvfV7~=Fp`K5V46h|RL54A|n<qaPjQ3*Hz
z?&0Fv`ivK{GS;tgty&%?QgY1ka;8J<sacOfEIe2<ku#6)$Xxv|a4rioa~oSvtsCNM
zyi=8A|9yG<$8|KWXW<}~RyUJ_)a1g16}>I@0dEyP(EHV*_d$fmhGm<j>FIj(zL8*X
z`l0EkQfGrN&%^uAS<tGifpX%(T*oJ$SW<^(7=dH4w?hQ&nkshNX4U2heBhY})KG-W
z4a-bZuqn|sM&d^0kVuV9C~@v(M*vBHO%lpvpsaMyoNCAc7v1x|`7oDWCP#_O#dmOk
z=u0@WNv}d(ac11OUHJJ?ua7@QTh)tdx|G3zQOiz6{OyKgdH__JB`YW6CjPsK>gveb
zY~Ul`>*gTiidSo)Ge0U<vX1aU!5-nOgIv6Kcb78GzEQ0Yk-ITBMekyi)J<+_41EDa
zDAMN7V&{I)t{)w}+*@B9U7??A?volpW$&_ziXnY6Q-x-AS-K~zO09X16)k`4N;D3<
zvf=!Ar3ejvXTU;ZGSVhH?5&Io13XH(>t?=0*8!sY5k|m>(oZ@xy>aIZrZxwr0#(pw
z^={AhptTGx^+T$(9fJ5-Z94-a7Ve|Wy16)X<6hwksGR!j0i!8hmZmqKVRN4d8kK5l
zbiJ0*`t+bNm7%vc+yn#9h~H-%#ha8?nR=7mW05uoqsTU>QB=<rl*BkKjNv8cpTdpr
zjxJ{}E~7k~0Jn4Ue!HSA5#)Z)Xq^_lloG_F*!d}-H$JCSFG62p%9{GBFOG~0w4Ogu
z=Z?@vCzggT$}S<%F=ZEd@(T#(8a^yI-j!CcZk64d?5t6gqAEp;q|_c@D3@h)Hl`Wb
zg>$<_Sls3k3x4p*YW<m-8D^7Y6oYI$ysg>`Z5zYk1HB^GoGFWPs~*jTN0OfF^*^6j
zci&rL+LqlB^WT1?y*I5WlzShL{gMP9gR-|8l2ftf3a`$od_ZN{N_`Rm*}?WiS=!w@
zVdy=#;Ro0g>0P6jnqn*gXAPW{>dx^OrmJW=>3nWrP1WXkQR21@0}+VpwFAv-z8#He
z9?xBhGVzVA<@Z9@9<U*lOXJC0dVKzseC$1+E}ff`-ANOPRBT`7L;Hx=$?w(%ox)Cw
z9xXn>^BVvjpt_ggqQA6jXE|c$N*ajZ2Z!dN&-Xge<Zqa(o9U-L?C#Y9lAulkJ>d(>
zd~eR5cwjHTS81*;+@SLml^Gmj_D#Gdkw~c(3F-uuT{wA`%cs8Q?)NFQEiUbL{Tp_b
zf{Nbm4M*<Vsm`X;-DDL!mwJGyptbSIfjj9trI{nY0i~5oKi*M8?HA9B`qR|CH9ZW?
zevXqp%hl4p%&aA{7jel{#=F0tKL*kFo9>jtZ)=*kZf9<X60FTq;%2lbN`h;TLQ5cL
zOd&aR-)c*9APQ!+(rGZy3Jn~nf8T@Zub#I(V(!b<z9A0JttdZB2o!!gZug@+ANeLU
zgXU)hsEvYI_c?@Kl`An)G1QB%%)X@!=-yrTAPG-`2oS>i-rDHiN_pXQ-%@d^rBYmu
zK7>d5c5qo>We~%U@VR14vcAaThmaSNk6b#kl-F2`G<TysA|UkC**?Q&nodNaqcf|9
z?E=X7BXh)hpEWyw?&Rc(`Pmao*YIxnGmoaE;>sT)dApTpH599eAY4wylwA?N+RRG1
z^r|&KLW#h|=Eb_ZWr+ZA(G>|{aDK3`d^osPWntpF5C~<yFmY(WS!o%5J+#~I;=1?3
zkEFKmEl+)3HpLIiQO6IiJH$-X>|kJ*0vT>|9$M3Hau?47+o-acHA4-OFEZ|8dU={a
zSn`488}FQ_UaMS^f6TWYzXzglk=i#ZZ}a4rt)$<1+C_}G{XxV~;ww$#7=rUP(HL&E
z{ZV#f7=FWuQ;hgJ3XWAAmS3S4*xc-AVw?AG!8A2F)M!*SX)qRVDl~uU=Qd$`wI!8X
zCQv%eGPhEEb6|i#s@C#v2`o%OpPE;qh}d>YdWH<4Z#q?c$~+#Y9R_=a_|-Jo^F_8b
z==i(Uyn_wBPeYGzRMj?a0r}JC`@afp)y1$d2){5)qJ6+H7l~xEH?*E76|sPjp&xwN
z=)TU>&S^mvNJs5~3W|#Hs8U+BTnw;sctcUSlI$s4WK>OTy~v&gLxCj`uu;nC)-@Aj
zm|y&h2+l1_ik5LwagGSOC(NM*i+hn^mt*BvG%<5>1DU;7T`y31>IMAytJG6V9>AYa
z?yuorKGUzKHXBHa+%}Ys>3m9|m0h7kN$ruQY%b{I%n`sfE$k7kM#p*tyRN+LOijSe
zgGUt43ydbjXjr}C&vH^|v<!N9uIuc%Ni?Ql*)h?gqkDQ;an$Rkw)}-p@<i7L|H;jn
zs7H<-s?D1XAkQ$&%8dZeDOi2gI<oF_me$>onwjLy{cg#$j1Zs|r%L=dlwy6tct|jq
zameLy9o|pWe6+*~Yh@Gf4G4vph#B$~dV<a_elgap&N_S1X25nvlC{0Zid8nrLc+H0
zp&-e#b#*Q`uc<jxUJ}>c7Pg2*HA7obE@RFJTklUA8vtFQ))U_;Y5Gf^nooD?Vm-({
zRhh{r&>ZgrbT`C8FTi&#{YfKvo28LegENVSt&;24l091O+Nwjpq;z<#y1YZ3V3VO;
zcqV%o=$TyRL+&4gAB`<+&3%2vG<v0@>;7O+L&7y5(BYAld*n|jt3GY1yrBw@l!cC{
z*B94L*%Pik#7l<dOpKl%qQ1g>9(^>Ka6{JWv{DyF$GaqR14*PD=BKbD?RT;x;LmOA
z7^3Ez2;Ub`sSRF`HjD5RPHm|NT`olk<lUlY?w@-u<K3T!7BH-Vk(oc=PlRU;kv?2z
zTMGTL)|h49vjI<Q59)raNcr^n_?O@?yKW*fDL}~an7dHSR8#)oA)Z?Ab$u}osYjC(
zos5yy)1p@}YbE7(cHz}CwTnaxWokK-fIHXHa&u#%Alxo`xC^9E@Ed@z@B2Dxv5<X7
zhc(^{vA{hgjwQ4KdyvHei_k@|k%4$@<}1@ow>)GL==rk)dED);yxT%!Z8K4=r?3kV
z)aSne)l`4{)?t&{&{m}D1);~a@k5?0nq?OP5=Va%xr)I8)!N!8Xi;=A5mm^r9G8Q(
z<vqJJk=Y`QmO5DWS+3p`8^V`B84^WiPOD#9`!=aqyXYv6YXU5NVTR5oHK}fkM)T{^
z?nMe-O_a2`c-v}o4`%+emv+=*=!<+2r9{2*z_VFsp;!pCXr(wkw--8c;y9f0>4n50
zZhX9F#-j_-bUV6pTI2Lpxn(eS!N}~T58{N?+K9QF2j5R_8gLKG#5f-#ybsI5PJV?K
zBJ_IQ%KY@L!TE4>^ASR4eU@tjNoIwz5$~9?Rz4~R6S!yiJ$BOe%?z(GTw(XxK`1!-
z^ZqmRgzV8{Mah9HrQC+!S*ZTBCO*6J`@Bp-syy??MN?j2s1COI8%rG^&J6o3+{|1m
zcNWp~*lOmjFg;IY#8rwjEVfmj3h==D{J)i1ClZ;YTm5YaIkc-qu1-#kk*_^Ps4Jwm
z4LdZk$iwrJ8a~}suKP_~YA>dtG2>S{d8`jqbI>(d(VwXXb6PPkzw;wTX@3_hZ}<&J
zH~Vq<HZcA8WfAzSjoRrIPgYg@_)7VANO96bw!XIAT_syJi<hfTa;=kUJvUZxQcWe7
zedILyDZ@c0k*QZ*Pdbj>xYn#bGnfc!3;teKBVIOLHrX6>(#0Q-_fa)NCHC36`!$h1
zI}09`o<FwnV#*`x)w{8bpHF{OlY3qEU!tp@ARzcVr%WLc3Q8l7SwU3)=d&x=C#R2b
zN;SQ!n8z)U%ZL?KCo;?fx+HH(tmKs?2U=-S(l)zm>ryZSQzmZ=&&QKV@$&Y!FNI*K
z{bBkks>{(2E6am^1H5ILUWOh8$(q;qxAHab7*=s>Dz_@<UwF1<-;{i$F5#t6wlI<r
zSQm{QPtweYCqK8DkckhSC_lVHm%TYr5~$-!;S-UpKKWHouQs9S4ygbp@K&QFq@kYS
z<Ko&lF1f_sam35nl$UNxs29riu01r$T*IEr?4d7%`@v`XBP$&}+HlZjU%@?kW?x8$
zG63#pr>pFWzczEnNSn7NU|#+@?-d`TfX1Y_UO8y5@^^q*N60xs+>>S2GO0_-UqRsU
zFU-{3toI=oqSNAo16Rop9N(-u%n*)4=o{GQEN4UTL>ReR)zq$WQAJ=}L^#ZO%fUoC
zTaUKm#*ldMB3MPsj>;f-6I|1G|4MVS&rjh~(r*!#Z|xWT?#u8YHM)EV;P%-wO%`||
zlU0`+)$=tkLnxL;CaYhaZ%YWM(MVlQ2&Sa)Y#!4tgPO6EIw^Xm`b=!j_qtQpKG#~j
zRXMAu%Qi?il2C_)<*9x?V$|x$uooLV1f0xbej5{sKWn7df<zQQlr3L%C;!4&>sw0*
zQk@mVnIvq-<(}qqPrs7idf$6cTFV(7rLfumHLG;TkvDSxD^e=v8QQ7kM4Qd$2ZI+!
z9Uc}75nd{{t|A+)S8XlyZA2>TV$jz;ji9(E53{3<NkAg|bxN*V3}19IsN{ZR6%{$#
z5UwIyS|RAs0zzM0YYWD^mktN-uh-yeNqnO_-y`28N}AiDL9AT4`Yb-BJ4mx=8GU(1
zMq;u3aM@VaAWsv12r4J28hUBNHI0>^tyim|cE-dU;>IuAlBtF4i~@y{lT5mqwtYT|
z3!W@!Kkb7MlX3b`U5Z?Bj0l?@Ljg3Fn=D^kf>1kF;YXPH4d^Sqbm}>|&C9Q%SFUxt
zt@Q5pc_+i6ow<(5w)SYtwNU+c5cus53OU}bwYUi8gjkQwwd62oxR7K!E=p}_Y{dH=
zWqBocMgKT20Cl!w;T{ymorhf7{x+pXAcQqQNM9!=-+kz@>Vl})IzM=nwsvT%nK(CA
z!iwOn^-+DVGjqoAoZ)>zrxmfua^=_)g!2L88d(xv4})gpjOtw-noPpNKkEuY7euBn
zl&9G^-wS|BrAZZ5a*5*2_AZF7x*{!^9%O7%gxu*1{!C|AA__|8!C8{Ei~~p8Gmf+C
z(Z|8LuBNr9=tyuz(j`Ju2%5-riGPm23*!&7dBv0&<ns{#1Mb%dT0sr_n=2E7`Z&kh
z)I`H95JzeJm>1X-Jar6~a5vhRin(r$ab51SGA~SPLPgucw2#=WRLX84IojxBb)xO~
zq8x4!bAMP;`-hU=&!S{2W%+g3Z@^a`lV1y^S@ypHw@A{Pd{CXY^tce#HeIEloJne!
zH@m{BK8C3hqg*<0)?gB4KH^&)`?c0r5SaSNs(lAOvV-$}ACu*ddBfO^oD03+*WRTy
zIFq`^AofdYHGN=3mtcz)^SPniqzdeHCN|Af&>?k)fi(@#F&iA10V#V!O5rx(oFuHJ
zUO7I|S+Ea_*QfT13}!eS>o4`R+H&0)mbbW&N-+)yTr5vcfF|T&Cl0UkAAwTdsYSSg
zCv5H}?^fKBv>E_9N+E*x654NoYmI&mN|;yNU>kX6t_SD6RgH5Bx-(nEeM=b270Qiw
z@qJgR5Di;WjGchR@kM2332PCOPFtKCEAv~Vuu&)Qg>WrHv`~Nbb+v19YR)>_6H!hT
zhR$eV1~Qcpq@RxUCCO#nXC9%J-!<>bPX7e^nttimuA}PkH`zb)G;5gRKC}>3zDV6R
zs@movHPma}NM$_(tXtf=GDAy1)mvuR3xwBMU3F3w!{H8Q_u4_CqO9IlrIEv>p2Ds;
zUU^pbr8@Hr?u@y9HNQi-O;@#?nO);Yf(-ug2vV8gKtx`(<IXT$f<^PrI~55UddJ^7
zswecL;oJb-y8J_2)#F+d6wTey!VhO~_End2=77SlLxfP`SP@&Lr^p~e+ySD5&1VG@
zjO)x*I%pppvK)H!nqW;ZikIQYl`vTEx;b8aa#TA@;b!xA!44*J{94Wgci2!K%#p+I
z_WoBJ&ATYDjt!zUgcHdLu~VjJBTEX)MK`^J%R=cViKwp?Ey7Wnz=bDO(G^+GVQd%k
z(e?M|T)j~Qmx^Q6DrJPgabRoJQeZsFuXQ<Xx_%u+W0*C`z^@vJ_FNh7UMv-4T8`c7
zyK2XD%=IpqbKPDG@7+%tF!%#QDkhEUd_2N(uE~6=L73uOR;inzvE`Un0$AI727hsL
z3O2U+G#Bn)R%cLi`h*LZ6mvg#udk<WdBdCF-?=Le$f9^&fL=bj7l<mncDa3mSreY$
z?QH_PI-L5khXu1X5DSe#%kQ|}x7Xdm^n}N-V3x`&3EIlsc}+5MhK`a9BYMrIX!G@J
z^YKL-lR5I-G{d$^gc5ul7&c%|mS7doM~@Q%CsP>dOyDfGgxkLi=%ADTxcqSihjj;e
zra(vr?Dx`t-WWELI2Uia2NYq*-l0^ufh3uFLW8Kcl!WOw1V6nx{oA|!=<)a}B;7b5
z=Qi%{_+S{)uSlQD;i^+K;aVb<6Q_<BFliq_lKYcZon*JSk^mEJLEKMWx68~~12*l=
znx_o;&b@yJg}useRlG%*UL%m<hkJRWLCTsk-#aIDKh>>aI(H0^VU%=(785dB>O7%b
zdYH^=N#|!;E2PMlbKq$y46S;H8IJhM^W@}+V=<Nl#XgIunz=S32rKhPQ{jm>HzS_#
zNLpqD%d_g6A$*~M;MMwPGbI8!L2j}S%CflY7>fAm!aU$OErn=*mz}I!p_9W<J3^gz
zJ8}$<S7vI80q+uR=dF%{4y!9BmM~nriYBbhU%bGKLDNd+ZLY;tP@OB<(YS>v71j+C
z+B*WJ%O=yO>sW4gC5Hoag(^t#v^Yt?<yI>gL3~d(lFiZWniPtmZt-4g4VQ&tUyN^s
z*S%!Gx|8f^Sz43X!Uzt?HDK*45pm3+jGU7XPB5n;dQYGQc2_(JqOj};g?dJ7)l6%J
z&H`FcsVhW0-|E5v>cYes{mKr%g68(6qY}iN-my_K)LNW;5E2`GDdwJC>IT6^FlA1*
zf}3WdxBwYzCYW0zbYTuWT4Wn^_PAStd$5Y2!L4yHD@gf>JW7N-I!;<T5q<M<&GKlC
z!@4s0VC=-GeF26Svti5Ly7GM1JBaC=YZwk$08XWc);ZQNm4GF~wYKsSEP)Pxc3ht1
zT2DCnxf+>s&X%)K7Ku_Vtsnb9Dcra^nGe1c*(G%$cJkcp>dA>!6KD9f$C2ebKiB0p
z6!SI&{5jCf4-TD)?HGlen>GbTc-~uw@!(=Ha=^!?(ihhQG1(KQFId~fP3~y{bl%0P
z$Ed|C`jyV6XLN8lM_Y?0`3UV`1?37dR#kiGuxC7$cLtPJ+qiPiL{E;B1w+Qd&|baD
zH0AD9OFM><Y@{bWwP*Kn?XLdDiz^>>Y+kj#=3c1<j((^sg1G7P%r?J|gTh;e6*_Wv
z89pl0lLsY7ultmDUb8&Kzlxm^uABY>Dp5<E7I!<(v5IlC>--J46ZxZ>f;m{#8i_bA
zUyL7sxT@Cqm=5z_F(`6GN`J5Y>ZI;$P%^M*<De3T;+@-9puXB&PJ9Fznlo)EcgoRQ
z#nODjAe(7oUgPEP-HxZf0qP&VI-I6)b!bvk$u$E3(Yu9zl*3;!cv$R<{b!DjUub{(
z#$=Muj&;^LR*-u3H(<Eo7e)l{{AlO>Z$J+Fr(2ukwHFTJtR<UH$)H!;6}%G1kBA?N
zUZeplj|BE9!8Pr&uM10})1TCo?v`XH=&cQu%c@!Nz~uzO9GDE34NRR72)}k-jePh{
zK1-gJDNQMp?NBL#^QfHwUU5t$$rYqH<vURs$ACS?XJC%gL>sCmQ(vMw%!Z+6iD0yT
zMSvV(sCs8OX<nKrq;J?a3^>Sj0ed;A*!P?!r?CNv5i<SmQ?qbV1ZUbxS^EUWh3;j%
zO8%Pa<g2TW^-3D7y#ASfcy4-brmsm%^~7RwDtPFk+mWyI8cO%Cz34H9w=ZiCzg+Jj
z=$v?wE_&rcP2@y+?E(57?BJ|-WZT5g4E{iT)5BO)dFv)5lTh%%$(se->n>Z5N(uH+
zEmu!$YB^_h%Prl?`)tKXR7n$vBh_JR@8z?UH)fB{OJz^CNHQ=N9bZ;k(xtI?YP@lz
zGs@}SJ8tQO>l-G!H9NFB*Z{)$$-Ks%&#+(5(!H0kLWJuzQ?s<=<sWQ*jq0zT+3%75
zWOhRI*VAFE$muS+_?~o33uwCqTYjUbk4T9&eQ!1!7p^`d=OlH_kM7}rXfQV=+IC~C
z8MSzP-O2F&YR;+$jHp7yo(I_oTJci^7JOSWUph)uCrQ@=oSmzu`ENMk+@BSt9<*-?
z|8nnwh@&%Nw0c9X?wv2Db=Bz^y~POXzv-U(RvLc1cxeST@IJl|rUypa2vS1f?dMf*
zy&n^?L^vefXKL13;kAr~y)%oS<l_ablo`OJ*r4^xWm75_`k3(Xv%24m!l!J%uiP6}
zww%A?mB$DN;d~slXq1we_jYdsGx%2$nNWwOpm)B~DKQgOmV5LkHBX=iBp=~AE0-l$
z!M#-EHW~#UIFTj!tp9U`E)3<@eh7WEnbPmb=Z=x6sq(4#8aMHv@k!^n5)G{2dL2)P
z_~$9sXUqv*YdAbPpZ|0CwW~vI(2UY%`dH_;p`ESE9LI7lIel{$Fc<$BFrtWT{0U9C
z71y_Z&hhH6!za%P{jmIzA|8Xmvx+Bz=we6ezBJTU7chn2lJCh1V^b6EX37MVYLC1v
z5*X31yozdc35Faug<|A#LxEr|TL&U)i9DkNbvcjKvhRak!|>ir)Ffw*+seo>fbWsU
I(!Xc^2OJb)wEzGB

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_high_gradcamplusplus.JPG b/Images/deepfake-xai/ms_eff_vit_b5_high_gradcamplusplus.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..14a0b754dffb0eb8ab631d5ae8af7f2a14c992a1
GIT binary patch
literal 11970
zcmeHtcUY6lw{GY~Kv5zc5s(t36Kaqqy%Qi1K!p$>L3&3J5mAZsCPbvS(4-28RO!8U
zq@xt+ARxktd%O4kopaA0=iYms``kay2OrEgnQzU!@61|j)~sAiT+9QmXsT+c0tg5Q
z06zF1z{LXLl!_0^768!E0e}Di06Bn|Ko&rVFOdKM1ez~U0AhTZ0RO9>{Px#>03`rG
z*s;z6Aj8*D;h*o|Zve|W{?D&Z|K|c2yCdE3G>}9Q{lOz0p0AWB;=jvfG0}gNi6kWf
zfS8zn%Kvam5c7-1|KaqX#Uv^$Dj_2bmJyL)7XizF@dYu!e^v(o7z5k^NWiZNxL6>j
zxA*XHkr5Jdau>9+akfSZBAgwCe5_oAL<EI}01$Z}7b}DV(u3U^X@_!><Jqol;9*DE
z$nn6yI>I_G%1C>Zx}O^o>Zfaj@N+;&+wjQWp{9iR$oM$AI3hi)*nJ!yI=RdE$Z`B3
zT?SwNRV~DE2jXU9D`R+9^>+*WlN`tIv3PrX3wnzQI=k5kiAYOJ3kiz~iHZu~TL`%O
zI(b<62spWO{uXc->5gzixp<(Qo!EZ~w6b>g^pNB5^hDXn*jm|wt!-?;0@hYyHUc6d
zHdX>wHew<IHezBnl3)=jQCksfj^EweApY*%#nbKKAAQ>(gpd!Bjz}jDcRU*+LL5SW
z4*I{VPdr6`clw{wkH?5_BctqwwDLgS#ou>+O{jn{SU^O==&wl@l$4Ny2>pk45TRdW
z{L4W9dq@6XV(SjX1|fs^HFjsWKf8w_UH|W9`yuMiFVpcSI{d?*1-w}NnxmkgfC&9-
zkN+zLf7IZY9R9t(7byNs_WuXh|IeNLqvS{@yk>jjRrq2WpbQ`*At5CpCL<*wB_|^z
zr=-1v=a>?B`3e;+1CWW40mwkl%*Jz*ne`ScJp+dz$1Pqy{@eUa>_TEfAW<F=Kj@bb
z0y1)PN(xH4OPA<CEDS85|LyCd5kN~$pigj}h=2n?NJ~IOOK{NwxPf=2BzPD7%S->h
z2ndOYNl3}aDJU=D8&qBa5E2j(5fT%TkPzejGeH3Ua{w_d$<>=8ilo;JtjIWA=|qE~
z)5$rNN}GU&gF9Se)^1NJD6i8qFf!fZ=HcZ7iGw91rKDw)@2aS(scUFLjf_oTre<)2
z4bs-m9)))I@bvQb@%0OS7V`YX%TP>AY+QW8>%=5%MrKxaPHrA9zpT8X@_kix&4=bs
zEv;?s9i3f6!y}_(;}erp3yVw3E303=uC4Fx?H?Q-eLp@q{e_DFK=dc9zmfe1T(o#x
zgv7)|#ALs45fFOgZz5V^lA9u=R}~G&tX!{ghz60<DMhE3Hc@bj8SVhB-3BSIbBQn9
z+WiIX4`lx}U{C*#ko^tpA6!!aDk1{><Pp&V<N?cTd=pvkj1)CV*nu2jLzl>L)w{6t
zXC%ZCXkLW$8Z2IcWy}qkrRN;u!f!Y=dxLoz+p*t#IkrfPdOFh%+uBjftSXfb!8)uP
zWXw}x+oOvGxjOnl_TM`6&5LX%QP;6QSH32)RnmGru2^|Bj0sE({r=pZt>u<ED~S6$
zZY43bS{TGyvoZFBQBiy;NIOe}T2)7`j|3XCjFTcwQcvB`mQ{gDeyJAkfB&TzMc~m<
zk>0ds6w*xlIZnO1V@=V)a9a_JdJUw5f%;znk?AIk@HBgS7e9<g3@C3}?>Y4uQ_sk<
zgYQ$^B2J7uocJU}5ENfD52`8XYK-9b=z&ui=(e)*8!0@+$PQxdqq6DZGJ;IModYib
z9i#3SfO|R@fb?uSfBp+VbpDs~k8dsj%9{L78`|sy`uWl@d=53(CqQwOL0*B#Jf(Jp
z_A_3_oR812N;6RY6zxF{BgklC2{jao@riDDQDoBvzP{uPFxu@h9cA`mnD1}_bz5*b
zglwqOy37l@V~U3|Qhf7uWTiJ+0kI03X1?t^n3o;BN=2H{t3g?JiI}{oElj1de@N2N
zwE0>LYxDXQ-SZihSxnkvTTEmeGFOSXj_Y=+S#UrtDM@i;NX)ypavNYcL65&_=HBpi
z&_YM#^JOC-);X=5@*di=!YZZ_mcc|=|M#->HiU0jYjZyXHHdECYP@@=@>ReG-&bt$
zkDWl}f@<O&Hdoq|b`8zZgt9>_^_38j<;8MObMB`q@c|C@i4`7cPWs&g`#P<#VJH7-
z71qWAdHARpH#cvqAPTg_o=I0xK`?a&!c&;@7lpjXunPdEPXA9|$@AxNFp|KqHuDrH
z#y|sWV7WPZjdwAH8F)2?Z?Z6((=>err`VYH-8zn6a<-5etR7`^fXlo|<I}4|5$4&;
zy*WZvb!#L{SPIBg$`h+!tbsGqCwYb&FyeJJeBMfOosJSKhRJ6Ulw>iwwr<dzMHH(r
z1fE)zioLCZ{Ot2os*G-ob@bpnSH6vH{u>SEpGa}p>%|bIE#1%Y78@}`0k2lpw6Akr
z08Visp)got3bk3Jh3CS0%tREb*Cv4a<|I}bW@0f?`Z+6Dy3n)o;Q`p1UxR>wjE<TU
zF5Bb1SST>#_{qe&dEPo%VB&O`H2iwiH9?d7x*8}Zpth5-sT1g6@BSfLg1zbmY1+tQ
z4;LGHf?M(^XNKUmYK3;6f#^Zi^iB8#XPp(Of}WAZ-MmCHiJWJsQoC6lkliZFGGtwJ
z)&A=#?C4qcvea|2eFW|(ALatCoGBG<GFeL=PWxD0Y50C$+$utybzR!VH%nBr_3ek)
z&!u7L{9?n$+jk{KBf1yt&cf;C50~donGCdc%Fk6_rJn9P|D@(E<Ug~6R6$aKPI5Bq
zTK6l|XzhP0cAkA07D#h=mAI%tVcQNK-%<xeumkCzkfHlo!#tS6jHk~cy2R@;rAeQg
zK1YTR%SrMv=D=+2RDUd(vx<rt^d6ZFKP<@vxDNxrTmaDWm6uyKOwBx`X*KV&^5;l=
z@IL#&v+BTXFic%=Qi@DNG@Wmk7UeDBMqdjSjPNMAz>r-1?${`*A<#&E<&-snt09rL
zH)Z9OSe}$PYK{pW7HALdDl!tR)!#Xlj7=G9*2wCa<rU!IV>8eBIp)H5O+C5;MbYat
zJ=uQcM8#A+uf|K%Y8a@{v6>YCVH(~0I;r56BtMzytZ6(v;t4ET8C(f7aLqzq0+MCR
zZfdQ;9~-k-(u3O2YhTT`M&n8q&%lmQx*gtX)2xTF-kQ`=L_x6TeXR;6rP8ReT0KpL
zHx6c*#KvCFY!eIhZODpD66yM6Ov0kS<x2|eMH?lFs_KmQ?$hwAX}EHko9=;P!zY2<
zZ-|4q-ijJCkTR^L9Qt?3CJ(8mGq{m3Wb-8dFf{qv^qjX<0B9{MAM%NfxB3F`PT~SE
zFod}uGPfr-6FB}R@GDBiW$d^IS<~e6_`0NF5BXGRMZ_6?9p`^*$x=sGz3AioZ~=I7
z0f3vo{<-juhjH6F`*Ffq<Rb#QC0C|}eC}#Ll=w3kKSnO#dZTS1e570+6_gfx%i_Vl
znN*RS-23W6v})JaBtf5wo)=ZZy1UZEnLsWhdiqOO8l}Oxiy?a$i%;!Id+$ZYN;qD2
z?#px62VXWe$;pW%HCTz&#)=M6Rl*}Kiv``ia#raCo{}n3WPds))#tize2~AkVZxC%
z%F$z;4P^0T$AOv$wG`3{D0#E#^sisVSQsdf2+-+^DC7od)9or0-dr<~)ATbEP?WeC
zWn<zva(8(w7F!_6K76;rx(vvLt;^+C%eq%f#UDQv;DTmO-`fHCL_h>?UI6Sb0AH1A
zb`8-Ymdg!dKvpwZL^m;1OoN%MQk%)by3V+?qn#42sA5Luu+wm^`Y7^Gt;k&P#_P+C
z!_{B<VgarfV{~%$6IaJt#Dr~mH=SO-09cwNmGc$IKoKhPhl35-hEeA`<FW9qkJDQL
zY!vXf%@{>pV6fb!8D+zEH;%8qc{Yxj*qk1-WAIg&Y83p<E?s?SMj{~tB0AXw*%mDq
zO44ZK6jlj-R&8ZIcu(kc=}wm_xg50$<*4wGfBxEmbiw?|p_+IwVK#jFjn@0xnXxZj
zvBhMwJf|jNEjwL^c+IlpD-KiXp}CfiYL6gmgYlB6tGCQv8haFojBTpF)lsLuTgh$D
z>mr_k<!c-eX->e(46iF_^4X_ZsnAgZ;QA;MI)GE&Bu*nQaX^t7TO_0m?eNJjlO)Yc
zfY!+0#N~uFEl7S=HFaV*;4-?W{aTPCL`#hAY2n68fdf8vL-lZDMY$7TQ&?Wy+VDnC
zk_cSFc$jA6f!}Dx%%~3MIi%HmdaQ0@(|2!138k2ok=OIBp8m#hzM%X@Z^MCjM(l43
z^_%f?F+L|3Kb-?tTt#(rCQHHT4Zm7J>}RXS#hUHKf}_*c*@{P36}kt=7~6+RTIW~z
z+=u+?`eD=<kJtPYgG>&7hHs^xteH9If9~XzpJH{N)CJ^A`mIb+H=tk3Z^JG%j#srr
ztVHgCi#Mm8Nz;OL-8)xnmI^+1%*<+C00NjQodf;#=nXFbbE>)P;0r(!UFmk|xY?Q-
zBgZ?z6f)P7aY^c*+wqmf+3aDoANLJ{mAcpEF@ZrD?fn@tAE(zY0FXU`Hl;5KrzR}l
zsXaaR@3^&x9kGTOkYTYB#UJln6&K>4O+IwcEEuU`)vZ%8c{v&tz+c>^o`lv6r9o>C
zmBLu~niI)$Bhd;n8113epjTLeoWWYU7p{|OhEz;&!o>j{Zu?e3CKc6DKHxR)9GwoX
zHKT0lUdz}mZ0617QPbyoS(ZOj4%8}y-%#qgYVRs=YFD?Xom=NgoxB*U0wa6ef`5KG
zEBMsj1Z`2uKQFRgC}a|;(Dm2}Sl^&Q=R^@s5NyMA0`7W2CTSO+@J(a_ZM8S+uj4mP
zgTFRTj(=>Nw(;gj+~O~DH2%vRt>hMoCmTI+hju%~OrJIQ)pB{XGLDOFKHG$hiMP&{
z9r~HRwFNaU7U<TRlh#a2sM}mh09HBBT)Ad8mave>G!yRCMxrUM6x=ExzH9uY&h$AM
zTs5!4Lg;h4Qzo>t&^zT(@Ln(`tjXnb#U{Os)HdRL=WKn`o3@0(>SUbt=urqGW7ZFq
z5!P^7GLg@md@RDb+eCFPB^$3P#v#Ue{z{f5hZ0QD6J&G?x2&hu9YiD0dtJ;?Xt?0Y
z{6Rg^$c(Rb-%zwYN%gd-KK$TGWj)sBdw!c^CDm&ydhwx4SV9BceIxH;FGn}2JxMoy
z(WMrZH{|FVbvIdwnlKvE$DVYrT|&4@&Nf2N2WwKF^?^*6KEB?zEr?4&<UBgQ%>3eu
z@iJCIp0Chz*bcHhY*STJbyCwROVX>#ScGcu%~i;*w|hyI>T2pW%xDcbEDFtnq8@fl
z8Y?iI(vE!$8#;@2sL2+XRMu0b(qWXIYOyV5DzI-Fozyj|VLu&B9<S4>0=oAnENJQU
zG8Sp;jxPkBg?;~R%G8>-PZ+m;t6a6inXTx|-hS&HzLJ>~z6eo}xJ9MMi0oQM>j8?3
zWW*}OJ?7f#{I~{3D*I3Vlo<Jg;KM>{xRB{hE|Y2t-ohWm^CpQm1c;-m7Yo2;V#p7p
zWANfn)QWk_djkmjEU~uBQb5B+X90I?bL~i<I@IW;ouXhm<!-BCEcDs(1DKTHXP9(n
zZYHu6li`G}T&b2~ch|l|9G=(@)*02@9eZ=d#y?mc^_C=eYd6^Z`gblWrY}X7ePTnE
zMSEz*4BHmB1h6o7zBvtguidQNsL^JawFKnKnE#p!qX0_@sM`STZHBNj#T+RSaFpi0
zaf_nSf!~ViYnqnM3yK8Ol_83hVsX|4j77B1l1iQP6bU0p9ZcrnA&o?g#Y!HIwQa7(
z&wA+n9iM_;Y9G9O7q`TKV^c}Ada_#mlhX+0@~#s&q|Chn<bGKaecNuOuvd3)m0^lu
znV&gD?by$x!XerWTcI%jV?<LVAqK`*c9byp)%u-nc}4VjBxb19nJB?LR6|R{oRRfi
z^V|v3UxqL;q~YId)UBG8A7!As_YOM7S5|Z2n=v)B1S_M~vIQI-_r*AC3SjwL)tY(i
zvB8+tgcnL91LX_`FvZS%7uGyb7k-0s(*hf^z_n`vD}tAVa*sUFAYDcQI7e(V3=MSk
z((T(uDlQG@GDm$3bU!ZLuy8Q#u#qM!dQ=l~&n?S!sJj1(y#op-ANK;m+NHjcQe8}S
z9XuZU?E5U!SY}788f*3%!Sq!s+sqFo4h|V~NQbmNtXN~Kj6#f1Ujx{$GdVd|gd{m!
za%61o2=vPG+mi|0RS}jqPYU$JI?@n?ufp%PRH<PCvRtuqRlzI7My}>$kRsG4ww5-2
ziC#l;dTO|VAR5a{4VU-?PD;ywm;~&(xSEqV+*uPv5%k?xnzKV-+Dpn&IkN4*2`$PG
zS5yogj9(&E`R{6}N#jr{T1XiCnEKF>R-Fv?+pC+F<-E>CMhUV3lL`a{BCa*6-8$So
zCu>nEh<j^NQ78C$?_(@y2z$qnxe0jxtq!$+DHa+#EFPL3Cq~rk`TnT-*>#rD>cZwH
zt|prPbs%POgHmG{ddw@WW*#%@iYd6Qq+7AbM>Twn))ltfs!06Ua2L8-g^{+Aq)W`T
zjs>QNPWcPs!=#L*^N-WNB2B)>^p}ru|6U>YRrE>Zfmb!64V@Nilg8=O#(^BFl+erK
zPnB8Sds)RNM4!0u=`UuAZlR4KW2%MKFR%^SfgOg^=O2Xj3p;ur+UqABsw=QRls!3G
zEiE3(=KdklgpSC}ZpwrXMP#FM*dY_AcE>NWiO<i*;xmIVCsIc7e0-lg9vSoE8s~47
zAfpO$$fY=eL=!|)b)bEnijgb6uS)mdekIi#A1f{P7Xu|v%XCr9UwcXtK*G`$HeP>n
z|8QjQ%ZucD?Q|Z8{hHCY$+8{0(N^)9(UkaYiHc-5QKcWGb3M#gyQO@3!07pH<DbEG
zDjX4`)N#?qP}1UQO#Yoiiz@MA(WbRtSc1SPqUaVqBW$w<$nySXLIsJS94<%W*_4%Z
zN8SaXOyS|;2iDA?*ly_5yF?RJrb)yxy2mRE!HF5#8}XpC&Mut<evHP1E;SZBGcu&1
zL0Nc-I+0Q<y+uE2sB2ZeI<}(*Ip8Kt2s5zcvv=)Wr!!E@QJg#z!cR=fUlYT7d><db
z%C-OWt@;(eQvVUZYI1RA%h^8MLo<>1>2jAwyGeU9RWeb?jCt5m!w`?93b`Y&ZfpZ2
z(sVsW=|MTdg-kNet_X15Le2#sH2(rnJ=ax2w^$+mD)@Rf!CZ+mL4<3YZ?U$vMP8?1
z&sa)~_Qq^vGJ52ya4W10SrGM3^MuV4@ikLwFP0)D%WOHFvmc$TnqQk7LWJW1LUy^K
zLGkbIMiRrQG&{d5v=f`KfzFHa0?L-GZPe(yNc~db{j>M=f}bs@@NM~2!sDhD)PVP(
zhFXFoP+vEPtXCAZs|8+%YLL;FEXz<Op@hYx2CPaqr9&7K;H@cBoG}siR0?Ai%QekK
z`v#Ku-mZkIkSxG_Jdh}p^Gajz4AXiA`B}J>NFUO%g*MC!cgWBCu4s<8-fJ$tV;L5=
zuR^(%obbA`H{g|PP1`jLN<3oPw1AL&{W3aeg|P0?Ojk>Ge&HTMJY13$oi&8D&tTMD
z&E?oNAgak9un_PVmIhJr(CMS7J$cewzF)^^!YS~v&;{V1=;y|z3&677KcUh5yWJ!$
z_YQXzNpagc6F3)Xwi!8)FnJ{Lmd}Qn2D3^`;mZBk*89Ye2bd+NFLe<=oM?K+jdN-q
zDW+8Rb9uBW_tWK%6!T@2*si^Oo?2vn0kCWPZn&tpSF)~G&)2^gCL!>EjI1CfqlUL1
zjHCdr+^(X@Ssv|6b(AP{e%}ztS$_MoseE(3FsSCk+t<|m@p6iYUVT+uux;&4rMNo`
zS<11G5@K|DQ6Vtub~p;`2q^E4*1y@Am3Z}|F-2|+m#nIedaq@?xD;u%v}ciypiAl7
znxY?;CaUkDyq`{uxyba(@-Yq_MW0Jx1_<LX4b`?SX!{b=xu`G~G?%St-o31O6J+kx
zeNt`A^zZ{SPll`7^HoxL$#Lu}6DZ4xErBog^BBn=T|s&ylohddb%>7>2}Q~h4vTi)
zxb<o{?-$RBF;=*K*4uX9e%xxA4Y+jCh{8YSd4FP^$S{k^waNf58vg_Kvq(YokhN`r
z_RL7j$mP!BiKLQe$oxrp0Y!3LByg12)d;}=MYRm`d$>t3X7@^99(c`NYWUZg(hF%n
z2fg|oYZX&o02srnAg?ep>Q;BI`9dn@QX4Alw*AJ`qQ6;kB`94ckMH}|bt6|h%HiR|
zqRr9oG2q<sfmM1&95kYWiyf?ewYE^xMogLIRZQCGE8rv}E3PovXlExr<gN!aYLzjI
zFj<-ofnBz<F<oXE4aqaffp-jE%UevL0kNYR2Mh$)_1z7EbxEU@z8z-m8GOL_i0@()
z6dumY@AU?PaDqv7=LSJ=1mkLsp@>mbPlrcc^(-u=3w(YB?c%OWgz+M%hlAz`;C|FK
z4mfUJN)7~1sr>R7s@DT+BX2v@gd|uw31w|Z@4Hhb$1$xRNMy31GWgzHGaBeL<_Y{f
zyy$mQ)2=+395D1QG>X}WyR*YF@n>MRJF=j({?#{`hl5|poD|#q+%3$cz*{Sx(s_GF
z5cv8K9Twr4DIVr9723_M#cUm)QsA6RR-9NKVKbSPL)t4-DV-RX5jChyGrr(_7A`6O
zXE3YQ+>hUD1pXOx{oVoo938nQd;xf78aU2%9^YD6*aKMC<u9x^gMI3y8M5TmSNM`6
zW*iw~%y<7J9LIxVAPA8X&cB1o+7I`8erwgru<etht#8&yskX8SRJCF<i_g`dZIUpn
z*Hwdc05O&ge@OdrsL++q7V(6{lSw!1eNw_EfA@fpe&v@%SEHTXk5!wV%{1nWCRpQ@
zG0MBf6iv2EH^wR!c0LpQP-!HCJY+ogEF4Az89-WKFCgu-(1n3-&l%<V+)mrHD30WL
zA8$lFGxi_XaxvVR`6Rl`_M-Ybu<73CemhSt!H?-p@1P`@qm6|7SA`37hHcw6y)*j_
zM#g14rQcMlI@?0VGn8ZyzE=;?t|RO>S}k79s?WM>f8LZvn1~h0WGTz4i+?@TpWTQK
zwd|Z~8J)c|w72dIBH{WHJ8>Cq7>8g(J-lAu?wIv(M6l5sQqlDI5lqX|o7ttvVz%h@
z#O)T>o%U@>FR+O(kCM7AcyA(q`<lrHvUF|XYxUMcEnZPqz0i|lKjx&cRIr1&rmmRj
zNW_awsjrR-lf_y_?<G#-Iy;%mC>Z*Di*sfRgjy1|J{x!KrF*f-;DfO7BGL7fe%kJt
z-BOXEFH$}^vXqaE&E^~fZ1g}*js3x$*=nRXNKJ0*Ess_fx9GIzS-mEq<xtxS5k$F`
zPIeCG`(v};$F#rPoWZ{y3>Gut)iSOMlHW{HUH_cJYs;!n28Qk`b5{M@7;KUHV$f!k
zboVXB)O}!FYt?Gw1vu6-4tCdTVacaFuA$6<KL?{2G^*=|!M(3#p5_HJjA-qZ9e8e-
z()$U_z5~^}d61Cmt=x;a#;RLrWc!ViGH%A~O%-9hALa%~-Rda~xm>=MIjT5&k{H?N
z+pc&hS5Ti76Fxoga?A3`E?nkus=&d50^bc2HyO&>9G}PH3g6%1JwZPh@|gQQ*66ti
zivozZk*LytOH2e~QQ;YI%2v<zEW3Jx+1}$AQeAHgDUa?1OJmFHhhxW*xqJO*;&OTT
z(34wA;n6Og2DPi{`+Ul_*zoX7=4&nd*IG{Pky;h^9&cO#^ft_+MjHffg=IYvEomaF
zlQFiW%+HW6P);)4Th*4-1^ImZX!WG{6>O(_`UQ5>`_nk<SWPHxrW5~q-dKqJzG8}l
zu3*oH+i8z|=vIa~+$#xq2B?NN3vHvlXG%lF!f5xIrm8+xmRKmHNqRxBEAJ8wvrZ5N
zlDVo-_L7~1JC72syvsEK5460Opndp06dIxDepsmCIXk;vfPU<(xw}4I^z!kK;nFq3
zPh1aY=J-AG9eI+Bh}$%4$Yp)A8BDjs=uFbCtroaQP{TTUHJBd^r{ula=d_(YH5woI
z`A#Q+{gT8?9vP06Uxq?%{Y(r|=Th5@)=K{Zi+BO#R&_TUH}o_4^76s;s*vHAD<>~X
zw=G8c6TsgYd_ff><IvUh&YVHEMYQ@fkXlND+S_#LVFOjXqd7-eW-MuOD>rK5&FkxL
zwUw~}?6S@wsrXO_zl{Gl%)D_)e%Zd}?}N?-yf^Q5@AsFyy1J;;Ao>K@02wQgV9Wn%
zQyjTZD7qXLQ3rxKV?P#J)a+if!-YgvYjd@#moL>pdl`!#-}R{GhR2!rg>Z*6>glVn
zx$%$_{v<KVn^}wvA1$`vZC<UQzjqKbV+T%k(tsHC@$=}gwKBP?>&a7(<!FVCa<dF8
zKN!>H-O^y{7DW3@gQHnSo%nV?m%GwspxcfM_ymyc!gA})TD{hbR=gG?8fC(MMpscs
zwg}-M;ui(h>EY|HJl*;kV?FJY+=sj~6BPNczSe9xSS`0I3_mz5O4nhnSj$4Zf5`{c
zy1M~a^&fF$2ajZ6;&m-)P0dx`@|jag<(3l}ub>N9_VQX#(rFqUd6%e$f(;kxmQl(=
znzeMvo`o;x2n*YJSRE*!(z{pHA6MO&I;a5{_N9t0ja92k7ay#uPb#jH1DkayFm~ZJ
zysa8+qua}S-;p%!J&P$-oN0!OF?By!ID=aCoQ3+0*9PAfVWYrHi1#_9I7Mz1xCxGN
z_K(b1#$Rb2zB@pc<G+(U{Lg>o$@)wfbl#^@-QAjci!Mk_9D8;qNTwjn5031=$6_Aq
z7NF;ZsZgV|&<Cl|eSRLMP|2L50kjbW&8A>&okiM6tXFUz=N#c5Rlz4A()mB9KvxQ;
z@+#@&@?~ieV5-mv71M1U&lwQ)FyD4lZefdehPrQFq9d*}SK3cD%)21k;xfGDa>T(w
z4gcr#7Tun;tMjhCI$zG}98w~Fd@c;AcHA(lj_$_g>_@_jhiwiEzklq#JAN;|(4cIP
ziQCyvPxf<JZ)Z-@#MsMb)i*lZ#>-LId@h6`w;nvpz4Nl6yWu;{a5b=`N&%~cG+w#Q
zvH7{!Qu1XJH#IiC*Neq^@X8+&x!gpK?jNtsK~l#)R2N9Lc-`~NpUD-VCtI^j^1N^N
zQe*P~LZ$RU+K?}vlhO)!Xa`8j<-JuReJ4)kS}&;mYJ#m400zyq-<c#%xS!%WS5geo
z1*NdvIr<rXlT0$5CNhKm=gN+KBSdW=wa8)4UgI|U)Q9}j(`P#YYemw~<k4x_k_qDR
zJ}qX;k)ziq9<uQyUH6Wmc++m&aRJEX`6F{Mx+C{NZAl|1?17f`k0&fd^)mcpNBLz?
zOFfw^UnC-=+JrMgMPEePy%k}OhM7#zc6kNAE$RT<onK{6bL+nRa~H3vh4(DsOe#6v
z!>wGEsu?Xy$wrvM$Lq@%08sP!;D`|S9gA#bHjFUI{mS9995D;^o=rD5zY$1%!Esh}
ze#@MV6I{iq^yIr(dkGt+F}1^_iH}RV5^1v*FI}t&16Zk&D8+VPA`P8Xmu3LR?TtEn
z!Lh@IBgKOQ7D~%|pf^fO5M$HI+Un!zXK5+k5e}-<PD6(N$4nyAs97vNr%0=p{qr(B
zLiBn$U*fbJpa1)VBA?vh9LvvcxpFKIhQFiQ&+lcNZM^Nu@Nx-s9Vr;}n`yE)A1dg+
zgvj>DuGl^~iQLaiUc-Bm$h-y_i{W$Fcj2RhN6EDcOO;Aify(CF0l9=~+ccN;>(FE)
z;sbv8!(#uy!}Sw~Ku(?wkJnIE>)3&jbgp*Ilz@A)Tpawio|~<ur$#)MVS(ML=kgbT
zvHa8eTkGj3*A9;<N_U6el<6t!GHnL#<MU*Z7l6b3-uibo>yHPowYfLme<S|$IR|Kq
z==7%q%yi}m2T3?6Si`NQ#NdymZ~XMDA)fMI&*nD1?E=K76^yNm=i5q+vTU-*hO%jC
zr>ba@MWRII@)>U*9p#1Fmpi*X5);1J%TEaU66KAc6K63g&Y<UbT8kth^E7|P<kqzB
zfkXNf&Wq`K<(>8360}3sEHCDBw4l+<MXFs4lY-oa%)DK>JagIZ-q_yWNu<eMhII2g
z*KH$TIbZNM{jTZn`L%v*K+W$HW-YdMdy>AsC*Gc|N+#K@;+XG4Wu;vXU&SVxQiV6G
zAD;COv$a<`mOCp9Ms@iTvjkgc(V%u_^_-^D<D-^XnD$(yy({0~7#&tUphcb85ji^N
zaX)N5txQXNp4u!f-ijj#3JyNR>4gf1aaK}J%)joF*WADLq$L-5qh-@H&-73apne|~
zcjM%{{OQ#}i~eT*wQYZtmt8=kY0agH$9ppT#cwI}Be|*|>qB4O_Vb#hs7^A)JdS^7
zm!akQJn<XbVL2*mdNQx;shVFY=a-QAN3a6EH*LiFiq3XyBaBvabP<h1Y#?rMShl-Q
z^QU_mR)!)u)@6QgI5-W}bU2bTfwvM!q$sUNg0neVI5LqqRJJ{>WkQBexycb!Q=cPP
zMZW@^(ZTmz*uFY!mki8~DLOIAkVktHl0*V!1cY82f`$lZi_<+FtGeRf=QSF-ku6wK
zr(`-luDwsINp>@W;1njDflotNfOkAL1p8|WL0y-XagQw2eRAZgSV3<Ss!JxLJ*diX
z%|UHyHHCG07PU2z*72dfeN7{<cq&p3=*pyEu8uZ*<@Ox<{d+3tPn*oA1=AZAy|bI+
zWeG{0j2#v`4Vn?!<uNU>GuVjwShC>U7chAJmvFdS%2H6SkyW`~uOBK|;1X8JKBG|C
z)Ch@2XkY*@=zdu^9OGAc5`1v(R*QG%WEX(J=s~rnz!R3@3&18DK0_%g!{689Z`1A0
zRe86qqG932Si{%Lf%s(QyK}m>oujG?fI@)s1%TlG8OS2i!0o%}#4x}9vAenaKIw@3
z4*t6v%~*fw;=loB{R==N{)+~bZxWIpC8=T^Opl%M`-wD(^gm_I9jjLMpAw;O9OGlm
z1>L39NO!KGB2+51L&E{dZ%Og~V7w&yza%NIM!I3JQA{7}naA+CL+ZT1dHxGP2*k%5
ze;$x_=eJ;BFWU-%ZX0^<_+P7m>^n!5__05_0GyEo?xa}^ZrMG|vOg{|Somgl9z;KP
zOobl?-eiy;mExPd+CC!<+{yZcmFu|=!Xk=3EzBbRZl&!^;~;-PUXKT8hX+XXtC7%D
z9D5OBVb<aN$@dGuAHp>7!d5E+SN~wI|D29Y{85rQ?{upJo+HVkzyXd60G`LxGd{KF
zhpDH;_(QFEF+3dp7x;FX6`Lk`J6q*{4<-GhT;b6W{v2$F5&wFkyIM`uC>+UMb@_$o
zyHlb${JX!XxO>h#)_>MidHEN{{8zWS<oaAQcp%2US1@al&%q6U^R3?y;`!+UAZ)3I
zAHt0@{tPh_VXPjmH>g*NY0{Zu<77#Bq+#c(-cC*`rOy+Ofl8q>;@C5OV{x}G-Y&Yz
zoO<9cjqk5oqS!Og{bQBNv%rqa_>^3uhJ14+)Usz2F$l+x^%WlL0U=%zjowbyo$I?~
zKLLk-zEhJqqbiH{w-x>Zi<eIc(D#qYc#ffZA^wVZ$O3v7007=`QCf6Yv#K>k;#X(G
zMiX9?jfXKl7M|iR$T==sbyU;&SP@$IR=J;uBesr4RPEm76&2sZ*;VD4i^+cjy}jR=

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_high_xgradcam.JPG b/Images/deepfake-xai/ms_eff_vit_b5_high_xgradcam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..1965aef96ee4ce7608b8536ab93b4747462cdd14
GIT binary patch
literal 11378
zcmeHscUY54w{PfO5fVCr6d^=<LNC&rw1m(>I)u=RGz9@g3`If{0s_)oK$>)w-i%0#
zRFNV`Rob@0jeGBJ@9+EW`QzMs&U2sp$C+f3_0Bu*tXaP`v(}pRp3R=U127xv8t4MZ
z$jAWUqz~Y1g?wHo9OVH37#jn`0000jfRaoZKu)5l001(>dnf=Ui6$d`&GH}p`UlVh
z0OZ5lJOCO}8zbpjjZ^_V+oYdgcmJmVkiqUjBp9gDD1OHgNW!Zijq)$LF&+CmO`#wK
z0Hmk?P5+H4S^6&+{|D2*7n6*XjGU4bOi5afUmB(agDOc&0sf;l000RH2Dk%$Dd22{
zlHDsLBtS`0(l1!T+0Ea@UBcDhS2El=KvG&lN)n)=9v<NAigpj-cX9Vb`KgM0?0hc5
zk8)ELv4R;(83$;&d!h6rgWM63CKj%dXjery5p^{tdX;dcaNhu5_Yi0Pa9<z4V5M+X
zf!{}0BGJE^B?Z(}g4{fm;M%%>_8?uU3j8^ju&^+RFj)!zAWunYMMXtPDH%x_8F5k%
z@!$x*5a)1lzhJ>X2hes8b`3%WgrNNW_<s%P?BX99qACy?igHu(aQ1+?xVgc^U7Tgz
z#HFR(oW-5pWTnO3WM$ptVbTgR9?~uXf3kOT{fm1*XpqnEyxm+S-F@7B-Tgv>NoYt*
z3P}Dj>HjXCB#8cE`X9nil9AL$Nh`?RImBI?RMmbdR9p%sE-h#A4<$<|NJ*<m{#!p4
z$zNdn(?tJ!Xa1i;OHIYiRmt_&-2H?8V2^MQ{I6--2c`DQbo?P5>EaK9G%RV&(b3VV
zNd7a&{~dzgTS!Ze^xR(yl=LM3|AX@XGn2nJxw{`}XNQrt@Yw=D3qV6fMNLIXLrq0Z
zOG86Te}RF7F+J!!Gvfsg5Emy0h=ZN`qR1s~USVE#4gm=PVNo#T3WSSaQdUw-MnnuE
z_G=I_8d_R<I(i@j15k{IgGcQDxSYKJT%aXGkg-va2>{41kWpM9JL?DVkz6Si$wmM2
z(tj2*atcZ+Y8qNPdInO57G?lB83hG7B?T1~CCNXN-6GuwP+p*7xg@PY&1&vUBM=Cb
ziN!vq71V6#0l_Ce3CX$y#nI8Tv2$>8311cw1&cvpa`Fm_N?O`Fx_bHsh6oF!rIodf
zt*e{6ho=|HJ2)itMp$@6Wc;1G3HK83W70D+v$7xN<l>4;O3TVCDyym+o0?mmwzjo*
z^!D`+3=R#Cj80BX&k*0t&dsl^uC2e{*!-}yeQ<bmeDd4pFJHg?l8X#L@rPJ{N%n7Y
zT_DLtPDx2YN%KoCGV(A|rMN&zbxE3<MZ=uNIgnLACYBbciGAGALnkN;{{(UgnxJPB
zg02W3{1WYV$^LVK#r<DN_LpFP%QX*Rq#z?HkKzJA9kBkPY~V74$PACY3OwN~?d_!!
zI<RuO4WO%Bc~FoQ!AwSQ2eV;S+f@O0<*4iORC%CmVe4lfZ3`HF(;ZQ<l+c+avTd=t
z;i4S1xQ>Voo}C3KwQH;SY0_i`O+L*@A1kjEhSGkn<kCD&;ymTbj6i=L2HtBRd!=OI
zix`pO(N(C_7au92_DI7;01%=a2_?`Ch3ZE8YE}*XoZFGp?iNVb=;pp~{M)p#yP07{
z@?{pRmw3HwFooXqwhv`vai(%5m=Np}6EUm!<f`Fn>f2JPsC^{Rw^&sEV9x?Zhnv6p
zs5(Q&tK&ILsd~YoMwDPjdBK4$XN1lOV{iVzc7mB7*!*ZK_*yn+1^DYdVy>x4$<Ec<
z$mRH$b%@B%sKTFgUL`%x4?!8H-Q%Zsnyu0TC%&AgEG;X+NAB(c#Jq?l=9VQKpexPB
z91J6+5%(2<li}rIbrG`K_s!BHU>*`(GcX01NWPv1ZY0xNXf&S;A=KH4WC&kRd*22M
z7~wG*vb~xV-n&uLB<i0tBj0gDfpRskt<{#sW$I;onYOz~8C1es=K+-IQm(5}s$z{1
z4e(<V7mZ62<Fv6)3bDRfIdYO?dg=*hkd00<rb4{2-f%ITTH!R5?CA+GS68h3yx}85
zYyzI|vf7TZly;;``G7#8_1MkQII-!VJ^PE^iasVmaf!w5-+QabZh&XZaJ|zu75N1-
zkL4*a)nlD_8BRygC0%7?jf2L`j}i!i635Yi`?h7Ho&A+~RkT-5Qw2qGYG%&3?LE^r
zmbrRGev6-Ix0gAK=9o~BCO5_&K5;em`^v53!<{BO{RONzXRK=Ux0n==-OBYJDe~$k
z>&xG`%<Vzxh}VM#B9Gr&6zAd0(A{|g?mQNlSrAJBc&@fg(E9QFDvi|=0-yBvmUxdU
zmUvwD74R72wBZ1aKt`a-G-Jo-T3u$BViwtp##d{N4U2DI5jIIO6E?wPsIS!^^kR`S
z-NrKL5AP5e1S%o#h{~kw0rL*x<9mVv?cbLy4H_FfnhH3}-)}vhCAJ+fU(uX#U8R_>
z{F(HA9ii55>h0{RjZSABI;xT0DN0s#Ez6@Yk-ImUwo7=vr=Q~qJ#N?Iq<&&ay=5to
zfK>4WlIyNp-Fje{Z_52C(1!j2CLrafv2?T7i0=j%g5)AsQ1wQm@3WG1AhWOTbztQn
z3w;S2MKjJz=LxUGwhP7E%N<0s1#K^Tisp70?YU&t={WV<GNdYbv=*zmX>`|EycD`a
z0d1>TO~Xtu%L1F*N`No+)YxK%CxXuaj=pDr$7R6h3^D6o?Pq{`xii4)NzA+YgPgJ8
zkpmO*hrqBpqK^Wi3h4HZ5!b|47zT|hj-8A21dYpOZ`?0yw}*eti?q!7bh0C=-pR??
z<5yVbxi~k5QpAMz?CNRRZwF>HdSLV2i(P>Y?^9HDQ6O)mZE=qkh`-c>zJ)pCa%`zB
zQ!L!y!uZwVjC!-{T}9{<JL6E1k1}j=lXRQumi>yWMGk7K9*gV5lc$9|ns_0>$)qe-
zHeOMHyo9m-x3D5SmV>wIW(Uov%NfAqN4Ka0)V0hvA|yf!xDQPZtf*84FVAi384?e>
zD-Z9LK^|m|I!H5w)PO7a_pGv{;jubHt_oW_ow52pSNCkABFe;=E<s(neJjdD9a^eC
z#Utm*3c#2FwyWM<?blLgqV<29QO@QzhPxSXvO9!{qqh7vHK_x1%&=PIOk5!;pdNdY
zX@g=L`<O6>f#2tTD+M)5ySOAHVv8&F+&{I&444Q06YkH~J=^}medTxDM~CqalG^B&
zBR@L*Gr*lN|I-$D<5Tol|NW=}6D9h+)kHn6Vn(rKvJc?Zo$ta$)yB)&!gjtA#D$+0
zko<MF^#_A>cPhlDW6>Ld9Jy-cnUza#wX{P`&8A$tTSeEfqffS(yw=Dw>(H14fxxP$
zM_Hm}3`4=U(GmQaLt<I(D)Ht7vHtFex{Q9giy_gWt*o)br=RLn@$Y5$njwPciZI{i
zk*;ug%?g<ChJM}TVWh<tSu_>1*5S--hP#(v+E^4&tL^X>#83*rv09|@wpK@@SSNA7
z5RAp7(Wmc%fi4*HYc<s1Kpk=ZCpjg+W|!Mf=xs#@0HQl415m^One%zW`#-A@W-XMn
z2^v_Nu)qR+yq*fupq&~gTmr13+^FeY@Tc-}5M>x}YjSbAwPT&%sAu>h!f8rYQ*Oti
zr-;!?bG8Qqo}MtmHJ*F367wVZGvCR^FWX`8sp||dVEUH|Kz(*=(%GA!l^Xl5hl~{u
zvP@#sw;W{fo?C3o9t|n<D~V&l1>WBcS{fWLd)kbaVaqQTHfLZ(ZNpa|vv}ll!`_X4
z44zI=K7SdzL7)54*pq?S<ab&5i;9KNH?G6$nnBNQ=E8(<_VVU?IH|r?o(E-?aW`R(
z>?`c}&?&qvADz1SW|;2Hw;aZ$3;Jejcl~suTar|ppJqT25**PjZ+a@PJZNb!nTF1`
z&`K@JBLnBV?d6;${0vK{OW^kgsuhlFE?5||YRT%pOF+cYR+ZFywPiC;=T>Hy7Bze=
z#xW_t$$Bv~wxIxFU^BdJ7ifO9=!)oSJ`>c;i_FO48p@zD=d}h;LXFX9W7A7ty=zl?
zs)aM_qE~7SUQHE7&b29O+5epSG5Z<s6g>Rln7HgiS7WnLwZ$`sX?@z^+At>Gr%*<K
zV#JbPEv%4G#T2idb}gI%v<$&&Y-fOfGBTjP>uWOE8|N}Abs}gN4m*-%GA5fD$rf&)
z<5~-h(08-h=EuIh(JT;6<X>jkA5stm^`wj~z9(Fp&wQ^lFZSUrv*<^Y*=i*s?SZ!X
zQLpJsq+LD7!eHB>+xWCq&f#dcNn!2GuK$vLbr)B0!`H5jcEuldR~oTENYm3bf|t#g
zboS8mtVODJmB0ckb4c-8V)dilkSDX!oWS&1^=`F3>!yeA4PNkjZhVb-<TVt=I7cU7
zklLxQP=|6p0Dj{=ZaQ~8akRNHW1I)K41aPptd(BIU=R6rl{-{ow^G$$z{wl;ab7RO
zyI)uD<q|`bWwTTZgrEo!;>fqF-|vt6wjxo#o-IY5F~l+64O5!&O0hv;z9G3G978&8
z*P6ejob$f++F^)3VCNd}fS7i!jBweDK*0{B_vJh(RZcB36WzLMZ%enD0d>h?g0a%)
zOdD$mN2u)~GCz5*7v88Gj-0V3WQlywdbP}`zUIKV1fl;Lrlz{R>MovoPiNx=S65$w
zSb1tE^pooLN?p!w&~P#aJX->KoS`gv_xT~z{tWOYFXjVEM|mZD=>ER?4&nJRw73oO
z!$W5AaE*wwEueuxiI4X_i`edBr-1N+7zqI@yo;&e=$v=EXbiN(qvmCs0xVum-wl!t
zYC)d^vOtuG%M!Hs*$f{i3qfrK9!C10;o)dq$gx7@IEbg!faQpo0n^FO9sf4uqDoh0
z^EwokBz~M;n8tOdr7!Yz+DF%l)c49<9jUd{Oo1A8Rt2b?s>Pl+p~?(7`x!eXOQ|h`
zJ*)EOM;`crRpGRE`S1oqJu?ojfXI>7X%=@5n2$k0LRoR0{j$jsm4yx1Xrr|F)i>jl
z>$kYKsKzw5$_;DX$xEWLkRugN?cw^Ws8gqtS}-iU<dwB$=uqM`&!y`3mxgz)RnuH9
zj&KE}&kU-ys@$W?(0^!^aR!J2*_R&iA{Mp7myp8lBR#d!RCDVY8t5Ea$#n3ngtrvI
zdHX6PIgsFEH@>N@ie|N5eMh!q%sKoaq&7*4$+4Qt`3#_K+l_lG0*#9`&_yFyTfSUB
z0|=@P{fuZmO~{1blvF=JEXMqyl>n<RFWuYUZ(_1JpZTy3Xs<!ci-g`>d$sL_JEfhM
zlv8UN9~~&$=ujk+Tjw*}bl3VV{-l-(CpilivTVDaX?ZK|69nNuDk{vZuv!f(!hVKb
zcdz7**Qms5(KGs1urlw8HBUvLaRZ$*w8IDjTsW`M-maboWMbFpa_(wfz;ahzkTJvM
zjUJ&H)?k~`mJuGt>1ge!Chi%qAXRVyir+Kk5KQi=<zcIGbEd^V^zaa!-;$LoJK5qS
zG!6C2ZA@Q3fG5D<i*7SB$v6V89I%01ATOo$ZW$RwF(#l6-Zp-eq}3#=a2)L`bG09D
zH&vM_=DyVCvTL$tcK6ykF4aKV{7SVc4Q_(NAWyE9?LbsK4(-UKJ9m4b6*gen#ncyT
zE<TM-IcQM$T-`v|G%47jHOpmlvwb=ak~!IOT?Md&jt0w%mU`tv185BfTy_|~zYBcZ
z`r6Miq89JSA8>1%K6@}BGv@UWvEHMpx$w(p^5#IhtM9W{bbwjy*!dsgX8>%$Kg^rx
z7YCBAbpF(p6;(`fuU*0QPJx1?x4B&JBdRh39n!6oi$>_Ki|(t|aE0G~sMg|l{A^Je
zF}sOUtV2O~Aqd@(tL9d*8&wNGD#b=|+OrY%xH$d$D#)@O6MDL$lEP`1j4@?&ZWM}3
zcUToY$!8Mb`-7PHHDYi#5I5RBnZ+H1hkM<tdWFKN=MRPVl$*K{TXhT%b{x1a-T6_2
zxjCUD^Zp*eB3GqClBw%~yb0ZE%kwAWb3b0)<E)YH#g27?B8}n>?wZHy3>G`h@2C1E
zD-iq_UfhT`*Ym2I8-mqPY6t>|J~RBS_Ar!oI|vT#v%K3z@bD5feD&-a1!fLn^BI*f
znZ~np>^ba;86ZU|_Y<8RGKo7vZWk=5a6O}pToWu4FSof+v8fvTMC&O>#nw2JNX9oI
zJHWB~Z(xQ-%ArmOkcWi4tk3ZT>@jP-zu{2V<UG5;g|sNEKt^nic2tywr-^08jknET
z?~YUrZz<h8Fl1n)a7gKHIE;mPa8uc{lpNL+STo#~Hk&I>!3^hFHno75dhk{pAb7ZT
zH|nvzUv^Fs=6R}EcWpHM9WtMIpm(K}3ngu&3qi0p{fipmKP346eamfNZqrAD6W?K?
z(d{L5k{YS)JEqnej&W&6`i;A<&;?#w<^7`UWr|EvqnvGzO<~v?oqoldr)>C$2p&jb
z&Y_<3A~D3H^E|_x&z;4#4HNb)Yhq<p&+seLY$fjdo}5n)>Ca30$*p1fr&XlpB;DjC
z)K*m9AXyk)@Ord@WjBBF^z@p-)#?=>{fD{78>yO&+GgJ_AeT*PS|XQ-xxOK71;R1G
zZ;omq<Me|1wllxQXEY?lkU0{rgerb4n%*nTdZ4NlK+w;onC#%j!}cy8ZW`m0A4Dc<
zQr!p^sg^TPO|IO1`dd|3t>x`x3phWo7*CRZbq&Fxm7}`vBXTU#iqn+`i7H;QjGPQB
zUIX9C#F`tswxiAH;q80KAd|r64OJbTpBV`wpedj0hFbKvPR#HNBj$?znYO)g&Mn($
z!_?_cz#e>d7xU~Yc}e&uu`Li|d^|2MbrAWYr_w3|v9uNjV$^JnQ`~$k3fzUm-x%8R
zq(>%*7RO(ExqHe^3M0n<3fO=*B}v;5_m6D|Nq9c!Dn5SMMt{rHF8*y%kyYs`nv)&w
zFrzdmRDB^I=9qFY`2?qg8rE%{4>m})m?KEvb()xpdJt<gT-j}X8!~Y`e248|5WsE0
zc;5CXPb;bI`7+<DZm#*OkO6|wjPfOaI5_*+)ctTgLsZ6qEA`8b=Dmi2o6UIFMQ@ez
zoV+ma8)P7R!;&)PR~K7C*5IZ~3Sy(_i}=c|gX>q}CbZvcmE!HSB#45q-<aaPw}8~#
z5y6m9vEXJa>COW5_LIOWuITTIVDb4`O@-=1T(*IfSbqBY8#{Scb(n3d2ux39CP!!k
zcOgarNcb_OF<dTBoj{ePSVd)iB=@;$B`*wEci9x9Wz>PQty|v)PNFymG$+aQYgn?Y
zn?^8Peq#AY2pwjV1za}lMdeX@^=KVi2P!{{8h33HhpH=U=>N>CLy9ygA64|80lsqn
z!|`YQbJT1PI+X?cMX-65ZzR8|9@W0RT4A6MvN1tz?jKl>aQDnUX{A&rOx8I;qwWT=
z+dg_3R$$r7t{UiI^5|0fYqsNBq12q=f`wDH)ic1Ym`7os<RG!PW6JtYznuY;_E&ct
zFM8CIwSFJYHrw!iZE)kA<;|EKS;WJEo84kJj~YH$)r^rZwi*F-)}5-m($RU(4ygL~
zM|^I!bza=4^qwzSvWMy~w`;xN9mm6&u-B~;dI$6UJ}a!4oF0#%!P^mKGMHL!3|bD|
z@Q8d|LJ#C%cRAf>PoWyV)GVe?BT(v5Fq(ZKo)8ERuiI8mDL`OxZ(C-5EV$qJYB!0y
z8iLs_&J(2GvMxcW1|vOZN=cbY=q{VQ+m`v20Y1#jmiD$hsvb}YkFj8}<dq@vH1@Te
z#`VcuSqE5MA*F5198L;(Z+!n2`q6YPu6K=cKjqn4jNxX5BTQ&D+1cD+w$Ao(+H1$+
zu(~@w@*>?<xz+GR3zi`7)%z}*GBd$nyt^QX0!TIOYULnIE`NG-i3gwHMGyw|-+0dO
zFULpIyUV07?qA}CbvqpnkVyJ1<BK|xFHtFGvIB;<!%Fo77N=L0@_@OH8<XH@mlCxW
zw5bEucki0piyG#XDr;k2;JP74DQ2zui9n}=+P-pUXu2HD_ZK%{{js=-iI(#X4nEih
zhQhqmd&I6<2R<|d_bc%-;SOcDlX6FvD;xap5*PTgg#0+k;Iyi=)sbti1{wJ*R(y}b
z73|f;HASrEVXQfh`O~+juO66CKTF;hUNXv`cAAPUH7g4?Z>jbFeD@l74VYl2r1N-l
z%zg$H(${%`k$#Q>m#ev2#V`GCid$XXunXs-RlQMn-aCK?@#I2M_(VVns9|MCU`n;`
ze3tIa@b<0hyDhb?!HJLUGo}mTIu-cs!r?-EV#)L?12Yy|3pLU;l21DBw=&p76586P
z7O8}%Htf#@+nrOht&pN*j(RihjE+8BC*Bh<=D85*3NwzOne%r`q7-W#%UEc!5W#(`
z?5rt4FFBmMT;kWfJ7pJdr=2k3qKtZZV)t@DK`ru)1?P}i8qUzmXR5{`SQ@M&%vnx@
z`qX-=n^*Yl$p0skXvvEIj}&(GKaQrqHW{a1M1t3FbwLC)VzApuf)No5GI|A#*tjjx
z021<iU%RY=%LWhoaE*G+=RtKO#GpQy4A_8Bm=4Y|a)40QkRU^Q$`Y|)-1e-kOUzGb
zskLa33U1rrPx(-$)~s&BdovId1~pH<wZruk2O8)n*BraJAA3)s#$cr?jOnOOH5NZi
zJv(7)R)j3GJw&^fQ53M$$d@QG5ZE>d3<m;a>0|Si6b{|}d1PyXlWi$Clh}?{I8svS
zzj1x&bD8nIxT)<aEMb$>30x87H4fgi_W01k>K$}30Nu5vXY|`0$6D|i0F@ZSQK#{7
z>3jZ^+Jx+Ju%+x1`2!JMf^}oNeA5*(#*FE2=WRES_vrV7gxWKUd3Z~z&H(gVyMl6N
ziL0iagHgD3JmxN}e1oNClB=m^(kb=TV%>?SYqn2xR#~^HCL}lUC4wKT<!G<93_DV&
z)v6?jPWT?l7DbR{fGE4s7njYj@40Kt)J~||w!UZR=?WBf1f2@Dg2S5_vHU%2w{=`E
z;C4bkuo+9wZc3+(Js=%RMdqJ9XHxU_S}Fd>eM1mB<ZS9)BfD4k0pr*EuxH1xzTm`Q
zouCAUl#8xqrclBzdv_V*@?eiMcyZw7<`L!VcMA`pi~l%a@iC~cqyEmek+MFyM#00q
z7_;IWCa=r+hbYZ`m4?Njk|&^TgR-{#%rGzrxYRY7UBu<(+)ADm{i#$06CyS_$^yIc
zG5}CjFrYJsW9G|5c1GFk7y4XiDx5mPJ8<`vctyRnQxCYP&b1{Nr@%^EjZr3ZyeP?Q
zl0Ib_3eU@HzL0`id3+b08x%ZRul0pUcyxyiohZn#y)?dNYM0pOTQz9`J8B80&scVC
zU-5X;HB>oZ+=)qMjzrWXAR|Uo)HFE}H-fF_>+*P)O8VcqrJj?Gw;KpHH=tf|_<{<u
z$#1ETDj6w_AI&WZ?F08*Oejwk%?)^A<%<8}<3Is>!XO{`8Wx1ESjsDp74Y2-SiH52
zzs0N6kna~Vh3AAMT};bRupYEY1%Rfv3eLqX!3|DP9S?NMrd%JG+yVC)Ot<s?tTBLH
zT6{wd;VQ<nM3f=tmD%_w(^((Q_&Pn+$iNV8*#JW~s-KnfJ+3dTvPE9JcZ&yxZGiP_
z1mq*F0!0kRk3Jv6yjc&AY$|yV%3SoPBl0avMLvgSr+gosY<03MFVFS}PmC&lm7vc{
zV+oi2VnrvAv+gF6C1e%8i4VlBd+@?FuQFWb4BVQ0c2A;Ers3Vd;j~xLMpIqsfL;3a
zhj<T@uda?i1au6r?3c?@4v<^hvvoB*zBuMF%b7-q7+A`C7ryc6e!H)+QqDc-x2q(h
zuKv%t*niE`{z{_AW^bmsg62^HwYF}Tlw=YVPgLIUv3$)C5X%-?qv#GZD;{7OjHfRF
zL|K_gSYq_VZ^3o!Qp6@VFYDj-P`uK`RVJBG6r_@SQQbae4&Pz_hP?lkvMk4ucy8<&
zfVQLXYlQ#zHBwH}mdIt#W|U}^G4lXn*-D=#KQa*H($Hdm<%W5D6~y47@4Eg&Tus=m
zH`LYpE+$Oy_E&3!@X<o$uH2I^Do~)G!gVZoquIB<)fKnN!Ti&^p9SRCTy386z+6|>
zvSHS4(QgPAz?1w&(KC1oHy`KpL2-j*C9?H$6U&JMMbpg#6_~9Qef$G4i-kq{{neTg
zr@FZpBfr&P#4IROh-Uz6l$U04or^(nZuE!B>zbiFTw5dbT}YJsK!BoB`t-@C<Kk3X
zW6mLjcBHjQOAEAUqq>QH>kjCFBXbI7%=^Bnkm!rj;E%}nH^cD23-w~cEI&?4eB}*$
zt$uFWPCoU^y6fc?<N|M~5)2#&Iu%Wh-gY{-Z&5H?cYk@k_WXsBGXP7)=9ha%^$18u
zy57jD@QAr=295+oa0Xo2qeKvPGLe$(2}vn}uVRRU<nfp$e@Oxob?^v}%ck!42}Dp>
z6lDnDAs(fk!JnE=b(8*d3Z-8xV^*tv*@w%&XXAg_hq2(7&H5@!mTG}bn_I+AfE0wG
z&+><nEPqdL&98*zx%uQ!)94|gLYUsV49-{u<6Nh(t;IN9bwn;k9%!rNkYsPr`30TF
z&bPF<)jRh2E|Qe}>aDa-H4cePhOm#e#0DwWD`h@t?NJYT1<5YY5(|c}Z9vO{ASmAH
zrJOG!3#n+6x5vrB)_3V~*vkuV+qMzyNHAFZTsikggjI2y<u*DU^YLxHMpWv_keKns
zP(wMTv^5Iq{KhCEl>NopWqeZBt~U@jk<XHn8*T8+vBeM-Y&LR38Y$dSM$wqjAoRct
zn&oJOt34{1SzwD~#`>-gSm5lBD{&ssml?#?`%}Dn{r$rb$LUQFmgDPPv>jh2yt-Mj
zL7nb~*7uIho4U&fkeJW+j_{uqQyXA8*$#36YswZ+zG3!8PoXzW<ScQX+&v{gG9%1)
zU~(RZ4s<$MOh)W}8$7i+ONe<SP-cKW+oS4rScA)KlTtBbdH~>~ttlmOoo>xdVO*g3
zm>>oruTMF0%|8BSxgca^#!>Dt8fC@UV1{T^{Q0ayoPm!%X5IU5wLSmz-IPsDs!iIh
z|2NYI{*M%j8*7BY4qQiI?^Je83kPwPr%P?=i^pnNHS=wqOeRr$*)m)*L1x4k2jXo>
z8!3Hj6&&NlWD}O)M-!uu6-KtA@~#Akv8SDnakf8x%iqy1mOXLBv;EcXG?CaU;}_Mw
zSQCU$u?*}rRBrNmo&Wjuyv-#Wnt477l}b**?vUTo2cK^HCW`g(r=IlbBiX$f>02t$
z@VQKk)-yn47^pymiC{7j`6!?-N}j-LQhv<7iFaDko!hd@{+za?*TH_SeNabrPcSEt
z_{`zxbyn$v;-1x_W>10Pu;a4%TaXf8LGjNh>IbI;^hu8OiAB}N;5;|AEJ{q^WJgoU
z8KBAhS;{Fvvi52;k)dj|{8J=X(}fr<2l6d?8h&`5=-M8xH`L-8Fqk-iXvp)sK3NaR
zH(N2|aU!$8oi1JzIC>btIT&Rz#9&!oby4S<h2-OqbREr}Zg5f<`;QaT?;ntv0WX4a
zG02*Waj7y?{#RPPU0=M-s^UP*A~O@@w!ttscS&7=4C9N#^JTDN!Ilt@5mQ5q9pjCf
zyKa#d2j(91U`7SJs0P5I#{K)NEFZf&_61_wmZq*M#$JbM(q9#8A35q6%dG`ch+Z8J
z#!NkGhmD4(9^FhrbFM_(ym*?m^FG9FJ@+7|D#|chS755an7jb8W6p@Ob+*OfbfowN
zaXZIJ9O8Inv9qAUNZ{LcIi6`^*wJje7kB)Osg|{61$b~I7a5OhRL-gbJ;n7lym?p1
zpuT^7IW2~3#%5w<c6(V1_w`$jcf+H(hf0`c?0h(j@we&all|pyEECAni%&zh&Hy`e
z&yR(RJBo3(4~;8$&(pIj=Xoa=8a_vt*)L(SEohe^jksCrhuI%(&uKC>q!f1>Tss!L
zhDSU5d&J-1+&Yv}1#^6^%&nm_**LhH@EpcKpOvDE7OCpc><NW6?Z@eAswzbzTU}Re
z1RYw<zut*q`DE;bH7o(87@Fal(-v75_f`*tvB~TnmV#OHCZu@p6L$6tFs^$Bc*Ycn
z35Hwo{dSAt?Lbhc->bGSn4!~1`JbP><qnuq)rlkuoizZhD*3K<&&8haNH*G^<wYYi
zR=M6|3pdtwKsBA%Bd9G>a;~@vNPw34^N4f!+}&N9j@#JZ-~8$GGJKGRZ?qB$E1j73
z1%*KLwBZ;#&fat0qay(Tauk3U5RCY$nW?9{f=+Gi>WniUQQqYEni})SXYNZv_|v32
zEa!F=)K{!~m9SPI)(SejxkSF>hrQhvJ9ODa`)ZDSU)Oltq$LxvYfphcmC0ug{f3#j
z>GZ3rDWz4$b&2GbSg=LKr_M*$Gjj=+9X^skthd~erkjD47pJ1CLS2;g!8xMW?hF5K
zUXjtrp?hN+n^hob*2;Eswx$~KG%z1Q_8iUUTZACE{^anuAe)Dx(9~vFwq*oI6KlF@
zB<5M`c|&N4TTql5azlX!j&bst>ZzN4Y2BQV0i}P;<ZO;ba9uN_w5||p)HX<&quXg-
zgG!ZiVGAlg#ha%eIdmE&ScbF+-4OD(g~-1ny>-Nlemw2@cm^mDEbf}0FpT6pto)Gt
z^ybOsVprX5SJ=60C(~ww-p&<V&Q;xG`lxbT#73UZC<n?iUX%2qy!iyi0X)zz)-W0L
zhE)MyUy_4EL#2S7C@Hb87JVQG$TjyO`T_^h)C$4*yjCa~+o(gA#RB+dhXkcJ_)A4K
zJ%`KfOWt?|f}KEdD8b&qd&5;X$eRv`ney7EyF}d~D1I*D#W)2n9Qs=l4F(OiozZ`v
z&(Vyqf=MXgGi08Of@gb*ra+0a2->Q|a;~dk%hWOJL7gPGNpl96z)l!_LZ9kE>c4%8
f`N_!(+H_Hqxw!e%TnBVJ4kKDb{$H%;&*uIYXbE(2

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_low_gradcamelementwise.JPG b/Images/deepfake-xai/ms_eff_vit_b5_low_gradcamelementwise.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..5c878a71c16ca48f55b2445df67353a4f0c3227c
GIT binary patch
literal 11845
zcmeHtXIPU>w{GZ26$R<N2k8k-5D-GINoWC4Ite{=EQly*=pZdLfdGLJq*n__2Ll43
zD@qpx3yP?;6W{mS-?y)Q_K$te+1I(wk2A@W%*?Z9t##jP=9x)mPUlYF1J0TmnHmA8
zsHgx@lo#N1nR)>dh4cmhEG+>*006)UpryJ5pr+*L001hpL?nQglBc4)Z3>J2d;^#O
z0P3M_0RRJ~j)hXbOt}C8+mw$#cmIC^*oAtbDKyZf)BMFFCxx%#blQLB8#6Qh%F}>V
z0f5ZRf93zhDOKhl8vj41|LROCN-AnvN?<KzH4$a77D!!7MGf$u)d2u@fKY%J;7<Tf
zmub0u!@`2K6cqzQ72G_7+`SY$f&vty+=3OA6_gYK+PYD}ZXSMKVIuBcKFB~F>Cas+
zq(zXPI?@hcOC`%-1212sNetQx7GnkXi1G8#^pw`U%*v!4r4<zr9N-n^CK46k9~i0?
zr6cy2b}dT&Pqm`hWo@*lw-(gU=pPc4l8)Fve2I*VREShn2txZPDr;(LDk`Zcs;J0Q
zB;-S*1H;^+<O4&+|DnLpE7Svx3=TsE1&aJp=;j_29;PD}9**?X@^<qEyL)<q<=x#>
zJ>`{^J>BHpJXMwDJylgb)xpXdD&ETOV*fDj>G4nN!Qp8CzifMYD0=yO1$YI9g;Lm1
zRuohG+v)!<n-q%vY4ks(pMsGhqh)~hatre^q+FN(1XNxLEU&Bv|F<A3fK}AB75{^r
zw&EW${+*-$y`KM<*t)Fk>7nKE$M+!g-{xUnA^&I5_D5d+^XT{+9i{MhhN70z=gyot
zqpkSwE&k6G{8dBga+G%e?4XpEBL5fC|I0@HHORdJDI+_QGK5bT0R{jDIy!nfS_XPL
zdPW8YMkWqs3dc-b>}Odxc)0j@dANAE`GurK`2{5fxp~AC#3W^8<u1zci72Wn0#&4e
za=<@Is2CU-na(hAGBa}m1$YF2|I6vL8^FOx1*bYsLnQ{F=AfeCpgQdX2vep~I?5FN
zXO{j)qN1jurK4wHJj29H5okFJpr)dsp{AvwqobwF&s4W4_W`sVbmv5s_36*sx-p1_
zaH_;*lroB6Y3SjCPJEG2bw}Sl!*qe0hnG)MN?Jx12m-6AYiMd27($GUO-#*Ta65Yk
zM<;}br<b>nFA^0R79J596&(|Q?|wq!gNK;RtVh{7xq126vT|HSWfi`<rm?BHrM0cS
zqw{5NU;n`1(D2CQ6nT1Pc5Z%Qd1ZBNePi>(*7n!EZ{PQS9Q^$C`wuQE0L|a9{)y~A
z;Nqa*qNb&#p=J1ki;6mua?x<m(upe5pVPNxa0@vvrV_`<c_pK?q34XaD)bANJ9>iY
zf&^$;^6MXHe<Ay?0lWLZh3uce{)KA+z(PYs2_6jxKo{`Aza_<e(LD0mw4!=A-6xlf
z-!q2Apcgi3sM*gu3c^RICl7+WW$NSa-Q4JS<KTTEzMR?GeqXb~AAa04p_5U!Bs)8r
z2_~6PM+y679U*)B5eJzg;m^!NuInGV6ta&s^y?V;Rb@x0uNyu7s{Y{2DIk5L;uH{<
z8JiNmBs|k1mlURon$1+=37BQG{#JK^=L9v~5-q<NreJeu54CN1K!lJ}vKqFMYOY2{
zaffE5R>Dree%84*h3j=2gHJ2l%`I=0r85_GS%)lil2jdB*^HK^2ZySADylc^tLzI7
zQXjJR`D}KiAG?q@PBdmW^P5ig(VaO>swRdyI`=)S!jX#y{?tM##eTWlN87u(MzEam
z&zFA8hy{vgH0{@9qEy=6V)eg!xaj`KDG!K~xE?VoyHVR)mtOK*zl$v3078E>JKFg!
zx%<XHQ!5{Q=H)41u&?}t9`$%e(pxmhOCp>a){pK?=#Y3kC_M7~t)=JZtcZ1Q+TVO1
z69gfMwToa=!tmU8?rCY^y#w{*BQOo>K=6t5Dd0ubk>7*N?G?MeEak!Hz!wElu`}w&
z-~Ak?Z0)(pxPA1c&$~&f2U(|pcT*>zl7$25lWRVdW<(bnOOaDRPRcK`2R!GUsMPh#
z9aTe^-^oN=n$cD0+9_ak=M-@8{S=TQhl&)9UHu(<_&avsY5R7S=DF6`CD7)~r>kZK
z2|aaA%i^-v@(YaFp}^tt35l6HqmbB;Q$Vcq5mI!f?u}Pa9Z$sN(AbpdCkp~;`@~bg
zE&GJTTN068%*vJ>J$1zpKVED4RR#r**Rd^uvpLU7PEJa`;!Cc!yIy{)dwd-9waywk
zJzsgYN5)lCr{8m^34L=wH-!V2)y1ilRn6b#F<CZUClEXMZ|+_WfUZc4Fa7~Z;bx3j
z7dTCAtjbh<i@3Bn#Ifn|%ifP!=(b<OD62zmHYhT4dm+ih?AlwZUDenWFI4CAh6}^u
z(~5R6XWQ6xiS`pMc~8Xb(&Iz~(hwz%QT?fx8UPN}{D>Ki>guVQvR9{opC;gv-^}k$
z0q@NzWH~l}RjX#XUqya`bGgpmtbEdvg;7RGbO}8w!#07g^=b7)$kn_9Sm_;_fBiif
z`%=F)_BG=06hM9L$nO%Uqq3}vX&~`Pf^Yno?#YWUlBa-yI%Kivv~IK5c~oRwa!5kE
zyKa#i>`^+Jjj=pCLMnJ8$+)VrBG@Lj!jeKdE7#^$t!o+GxlN%m%w9zfr1e++SdOTi
z@AZ$3pSs|KZ(TM&o&vUOYbIu5S5E;a)W8I}8SnP6Y?(PdRhZGEet~IxE_Fqb?D@Vs
z*Ri4F^C*=l{I_xNxRh^(N`nU};KS@L%P32tazpniz`Ne&{qzH#Vc<ais6g13BL*if
z(<0tf;M8(wSNQ4epVehwxvH#XPXVGXLsJUBu6Mm=$&_97l*^39jrqFaN3>8uOQ<s6
zw_=IwUIV6U_WX46-}nfU--I~7+UE2A#<JmIDPZ56qVj7o4@k!g$qLAi7kt6>Z`Z4D
zZI3Ljg;l5xr0+o8ih*78;`>DLpw#Asv&cCXB&d%@vj&au|KL}e`7U+;>P~@g1g!ZB
zn=HsS%iBHSx=-4xYu08<FFI3YxWtwVgY`cEOXgN|sy&+>2CVhtzT@hAfM!srm+(bi
zb&34I$~uS|H#^C^3Q6zR*ryENoWoA%LQM0{v3SOYh1VF8`(B=3^YSwh+;t73L7jMa
z+(H=W^bDf50~OC(z)%lJapVJ1Xe!O{v-Go8W>fQ+bC2p4rXIJdsarA>@TH%xjFR4R
z5S#XFPuE9pq6@G8o~h-_ev(x$6mmS}a{r^b8-chqo?Tc^8$k*)Tu-i&dQ;vyXl0v=
z7vGjmsEtJ_RQU1JGTFr(fEhPhEO3hssNN7GSzm&|c}`n%6WX#I4B*v_?&={W*!Qpf
zTiEY_=2O_0`|`Py$wR?Fu0YxOS}*Y$ne-N#DzFfeS*RC0Hs5I2J+qYL35~rR3}^3T
zkLkO@g??GV-A422)qA!Qus)ZY>|nN_TEz-S&|!g+IvdY#2kluEv$cj{X(<AO1o<JP
z-|P{RAU5(1Rz_APZ+Ao0r0l9!X~YPGQ0TZw%YC6MfRJ=wIm4Z6P)rAk5K_I^&G<%F
zYcptnV8Q-;)~-{-!qT=6dQCAx8siH^0e2QUzH&N=TH2ig?xoz|8|{3((P}vv5EMLZ
zykl|0PALiMU~C>lG!55CRWjABh){M}ECT6^Fr?mo-m=MO>t`PSLX`grSDRWD&%Isr
z;A^%Pg=01WC{eb8s}$T%wtcnLezR%@+eAPUUEZWzY`T<*9t3Mkd94=O#3)-W6g^KQ
z=u>xE^q|pa2)Wp`z<&8FB|p;4R?=ISzJHla)vtN|($UfE=FSXu=9bYiCgoMMbUlH=
z`0Uo6!rQ&<ajSi^f#X##9iA}oY&Ri1urc}U%M`Youxt>qq2IaN01XuM76-SpkBt>i
zKf@!E%!@_<M!O&URDRLcV<%L@$Sb(yx4c8g@3x#TCMVm)LABx?XX17w%(KG|dzp(N
zfy2SGDVu7Vf^?|-vBe<E2O(HB@|{o<h;M=_XM06i-o1X#W*d5=hEss>uT#KvI}zZI
zI;!WZhh^1F`GH;rKYh_t)pYWqq@oZil7>dwM@wy(u{5?Ho5|p#iHhQ!c8#jeMzTO$
z;N9*S^#qcU<nT%y4q@M2J#C%oji!#rly2S^c#mzt87<sjHI@-_eUp6bYml{`rtG@C
zAMY*{vIU=43>;c=#D6$bcQU0nRqoO`U}<ox%BlAhAhbq54%CrRtwN|~*S?Zx?^SkL
z@k=!;=_=WD*uQb1(a_oEe8^fHtp7PpUS8>sirkx|e8-M`&=h~~h@B?x`%Y-1qq%~`
zy@rFz%Irb24^Qi~pwR-*4sv+_3rEGB7}$%(9znU#;UDG0WvlB-xh8a-_V+p&Tk?!d
zJno?Bd?oaRvXunBL3*>_j3gO<^MTW{4)TI#r?lI=G&D4Zh<8fBvdMf`LL!xS!zRE3
zFMBS1Tk4ciAnvYkoJkMh8GzW_2qD|~v~6M{HMd{?c;fJQ*K4&SQ&%=VfI-4uJMp^6
zcOpYt0IYMM_anS$1WpzoNp`O1YLoZd3iN3MMAXi6s?#^`N;jln2Ic~L7`_q9tB0X3
zM7aw*ZVeEP<DK6aeMA}0+vb2O46si%fOA_mF8&5-Es;#-A*erf9noRchsh6BeJPb8
z$tx^uW=wG?V{O{2=&Irzv87+@LaZ`oB!8tqz};~wTGQeiF0C6wLXS8R_BV3+P=g;-
zp%>+cdPs<grAtnx8Bqf&3}Cd=Ip=sW5!<6JCIy!ttwjS{IWzUF2OPYP`(v!gS3mYt
z_<araScCv18-08Cjnsq&+@p@p91kvFGEAIH1g~D-?4A>6weE;eHltIJVP@?vr!K-M
zn_Lv3nas&awhh6^_8`lOW!h^+1>g$wQPtQ|l!7H@4=?A$?Q0a`d`<|6D~+TcZTHFe
z2!D_fmuiJ#4sHGJ03nW|W<JydKMkNS-EdI27SJs|)B3PCQZk}8oJBW+y+svNvq-d=
zf`rK@&;6FAlXs$m$a8iNPP^|tdB<=mVY|wH-}LFO{)3aV3^O%UFjcKHdvX}rZW}rc
z2UV>}Dw9cC-U@dH#CWk0QtgRY*}wzxq{rdaRZU(y0bv8gLQ$rC{jC=&=~R^;TE%8W
zGrqOS?-R5rOM#0aG_#q?7VndbEUtvijXsQcVwR7R1Ova>jJt`2fxUsXMW{EoYMg3q
zy;U)}U~2fCNts(RA1&z%WvE_a=;TH3^Okc$+Z_=yWq0b!F~Oh-^I{n)Q@kT41kB9G
zaS>x+=XF?FYZ6tHsZL*;VB2|^2#(iWm_HzPW#y4VXVGpApAmW+8RRW@^MdPK=GT2N
zg)?dKZnouwRP!}#X$S{q!%$k%nHWw?Qv*kDo9QHkRggdv*v7#PjQ6$-&4r!VO4EIO
zH-WqNMC#lr;F%33KQ;H|9Tr$q9M*!&&CSerV^&oweX*mpipxw^{j$kt8%wm82I=yf
zb13G9rL8@};-Kiu9GhTEpLfF>OPZ`5#?(#31D@H?^Ko%mleISUMJ!rX3T_`^#&+bD
zs$??d*A@@S+3>RKi4q$SXN+xp-XfVryj*U8%}oEK0ZT;qR<Sj+L{*T~K^?drW~{0u
zh{hyR*o2kS=28;&%t{src6J7jsMG;UKhAubh^bKi=-_$Sa>aGuUi4Aj<>flmDL_to
z^{T`%Yv{t=^-}=fq|~{&og+#7t(;!AGgodzRE_knc}q(au1&MDMr@T|Zn-%NkI--t
z!>zS>8YRQN8p&O8d}C|12%#nqku%C)WJP{O11q;bv}nBd4OQ+D4vQnHu}U#Ce0$H*
z|L88F(nkll1+y$K#JaISQrxb(G>;eGZ4PJLNLVg?L|J--I8BMX_uwLlPo`iy`57<1
zY^{C&OJA%>Uc`NMvXjSSk#QeygwMEb2Y_y3gUtL$0)9E#liv6|z6?QQsGLxMadaQh
z@sJ5WJGFTSasjyM%t@HTw5FuX6-}<II5oAq_pM5|O%hH40oHWAS0a_|yH|;qVKOtx
zxYfCxbiUiS{0h>Xb2JV}{J;Y+DCP+Ig$F)2qY9lbSboN;q2t{flZo$zufU)a6{B5S
zkr;gAWbQ#VuU}N<r<xNlAuo{*U$AH<A-esu1#;S0Dv55i^qVz{oM;PU7J+>X|0W8f
zF~;%DZe$p(b)|Qv{8qcPSVzM26Hk}^^Jk-Z@Wlu4OqH97Oj_9IFlhw;RYA^FoJVkt
zD$99C>g}J;mFB<qoBv3fd!#*~VepP~=d$Ou^gyn-Y8eNi4&q#;Q|VAGD-!dD>hm!w
zqcAngM0!V$@426@sqP0qhj0Cq^-=nxvMi8&<&_Cj2(_#qRJaA`ai4d5rtW9NPlg*v
zc7}{BJ8k_l=J#(NuMb?#F1Y@xO|QC#sy>#+2`-kRLGE1~dh?ZfZlJc`pjC_N_=W?`
zrO&yOoyd7yTfnBoJ}LBE^`ZXnS@75hoJ(e&?*ffPh;t#&BU(Uem5!gHl@*<z$2uop
zxM+yeH-z@2XI2^!S&!#QNAh0?3u|f^FS@Xy+i!l;|C1-Z4DIcznWpuEQJ0Ozz!o>5
zh+@=4VpE!~F04G^lbnVH>VdxIMw#=WbUXcu%&c}GO@@mnwrOOO)ix<Hq9>+F6MFAS
zc*lp`0HlP3ER@GyaxgKia6CwAGKOlhyfM|WR%~fgW1Rkpvn$B@TVYyElH;B@G|clC
zPx`#eT<>#J8md$}-^Ht+TrKU$V2!oby`EnMJpFcUq0ot4w;HS9jr`6m&#%zWHEQ!x
zbv0O05k6wu2u-IgY2O}DjxpesA6R(x^o1O65AIpJoMUbze<o4Pb4a1iJ5nPw-gd!Y
zW64Ou|L&XQ>1e4{qT{5;x4O2hzMCiDGGjsHUQg3R7@lY^i8PUF^T@JhQoB!zP8q0`
z=l8JSp2I#^dfWVUTyT_;9OnAHNGI^-5wi8OJjeEf93k&hfTZfI2IuHJXnsXG{;Uq;
zP#t3FaCZh`Z)876ggfy|sI#zZ)*uDvVt}o6&fMOqGiq)@ybPLe7x(I}IbZvJ=hnw^
zm->x{7Cc)20Iw5h$1_KswEoRBli;Untoz}yeU4m|z~mc*s%bKloQG&AL`tSoEP;5t
ztw~|L^BMn4wNZs5N6$EYk$hA4iAzxk7X7_;Bc*;@`L69;cEV4qRg#-^(BX<6;b7`1
z@?MSExWofCJ$I@)@7jnk(8cr22*C8i9kJGWQ!#$C)yw;~&R8R*OnhBp;|9T$cbI@t
zLEfyl6b~&VC)N7+?z-Rz7c<pDXPyb&pe;NL@p$9=@cgI6!AauTSMuk3v$iu^d+um8
zY&qNfjFDr`l)Sogsh)VuWKtX85|@$~dOd<LE?-<KvhYrG$4C_+&#~?<>C{@@s3d9}
z9g(du&V@QODVA3=U0r3yE?@>UTr>sj<Q*KjZGW8&Ir}~h;h_12>;;W<7Ef)I7!h|H
z3%N>r;Zxol+%1S*Y9H^thrKlEe$nQZ2?ce^PO9%r-vl%+6RaeLYC<acnu)XE{6;E4
z?{`wKEdw46qDaruq<L_vdF1m47pK!I=d5eq@|5xJOFyn#5mdP}ZwO_n<`ti5u5I>h
ze$67o{Xxz!cqK$=9MaKexy-?4xSm&>x5|Z(S94}qP!ZqJ@teZ6CH&A`60DU|_-ta?
zThwi5WVM+H%m$iG#t4)giEY*Dd3r0iYOWHYw`9MQ_uzrH-Te_k_+J0Z<J7W1r2$-;
zb?R5cJ}>7=i0_&ow8)N&)bQ#<gGD0%9(Pp$E7E=qbgnKIfi=X8K=j$1d4I!87D6vZ
zoe^I1yEM1Ro=6mqk#b&mM%c;X{g|=Tuy42|hhA~WZZkp0z?HU)p#@p@T%UiJABUpX
z?r%URVgBHXRGAy>J6lid(X}c%)qYXs1^K3-qW7Tgg-AIMrG8FZy$AbvE{O+m!wgbQ
zdfkMD=KfFFrppdu)jk*~B8fTNOr1x>X~Xa~jeyT&=b|9njGTX&qE*Xqn~ofU{Va*A
zBl(VVrC(-QkOzXi4L8-#0k@d=zj8}3<q=MCGa|RCTZbw~^M~LkI`CT6&>uGIUSu#Z
z)UMor8sCBQf0&`UW%lEx{#~HSEEvrbSDtsX;?h-FTubM|y#yEbP$%~h6C67)P=Dpw
zxmQm8RY#MGG#jh5B^99u7aD$zuitug^unyp16#DNDTVQd``MVM9Nykqm1@YfmME5c
ztmd8_aQ8>)P$yj6>c_T6zo$bBg5#{5;T0XArg*>ST9^bJ@~h^!jk+eaC1MWR;gzWZ
zEc_Y>4x2KIhjfoQUN!L?5-;VtJ?Vst1)>=TSx|aq@jNw9Wr7))D<5W$47lSM^7Q0x
z6LGU0srEKZw@=P%nQbv|?Kjq~1!j+P6OaV6*@q5{-`tf_a?HZa)_3H*)ov`9ZM9Oi
zzdck|EUNz+-k}<4rs6_8x2<L>`83imql<>VaY>tLGFqUoq3ImqwH=Zh<1#BC+ivt~
zsAL_imK`kejP9rX`^*@4WIU_g3NNv`oNoe}lmu>xG*y`Opt;}E55>`NdUs~NUNO%E
zike#CY9wTEWp`A<oqoKF?WTh}7w%V~#u~2)R%zI<LIq={+qb<Evwrx#HZ_13JLFH|
z`+}!SSqfqtv{xc0GtRVm1|qxNYcIT=$ux!qX2gv&NF0Z_><Vhno)dQv1?#f`$yPvK
z-^qtH-{mJ8-LkMRSmH6mSll)FelEszzL>&^;QHSd(8-M?OfMr|60Dudg?iMuMs7$R
zC)pDVu;nFKqVw<->XTsEu3wd(hVXOgzSNV2m|J!izt%KS)3C&UghA9!WI9-~IH7CM
zbHF-L*(%JH>GP(k-(3joQUq%$RkNw<cCs5J#xmk)5K7#F$t{3n&k@`*ZLwA(imN~c
zgcDPF0~LI6Wp}S&WFE7bWLlXH<Nhr1Hd?(yB<=ehP3xadL%=G7;(cQnCx>c50xzLp
zz%&qHUQ)n-nX*(_xe=ZNrgQWZnB`VzNLp`YG{_kQZ?)!DUr0}r_8Jsw8JEene_HCr
zIzMJjn={^O*E7i3=WmyESlRVR*{0M8P*Ya#`?+#$rK3#&BI$y0g?JCl5qzh!jE1pt
z(gJXTtRN~=$ATYJtFmpzsxhR4dVG_~LlcxY_{MUOz9#V^vZ;gFj_<3ROB*U3L|qb7
zO!duPr?Q1}hc?l7oJSB6(HVodm2?wz*=;zmGUdXiyo;Sy;9?`@x*}*eI0mj2&u_}C
zs4P@GopHInL5S6M;*3$q#=|!G(E|N50K6uH0H!Xh*-G8h)4NGzns2j>D>%phf(|{y
zWOhPX&OnU|J%aC=St<W=4Ee5gj6tK657^Zhq`Fcbxq_vt7-j%NLrGT77@{;d<chnY
z%6??ZQ#;=awGnllzRUQbg#v4ow_mA1K!Ptx4dX5d;k01@;}OTj`;V^E4AIorbom)O
z8mpfIzEHt$RuUI9eTO@d4P3#fx3znEgCqVv_?j*Lz>O*n^4y|PEy>~aZr~4BT5Awq
zvYVIWn$3iC*4lM*47q9@*yWLae=G#?B&G?8>FC29b=!kEQnX+xh6fHZx>8s=U0&P(
zYw*y4`=F5~!3;>t<#xl~V*0$a*(1J()|_8_1&1SVOxR+gK=}qv>UPnf^@8}<(+wn(
z3+bdw!gTYsKx;{3n7w49pY3}80Ys{Z$SB5P!9Jqg9*WIeb#NHZ7P9sZ`hx%AkUhUL
z%4|m>a7j;jJ&jxSPwC~YrCC01+|4(+xyto4sE+-pMfVQp7>nfm!>kBR9fF^43{Q+o
zS?xURP{fJOK@XRV@f%D9PF+&|l~+PET)}NSv{u}7&x(AexqS}BGniZ11XHz?zjB_l
z7%278nxx!0GohL#m58|kwm;`y^a4CfA54~zDlqNLZSl_^8JL{3#a{n21mg04HyFRT
z6#eqf?G|I|vYdml?7}aI{2ysCA8r&rS{<c3_nNvGnAcvutMOSsQohCff|`AF2g#gu
zT%#5!9^?O9W)L*(KlKq~4tstv20nt~s<v<N{os5q-ip4Qz8Gjs&H}whHI0dXw7E<;
z7p{SG!toUgH3%b;8t^Y#4%E&|6yJf#e29-*s$h?s={WQ&TVL7)VzksC4lkz93Lu&S
z1%uIYhRxy&(hBgfQ-Ckz@RjVxXbwTYgsWC-S?ymPh(e!C3=1@LR6-z0S;>(;S2qY@
z6{)J)TYLxFGmvC;+MvR8)tB)zExVuT+L(ESj<*q-z?E`Aaf0}ZhXJ38!J0G`HXbDI
z-|4wF>B2$lwrAACY8TWh7w`Jt4%O$axE3RO!Pqj?DuyVxODywHzphBqyau*1eYWME
zyVO~hD%+%QX<7_yw!{{ItK$VMITu`}84<FS2aAQIjF}#Jce+cNx_!%{N2;Xr4usJl
zBE#y60gb*kRCc562Dc!|ksqiEqaoS1Y9a~?9Rv+yIENoiP1n$(()reWoPfosD{tVM
zRX(5u2r!rAW9`keHuL-q8S4CrlQVgdUk+KPogcs6$pxNV^!*VcWp||nF_1c6-kvfz
znjk=4s4LSy+czlOXB>;dq3!Q&f;qo?g!>7r=yS?vGsy4PPT)+a`Vp3T!1vfFa_$U%
zw~CH&udJmmUm_@9$;gjzZ;n~-lV0EltyzuVn2dvd!ywg(Cde?s2pkYvb_*f%GgTn%
zp!(}}&LsZZ*=p%F@T)}mrRD$+%D=8sUKi!1S|y)4u04wuJFMctSpl!+RdU4JydA~Z
z+8f@PYU(%S=j5(-X#XUU(DZ3-Fs4nSMBdBg`GzKVn(@tAwTieb_#)H}k>SiWlhH4Q
z?2$*<57pv5dp<gp<$`t~qk0=T*vcXc23rUx%bJrDKZ}yF2SiRe;^Lrm2X3e?o~s^;
zQ8n_VG9iiZfI1%YJ>)Z-Xj+5s09(p$P6uOF@R3t<57}5Ph&J^))uj)>L*Zp}pndIW
zeMyTny-_pdJn0G%N~6aH<Z;KpapcxSJkJK=L7d4SS(!{dA{V<NX?{;yzF*9hczKoG
zjm-36N0Dz_L?&D(4fW!<eK9x#SmLr#G(V`ZqEj9`x2(I&Ivsyq!Po=a!VupgdXTkW
z*L#S1|DCc$adV2_vEyWqvP}j`QGB?k`E|LV%Xf6-F)+Cpxy7B+PLwM+e7MJF#n?l0
z2X!w6W3j*%aWM2|<#A+HN6ik>kzdG}I&Un*LNU(fX%$!Vbr0G3GoW?`Y^k^>x$!xQ
zh7@r{O{2za#YU!|t0ff;@wX#H-}dJ#hhB+zW=<2RK2+TVU(OP-%SIUpV`Wt#0Q3`n
zqjl%IT2rV`I857Q!1Z~})wUM6SSO39T)(oaB}oHwMFw&P>l%&w>;T%;@*CDaOy8t8
zcx{gc**j-$y+0sQkq{kIyZ+18V!@x2m1wF@vYj9y<V(K8rdL@#-%i#HRUn>F*^UZ9
z(1Pn{=c;P^V4e@-;hr33@2{rD)Mt7(_T7U&42hq~S6Aa|fL_ebY13Q9=DAQU-RUUb
zwB4METrei8@9IT9?!qRiKf5%xO7uLgjhXl{LKkW?YPbH#&bsBi8S&|BG0U|nzRJ7c
zOuf1l<>S?b^7Mc)O#+5YdVb%@iq1}P98v83(T*LleOIH##Kk*{v#bcSWwP#{j0rI5
z#iZg3?gRI6oHQ7F-iRfO>-Klmxl#lQ9PJ;eB)9@E+6IAOG$9btxlLR|G_%I9G*{W|
za>C*-hZ~RFTyD1%bAw-_VY%8cgEKe@rY`z?ojIXvOgD=6r+uHFccCv<f_q5b_mZ7w
zvn#c>oFCchv>dAlu4zz8ihMPQ*$oPKKLvV(5$S5wAVzynHGi-$84@(s{^pXhJ9X7X
zH8Wy!ouR%C36B7k^UmTcBt0>Emk)+ZjgSdAE5BlB$F9)@Av(HHYwBFChaL7NJvEvA
zaR`eJyhLi8s{Ab1)rVXa5_a|`4HkFnGw$n-2f<7xR`QHws6#}>i9&98I5wslG+<E$
zmT`SFv<gOwZ((_8L#A<QoMKe4`~9L1xG|VO-~h0tr<-66ll8ZGbTmLHZn;>s0p^OB
z(g3m{Zh?{0uT#B27&k=Tlu4x+%k>E>acqF-W2uS+_j{o>lc_C|Qpnm>-wn(dTL?ss
zJ1k&3m`K(Y<)1AD+Yk00s9SX~JOMIX8|WNWWJx#7b-8gKF^ZqtdeN0!eO)UKOCm>T
z$0)Q(Ut|OknJzhIe+p(4V9JG&%UWAKQ||cRthqGfT1tHOvOWpD`P`3p?V_o>sg5c5
zUd1h|qsvrnB%4%kA9u&KV%&7ZN~H((LGldS&cQj82jq+&mAh4nTv893%P}r6_|zdh
z2Shabm;sBZtY3*!DVKD}yg0p=U)dZstg;~R1CL-L`_`x>B*o>4O4ievHSl$2<jtvk
zK^lTw7{v|j%Mr9u)waI=4ESmQ&aP^u^1WKko9~3YkMK{Oc7tdM?-xszE5)aPnEb2c
zF}?@ERZQ1G%g&W(lW<YoILPjS2UBt~d<|yyniiBTF0Vx9l%)Jx8+V(TOX7|Ay7!|_
zexa+Y<(!fJ_d0cGlUnA@rvroWTf44sl%}B)o2*8-%q6&YA-T6-U{=SE1%<52uoTN4
z;X%I4pg1mNMQ6r_k|b%{*ots=p=E2Lt1rZ{&RhC!ce`M&=2v^j<PBHm)qG7`$QvdB
zV0+bujs$*))H^d;{q?E4@8OHGhG8!BB2{tLj$39ccSPMVRcKWQNBB7raEO@M2CPcz
zaR}za#pWlipD(rE|50t7U~Wd0CHH-87VcRay<amGk8H~2C7G~B2+ySt`mOuZvLpIB
zU%!<&?gczgj<IjRS2S$aT1f1CSEb|9sP7!3H9Qe3kILsEBr9o%wOK5vSbM8v7PVm+
zD=&z<*=E2pEX2||Y@H{3nI>c-E<M=fgm!AWDrDX${<I)e*jsS(6wtkRglaxHH=%In
z{Cpr_?e_lQ9>~^JVUd6EZ6qZ2r~XaNQ-I#*qcqA12QK|D(uwR>FP#F!hfe|hwh;Q_
znb=o-lr2E^-q`PKr+}fo*st)|mki4<e$$@<)}nQP*2p~8AzE*^$G)UlE<Xh*z582{
z@sWrtpYS)<e>Aa$%-ukgyc&#sd1iV1H$&`SEp?aazHA=Tyx7w&IkMxljS!lt`lkEz
z+E4wbTDlv8r+^H~$%vKe*k#5qUrqseu_wlqlNFdUj6&}M`NZcSmEx76p77o>MG!|3
z#8b+udZ&OA_M?<PP#~NPE!9WkpV@yJfY^UgGz?8rd{69g71r_B-2?qGY4f*5>z`M&
zbtoL10{(7kd?YF)bP9MJdt&srMVh}1CY=HptrJfHAB6v%MLEiD+IRMUD7V%71M-Ao
zi{jb8JkkH+kT~xAAKW^BayE@U_O<RmsD@d`-PDNLdiP7n`UHrP>0<YT{mKUkvA-t9
zops!OJ9{t|TL6FYB^k*(KCincM>+k#3F}hLefn8m@27`}PJr&7IHk_;g#C5NXUENx
zg%^J@AT`ssXI$U&{t0_(j_@x&it5L)iO+d5MxW>Bcx2w5Q2Zms<vMw5Zqjxx=t%Af
zbL?wMofPFP#2-HIhc93IO-G4y--|y1+LeBr9U=Us?k2^&-ixEWTuL}7%-;`Lw*E~Q
n`!Asfo!Y{sG62Ufx?4@BfUj^{h`_f`K0DQ?0N`QT>HPlyy!AL-

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_low_hirescam.JPG b/Images/deepfake-xai/ms_eff_vit_b5_low_hirescam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..568d4b5d902de14322ec8863d6c9c017b82b8eaa
GIT binary patch
literal 14089
zcmeHtXIN9+wr=P}kS<;6C6P{$E}*o~0|^iU3JFL;?*gJy1*CT(AiaaM00C4$=^YYK
znn(u)l%~k1Ufg@%z0Y&*`El;O=ef`Qac0(BYs|UE8e_iiEMsORzu)~{0Wd@K4D|qH
zWMqK*7azdykL2^Z_c1O2fQbnJ2mk<R0hDBF0P+io3IHI3Ji!1cFJ!Wd!#pql&lNxq
z0FV#v@&RZr>KHD{nim<sw|jB=bN9a!0OsfHdqD$L3dP@eWWV66AcgXua#LFB-((79
zc>o|S?O*c0I3-K_L*svQ`mezxFDtL4DhpPXQxcE^t18N>$|?f>qdEWp2Jizo1O9Zt
z?~jz6uKxbssxmU3e$tL;FDGYdl$VFheMfH@IcZrLfcnk*-i|0YXMX`F=X)4W4e^7{
z*Wv;gw1&7f*hJRE`<AmS#vsJk*)qh`3KimpQbCK~)MTVrzpr}V!`s8z-%;Schr6er
z>U|BNzgbtkkpEQ62x+SOqFq#>xAp#^aZ%C``bS%Wf`X)j6r{a;@5#uisHn)u%FD>h
zOI>J4`2~CWJKmS_^b`Jvf!of0C|``XKgP>b;EzE^C$9j14WWPl3|iI2(FN>;MuVlC
z92L+~a&l-#DMz$|oD^C?0eu53r!4Ox=OpwG_h{5Vy?Y1vy8q2L8YScG?(E_0>F;;J
zhMbI$%-<XRU*+?HqJKL559z<ac%h?u%h%b_-}&}M*8J0<QnFwvIVG$A*~!w%a!Tql
z|4vU`<_{VFrlJ3}E&orkrKyfasiOY0-OKmy?k$~t{_lF*9i#bY>G(Uki^AW93$qtv
zj*gB_UFP3<{9h^fTg}CgyLj%O1N7oaf&T~T|7R!vn&i%&7c)EPVhaCW0Neu5P*G7+
zQPNOTQPa}U(9*M9y5N|egNd1eg^PoSn~Q^sllQ8)ATPftKPQ)vw2-KTB<MPbM?gkF
z1}HBM1OfjTA)}$CrKh83zjTQm$j8M8{NGN$y8$e;WJY8h6l6jGauzZQ7P8-O0oN`<
zDb+=Y{u8DDk;up?D5<DvXzA!LT`1t00pw&96y%f?R8*7~@tG{_;y!?qg^E>BPMey|
z!jVSEhh08AwTM<&r?H0vI`u_F!O1s*j{XWK7dMZnn7D)_P!X(jLs>=j)@@xqeFH;?
zr4<ZrZDWf-p`Bgsxni(>{sDnO_k%+sqoQM;#Ky&^rDtSjW#{DLic3n%$}1|Xs+*cy
z@U3m_gpS@fZ~OWO28V{HXNaWPxp(j9Kd!E=Z)|RD@9cg(JUTu({r3Ir#~)l|0E)lE
z`X{n~hl}L`7da&*1trZNTx8@y7ny>El1fmHnpN9^#?gmONIsmFT_?4uv4>7r0s4i*
z$#;tWiiqMz(XW4?{SDcF4p_wh5wd>*`xmZx00RZtMdwkl0B!=_5Cekov?P8G4pVP6
z+<4tysA0?y)a_BG&7R~>x67TCki4o2GbbU@J2Nd+c<SK?l}y(|e**1=jpEOxMBdSR
zn+MdtnqHO=qT`~kRwi%kbAoa05b~-Wkd5-FFDFs%jS!Kbn|)(S-`<QldTsWe6T?1#
z(RCoqPS1XR+sFwbZgEhT)QgTk-zY@;evti?ip<b#8w*L+S$HkeY_{p5lef{%QQ@X;
z(CuyZ(kXXn#c~Z*zyyBxo`ee7{3Q`fQD^0$r@yL6Rz2de#f<ivZ9Vq5y8QW?{M{lm
zpz8CGVQ5n)ouRgUrvLe+e2CtH*knnX8T(KjIwCM`IIV)41|0bDX0um;n(YwsZt7pv
zPt{`1m{wh_L*k`Pj_znPsOgXTz&-%Jju{CAo$pz!%OG50>YhBA-xbZWji_OGTp?Do
zJ8v<Q3mN3+ovX)|S~qof_khGy?@l8f%mTIBQ(i6S?%5m?w_YB9cIF~NJJHl7;xu&{
z)p?D>0~_GR$+!`TeV|fGAlUF@JIza-2OH|s6k6jOfm^`KT4O?^Ov|g4wN1xWI$7KY
z+6FI)KC?MJPwL>7*WH<v>sSXzN7^s883!heNi&}}UghYs0rSvVV24#N33k7T%Ro}Y
zf|ylm(#otu>zenaSzJg?%fUJ#{7q#~-O^1hw*?=C5xiX?f5;rKn_6}Jeha_nBxdT3
z$Z!&YQf?rmIt}+^|C@>rW1p)vvE@or+fhk+$cyC(>|p(SFDdDvSV$V$m#_=sM9fHe
zI33k}z=$yB4&H?LyBQf+EpArWIoNXfTf~DJ`2``8lDYuuc!b0yrmt#7O)SYvTj>W2
zKju{8cOO_LH|i&!ZQo3-^|-R+86aZI=F5=&6O?=jd=db%>6gip{z+TQx_6JsHMyN(
zi;o3vZT?c)&qpZlkgOln5HGQYq@o9`kRi*fM7*jYPAe3V{VE5R`ph*$W&F$^4~&)#
zl1CYu6w7dy334|Ju+&cd5u{fI*AkCUxEK8O_FtC)Nkx#w{W;ZF{H(H8E2JoIpc1}7
z<*)L%w|O6ZW>bqYuc0dk2bviyeD@3b`ksqN!ap*7i|d2MC>NTK`?{u`tBaGT!e#K&
zLMw;ZPUP19SF3NFDq^MPCKQd5na^I{z-VBG@kM32Y`e0x-w#Lp(hWO7aohFKb5M1}
zaQD@yq?T6}^}hjD_{TfYv+BpdUk9z1mFuA_lzzfU#zCgh!Zo3WH?y#h<C&56mfPr@
zv8%Q<xXh)L_D4;LslNfVVwGX9q}Lw?OhZXXeZ7Uj>NG7xU2BQw{k5Yqw(hlvo2uM!
zg;%{l6D9DOA+XhMfAug-jOBN$cctd2>mEl8@o8fB?lSWaG}IJ1Wjt7N3)M+B4%hUp
zxdEMibjwmjS$7DOuHswr><GcE{mFI?w55WwXNryP%I>REN=?ptp@S_*zDGqG5gioS
z8;-|XCoAYdapqetw`EtWkw}gFZcv^}r^QD->2kUTwF8dmRmSVLHg9`7XqE=mNy4Jd
zq8BIf4e-vggjVZ=_;haNT{mI<+|&n6ZpG7C!qqzB{i9pyNudw%Gohf|vBUD=L|(x+
zRhHX&lAyw`Iu$?E1h`cw*LIM<7}={s4I~kBLadwC?0Tc6)oC<P8i*F?k-_Vhr)I7d
z^JUsn&sTo~c;2KGWhMkK6ew8bFTlmveA7+b71*IuR<7c46{z*T{+X<s7QX?XYL|e&
z0q2^ZMnMGL82+z_t3r6Mw2!`8XEwe*wo0wmoKa8jkDy*)w<@@RK#j{-Q|aej9(V3P
z8BZ9%mT6ymaPn6T7mwlHv}H@aVc38<M!`@|HnF5}&XfgFtQ%I(OV;2QIXP!?+G$v7
z^t$y73!pMH4+CY6Nti$U23dP+p6H*^!S9_>cvd{#{ppn4YxTh+X5?lwKHp*XRMGB!
z(#peg2B(HL@!N$i#RWWhB4t+#T!6TjLr?d2T@l2_$)LB=GNp$oVdVp(${q4)(e}3V
z(eTb6u<N=vZn_kdE$^Pr5KJ|(y)ZPm235NF0g=19&~lIa05J>!?j_>+Z2^wZ{sr2*
z-r9rJ3dl<juT?4wRU)yYzKs;do&4~c)i<J#!!sBQJ>9pI>ot$;vy9L5vgS;>Nc)sN
z<nMRQw17+BqRVH5tkd19l!s5jav1_=SJumDLo2sR351+9M3KDN&)xS|Zptjxv|i=z
zc+FdqjC~Yby&=jGI1*@=H$5b%o^DCEnY`avxA5Y!y{*cTqgtpeHwSX3G$}(;tUeq8
z*)GgJB8$L|@<bkoq@T;R`Tqn?Z#PxfbQxENuy*m#H}W9a#*EuWU$uO#X!;uO-RL-<
zKj3s<Jjv(iq^b_C43<FVS8}n2Ri`bmdn(ts%4=fHU{0dEHT+ty8qOhOr`<|S?P_xh
zbp|q?V!7626XYeSDud+;&fcV?YfSWv?I&Uo)5Rr@1S6M2Pr_<j`leQIV|@{w*e+?h
z4JkChB|Y$%DoD(4$OeD?$duhgSlJ<W&mt8rUAxtsBK6jpj+=uDbESh@KWlk=Ue2_S
ztA=aQ!nBVaaKN(Y-ezExf>zO7d6WuM<k$avbSYS_F`4pfq_w(*r+Rz&#I3e#cYa@d
zf4^U^3IR)$P35)hRM1>4pM8DlS&uTePLZz~&tRGKcwb(cnV%ud6Nai!Ksu#6O%l>{
zwfooNAvWXhIT|N7o9=0M27K(H%6>+X6~OI%5#Jf<C-zk1_n|ON)Jc6Lgjh_pyh~-T
zwOC$WpsYAz>v=cMqBH4PVsPqPtEc|e56n>c1Yb4jw;fLr+OCXSi`#%*w>%==wIbZZ
zgSiywH*`+Yhmy*OqSVF@Ui)9}tzy^$-*@N3h)TK%SF4Y!*X_J(;d8<tPq5WL_d3-}
zXc&4UgWV#zWS)kHBa}LP^?DZ4?bD;M`dN^T_9GCF<$dgdRqlQXsP$I6m|T42{I&OD
za>h9UpX}fvWu#h7V`=isx!irX_m2$FxlbGSO(c%Q>)%dC8|;V|UCG%iU!)VF;O7q0
zwI>CTRdzRC7Xs6Wi>rX|c-T)%>5iB2q?B223VdTPqxB2b!|SER8|CKoPHAnpJ-Yom
zhW6ovHwXF8**!$ZY)bRMW_NB*;I%%zXCJqrvU5O@biT+*I0tsBO7H01z`Uf)0TYPv
zGrIge^B14VS^r<05y%|6#AIzE{r2p<GK&5&Fua?;=1Q2~&?`B8E`NBx_ST2^cxac0
zKH66g4AXbYxXLW#EN@I7Ie007^c%oz%k+7B7(nTfrqKbfZPPT#edfl@sZ4>}@H&xv
z@y_eaB1AyZ^7QDr;;BANjg;nAbS1z1k*mY?2T^>D{MBJ@dQVL6N9iVg@#`p^lGdE4
z8pwb5C?oubwMC^(d70M+VPRaGKM{tQuCTrHQ86#i#hV4=yQE^n+(;o-)xr=JA*&k@
zWNuhAsyQLwY|s_s;C5MevF{mSsJ?ypN8Qt?y!;>BwWU^U1O0A#Ha$DGbGArQxuqWR
zQ#*^LYtRHLgq}~n0I0N4-SJnpTl^GXv*GrH?yW1G_ck3K9*l@zk7)+m%d06D<KE}4
zK9nEGBZb|tNbBa*y<Fe!tg-!?xc=ICk<#bIZHI_r*&3cUvqcpSFqQL%4W}}$(z5#L
z-gt?^OPZ1c@78_;>Lbs1HX7}IiRCskAU#x2C%*w=4WMUCH1u-D{ZBWMOWpM3`*&&l
zmg5U9VZ^QFq#xiHA7-ov*h|??+qwp(bf*n*=a2KM?7&O$_=07tbNPDayfuzw#}{UN
zQ~F915{_nE-6!=ChnuvuB6$V$Ye^?ZCp)Vpwwi9`z4GOSYf#OTmV5jRWA42Mi$#3v
zJT4`5OAeL#8SVVka#s5;BJQ`KJF~3F(Z;s~6!C|3)9Q;r_cQCRv1BG!!%*gSf|H)v
zuo|h(T0f?5YIS_>rvEX=<WAO(GdA|KY&@BG!tA|R&yTrl@nVaEhVK3xTfJqrPKa|c
ztFrMKGrM2Zx3Qeje6ZM0cW$-Jl((Q%7{$F8J$}H$-dbqi|Jwb_NpfY*3FqtcIiME{
z-Qee{iq^gM5&}&8Y>RI*iV@s0@Kz;!vc4<3bA#Mb70F7wa1C1#bX__{s{JmD?7~Yw
zT&*Nx6Pe^@U4d$A?27-vB6O`%L)RwFDn`kv8<>!023a8uZX5jI@d?zhdEQSDjZP@|
z5h(wo*-n1%w9Y3;aDI<3_@mX@NmUD%W7heD@f$CI6|Y}9LnE{YmFu&w1!cN=?$kTj
zY==;jZc19T@n5Vc{+%~l<dSxkx$-hl-L<7e>1jEh@sC>l3g&DDX5gWiH;|`6`TIuY
zi|GyrMHTE<X07GML{|+snkZ4~@$VxY@L@$cMH;a?-{-_6dppXgZpP+15>L{DuJ3ZX
z9Dd!<pzHk!Ot*?D3Bi}gzSq^YWqu_Gn|<%;SrV5-(E!yoK0(~tnXSwPmFPn!_gKd*
zc!I^Q<rtVzw+HjOcc0I}sy^M|%mR68X(#6;uhyGB<eI&2V-<y{+#iQ`)%wf6I+qQr
z<RgV?ddU-I%CMNPh#_y?&(UmNlze^;F~eKxAAkfa3(MuQt{_nlzOIMmg>_fniU;@*
zp%z+$UB<Kp48H;XzX6bvQf<B;c^KfngQ;dj>tE2w&pr(lnl1|8c?PDKU)3;EzOWpa
zCP)xZa{^?9^L2aElihGXvSOPVPva}szUu>K_6naGG(2iaF$x&?8VXOCOF@urb<~I4
zd+8U_cWgbQcXtV|ufkQ2_02m@n<3;0s#fXz;NHJw(5FWx621f5_d>J8;`Ag`B|WJO
zuZLTM3-(1qwl(YozigCFAU-SFL;B}QP^-hm%Xhq%|EkGoHO9ME_7cbGA){4L7)LIl
zW_Kv?3a`_H#*wSJr?q{RS!IZvJ{|SSd1iC??IA>#pPfGDx@E8Zhy8|MJY(8$W~V1r
z6~MG+aYVLa*6GaRy`i}0{N6x}`@VqrjLvrHnh3N5gu8(s+|gCEYS0#a8!o8xIh3Bp
zzm3<lCJ13d%6o_&RN6|r>;qGP?9>z{WZ-qCtnW_O?>ousU*+WHZJ(H~Y5?}xFl+bF
z;31eqh}3&I3F>$)Y>a9K)FtVqN+G3KiA#l&KK?e5Rd2Iq5PMG@X3NW6EtnAFD%j5C
zcP%FpfA4<$Zm<1oM=h?fRZiEMxj<{?g*fN9Os*AHQ1)vlu9<jK<eZ8q*D!MDS%}~g
z*_dlu5YOnON*lOZkFI|3GfK2_SOJ1jgXKQ$h&Q&-O1#5tN2fTqlJ>;A&v!OR=ALh}
zeuQ=7KB#`ST(TVUHIMJ^wf^>(q|FuZ8<8ISajgAsfX>q1=e8NI?8lnBNf9;N3_h*A
zWHC${n9j%R$F(C1iw_#}=8OzolVbZs(VCmv90FD>*KpNOq~F(^PVR4(l7u0>oSBak
zFR^U{|C&8qXg5qpp_IXhcP*C=6(moUt2PN~t6DR1E);<05aaeEzhGJ*TAlRZ2wK!r
z1(xc^4xd)17GQ_l*lkKf7}G?QEcVQ++5I^Zt3Sp0U0UeQwe0a5ct|R#FugY|9M~-y
zLcgvs7Th9H0ktgH$;*IxF1?o9y(9}j7d(cy=6n0Q6?^V2T<^Ln4-!-1k>r^Cn6ZBC
zmZeFM)XbCi6aoc>Mop^)6#2r%e@I2mAB?y6zH@86H`@H?>XHdMs2mO5EN5RcRc}Oa
zvZqeYe}6vmxb(}$<8IJ$qo=3C{>ElHS<W#BwXPTz1!4BIKs4&n%V-}%sY|8eC#TnT
z_wy?+Nou=hW)}%O_>gbTh;xwKzr9*<6^<mnfS!g;y=4wb9r{56=k-V5_Gt}mB6n4;
zd}nD?L@$>R4DSot#y(S;qXDcO3k%iXyk+G)j-95ds;c?qJy8}b#2rk{^(e4iLpNmh
zbBRzRo~qxlstlq?Ds~(hPP}=U=@?!Y$dO$ndoD+gXAIOw72J}u)SId+EsU0DC0|Fc
z<*gl8Zr3rO>tj~q1K7#xHl%)D?3z-TLe}c*N?TJ&F@bs3c0t03?HNPvH&>s`;5O6D
zw=K@Z2`|KDHm#n8-nkVLr+T>Yt{?YF!t$ld&^7J}aAR0S#v{3hdoGI}YaW3&{BFlO
zLTwn}d!+(;mqD82V9!#tKi4NaIcee5GWN(awdK@kC?N~EmjG4L&Glpz{6?rAu(~u|
zXhi3-F=q5LLk)+a8e&5Lj;!y_390jpsY0W{4kn7p9Utn{*hT|R=f5-Oq&ZbCS*!B|
zEfi~TS<?!%CnoW&rh4u>J~VR5mROzEAQNYhzl>S-@NMOJmgiMB=ouAJR$_z`GM>(*
zt&T9ood4XVzaE?soOvhcC2ynF70QelNaYdgB@A~gRN!hLjXC>%U=pUIJTOWcTKWVn
zNy!IBEF&F|r5<8^s1dmp*kLXt&{|Cr+-|K-WL)KH96)Spj!z@yYdcA%6qiqmsMFdR
z#-8_Azpaj|$>wz!@n%%q9TD7$55agphj{d-g=?C&*aV5eS*2IpuGx4*ezL9;P7|`|
zfbskI(%@-xc@W}%g+zV^!JGUQemy-V$CK5qj$=Hf*fU)4^8C6t6wQT=U<0QN?I*wR
z1GQ1$fR2$%_8cXTa!cJA0uO$I(yh*FMtZ~=K5hEfs5ZN~{ff>?A8iwS33^80brA@e
zGpNcL2eIyxOEb-OFycfSj8jZJB+fuey{aTWX$gV2CmeKutC{x~jP-6qnM3!$P%DYI
zvI0qOp^Bz!44uHPxm8WWa{Hl;ez)At&Y;wcdmOID@1NVAy|RQ|vq~N6sf~5Y!aPJ0
ze_7TR*+p8#5<CuEb|q1-rF)R{v042~75q0)@mEMcq6SZjyK`T734M(xwq+z`Mq?5l
z-v>8)mKWiKp`6!s?Kq%|j?ozXt8QZ0*zWS-z`8qaO<60c49=$sd4bhIt%v=kU7jVx
zD)z`+hkoa={=V9$HF;G%HQA3dz)^>aY3zs=G2QFWI*3B30xGsXAy!%JB?)O6FJl&I
zS$W$Ic($Q^9!aFQeuTy%jV#<!`^8`m({ypR@iYHa`Cnf_K!dXATrp$_PaSDW@<h$f
z#Mrd3#-#=7ySUlTC>CV&n4mls7}$xE#VFWh_b?9f%{nw***>ycGa<x}NsA;zc}+K+
zD30EykD7V+V5pATHSs(GUkd^mASq7j9>Awm%G&28F;wp8KU+{;Da5H!O}~!m;5gvh
zNp+?2X2;V73HLp8XN-va3&Ge4@Wh0Aw!Bw1_kflxm7}S573If<hOx2{H=78<zhZ~L
za9+%beE+=q-mV#WMf=k+B+ty$eqnR-p^N^RonT5p{exS10!!e)pbxi{c!NEjxepF5
zM+MXveD@An+2RPbBsq1oGWRB^47qUixd>aBLVZjjqZXZrlJ*&}QRq7h*(zyS%fBRX
zO9_l9g`8|}WXym{J)9093B7$jtj$poRy*>D{Owc}V0~o^JaY07QSBWPAm?vCoiv~p
zfZD7gZxEldbrwlt42;iJpK-z4a&pzLkH;iA_4i1vGrR$dm6X#h8qp3+<IxSE=^V#Q
zPKs2k90H`_XTE<rcJkNvuXzvgSdfil%Nu|xkd0zX4XYcW$E8wQ8ygM2v{02ZOEdA|
zn(^GNlFt!VS*}sxAyt)vt?q}wzkEJBWU?x4nAkoutVK|-_B;}Vm7Fa$a8ahwltC@V
z99low>H*a*v#&8)knTTRo#*AW<%SVAng=+pyr{Kt_ZKiq>PGVn?Dbg^lELF&CId7|
zUJm(tp18nk_=2G7DY_ffcRCM{C$Dm*7PEc&-D(fx);K)=lDYq-MnKC3U8Mo4!O0<#
zOcke1@-p}Hq>FzJo>eKv&X}h55pR1Aj8eNfof~z|*aTP0F3qLF1o54(^K=<(HHNk&
zN771pLKP$>hHr=z&`P7(?7l|#h@b-Qe!x6(O^LvF%-(>fbCpADnS^Sk8XN~VtFI(L
zlinn7w5W8NF_og3RqB(bC$-{jZ;=tNp7==VRIV1Fx{I0}EK!yuk2{8kMQS+Rewr^o
z-IBX>Mh;154=rGWi&%;2M6he_;bF0wk)MZx-d}5cq=PcqY+r_`MhXBfAW)ZY{O4`O
z`2emuhNz2oU|KvuCa<-gYb2l4V-@8MDkHBC077T`Uq48hFKfA7f0Rb!V?*xq?g(ik
z0a1jF@$0G#`@ByZj^-+f;NoM(I0+Y(doBB#npKr&YQS4@s*;U|r~o;$F^l|jLQNZw
z`)Nv?vxHUNgSvE^W?P74>ATc6a2<7W=yf-tzQcyW5i*;p0=Ijnq<X1!Ewl{l*M=Wf
z!#`=Cx3sl0VH6qTP|iL51}ne7B0ioqKl(<k*>u+p_yB&9u%gnkBLBP!DoGH}ymicK
zXWI=<TlxWP((^<YY=|)Co1S<NO{QjXtKv_r9oK`x>YX9L33au@veBa(xfgE_0c72^
zf_s_p^`Fa+OQgL%eF47z?zDO;Comry3=N&l{bED^fROa7=W(BPD^c1*Rxl-+M$PON
z!sk8?mdx7A{nBc+xcTkT!<`zI5ygc^(XXMUnwQf@$!)E@!b)%zb*``9PfI3?DV=V+
zRm=qB9H1peHeTN~^@DX;b?>aD=C37>6G{*r^+g(S3Hc`<8vCMqYT#?Q+J`4L;e;1D
zTaImxWiR)-rYTwk&hGEJsmzGPaB1|^H@_+67g<m4a(ng~W^@*z?9-JO-=A*{OiM=5
znh;tBe*>CJ?JRsPS+)S@q;n64RZ2Pm*2?aoT`G5bZPxGpg=!IE*5WVy1YD}6!Ima0
zbigWt@ZC2|v8emx>Y?6#edca-4^+gK8i$>?UUU5hoRISA9|=RvC-BI-TZBD(>au%R
z%c#dzWK|gFIMTD5aTQ*^{qh?XVr8vQ*c3T=`VK#GK9sFi?$enM)mm_d)~d0@PiH#M
z9B^BF*_p~Jd*X2%Bo!bbUrnAcz^`DB5G^0jzX9y{Zu~NbgRLLfQ)k%Xny!0pm=dS6
zufA?Ij8%GVX=|wOm)^luS-ZEqzv1}wK<)r)-|9M3u9TfRdNb;1!1IK}*dkjO!09n&
z^nz*8?)J1W9fzK`>a0C>S9LMlt|W3jHsKzj7B`>vT>Ld*t8ZfOa*|R;N~8$+z_w&d
z;aUZIX5pVz{2mzmwoQRs%xM%WYw8yI7H-~Kr3m4Oo7dWQqX3&{V09s4ie&*s>YhQc
zZJT?zF?b&$2@+S*l=n0d-$5Up&Hhw&9H__5%`rYj%<nJisgVQ2cFgpJnp|S16VGZI
zHC*236nvz5lacT|_X(CD{5G4Rb>BhOyEb&z%THG{l=B-j(rOlxD|%82kB&!Lxrt$n
zRwMLd<Wh#cyVqIf3p?6_pZV91zw5Wze5Tk=U8UHWXz;m>ZvfXw!9YS1Q?|T|$m;ls
z4K>&0o2buwOch`w%or3m-}GRiq*{Z%c2Y|p`bphPY**9X#+I62KlZjaNjoR}w&k89
zF4eBF%*=2jmW4ebydzq9%-Z%AL!`kSjv(=@A9CwN_v8MwqF{r<PBv~prKgOj>w|Pq
zFsk@o{Ns^45+-Lgicv_#&d+kq0qAj+=&<XW=3BCPH*^kGZHuVS79L@aEjpvWY*u!&
zRxjURCbT)J7WcB_VTAxOW6cD-Q<r4j4xNThYFX&AlVz@6dbjVH>g2I(BN)4b8ydSo
zoxn4ts{@o6&I?94uFSl-YjyYP2{4HyZ++uwBV}8oPPf8iQVmLVf>uM<50{lNV=St7
z6Y_2<Jc%)pXsurC->+{{x%}RwWPr~oP7gro+g_^`T*YnZE-@tjSkf7fV=Q-le<L9w
ze#63iau#O}Z*RomxvO1OYJhUk@hGJJNtO0$bs3~iYeeFE;i&FyUiK(rV`3K6YO|)c
zT1BipV9BZ+pW56O59hwSs}vj9^7vXvBja1+&eZU$q(S@d*!|#<+60xIj{1dLK8`e5
zpUnv6Df~8F$h$n~?5{7D)!~#-$h&?&v(QU%=P8o!82uix=Sh}|tH;-2iig-8+(4P*
zH&6+r4<PJn3hOOF^0FfZ7`3jlh7oyBNZfqR9mReuGa1;y5|@IDJF62WT;D#fXqh=~
z-{(D09zms<m1);t2h4%7v~j`QyC$f&h5%M?Rr<^%5Z1rAZ1kIY8)Vwoa}AM;^HCiB
zr01@nhmk0!s4P#{1^Aw6#w0w4Om;Md8bz|VFKVa}HErs?$t9#)>yF*fbE}N{#<c#X
zfbLpI3tR8F9;y?5b?7%>t_I;emf4@rU-pI)eXXw#sru@<Gw*dqLUw}0CkGQq&TB03
z%JV17(%1*jUU|;S)px1+v8$!Td8=^s)OK3Uzdf#x*m~R79>J86WK`hBD!tkEq}A<P
z)t$llgfe0Q4a<_=-V~DUY|grR;Wxk)vvMnOQU3hgS^!k|Wn)G?(L;EhEQ6-4&U@7<
zv-&qca>fEO)ZOPQ5@C`qsCNt&I_BtH1t-=CX40sA&GVa_rJpW=!6R_H0Rzu4QigM7
z>O8SL-ry=Up^Sq|aRtL}K#kFvp|FZ4va8TpGXO+Q^>l69CM!rrL#&d3L&NSdI3oh*
ziC4<!6~Wfl`bG9LN=o{kKTB7e|Eln)d6te|buU0!;It}}oxv%~E~emKnuG1Q3H@*=
zX+bA*uWBJIM$$jrlNp$~hxAyA?yJ)DH!e}0K-tON&s=S`lG{)5DR>+|L70;fTfQU#
zAM9U_Xm_=&qmd4)_mddENq<5}d&5BoRv-m}r^myCERCRSW=YGoq<N9+%+8JT;Yw+p
z!b>Jt1qAE%99?->?OHkHUhxmok)Begd>%1K0_0Rc7f9cg=1o%vlNnQ0N$>?#EJJC3
z%ppB}iSO`=@^-I=Qd1lAcw8`^GA=K33{|Ke4^Xf#&Q_liH)iMespE~Oa7FR1e+G>H
zs*-L;+`zK4iY7jBi#~pT*Jvail)%<a(NOsi``)$I*CH|+e8ui4T5W>EEY(DW#FL${
z#{Rv6AEeA?1{Z9v-Aokt>;}45X4}${HvZ8iD)G(SON;Pt5>?Gem{7r%djMR~E5Pss
zP$L0@snC}skOC^Lh~YK)dUEji$(?vZ)_89)vnSSg#oT<aJ76_gg~sQMWNTB9`OJr}
zbi<CF-GjGs;49O3v4=KiQ7w}+@{sryBljsKY+H(#72BiFN4$5Zo2xbupLaDTx;484
z*P>wj=^Xd`O1;f<TvJhNP&&htPw;FbUsHLqa_T1R_fLZriB3w&V#TV)N<WZ(VTI}m
z-lzdc<!s@hawS?)Qi{<cm@hKY0Lj+p(&uTu1@=_UdfcE=QPXq}CBIdF87PZ+Uv*`T
zHz`x3BXg+v%9!9`&rc1YxMn7F)tG)(&diTKYZtuHCuEck<3ZT0Z*~+XsTj=YjTzx8
z7&6gHZnwk+HnSDy5l-zEM!FO#-ozDbnCUyj9aS0nV8G=9k|p@nCUaVbRjM?mox6Mr
z$2@Ra9?vK+jxngjT^R_6Gs5sHt97#VD)}t~Aj>|jx=3hMJv>n8F6RB%ZAKLn!KVpC
zsGh-UZN20>>hP4ED{@Gz0Fq&k8zvm1c~VjnoAp*>y4D)7D*O5^FxNP2nW?gxv(XTg
z<!f_}+sfnwR#zxJEe!gVD`L&Ech-nYF=IKdwT+re&VYGouK05hu}KMIT<3*WE}vhJ
zH$-Jiy(4}pdsCoYyryXK=0?ckq+-#-exrmlqi+5ia0aF6JWmeNn`+V4Y(Ayk+eKO;
zNi;9-so~Ln)_xl;H_*8K=gw|CTn}x{zC;q?EXb?Cx6@tq4shv#DzN=C4rWui<{(<#
zKHDUk+3Yag4jj71f?$#_mi#kM@o-LxHX}WM{Sl(~tH`ODP+h^&uJ-~B*Dn~v(3J7s
z=y}d8u@YfweJ>xX`5OQp^cypH!od%i3rMh&))iy4PwWRt+C;)v()|GPCRaoz5n;OA
zRobO;uEp`x&fg<eRrX7420fk*(FkldRnL~7-8~b6MVWvGpJ1T-rCCp;C^Nw$4ZD(r
zLYzxURd@Zg(ow->tol^CrBG@J@nf{TFgV{O=q1LQZ+sUma(D6?hp?2^88yMQ#Z;t<
zWjE!4)l&)f7zgSCMwnrE4RZ_*mLx~WSs~MEC8h>fd}g#hP{h=Fz=HxD`&VcnH6Fem
z<*X8L;bw+NOYVN`*B~lTO|eX4|JEjXL5Bxnmu3Sx;_H%G`hYxCqMQAL8vpK7lLZUC
ze86^bo|Ld@yks$|pyfa*e38L1gg!BeN-}-9D9tQ{&fTpQOO5qO2lMi<c8WIM2qAh1
zu+O>E*H8P|_0y<tIXb6<VM$m$5uE4zZi2fn#9E9eV8JeK_KI9;ivD*ihT23tvFccu
zV@f<$AbuEp_|1TtzOtL+>7$UjnX`oW_LaMV--@L@{~BDYP)%|;oUFr5PQCcp&?!Gb
zm$J4YjK4Q|nNda8(Ds>f6-dF;E}B3HhLKaN&`a*RDJ`2>TBo*`xiNx3LRd*cad>7I
z<mMxq^NJsEp%S%d?X&?v$m~JfW44}OJHIlYq;WBO>kGtpn0|pPZ`(t44B2!A4LRLp
zC@mS?Tu&`_*&v=qa^|W9glbOjFx;*|_#*oXvmZhd^lqAMGh5usx=&ShLRObsJRI@k
z;o~`<;PhrarX_%%y~~qgOat<Y(q5Nzkdnbqw}@A&i7LKrRWr2XhQsuyDn}m|>NMZl
zn>EOA0}d6;>pa{#-Yj*__ZK<A2M9g2=QUz2XolT|Iq#kwi!o4W|1n=J;TyWcC0ndS
zjy|yeb?{4|)Ey%a&9H<Z5JQa<zDv$qxFhX~01Y?yewgGh$rQyY-K5CCN06f>$aCBE
zb0t;cbMWgLo3(rc#rdU%>2&i{QtA7ai+_nt5EZ{Q=JkKB^~tNMPjU|B*b^QbB33pz
z)h;Znvt3VZowZECrjrBZ*LjMXt6qL(Z##WTAyl~1NOJ(=z%^2SW<sRc<BbJ*uv2gy
zNdG{-;nw-AIRUktKB0+sUYz9w1eKi|@P869FblcwqOsk;a|c6>Lw~TEWV1d#zZDe3
zC$P61GW`<o<T?ELE;~z~7xwDabI~ds_oY4VyYF^v%(^QerC4d#&Js=ZY_u(+s`0FP
zbO2ZD)<navXGGQ2b{)<x-Hi^n)U7~{cstq*`Xs*JJAkbPC2Zbs)~xoi><ji`c;y6H
zXiP}QZ8s}d?148LQI=Ely-UulUX&~ADC(x(=5`qOb#E$U$16e?jFa`srAZVAY_k7N
z!F_?Mo!ed>x|aiZO?KH9Mmz)zCrePuian+{j9B`ETHBW<O1Y(89zi0BU)=OKsoa1S
zF-{e@EZa&Zb}LC~czjyq+oE4nmp6K&l{N9!dNb+fgn-Q0+sM2CaO0{WN=ZV*@%2M3
zM9s$?&d=GJ2GpLB{66vPKi=4wxn|}kL9FVqL7&_*PeSSZtGYgEjIc?|GeO2?-0Q)h
z??tzrxCyAmp!#n{(+2^xTAk6x_TPTM7sT?57gD(r>6tp(f5C?IWdk^c5#P4`NhLr|
z$k>lvc+2#|ekX=pEfvko<m&Gc`e_DI<fV5r7gcdFN644$46l5w#1p%C2E0OslUNlu
zw=`@Q*f%(?)TYdC6&!DfCX3%hB;Z1<QJLdWsh+bgc!FUHx@bbN%@R*~GbRq#pur~U
zD$`HYjcizN?%wuQfoh*tis~$;ryszFiJmR}qdP*?FRp<rr#1`IfOHD<%oyOBu6c`I
zev}?h2DJA2XAbt}tB)CouPg=MNKyt+S^^SyCTD8LHrBn58G^UkLaLu4aQ3|jd%9AF
z1${CM<a9E3N%wHyzY(J(^r-n~a^Y`)t(CNg<!z!9PWWYTFWUrUDeYjWSs_ydxRrz-
z`LGN=naYTu(dFI-jraC8dChm0a6rNH>u~toXk+E>;fF%4S4dR#mP&1IJK`B$Tjk7c
zGQwhtN20PuSkv;zN~|pV!0P@<{q8+col!Vc<LT067Fxb<m2Rc2cKTn=OFa%fMm&`T
zegl4dddATr8aSbvuYUt{>==*X5EbMoB2dsnQlk~P)uCi9Ei91{vaVI!Kw~Ql3yOG2
zo6ZFL8y4-)+-rn>177FN*?;UbP;^U`np~`vT}01WtXg_7*821osys-z{H`9MFE!gP
z46`l^2Hhzt6`bHzaQY3fYk!&1Q(hw?;BPU+5h+r;p6xV1#7#{HC?D^L9GN$@^Q$Jn
zE`5q3bfwv(J2kGreOkia)d(ZJH!|yvK#gNMi^N3ZYSmP3_VtxX47{{yqlI;`-GI3>
z<TCNBuB<g3y*;?pq)Ph0wDiVHq)Lt@8+ko{0FRiH@u+Bvmnc0d)$t{b?wT$5`tB^M
z)ct%YCMp}B!Y~Q#ptc4TWUBMP#@>hth9x{aRqnH!@l5Mq=}L(sThkzh!ifa}EzF6j
zuU473_~0;c<(3lOyc1CiHC>K+|2@^^(crJrg9DbtIq?8?O~0_58`%p%2x3Q;hv|`^
z>E=gszSTlPLpy%qG&bIIr{H*n5QLs8|M=C&db0E5NT>yG+cnqU0JnzI(yx?(9xeUW
zdw`FvzX6C*+E5<^k7uIPuPqgWfv>2l+I&%-SMP^@MGyI&Plr4d(@N$5N|}f-vM2H;
z$mg-ZKYNzoX~LQy?+0BpKov!$&mdokJ2H=XF8w$Q7z&S{>LNM6j&nThmG*n=hjSj!
z^~10l%0H;vS$H?6p?nCs(J5QPNjo2+WR=7%mX0>Ce+j6R5lXuv31s)u0ileDB@He|
zf{R>-YcHm&-P#)OEYieiJZ4}G!7WQ8XImqEpKH_=O&ETw$^^X_GjF5ryz8u}u0SOq
z2RPyWgkUqd`y(XQQPQX1>}H}f^=2t`+v3MO=^6e!sD&<qD+<09Q};1t^iZgo93bLf
z3(1<_wKIG$$m{0+V>aOKu6CaqGOm0!X20ECFxfqoI_3CM&()KWt6f${!$-r@kcHf2
z5uymR%<O%i-Zy%m4MbvbHak3~qVRj1jYCP+aMKqJIhzDGzqhYHUChHjzi{_p4+Rz?
z)MV$UZ}H@sD$ZEtk{ru%x;OEPlX6k!k|vjC!_%Y;B%e;_Nl2vd+^>u_kYN5oC~;Q7
z5Oe4}K=a7!Q(+q?{Mi1wM?XtM(E5$$5AjrT(Pr1}Hg}C8d^dNsoAAd!z2Do#I!Fzd
z(CZ^c6eBX$xN8jQ&7q67>o+)=Kn$GjDL{=(3HB9WJY;y#T_s*Opx<gM75DD-9OK55
z)b%}7;=s+}rfA0LvR-+tiMyokjX)hNouyFi4Mbv%O{AnEk`(NCTt^NR=jd`SN+@>D
z5I8b|qbW3ioD>|@sUymQg)?6*ULbP(>xxG`9CH+%3NTENM!6_UdSXSH>@@S}#qJ&O
I==b~o0YsylPyhe`

literal 0
HcmV?d00001

diff --git a/Images/deepfake-xai/ms_eff_vit_b5_low_layercam.JPG b/Images/deepfake-xai/ms_eff_vit_b5_low_layercam.JPG
new file mode 100644
index 0000000000000000000000000000000000000000..2225fd2f901409138413da8fcdb2565df93974fd
GIT binary patch
literal 13648
zcmeIYWmua{8zvmQP^7e^cxiD$aRLOlAi)y|5Tv*S*8-)bSb-F2&=RCrfDqgYrH;Eh
z^l5RYI8~ND&%5ut$A0@`zwg^)kNvTe+{w(`bI&#BIoHf}&oy%~bMYB)4W^~51t1|I
z0fbzB02g0K=b$0(E&zam0RRL50ImSYNfZF2moW+efCLuj4j{jblU#m{bMyZE184vM
z(w;320M%t4?PXf|@&a&dUH<&J`+oz#EWp|Sk_L)rWPkI>bjer#GxC4O>r=6R$H}B5
z0f5xh|HS{pDM{)d8vie+|4}AMaY-3@aj?9E%xwv<yp)W*6a?@;$^!t*0097Jz@HAd
z_(IO?78vL&FDB+4AZqXA<KQgn=;I|8V(%*^Au28gP*e%=wRiMz4!rH)?CS2VB(UH1
zPT;n?lahcr*g)LCSHs!OT_?=n84+e^>KNwX2yqfnQKq9&43Q7<^7V2Kw7(tV<>?(D
zAELzlckS|*@juyO+{%jnPA>9rO|5^GxJ)T=|En$c@81`_FD2^Z?<yt%fk4E>CB-Bq
zMJ`K-1cZ7A+J}gE2k`u>0!`-tM}K$UKzASS+kYywckl@cRN@W_a(9w<v3CJGI5~ku
z9PFi>L?k4f>_zOIq$EU~q@<i=!4h(kE)ovh{}S)y_;2aHLH?e9%XV@UbM|!ha`p}k
zxMV{@j9cuVM*pvpc}dZ~h5irezr=W1Mqb0;**?%&^YT*u)1e~bU=ax!)BoDZqEeEQ
ziemp=IYqHQWc)V`{a@Si|0%YV6`dUA9sji5$NwMk2xq_lf4%MLuKZ`}_y^r(;-AQ6
zwU;_aO--#R_TNhUe^T&w&ZWv-KKE}0y?pZa|Az2?vXK7_a%b<$k$wMi2w%(tGyqf-
z6qFR?RFo8ySE#72&@f)T<d_D?aE+Fc1<1zA0%T!kza?;oos*A~nT1=Fn@><k_^vS9
zZ80e^kfZ=e81$zS5~?d#XsBtJu3lvVaj<ZJ{$GcS_W;H#B!(o+WF*`GQbrOoMv{vT
z0N14{rMNWFf2{Ps0tqP@IRz!v6>6HRmj#-x0Z2*6$Vkb_C@9D;?K8>4%liOwMvCiq
zB-AKxnAlTs`!PvIVhgYEsMmb}!biXIN;&vPQPbRHW?^OH;};MV0!f2qWaS|88k$fo
zZ5>@0!qg0DZefXXbaHlab#q4t1O^4)4+#y6j){E|7oUJhO-s+n%zB=UD=IF*mzI?i
zD(V{=o0?y?w7&Y--qG3B-P7AQHa;;qH9a#s_ho5$Wp!<RV{_}^@aXvD^z8iG_dmEu
z0A&Bb`Zu!w4j1DkE>d!GGIFXvxJXFvUtVO4<P>)#D6gxTP}%$4;FgTM!laHZtouOC
zBL)8obnqXgxydX2h40`Gw7-%4&jE}2|Ag${!2W}44nRvra@l!gi~tot+LF)DbT6ad
zvk@EV71)p`a+S*zS7DS_IA~fAWPig%%Ps((?~dW-LYU)Xpdjyn01SM6Z6nb59Ip|{
z%corqsrX*{+Q-q&-ljaPZcmt*X_aH5f+@XqY=TGq3h6p#v@z2uzJnvar)$B?RNc>f
zNng~!G%q>X^o%$DnyRfZp<xQP88xNJdbLsJ!9(G!qYQbojHs2_r>`IF%WaRf;$2$|
zuLZBm^r>;!w~#GNh?I9TEPSZUO!bKWNQOzhDk-(bG12Ue^=c9-QMnCxFm;k&M=^;(
zZSh6kTdEF+CJX(9?psXaX+~F{pBTVI2ie&sJIsCCeO&JELRi)x3$w8&RN}>ML0E22
zVDIUY?ZxWI-9DPC;2h}a@i(fH4_r0`G0w7<zYG%b_JA(wn)&gnrBvi1h7*QVCLT4#
ztD=&U!%TXaM~1f=+t@Wa>i4}yh3<U;9D+iA>iXQ(&oYrD9Iu0ia5+EtzZ(a4F$FJ#
zO$7;#5!emf;ISaFY-OakcBQ*<Zc*=fmBY3?Zj1R&hHt`Iz{HeovCjiLZCp0ur`}k@
z;L77;Fq7{@`CB)h!of+jx4CRp<2*5A0lRw-xs2JU^0ptpI|aGtX1i;Rufcd1#uVRv
zBu>3+ChAl6?%`pp(!mTeHBM<+B-tHNg|t1Y*LL(<3ene8uiV{}un<R|-JGuRYnS@9
z^BpPI68QNLbZo@2xo?=k7P#C`;aE1)HMixyb?bKOZw4Wr!|C(-0NjwJDgrx5P0zdh
zMr7+QsFI%k%g=2$J_@(gLKYFahwN5S>{?O#ykuRYbCvhbpFX-drc4!l0iapqu6yO~
z6oPqms5;DFf_d0ZE3Wo|vO?PZhmpINp@Mov{e7#Mi7d-kILi1P<96~$Q=zc2r8Y*B
zG8SQ|MO~RKJA0w1Bm>o^jrQqkLEL@Z8>z%2Yxgxw!dR?P0*h%;>HY)GT=`$c)!`%g
zgPo9*Jh{WMV#j<sKhvf`Mao;=yYx)!3{O1Y5&z=*wV%UsM}3!dA8K6G<_KgeO5y<m
z9i}^mzX`s2@vHs<Fsf^f4|;J!Wu*umaoT?Yfb#igh@W3ut|ftYWlO&Yg**|;OjxpM
zsw#7jGfr5{Evg~o+!(36^9DN?aM+!aHI-dq!>;g-yGtwSnm==?M`}QuU-DW&E?&%)
zu1^-ZHnmykLTlavesh|5?rut?Cp|viw;pvE(7Ud?Z)4DydF1B7u2Xy#0@;DO%6fIi
zpE**VG)Sg;6?knUN`@9eG1nC8`05M3K16@bR0w(xkskWEBtCzXYbx3I_-g=sQ@opC
zBc5GL92yB54<gEW?tOfJ0SIv_!@KVC4EIjt2Mza`%R)G4>_rc<dfv-*h5oDnqp}g;
zLVSEwo%Gk3Afkx|7ID?{w8SiNskF*!i#hmwZZ$pPY5)$qq`3t(KjPtVLexpdI92hA
z{MyNyxgiJ$24bZ_0kOaer#P+<Qx~~~ukS~_XcC!#MTuI?o?jbRlcB%aCY@+PI;HEA
zi;6mn!T9;JwBxc)?jS?ZW~_j<FX)C{-(O;|k|$O;*@nO)RxAHK7A0X>ct)8Gajdeh
zl6h^7Ai0_(C7svARczDi5qMpTUq((L1-21{Y^Xh)d*J`u2VVSrS~}qfIe<(0P(SyD
zm*$<Q_U$RN&B<Kz`e0Yt`ZBh+V<zuD+u}Y7iDa&g#OgZ{IUJ8F*Q_~^N>2zueEod=
zE3UB|Bv#~mkX@T}(`E9~y&m1FuB9xdk}<XoeezBge1S0uM#v8eusF}Y0Pry*E&$E<
zz6%Lw04KxsRSx`uRDL*L0E$vuO-vXfuQ&6mA#W!R`eF$m0v;P6ZH2Y9Gk7MJC65Kj
zwqfvkQG+g#!g}db#A`jCJ5*HeJYdQti%f_Bfr#qp6~*64Dfs*g?Irr)$``Envn;z%
zB?;@X%q$M4NwBhw{z2~&p3LV{`6;=BhF?C#?eN;go_$PO$!J4{?H3-L0XD6d$X0pN
zxPsmCK1A{HY{mVon@6Y*oQ*D$*^4g@Dz-rSOwJIY#TxEad`I&b^=+>zM6Yy~WH!Io
zzyn+v7tlr4`2Abh8Q@5NKI0y4F0Gt%%k;F|l$CV=)@ka>sROjN(GoP@4^VTPw44He
zOBPl>yQiaeB>+=(tInOT)hga{AuUR3tiDEMDc@jfqeo%oq)?J+K1)nkdjEvTU#Ep}
zl@uT5)cqb@si*hNyE+E0)7w@`g;1aQJYMip+=jQJ6%>Fe1o@k|%gk0F=dK(D2|L{v
zV6A*I@k!{p#+{I(!?Xp{vr+JBZyLFnvjG0ysK{<XD0D7SoL!G5bEn$%%0wA`sLWrE
z%1U0tp9T$Xm2=ik3~G(Biq7uxIBQHRH@UEQT>$EFH%N|Bw0<|TAHjO({rZEiI@4L|
z{iH`H276<qP$Sx9_?GMM2Sp*hTj*-jdCc4rZ{5NfX;J>TE6+0zBYhjCYNm%T`oez}
z<YFFGK6^TbrS;3?Y(PBpbYy!yzT)TBnM~4#YZ;N!3>I=se(Zuf`0VUa{q;~auvUk*
zTOol9Z-DAcgv_;Uy|oI>x7%K|UGW<X9G&Y6TxsPw^Yoi4I^f5&TW~q+^Ohyj$GtnB
zHgNZ5hkCelQ$O|_nHp+z>|gUWttzp8<O)t4TgumxhnvG(mL;Ha7=5#N*dJvzI1rEM
z?3lO!>}LJOz*ZkuH!N9P0G=igE&%glznC)mmKsvtCs+Z>#*ulJJ#%;T+I|%Cr<Ozo
zQ2wf@dIMWY&WHwk&8y$Z^u(ppphXOy!Yx4K-Ap-AWcf`u+Ppa!vDkf8i&k<>Dp5WT
zu<kB5sLOlMI4ES6r2B<R&GNT9Ezq>#761sEF_M1qnnGfNQB7BPp~;J=C`)Vq^wF1>
zIIE{VFQc4ZYPvuc7{I}EeS+!XuP2A^IypQ|nd$6*K}hi_DS22iJ!8%(tli(qE(l(g
zVm7osF8-mU5m=>~{??+*cU4Nc<4vq!qsd(3>$h%)o7P0uzZ?)h!HU6X=Do>6E+wHa
ze$Z}*)<%Ojf<M?!fo?RxLe#}FdWQZWfuyII;hBo552oI}hidR@jAhG61%IDxQ!=z-
zfpJs91EyJubhPA3MPdg#^uKCZniDp)lF-gbalf-u5t(-$qTvgLT%Uj(w%iR#+2q;7
zs#K`8JxRVhPw0Ksc1=s!oy|Gbif2B;;U12|nb#TL3A-Bkk#*C0_w-duQgfm?kgs3m
zn%i<F8Fpnn?xEVl_3*!1u9IFD>N-Ty!k5ydccx;KxG{(+jGfVeU_|G~iFf~Wq@Rs>
zN6j}uRMx-jR+Hu<a@Y+P{2Im=n4452P2-C)7W#~&oxIA1Sk1VT0$D3gy(dIxY%t!V
z7g$j5oKcph7)y-D_s|VdQq`8KG4POK-f-9(aF20$Tzi?#$5Ee~IC&(5Lt+nFlMO*d
zrkhjP5{DZ935PNuZIB!RT)U+Y#|R~V#GoidZr7&6M^yp%A`B$}(J7fPNHWPjz9n0t
zr+F7NnOW4_+yE6QMI@h?Y&GVN*P{oz@L;%NW%sM3eAnQujx^Ki!ia{1x<l{z^QqVE
z(H1XbevKK*N1{5q{2tOh6ltG#C>;jBDC#4XOdE9bs(!u`(60uS#|Z@>UF0$Pqx#gp
z`67DN|0^@Z^7E|A`S~jRU-NSokn#Llacu$OaS79brOqcop_<!jgNk39lp0Y%oozXU
z_-CYkxpqkX09;g?ByV~zzdWYZ=NdPmLhyNff^a@{m7F<x#D`P2*tkiq99e?WpXClD
zGGZdBfDZDR)+WyKrtc4ot8ALBT%r&P;?KnPx*go<jWVp`*-&<6WNI3*atNqgU<nD7
zsWHx8$nCp{s)Y$msBsHGUA{<I_IWOw%B;AQ6SPV_W0mSSOg^F!HWN97Zop+HnWax&
zM?OEPv}(^5wsl#Vx>Yc5#a*rNyk}$QxZJK*w4{;P>w^k`AA-b?a;z1<Z%h?Eix;D2
zvA(Ie)*PhMxH0h%`&1|N^H<@p`zWdIm@O^C8-~C7?l-@1OuvJ*#^x)S*|`tsN!;}}
z3ec|_ZA9pqmw$FEUBWnuhQqiHe1pXPWzu|w3Fm2-q>}xc)XL?r+?S-n$6Ut>TU)Lb
z@=n2J)f7Ldu~rWLDA_jd4|Mj7Hj2Dj+v(z##BP#V1^v+k8@n4+#l;Veff>j5cI_As
z@;FHc5sZ(^1K`3AB@@kgP)oxM19}~zTL|9a>D4(rC+HPh?k$j~SvMjM=HmW-?eQ(4
zu56t37)xuhbdVhWa3O;1k!jQy9Gx^+NIK5%BSt@38)bK=d=B5Kb==!ZHVH|x3B#*U
z61Z7mwkJvDiP2V)XO+Yg1v*oc^!RYdi<gcI)22(UH4LDq^^cmy(r}wky6-+=)W!G}
zHCeILwG`z+CG}LUL(U@!5r0pY0}2h?)s)PIS20`4yb-->|5Cd3-DbPEeM{}nP&M1m
zkUN=Q>eyZoHZA4fysouX^z?7(QEnmzPa}(E>aRUs2c`uMi#li5uU9$EmN+8v_Z^ie
z*>N+qareRMpTu_uy=OZZ<;vw5C6kIto9Ju1^=>|8TcN|p4+C-9cn#j)#-E+?qwvJ^
zFRWHMg%S!{MzA{In~tZg+;V4fl{w`#Kk?XNP}2BZ6>19=qk8wfeB`mT!o0dG+{-p2
zyX5vOqc_(s0B>WJ*`sqM<y7`?LoX1oeaR(aP8J7^@cp5|d-Wu(ke+?r0KQO;!x~%0
zwN#|6h{$RyRbS%4cJ~pn3cpz5JYX%@+fHV^xYz1eBLAe9_iIrl>U?_0{?-!9!INe1
z^?*-;gAQ9wgZ+geFzRl+*mDJBShsZYn0lWEM@*N59%t8bQ;YovN}_h%#&{JC=D;OB
z4a|KrnC5Y&0Dz`|x6=sN-p<|2+Ah>1o!#I)#~eLqA6TkMGO<o8NoT9fo8x{<KOEE5
z`X+Plo}>eH7o~!!LWS4E0UN;y-3e4?V%zhyiQKOe29ggZ9r;20uRfGwn|5rjFO?(=
zq>`&;Iz9$9b>`No^17Q!aNP&?Tc4+{ecWH6;7vR$Qmj9>jBAF!dOC63(unww{{`qb
zBtP}2iYY9Z)9Jcq#M7}>;yw0XO-O}UCkMY>ldb7-TTPKF_CVKtad5A}(wn9-V4qWk
zg2+8t4>$COz?b#)A62hTroe_XfAc$(J)t6J!`ss>mWIFNk~RKVn4_0HTjIi(&<ozM
z%wkr)|L_&_OKJ$Jxcp~r+OEl*bnRf>5hl%ywL!TQ2hv0_we_f0$kai)1rDk)tr7GQ
zRrY%xrBs+X+SODY1e8z2)1k8{2PYKYdcdj|u0El<6<(mSKV}wCpe|yX0P9W690ocr
zsr*c*)BEEFD5>_Ush*Qy$o*t1fmusUiwU)872s{<OpZ1jy}2jJpAB0-hYb4v293U}
z-;>pxO-Kd^U{N@;pHp-{-eohGf%BHy#xLhFV@<Ri)l{n&V$SyW=Bp$#C7aQwE^W`v
z)x4jp9anm=eEs;c$XW#LEscrN$w(PS_Dx(Dp?qq0l;c2DkgQQPf_bLx7+cugisMAD
zVacUFVIf?CeT02v(+AIyZ=Mx^&14CSVMdw&<IJ74JVUK9l*oWY*ez(XvSpHwdjmev
z)enh%h5bp4jr`p7+Vw$WFs*vxajETgmWZgx^CJ55Z7b8KxuJ-<emJ5^)WXyIyv!8Y
zmbg3I=Rx7~1si#rNcQ2lXaaxSv1Lx7Y|HWtQogrim+(|0u=?FcKra#2*ELZ>z#HrZ
zp=28M6PRZ1B$FL?21U$%wGnh|f3d{r(eK6D!cHu$pY$odLcSZ@O-|Ei9O0^FME|Xo
z;O6s}rBd?vYkAp6C)evbHAH^C{Gj-`*nJu3Al-*S%_H8XsBo_@n4nND>2uK%lH}0+
zi8Vreg@reMeC3FHJ@Hx?TLS-dTdNtQ_2FXwQHU24*Ky|MAcn$i_=R;}Po?*m7x3V%
zY{EW;unz<!#2I(B(z$kY=#0*B8C#^RS|3l{Y17RQ$D<xCwYCKX)?~N<6N1X-Y<p<i
z?<z!NR`$I2Ec;*AWJ0!0-!-{5Jr~cLd^6+`782p|f@d<@hylKN<QWX#l*z<qC1%`q
zg$z2)pC=%q-&Y7#d|I?)eM;C-U285FuPWk}jaA*zsp42<3!midMG#=t^;194fwo*R
zXpeP6SohD{w+sR!X5~Xaa(+|UWm~@MlN9?N8ZH_FoSpsjt&Xb8P<|R~#h5_k-PpVl
zVpf`ST62%z4vkZvy9Xi)Su*tLZl08FhtW#7iGf-rxz{0|Cq3o~Vm^5#8<l3-1?AQ2
z*_F7P8HuiWCDXO(mDML;KZ0QA!ut`!`6vcLbq=|pKrwByNA9?}*E(>teAB7}u^<H%
zgBhNTB(j)5nR-RSb$rfSP|vxJ|6czG=*XJu9<W9%Eu&q3w_bY9szag9T#@YXnJ{rf
zBUo|h-2jPh^yrhncBM)zzoe+Zs)(CKS9fa@g$2Qh8OgD5#Vt-)rpJ*JvOZr}W{!t}
zR#=^k-F496b$H83xhQNd^#rB;lskhSsy7vd)lIdq(50Sh$2)8X{AF9w)sDUZFkU*g
z%3QXbrgXWF%bd=nGUR#}i=Q-(>84B33nPrjjOT?_mAy|YrVO~$I;r1^YluQo_I>_U
z&OEz|Sj_ymbVkS-OpbV6i?yElm{2Gc>g)(t<@g*}<|jBGz<=7_*D}DVMAS2k?}_$Z
z!|P<{U{N=xzCg%;x!Dz}Sg&qiv9}Z^nw1#*3+jZ%+G%&gFssUj8KLH@<)Y@osvJAD
zH~Cxp+aR}o>>c2C6<<3!6v&JwMC;^aDMl`saXZ-?6o+Q2rAq2>Emntli46!0>R$_4
zdXuczrW4q}>lwxQ+^ie&-JDUeuJ&8;`X)k)Se`eIgcL0da0=)?8*HJcaJ$DI-;#Jp
z3E+_<X6}u0bd>>pdVc8i(|bHfHZ2xmgOHTov(eEX*J+1*0B?K(<~fefDNMI($zvzX
zuH04K0?k3|d?}oHODgqiN7B7KfF)4b>+&fX&#~Vo%#C|beWV;cH|Rm*)lxRnsWb25
zeurPlmPpQ!D*xgkyqH}13#S`#@2inQeabIkxlZrM*b3}>;28-r4z`1l%Kz*xWB?>Y
zUmMlK2*tM@Tw^%5V9To6yL6%=DcVX9lp!4$c|z52{Fv8y9D(8c>fauM$wwgWH=P@%
z;H0F`pp2RD9A~W^BXIT_YpHAco=d6sp!ABrFr*mG^o1ezy3S)$jc9|==O+c%eLcwP
zc2J3vv1^1vqDj6xHTRepX_@9?r#2V5^%^VQztvm)G$#%gyQA%A1uFKEE>J38qesxr
zFzik4f-`$<P!w3Juf%+d<Kq5X=?n20a)bUog(CI35+O?#6UQdnv!|GM^|;Rtg*~HM
zb4;`$U`3uarfH6eXSw3#t<J|>V$!|LzGZ}OP^LZZ+UDLt%2Az7WM(d<`vPm4XkN>~
z_&E5gJ1h*d(8;d01)4}0X}3|l!(zQu)$5t6k75bOchfMyMrM?RBy?-*-}KeJ2l_&J
zpqNIq3=7A^0+IFQ)g}XzPu;6`{Wfu<ZT7WO=)^Z(`dx*Rfn^vd2Lj8w07}f1BYm~h
zC|4#-4Yhn1-P7j$nUn+&T+x*s<2+Cug%K<@5$LeUk4Tt{kvMU@vN?XHYv$8vQ82Ct
zO#{)UGVqt3r}1bUh?!vJoW-#hfJdo>zdS}3HlOe<zL};<&PXgaKl5e}`jM;)Ge^Gb
zKaqetBA2~p23-k6<KD5RRHZ9=2-LK`$2Xdf^0cX8G*L%J`JIpR#2M5fCC#r>=L`-Y
zh(!3X;!A<Rdg-`2{#Mj|`5QSeur?uJwA7$k)jS<YNczZ6Iv0zLK7%^*XRZ#FxD$I)
zoUowA_`a$UY~tke{07{fu6WX&AdT2wBBNfU@ld?&T+TJ&fo%?r09iG|j9aWIh@7(=
zSPL?%<IKz0!;Qb;?U|MuhD&|*?Z!zl)5uaT>M^Y;oP*3TR2-Dvu#yY6x92!Q=J!3Y
zDRqAD9m6q6`i($dJs>L~$HFR|fHp_U9`DL~TmViqDX1t?&@W}<s!epwkvs)mQkF>O
z>6>NZ5Y{aPwpN*j@PlVgWJ>P|;b-HH)D=HJZ6iVQjn|*AvRj5>CZeIQ^2djbDZ_-e
zpmvIh1C|MQ4RygIbK8kH9$AI)l_9yo>;4;{Ljp@RLN0kC27xqgzde>!CF<ouiL-#S
zs@m`BziP1*<qi5#*?|#NKvNKkWLq4tw6(0ANcYxlV_k)EuSzW~dL=>V<vXo#rUK?A
zxvw@}jcpCLXEE_rpdYtUYHxxGVTz20a4;?)o+E_o#WbQ&b%#1*t?n!hx7KoWp^_r@
zJ&C(9@a_>4gI-R-0>fYA=XsaQ8+Oyb*Nl9^8BWf|10|FjI@>n)I!ca|`O!kT-fmhE
zyQyYg=mAb4HZVoe{_dw-r*!VN3jnqs)OS7c!SidN*FDKUm~w9_Rji&gA_G7gDmmD+
zUCUxHXeHmh*Lu@@&2V*FShr?2#*;}5Zs0Yy;dwgvW5n4cm;OoJ4pThuwcciRe>Ej&
zAnvMoUtegAXp_5o4R>;0PaZm<8SfDY*<>O<dw>`_4cQ%_zQdSbU^<ts2Rtr+Lw3Ky
zVh}xy6fqR4x0uJhWamPaG<nnByeIzU!80jKsla~h8^f|s#aTPgDQ8!Pi)-$6;c75Q
z4(A~MF~~3~G-rxPq)p7FUu$Dm@KSU?r-_mL*>Sd1TV(<_Z~8@@2ffOX*x+O)GGFq3
zWgU-tZ85b5eY__Am(Xc>;CVwu2ZwV2?roOP(CC`Tu)Rp7M6Vn`?qt)rba8=e>cukS
zqdDt*)gQ!+<v0E)+E;1%b&Elck+^O;OY=CTN`2m)Z{Im913)z<@X+~lV~g+f9{DZ#
zdJoFtf_L8>U?xYZqjm2%{`e?fki57Zs%)_!U?Lb$q=i%2r`hoQ^2CV#n#JuPg`Drx
z66JH+cRM>`osNyR;U5A5I3t_I4!MpnPjbhapGdw55#R)0alX9_cIttxH~ZnD4HNj@
zVIUz|o@{%ajki>x3F?vEz$ZL6Xy`#U2CMvym^V#wQ77yinCXeY6yAR2U)W?{X1u-B
zw5Aj8)fLkuXdw9QiH5MBp#U!FTh-0gdVu6)uu2BHo&i@y5?fv2wp`wUP8^>Zrwio4
z7?}tT-YF`_x)sY6MLs<MH~-X`=#k^;y?yD4<1H0!y0q|Xvb=T#j>(h+-@Qv~(X6um
zx6dv-a?^~3fJpk83`%!IPG4Ws;canwsgBLZJb(-SBz)32CS(8r$V_>jqS{FwZ+VAC
z=_<IDUh&#q1_PlNsE@xZ{G<}CIi4wnOck>=I?Z{K=DoP66<~H?_*>Ks+$}P!o|jrr
z!7&-LX<Vh>^i)Q36X}JM@PFdd`XWa;!^}|V3xRH{(r~9#>+l2hi*zm+*|RI;kHb<>
z@oc`*H=DT<$I5!P-ustv<LGos;H>!at)~w?IJhMx`ma;**N*|M;#aQFkLQ}^x|C=!
zX$c8)I~q!)fM#lf>^TRwhaOin%c;OJg>^<qB5^q((g%ZPm_$>!4isuKg0K!VjbU>P
z#5=@8v}om{aXW#78oM|BdVcvuzd$X|mE^6Z7`+6o8t-lR4H*QO7qgA7VlPKQOG}#+
ze%j{_TVT8sd`BFxZS9WM!sf0%uR9Tt{uqK+#}reLTRTk7FJcTNl2}UwxL+;NVP1qd
z)}S|TkIe4SHb`hIK31T4|I+IvW9yIdEsO88)0Q%CoHApm&5|XW4AEuq66;W&2XBm|
zZV1_q2LEn`z}<k?!f_$sg?NYjMjDv@k*UryR94zpa=#<`0>FF~PWI1h0o+YCRkPln
z@=NZG9^#AZWp0o7w?yB)x4#`R&s)}CqfP&UqhsCon`sdEdWOz;T!XnB+R(&Z6i-rU
z%dqz-@ZHS){Df~4Twl`{_LunPUxY7}UI1b$BZhDO+%`nN!9LC83;g}|&}>AEB`ERX
zrLQ*@fg5@t<R|HuOUV(9)P7nk7yC}W@L`QXaJ$`1VKjp^X7GWlz)(igreV|MXbH|>
zem=X|j=tj%{Q}kCvm{*FbeOz=O`2J&h{k->JyDXsUcNu1Xi7AWh7&4T|Kdzik~K&@
zwSR_`ILqO5oA5++tyHT7KiV7=scsGMU{Py&|Jf<9+`Gp4ZO%aUTZ}wSlFy1G&~l<}
zCjOStdraoA5$<x6Dj`GtgunRp$jKE)mt?2|72tk(04CUTaw^6q2WYrP5-LytEDC(k
z7(JjM@9G^|EfU1O?eVN#aAUC>Hh)X3B_!&#z9FRZ#`>#On&hH$Ml>t!c&JI@%Lw$_
zW<<rx^9#U!@o!)s9croTcnG`dO49tDcA`F{boIqb%;ipiTkfRMoWPz1AE@dFQKQ%+
z^oMllWvedPbJ4A6l%*6>Msh;oPdi>mN{h7pO0%3U_2^%g8)5*2(fFQbg>M24Cf<}J
zCxOl67CZ`@8>4LaN*sTzvs&k8te>8a=VnV`wkt(7U%5(BgmkgE;z7Baq6kG(JWJVi
zLLLJN(a7SqJJKnPVbvN<J*5%lhZ|=O=qJ?>&eZ%-a9TPV5S?B$t*$?|gnJ|O3vHk!
zPPjR?rgNr!m2n&6h#0E5Piuv=1-$Q}Y)lxo92E2aOHxd++Xqa%^Jz!oWEd8^yLZRn
zP_yrC6U^$R`#|Mv_&Ps6&Ahecu?BCp^~z^Q@FY{{Q2_kRt9oT?mb=V4p9-ucuwvGO
z7u?$L?eiBlku8lGZ|8fRZ(U56I1%J({E9it*5P=LYN}$ARqZ9yYD=Rrq$s=K0+29s
zNg1}{kF)H0Mt@`S24*G0<S=^fl@EsMx2V?Tu7hX9G0g0tdLQ5DwHiur57;Et7$diC
zqSgs+eN4Fo;K&!M*D~p={lE#84Pvcgto>Zf^n5Mxz+jNpt<|H`b{O~q?f>4~-K!AA
zZYHIC#>4w$simS9YLj<sPI`ABpkmU&0tHOWPZj;r>6s^N7^RAWlKYdXpoo?IJ_1>*
zFxK%z2H4802khe=gkCMdVal?Xv3tCS%UwIJX^2gE-*+yeKg<lp6((?b74G8#vn9@V
z=~ca<29rCU+uAQ?tNNy1j?Xr*u!JwFaUj;GYMiv-rcREx1I%I-?Dn-m-lx)mpBLvO
z){~yzNHz-d)VBXdw4NuL8oLvN!s{A#sOGDrwE2L+&}$g`BTOA-9_Ij&19ny7{V;6S
zOMa@0q<+F6#RO-iaAGke@^*-1GtoT%2qZMi-``}eJE3be(f4{3`jkwbK8&%4aJ;B!
zQ#T)z=+}Z7$^0#xA#@vgE5@bum!8D~tBk1bm8(ZGqg@`&dH38Z8(ld(CUC1(XY$2`
zC{Gos<DHMK3F+;jcLir@{9d4vmF}BhVOgIOAI@31%`~qK8Cg$^Cu#nrSnXpdb&vU5
zTID@4X}rua;Q-^vHntjH#`J5FeoK9+bQ;*pGO3U?S7MqZ3gV3wecjI6&!YBTAl0KZ
zmtKC7$!zxdo*vt^Bb+@1+=}14IW1G<Uc&qGYjmPH5^L^CfgUZb;ZcS_fMEhXRobIF
z*+BDEkoXlaY}@drdy667Z}e4~p@^PiCQC&?_!~`xQ8|w%Cck`}$wam2G4s;n%e}0W
zOLgPARJW9eLBj4b<=uCl=zd3bcao(fu$BF)yWG_mvdEH^XioCEe)w&Dp6mCkR}th7
z>E$tp$$9!;1v_s`yEMsDIby_x+R7K<eb9J0d(=mF833=H5*bpOA&a!>S$I?{h%<e1
zr1z?nbAk#&{jJbWYq=F>`9U5QsWlW=&ezmVeGe_ZKQ29JxMPE+-%>OAh+RM`|762*
z@j9j<yu4v08Ak3Cujq8>%q*d>TiFD0>967+(x0_IgKCw5K{!f_2mQTwuGD*}eCh@x
z^%QbqghkqGmBMIhD}}bc$=Mn;Y4iF#pjN+HsOt70Vzu}J@RwHZtGnn~9)XF;3xEn0
zCH{$9V)XE0!`n}#m2D-}8161!4iOcH^3TY`arfVdP?P?5&C8vxwJvZI-CG~AH2BF~
z2LU@!+Z&~M^5_rc9HVwEKBdfOr40_m*SS-5k%@Tw!`=;adQzIv{XtDFk7~xcZpeZc
zsAoewxlf6tligjy%3>`nRivXr?fP5oA6ljPzpctNM3qgF6(&(slo_%8?4P=-UT2e5
z5tiz!Rz7sPX>}YE?$$v`cz-WHGBp-1ya3$jdc=!<@#m<8Cf;Au{N91WZ<x%VN#flx
z%$(N)7D0L})zU{N|7<E=mQ~9yaSWF~G_lx6scalJp<8Uxkvc{h{7D{(n~Ww89-+{8
z1+ujsHrv@X@8wQ&%ubV!JTKK{;l`8Znsc|_S`Q4C;s6rmgAOf+P!=&5?K1AXyMRfT
zr(QaEAYb25-d)cV9r9<7vpJZSfplG%+mZPY)@&_bZ4|K(Jv!U<bwS*0$P{bkvMXR+
zOi4A?YQ3#hRkfFIy$Yh~GYZJneFB1}BThKQJJ}69^64!jKGxqn!)zs{H!EyTngK`0
z_{&Hw@7{AkhTApVCwQKei^Rz!eCXjx`#vppI&fUDc*Cq^kwto3R&*b}w7UwOsrm_i
zU^MRBMo}{0KF^b8-N76y$`ib5773oRI>$_OP-h^LpEQ)Ov;4s9^fj)M7gLly?2`>X
znv*y(-TnP_)4Vch-pWRy-)SIPiRx43FYOL)r;w^~U;C&*OX@UrLzA;-TB402X3YEE
zv8%mUN5<&QPsF7Rs9Yk=EIV&eC1&iJpa)TsD#zo9BqbxBX?76)$RP0jQf2x}KAhV<
zZWdDdZa@B_PHCDD4s*D+XpcKnJAy3{Q%GAB4F%{VHBRN_&O2!<?%H|qG+9JB#8DO&
zy;LIY`u%Bd;TeFZB7ef1&m(wLX5%|H*{yt4`^DXxZ?{KU4-_LhyC+pHw>cAFE8A87
zEX@8nY{4g-#!G*KmuY1#oOO}LB;ZM!&j`~}GBc6<O*?_609{|mHU^%NR7fT>Y>3;8
z^lK6>XsER-6*gVylMJo$k}|Mb>`UZHR&*sxvwT^8%Dm6v&q4SSw{D$UQR9(Al<nv2
z5%pSs{QDd0!yl$s2LqcsgXn-ir}_aJH@2<EPAVb$h0V9T@550=c5-D7F)N4z<6v$X
zRFo#s**&?DB%4>O7_;<1dyy>TbO-$dZ=QNfklO@~bXfGrLd-v}y`f}r%Y(mK5lp6?
zJfS7zsekA4kQ13s3IAvmF?ugR7w7#E1Vt4!0~x^f%+ft0X64g!IV^SE+eo;)CMr)G
zNXBT+l&p+^)2H`7+N{uj;yQ#T5O3gWYyyWs@lb*gM(Dh{#ILr?oo2jF3lyWel!K6*
zaDtOGl40ohbj|O!OW4u@Bh;Fcmn_+g%iKQSfNG=~4y+LplOX;x4F`tWTsOebBt5Q$
z;|oBwkVU3EwDYe)Q8SN`*FL55FH^u*vjKO1-@E6Q9s%odrD<~nFrd4n8W}e}GV2gL
z?7i%Y$pWUDGu&_C{riPL<?{3=&nF8DZTYH>LMwsR50`u??2-Pf;cZ!P5w%8L!n#&Y
zf4xJI>U~$-`U>An14pXA?mh~)`*8scE%x1Pw2Q`Cy3{8vn4N(OnF?MYrFFT5K@w&@
zuSvG`O`?sYA0y=90x5VM5ECJ@q)`MMtHKl;pn6SmLO+`b6_Rm?WV$69W@<p^CdmM^
zPnmlfKPVnxXo23(Qcao5``GFH6e!Q9yE;B82rwf63bjD`esD4)P7ox_w8IqsYpy(v
zN7TEC&PaT%z&`qtVci0zt(0J)&(rFz9M5(}afXe^_#Be8VrPP2fhM2(l}^|qND6Lz
zF6!LmKhiv=MF+eoSP*A;{l41?7i&%S@`IKSU2BQ1zh|qgTSj;Bb*6{!@A$F=S3D9?
zIq<Dk*)CD}PC9iO;jf%=B)4N)E(l)OJ@v+?2G_TCe<<ibtvQZT#4EkSlyA$py9X=c
zL5V7;T|Uw}a5-F0_i8naRPDGLBbVYq7-pN~u@q{^N^yOzxUjRI+;qdPZcbe4eVJ)y
z1=}Ie{+W#$)jRNbz~rO70^Gv|Y@7~f`mtO&n`{!M2d^Us0d73&cxwc31j0_ISc|zW
zCgWN*Z6DqZM4JWf)(wNkW0+PNUD-W*8Vb{$$>`cAkW+O5j^Nko)0C=Vkt0Z9LCZXC
z)ef4}JK4s!a$@f{RSm9E@R!XPyD#G=r@ibWx!oL1c|Y&6R6cey+z6}yJzLW_T}w(|
zAu@JVkymO9Cq4?4oUhaG92D<Fo{n3((-KyHm^5&r9fVre9`np=%{><pgqf29QRzZu
zvA}m<epomqtzyaKJV~`-?Dee~C`>LSJB}iK`Z@sHyl@!i&{Pws!&^_f7~DriHN9<f
zqpL`n!ew4TX=P&}){bh`5>Z*2o24;OWm}Rrl_{st&stS_xMt0@ZaPQlK$@5~LbIo6
zLdtHj(mhc1_)$|*kyQyUUR0_{7dWj|pAJcmQ(1((8SQ`VnQRtG&T-9+1KHI|lZ5h>
zTZy=1y7Spmj2pRfv-<haMkjH6{jk+q_&)My`L+%5V-k9+J*tFl858S5R>0qgN96z3
z=Da+p_T9o#Q(Z1mGm~eXs;<1wNZJvzFgK;?<fktT$JrBDHax2Z;3FMFN~%R|bQDaR
bCVunZM{RAwyQI-_R*qNtDw~p0E@uA=5z)^0

literal 0
HcmV?d00001

diff --git "a/Images/deepfake-xai/\354\204\234\354\226\221\354\235\270_\352\260\200\354\247\234_1.JPG" "b/Images/deepfake-xai/\354\204\234\354\226\221\354\235\270_\352\260\200\354\247\234_1.JPG"
new file mode 100644
index 0000000000000000000000000000000000000000..85a0e1aea7cec9a380ae8d050919aae08446c415
GIT binary patch
literal 14630
zcmeIYXIN9g7A_pB(nO>P2&jl4AVPr91S|-l_aqRS5Lyxly-E=g5TtiPlU_nVgwVl)
z0qHFybOi({(m_BJKRoAr=iKMs`{TCfd%hocviHu~v)7)r-gnl_>^*z_O#Jx*xTvk6
zr2(L#q5}Az{Q!UFsi)Nak&Xa>o*qC9005i^&{Ew6P@hS3005PC6cRvtCR3elhB>)^
z{{f5u0QILeKES!NJm#}=)w3PIw{~{;d-eY&0L06|^Na?%M4Er_$Z^J3ULx(k<u}Pm
z|Bz|qKmb5;@_*$2a7vZ@m&X6g>3@PrT1r|@NlI1;2)ZVvs3aw)1X2L}_woP$2*3;A
z0Qg%0f97c~J9~S(D@jUXyzbcAyV*J1LAbd}`rEoo0`Evk0+jFhyW1kr4&K-79Gs9C
z6_Kshw<6b&_9`MUke-yDyPAVDQZvxg!6Z=M6cLC<DB6qMQ)OXP_E++Eb$4~}w!P->
z>Vol7@>dc3N9jsu^51MpK~-f>dq*YXM;iaqI6G4j{Fg6&etviSWbU|mI!OW*6%{3=
zq$Q=LCC)S?yaF)Zw*C?ruN(g=;E{tD!V~H4jda6Y`&*!`otuxhilC1V(q75Y))8c9
zZx52Nvz4)z00Ql8C2Z|wfD-mHGWPNypn|j`&`$7Q=Is&xw(joZ>GBWT_6SJ_7YA1d
zjJMYr8$d}x$$vWif6C?=MgKPX-=+Ty<C%_<nx})Ux5J~eUG;B;N=V5{070hzZzbQ6
z14$`M{#Sa+l7GqgFC6_p>-ql>TdK<T2qnZ{-`za_Y2L)a<A1NVU687OyN-XNJ3INO
za8~SD%rP)9C`<mA7XMoc{*iMQa%cDbJ3!Cwy!QV=`v0?$|1@$3%vsC!J8Qy!rU7aI
z8fxml+gYML+vv{I(b3Y<F`hd|f1a6<nVE@^iHYUHrHd>V*e)<JUF5pR#?HaX$;r%m
znVXA)`w|By$KOJzXwL4TrDLF@W8h$6V&V9oPJiA3*v?Z~Q>oBU2?D6usA$-z{(J;n
zJsV2t&W7l}qx8QL74?~q^ykhqFfyGf5H13!sc2}ZX=(m?eWo3Jb{#;=MtA8t@Bux$
zk?lD_4-V<Dq=NG|9@f6+G#=R&lCkrA#=yvRnVX0Arm)B@Q88JNoV<dflG-D64X~z`
zwuvbO3Ntf@BkUa<ot%*<FK-`TKmUNh@aGYcQ7@unl2cOC(lau%u!Tj%C8cHM6_s!5
z>In^vP0cMIIzJM-x<B>wj*gMW$&`u7sriM)uS?6{R#w+`cK7xVejFYh|NM)K3PAHu
zSpP=$U*Tdq!$nO?OGA6^FD@!-zq6f&jh5~@kp9vGqjR<%?1Iu^=Q$oG71X|GxFKV_
z&1vU3!pJ2gJAZTMFSLIk`yT`L?0*T_zk&S+*A#%6hU%>HXxISv09ijxCFV*)(r8S1
z_blUD#^8CRt_?LR(d=4%5l$wMHm`t%?9f7KdMVD!)JT4N47O#wS>z`eNtffGe(Tuu
z>z7~;y@cDvjNg>?@LFipN|kAX_Mm?6a=qQx?{Up{uCXL2-Tui{I)5eY+V2Z!VhyMw
zXm?p-0J+*Bm3Q;1qLBB1)5fuZ$%&GI$#~nFS>zjVJ@j6bvM%<yc8a`k3qxWkL+=lK
zkxAwvE-Fo;mr207C4rT=8}m)>*M}B7r~6COA(c(H?n6;LIWK9ft#aVI$|rJOk!%M<
zH))}OqsX=(&=7{#Yb+6-DvUV^sNh8U?ioNTEs|v=6f;V+xsuW=x;|M_u21Vj$~H=j
z!p&nnC3Kd)8cP}(3+$5@9FT>_=ey9!JM`Jhn7e^6hc#4W0s5PUzKj)6XxHi?r4P|1
zE|K1{b<unybth&SgP7`SvGW}AZ23h{(tiR8Zgrl0CYD7vB-<ckjDwhB#QH;}NmQg3
z(rqBN@!AxM{-m^-X9|j`CvWJdY+}QnkfNXGinW?<jlM!L#(yyO<Cvc-a%+F^sQ%SL
zQOPSN-)|C@u(&HsB8yh3GIq`RmnV-$TEB6|#`lZYSNs27580_~lh3p8AzHkB(KQ<M
z_9LnzWj@#7BP^OFmGzD$w^1IU|4VkODi)+Ndm;Qjg<3N!$jcgEyw2puE4;CX*hDB+
zhtEfS53NMysgrB6(U#&D{Z-KGZcM*;x80r>X<dJPnZ5g|eIk7iO@p(E;l%C^CqLLR
zD5>no#)>ql-jL?0=6zjS{O$JPhqvsX%I)9yKl=j+2HF_()J!6eb%N!%Kdtyu{lc*u
z?T9uFY8pHS8r@b;5H(tMN_nvT*)YySoKE$*UHBLA7t_D(mRtMq2429~mN$2NEOCrW
z>9^M+`4;aeUkR%6O$+Umo%kgg5@Yu6<ApNr>iiQm{vG`Uh>r{A3mMpTCD?MYw9en(
zK`eN!F_q^9KdvpnF5CAj?MW=zVWaqbGlbFL710;WS&;hroAmGoCp+XM!W#Ab<?pGD
zGK<XRG*v^B*s(?vgh@Pdg6Oj3m2SkmoE#GvyO1oB{sFJRXEKg_k0O8h$fTRB(ic<P
z5bJ}NW9wR<g5?zdgwa)41Q<(GYOJRjIG)K^do3`r^OLf;@kKWj*E|gb#H8BS_G>o=
zOdZxOysY?=<DNs8tGZ?&%8yAnkqk{j?VKJ7;(J1H#6(s$YA-vP(;Sv#OG5~kLCp3B
zXOU4SGFxsz&{6{oB3lE>MrI7DVj1IFNE0*l;|UVF_9zcSOZe~BL-Y5llQ=xjcNr`B
zh#-Ei3e84oaKrUaRVA1bV*O^kjn>g+uLb%T&6LC2GdXr0(MRBInAWzb#n(pjJ6E3d
z%fRzE2xv5iu~bn+^`g#kkqv4W2hn*HSihz{h24&b2gT77X#faLeqwiOu@i`cLd{pJ
zQ0-+jZeFZ_+Wlt!Sq&oKR6Yyc{<`q`jvG&L+3XzGnwJ>`_Cb~KvO=8K^ASM^51G~R
zsSL#+FUM4P1uJWt2mYwCtd3jc93FA@8jyGO(p$hqB;Kxw7!qzMBF#6wkZu(6GK>q*
z(YrHZT3y;Z3hMG2M78El;ukh5a*90LKfVQ4t(JB&w#o|aiZh-BN_FG?MsbMr5)~c-
z69#U6*;lOnrNg;V{F_bC_puCVWo*n88|_)g(3>WP_rb@~SCV4tztk_mM_lX<FXDx#
za&F#&Kkk)QZ(Qy|r|od%ev?rJTH!RVC|1g+3p$%IGAO$l5hiaie!}uMRgZn-642dc
zT-p>oj&oz@S#vZF!=^1xApu0^*Z>#56p8ztfN^qy4=h>iYx9f|rnJyyJd5kq4UWta
z=6v76ieeskG_<$?|8NYeJnC9{C31FL)^}+FAdMUbH`m6Jf25S9X9i^rAQU~<klg9@
zHmu<0oB7N9zLbiikm-$!-E?gazz<SlVd8Gga%&yPcl<bg3h|PKR6Hrw=-0!xKmy^e
zt%2IuMv#OTrnW7)d@FXjf&2%cH|DwrdI#h8m}~1j?BeYxb6_E4+KUBz<4A{t#y_GZ
z)-zH>#Preq6(%OtY<Kz9h=D$H>UFYDeI2_00IFY?wMc7f(t%%9NbK+NUh-Iygn&dl
z-aM08dN_1zb@wq8y5{=C88M?hbMOn5B6g3~v1Ul9F;@2}cm+U&`j1LIP?XK=t*Xqq
zMTF*a^WbRhu|4B%5jpm068R6XDbCLuR<B))4L6O5dH6BvU6sWUPo|CA<g`F5TFHQ0
zs7+ex58&zWfJvRwAArIgKyr?w#8dG^D{H0*YQZl389V&@1%}&QlNlo8PXb3Ra@j1W
za@nmlazOI-mmg7k<n1^7_|bwdPiMIB-IK*~v|h%O*|lpp#7%oee=6rd@PVv5@_Ul#
zOm{&Ab?&RYgAe{MKc*lBZl2tbTN!@J;xnb%Okct!6KOKEXx+|;6;MGrEWv%JGH+qR
zsh<@DQg95;?g|s?$4I{?V>FzDNIwidy^9s9iN{)ok+ZG_dfi!LQ!ZT61MSp*aZH1`
zp5tGO_IRVP#kP8qHO)QX?w8(bMyZ`M@AhD0mKJ}uL!W*vts$1i>Y~id{2U*<VZA>-
zOX36qj-qdnN!2gf(7FcjCf<;mx?g=xZHKmM<z^7Yf@QZqUvvoXQIE<iek12)_!geg
z{~N_sA<I2m(Kd!ne#*?QXL?*QF00dCYpzLF%8)&=+GSj=nlS%xL0TPQ7P7KXt>ZJ*
z_Ulo-3Zuu(^wmwF+xqIxX3~dM5M{@*xXIU*Pkt`6DJ9=E^}gH3?+!1la^Y-UZn2Wv
zU0b}QD0t7rQaB22Th{p_yQi@)=Qv=%x@D)xl2v$C*Tb8;iqVL)DS431o#LFD#cC+d
zLpWvZ=Ke{MU<k1A5K38xlN-M7I!mTvEn2g~Db)F#vAfJENwDu5H1Z&l+3M@#$6s0T
zA{B*Y1zn)*h%On(Aa>Nr+qQ!+f%=)ARPv|`Im|{8;rw6|WH@5m^th32;}0MZ4}RGE
zsAZ(k!Id0iP%UGbS28~$W2e9&@Wif#LG|{Ez3c6`P2SSRosg(<VMG=r;l#9rpC#4B
z!^7Y4bSC+A>=V;QCO@YO_HQmI@x@ZCr|QK|s~ueSzsbHVe*m2Hrqb9Hzsc_W-Sa2p
zm&a?l6+5@hq?nZ(J2b05hT3c=d~iueP6k*5x!LU}j6Ib4X2^M(YtWen4Nt$}y*L}a
z^b-?w))g-qr&WFOojc0+>@wU_(v-`$tge{JEnPwTAB8Y&T+pmW%v2TAtXgIIAqH`5
ze*mmOTCB14rn&BK%6i{H*Sq%wJkdQNJoOzslxi26kky$lA)$-b&5dqqIeXRviUR5n
zvPF4Br*)h-I?Vaec<q&Ub|s3?9Gp%=X5>bZtFnZ4?Q%upfx4r$IF03K+=vh}#c;$p
zI8h$n^a-UT01GY2)S|U(UVriWsK#R4+t$769DJ``M<aP|DasCADh*6ExIXrETAJco
zZ0W!OW4*N4Azm<2-JrCG)mOUMaF^bd)KwvIdGI+%rEuDOcVgTHRy-)Cksv!Hj@RMN
zt{kt*wgEv9osjZK@yOzhf)SPhx(r3xEh)cB(o%Ow+B~eH;=57&^Yz=wnw7N~0<$=^
z4F<*tb#aj(vUx``X!pxqtu2#>Ot*Swg3&{stx7LHqa7df`0+NQb0;~CnY>8Q=jt#4
z!nL&7xM}_-JnqTUHc9w$LDs2ZncInV8dt@F=IzvJcy020<eZA$IbCFKD=N{fd)ZRq
zLF`uw)uLn^s~dCr6e8e__Z9TbwyO)Q7elMF7NUk>>P5dqV>qF?kCOiY{KqCHO@?>c
zt^WY5c9<iU((S=(%|zQ-f4dchD@#H@3BlLzbvhvFhXFqd4I-0{D>Vnu0|nZ-lY1@#
z{2lT4rN2JO8DAdRDmYG>WewtkjojO=##62eqe72hKZx##l78fOgM100aj93@Uu$%8
zC{`slYchq=he2(O2eWtQa|^PCm1b?C8eU@V>zeMFjTqH_W&4aeKM@zRSc=ANZ$-n^
zPgK4?sC$o?YqdGLZ)^R~u?`a?=2|QSCJKyxRSL`;Lq7hrTNvxZjq}}8znP*QE3XD4
z&vdmKXcq(^2<8hX3xwz2y+}m`Ut2a>S2~0*Gq!fPP%2jrnJSUB`nl(qn-Vx3=lQkm
zAm2c(+ttTDOd^DItN5s}b#l9s4=a7UihdS?@q%b9LG^R)&M=>2f+T3oB<Y)u9^*6;
zg6d8IXT(sj%l6k%5N&tuL<vEGkG5j>>mve9-c9RaOf^fQ0vDQpn>>(brxq`LFFTu7
z1C&-(Jh=;v&t|$E9!r50ulzW*HlsGsr5KbPVLJOR@mG$oD&zo@k_s}|o_5j@P~57P
z<hu7xZclko^2zkZ9O3)bNPeTVey!`ohjw@#4va#%CScDTE1d$43DRi<$<W<K%EgZ_
zk=9&Ux@ls-ig4xjaas3ooL|T@2dggXu9pWa<kH{fs-ZE=;j#lJ9}SV3<elpE@L`xl
z`~;TW8tin1=5W9;ep8FPnYPT5v3JyUvu3kxz2Ky4r6X->WlCF)d4-V`p6>lnSR)r*
z?jMIee{s5dxVh%xc1@T|eG~hTk@E3UoL1SZI8(6*$ZE`L{Z3Q3;Xb%XKQH3Qd*!ft
zobi<}E*}9Y>yj!g>$+cVtW>01)evK5a{8epiU-F?xTRet_&W=HqT=aKl$<zci*0@L
zG)<Z4nb?V3iqFJ}718n<)-Q1vWmX%j$8x^oDYY~!jy^>7n`*n)d+}&z4Ok&X_v|14
zsD{BKst}$^FdyFs=RkvK!{Sfx#x>xX+DZcQIIZP5;OS1)5ToXlTTDzvcNS%@y3T5Z
z%p;Wb1a^+|a8g-|_U%U4&zSx4K9g@AM&o%S!08(vaozV`m10UdGCbdXp3^-gT!J-T
zh>9|n6PX>2fUB<>R6l<#JOzP6N-9M>828I6M@?TIA`6>ct|x$aDk|kFX_i}`)bYmh
z@wQ+;wz^v;l-PH*RvuL6jlHJXsqx4&^|^9$CRyunf*hH@-1hRB0l7s&<^!;gx0&Yv
z@4U|~(M&_%REhB$i;E3s*zH2qb;pu7h)qmHd#gJ$6q&TihNVl4SPn-lYd^}L+}5wr
zMM9JxQSkaju52~hIz8!R>48;ip#vxU`4zQtR4sD5R>h*kZq=71)E|?x<-2MwpRDEQ
zum7ui)v7FFBd7^d8)oh~u`k-SyqsVscs(niI&x;QL0otiXG~cZe3P_lQfBWuCFfx!
zm;E7p)pF@YC*$hvK@RApnSIf2HBUw9SESZD(?z#b)0J`hPqjQ<D0eT&qRaS;7T%14
zha>u3&Rt@P)v88)SW9*zf(kMr{o@WuD48Z_%qylsHgiDONw9tK4vQD1|AGfR<C5cR
zA1&WHXsx2XCw7~K@74Ztn-@G4K6m`|%k8HLgY9U`Y(K0Du2b16py5cQs<3tYB=61*
zi~I-9FoSX=*I2LTG@XUQnz-cq_0|tin}dF63b>&oKJU|UAP%aag$=-@SyIcysG?@R
zk>237d2!Wi)?Nct6=*;0Bsfk)JIk5(<{(m(EAmxT^PbSXbvmC@{@P0$2+QD_pG!X}
zCH}s$7CtIRHHA+E;4`Z(g=VhA2VT;VM6H|e%-r?G1%Isi{oXu#WNlvA=b8^YV>@XR
zo>k4{qgZ~^^+U73a;;3dG7u+55ZI`4X2&cv0Y!W+b%t!{2s~@r;y{no3k!K#C@d82
z<i}p>m2Yft9rFggpSJ7^_T+i{67OF(b6{Jk^qT?eXq2p_yxeZ06AMNab%;l9=S2ef
zqSmc`zK%QV9kM6(ksAGVJlAn{?>kWM%>$viu|%8j7~)!jI)4cEU6@6B*7OlYreuaH
z-V^<*F|iJcQM%HlLAvK#n$Ql>k}53KiXjIzNMe^JTCh3Zl^%m_-|*TgG<hC|KO`^)
zt*&B(hYmKDq!9OW;V}bPFRng&lavyeMOeH(l50bYC<FtdeTPwP;qD;87QzM2)N#CS
z$qj*($Gpov>C^dvYi@#K#n>gnl5257SIc$wbkAlYsykX`y|XQt&4Of0>A%_!Yg*G;
zrnf%$YOPF=`uxZeLw_H56th-xMW(N&PG#&}Oz1qz7{i*~X=#0NOx-Vj1vlo^FDwO!
zu6yioyp5WUeb9hwZ=tedq;{RUcaEgzo07W@W^iuy7`pP0@Kg!mxiqm@Lz?@<8>Xq1
zt>w~5dlJ}U7ap9aBpax5;ZE9sxBzKm!+Iv3*89>?=~A-@PdWI7ux}q9S-KWeU#NT$
zo$h2Wr!<ZJ15l*GIW*p?exjGcOo!=eeS=O0n`%jH5J`PJ3d|nZPdj-0R-|QpsI1vM
z4pMp|LodQvi||0>;UL1w=O)CewW?T|=(=tBo_HQxrN!?;`Q{`~`)@vXJC&5{W;I~T
zy5zEW{IpJ6@<=Et@)&LK4!R}O3!dfts<np-<ZCM{Fq?e&MByE`!g8@pae%ZOiC@op
z*S9NQyPGV#)u>@7K`=1UhUY~{#Ks}IxzYn-01hc6n7m{>2)cYLleH9GHTp}w6#32T
z<4w0M-gcG2G|XpN=c2hvwx%VM)P;=@VKmm`!bpy9FGZ&`WJC|O(kS{Ya3^M9yv^a&
zG*0x<-C*$-<RHEI_06EL-M#L>>B+1oh&h+H&U$yYD|x7!530K#?GUSR7lFFcfeRgq
z$*!!>R29PeNb4|lzf(r6t9;+IyP4mBt(dQpt%cf}*@5uECmgz*-nt32taLTHZ@RL!
zr@FBBxa5Vvc(=mD=JV#H(U*IJ`pA&ECwx$4PZ<xRIZJMYqC%BxFq549cp8ju$2{Zo
zHlbEA{qb9qm<69Lw%<AOL*dq-#kU+i<)5~*VrpuMZj8uHgmIae%9zE`Y^iha!x>!D
zMh@j_=!u5jb#a#tgwGWX^fyK2X!}0^&8)1wb?oXd`Gq)EHHbV@$Nibe>WL96);s>y
zFvNKAQTu{~=k*`{JYb{UgYwXM;Eu9K%I({IhcCLsLNEN%9wxBhTx!K1+cR!#1gRqf
z+`=wh^8Zxnz0ZgC;v4!q{W#j9TOwlAi{0CLd5O?bV;_DrYjCPeSVKIlk3X7w5HZ_B
z-8Q{&(Cn%)Xw??*a%e&9a!a?lmF15Fh6dTY28Z`5@V*#y-geOYUn+?39Pn>F4A-jl
zoB#|w0A16@>$JR@-MWw3P%w8e!@{HCB_>z~K7vs2DpyWDlM$)1wW6J<nR|<7Y2PS3
z#%;F1Yw&_-T;4a(mJXutQTZ&gzmsn~XrairS}T|B4p;^po&#OZF3)gy=w%r`%ee}x
zpI+8Vr3ejv!uZP99w&YA*?{g_SXC*$Cz=vi<hoT@;bbkkN}hA-JT;du8)W{p<f~Uc
zTyRX?MQKj^<87jT+%)xwh1|R7pXQV0*dJ8z4m9Gg7}N?0>1y+>`X0!6Jy^f9m-*F&
z4|P80RUe`?My)oW*8O@6L-<Slw<ZNb@z%ldlb;m;u<)b41^)gF_6JOf90#2Hr_Emu
zNE=w-c!YmP>$cVgK{&1Z(*|X<G4iA7u5#c9R|DgC^EtUBs*`bC+3Og}K=b{J+d%7s
z`FrPATdcEH^JTS9JvW$V<2-rGKKRzA71{|!`7WthiIGW*$w}XsO7w*KWR04wUzHDv
z54T;FeIxt0SLXs_Qq2Ba+9M0C5pQtv7oj&vhkB16d`VBRkZ%YT;~L7?Fw-Xwua)um
z+?8bM>c4z1EifbuB}_0B&mY;p-{5@IiAb`_h;r?XFtPU0vOEy&>UyqZ^T2ZJ<qdz%
zuodyqg4*9x*snpZWygJWB(m_gL7R(`#dZDtn)-nuoJ)bE!^^iY<J=$pSFt~77Qmg8
zgg6z!=U+Czo<h{Lw)Jw9u8(dS@@T&ETDPhn6l+`wstK<79c8E~9C!H1T`CKQjSLH@
zMAke(v+E+(aMtZrzqs=zb9$5X<sIx<lIi?ZnU4IB#r~cKVz%EtU050LB~E*Fr-|OJ
zTihpa$wpsqZ9qyagF8!tCAJ%h^1MH?X%qJ36hYc2Ut9#PjmI_u!}bsWC(zp2+oLMy
z`+35Lpz`+1+*M_=4N9V8875$GT<P%(qGvuw{XsIf{w&cFR`&Zjye<)F85JwL>XsUy
zrb@^OUiOAS<KnJTb%oAfZ8-NPW|ha6?DTwtSe?O#k@~_o5%|Mdqr0EH2HkYt6w?M`
zHx<mzMtEc;&=^mf0QVd<##5857mD>$d-GINS*}W$x|(b(6gkv*@14hHe@$P7F23|6
z?M%lXyadI+=vPAPS@0S}#V;X@yBn0DaqYXMbhfv0*^-*k4oj~D((KIC!1b}u#F%u5
zNnnCGGTfeV7X|j3DY+|T4?lvE*R<~WP-@!E_2f@vUGzD;ZM0w4T~VY7OPYJtG|shv
zp=Sx&Fq_i-8K}C?BF+$Pi=At9{Y2=Q)P*!xr|746me}>g`pDy5J6(kw@z(yn62>a=
zPZK<_`FgQ(GZd>L^P<OW*nDKK>-T1Ya{BuG&*SHXZx+kAt6KA^XXPGv-am^4UMpd0
z)h&48PiAI)XnmZW75n?DU+3h$aDM|0BdwX9IMT*aaHUbS0oPFuwd!j?bNz<Mq!L$s
z_7jx?LtrelI)Tj`rb;NPu&4@{HeB`J%TOhBZ-A6iUE&8P+!_VOBWjSPOVKo-;fG&7
zyb3fi*O*Vj26cj2<Rfq90vnoDp1k+dk%P$Oou+N{@HVKz!F6#RyN(SJdb-6aBl!w*
z{x@Km^A3sNRhai@nm)8(olH8v*IfHXeWv{>v#HHQ!${6^+7t3Ft~4@aP$YL%mFJRf
z)kwrOAO3ZReIV;L#p;LI825eEH0$MisXJkogGC*smDq-;9JG73|HLBakYIH2aSP05
zk>w>rw<SfSOv5}{m9RRG=nwF(aUYGm_LOS0#Nic7yJ$d7VeIV`RsH(SDV~N5?h+JN
z)e>EWn4BHZ0Qw4^j91EPiQA_yTK(jKeHnaZg~cGVja)tj#YJ+0t~@LPZRQzW>YXt?
z^<p*_5;xv}53IRfWCxbr?#izBE%JfKV-u(KS6(+J7jA3*g!H*WAv8kyc!$j|q{N<Z
z>!4M<OCU$3_Ehd*hV+cfp6cp4th#>MYP#@bymuA;B7sp~97Tc0XwjPYfcB1JLUs=P
z;d4t7{{1d8<ObQvkPV}j=hatg%jcVbuip~`I63Q8VglC!FJuQ05@Ak7v%yol!S@XB
zMp~5k@>=5}jau3!Hv<g`1eMDhSK_qsEzK6ehSiFLX!tY)`fgC{j>*y!-VQ5i2L>LY
zg+g>e<F(~l!oXu9yIq~<u(l8{;oa@~ibUhF5*7v8MQp!(ocl*o!%5qF!^SJoE9r++
zwuTq{%X#-)a!9<oj+z-B?}+z|V!_YK1o0aMa<vuc;46>W!{Itn))&xEo7<lF(KqZH
zun}bi<rHj9c<!z_+>)<U5{5t1n1n2oIJ_0B*5fqy;Cr9%$H|jgL8Ctus@io7Az>Y;
z-i~p8FYHHB4OVlqF8-E9_RZdECp5)zKIR*UKd#Aa&s^_@o;yd*o|UrHi;c?yX<Lu$
zRYR`|rqs^H=uMVsN}<GlLd1)Mo|Ewjg<0w{!2Sv1oz2u<dAp~=EIKVk5~C?T)Hi1}
z))IW3i@SoBmK~F@Ij`|7jFkEY<(=&EJ)rh*O|_^p!DgY9T6P+#yg}NMU^T%im#+Yy
z#O@(`<YhZ7?b4u1OL^Zc`P2H`f#3M}VC)NS(W~>n16gFJK8@CLr!*cfx3-%76!3^I
z%&PHs(jym|J4C=u`R;j~dRV)n<n{781lqA+k0Pr-0F-~oobL(oLAkv#Z`Qo#6h%!l
zKE#?H)z`h}*nw#aFhAf9n6*qFw2Ec%Zjh%5w_K{XQK*h2*-HeN%p4TlD5+~xrh&W9
z@V|?B_te!0b|c%lT;7;8fj#a9EWGx4S)m+x&wrel;ZfQH$hfv&#peT3&_geP-afAq
zYQ+kkn)MhmoI;Ymgifhx?4(!xc=Sxuo`2%vJ@`J;5Z!8Ep|q7nwBdoVj9KGV23ZTU
z$@)HyXI{TFf%-sFQE~XfBc{ypNUHaObw2sAhN}Y+CO1O8naPeR+YPH0(yAuP6kCVw
zY2EuZw*@u0dsJ{>en8(yknO$bG_9wFjk>YW+3f6qH~QjkC%mJBQm?x>Zf%%TxaQ4X
zb5~6bv(S<J5X+?l0rAIG;OstyQV@l6h{Pn{Xz{cv_o3@2byalSxbcE+MRiMHJwlYz
z`$*o>X|Cn=h9y)Y{QTD$^us6bCWyhRTyo%^#S$FPU5KwDQ8C*GEP^^9clWn{$g^<f
z<Q<hax-Y1=?^ZhU;86^zc|hH`Mw=?3Fy@xfPQ&n`&HU9kV7<a0K+8Qw4GR9~xFW9R
z!6fc;^fFst!Sxlw#bJ{iyxHmO^FIJiksn|0&Ee_qPcetiUX7DKp})6@*PNQUH&=W*
z6?HN$iBM!d?TibJKx#x^J_OI}pH9529iTCbZ=PE!v$tt{Z8i?$CaR+EU#70O)66@n
zRo&GW(_7dEmn*xg#Sg-!))^{6N&#h@(Ix9_GuI5Xr{fx7pB7dp^AX_{n}rrevCx4=
zR8JHh`C+2wf(Pj*V;~MY(zhFvd$Vd~s%-B$N^8|z#(xtDmVn0_%{4bGYa!4dQ!20i
zh!?AxyF0qTXIY>+VA4FsufHK=oaQ+Qfv#eAvs6Xl5B0lERtM3bQ#nvovTLjyn41Sb
z&+fy|>T6Nr*gcE$C>PVTJ)#V<|MG81RyV@*s+ihUAw#)ursee9xCT8H8}ELwJMr_9
zL8`!mpHRKN;I%#f5J_bt`eec5(`zy?i4YO9&%xij2G)R<b<wGDClzwFm2r~VRa3>V
zOncXNq}peb!p(dxJ%ULwwWYFQsbNqdqekptOInf1s+G5o0`9JOl8)a}gG@}2-|2C6
z;2(eqqt>P<Hl)%%1iiAoQ@cA+F}L#ldUk(>c~ey%PvbR^a}V}F(BEVxQywmFs!e*e
zTz`|j$fpc5@GuvJVq{DUG`(3H6G#B`loy*|;SAvZ0rcs(6#U@0QX{=^4?PV|zWe>{
zhS3k39~B+00<MNM@crY8zB3*9fuezOY9GQ`(=M^FRy}~3*BbvWtW95Tr`KWsO6{fn
zx<2%*31YpIe-W%`8}B&pDf_(u2X478^<Vo*$wiL-0TfG;+o{J+RZLG{VK3vqjXu8l
zgcrK&Sfl1K^}3*GT3B402mN&oGgMc$U1U9{D~HS&lYId}z3GDrT6NWez*iPnS()Au
zGpkiDh$p5zh~#PDavbVFPu?7ZI$S^dioT7lzq5YWY65;_@8NzVICaWt_D%Hqca-{(
ziFcEr;r)YR?7|{A|K4nABk?+B8m<q9&xx2ic#gmAyt1+d1t&4zC|6ncyc&gtFKd-N
zNJ%c5YA)Kb6TR$g_gO8*K17-~{9<<d@qi7mV9LPBmh01k7=nf)@dz`8WU8cEjp-R#
zeA1Kn)PUU1H8>2Weg-6Z4f^^iR=POAr-L(?VlRx>F=rVWT8brncxCbQ)t80ln_Dhk
zUcAOy&#OaW&AVT`ay+1pB<dOo%f{$~`#Xue?YD38=y>&Z*(pNeF(;-HGEcn*^h!Tp
zoCHOQW0K;5#_v69BZ94zENeB7D#sPGRjz81{50rXyYGbZ?hV{9-ThSEr`NKV$V*9{
z%4ws3h}v(xO?8G%afj-_f<9wYQ=A@GGy_q^A`Eh9mxnhszRk6&)q%YtYwRGiyd0A$
zxNaaKSXAS(0}hal<Gjfs#VVF&x155+q!yT?xG(H(leP(02Q(JiQF_pi1;`~ZPvwBH
zsPfoU%7n+vTU5~?E$kK2bs4#$BHq8)<oy_2Ek4?&+Uk8YbQs~U(77r_<ss0cW}5h!
z?i~n&4h0(UfZ_rL7|!3|+?6@X`eC~0T;zL~=mO1cD{>}!NPP})TE^?>A-!V4UG+aG
zqWsUkCb$Pe`5Rsdb%{I9J$}VL_Y}@H=MVjCJc#&uSe1Wdtb{4Jo~#z>76dbQ^RHdN
zS65Yh=HLfr^>%L4XeFFeJc}RWp?Up`(%%zx7QpqmCflZpX|tKPfk~KxhboMNALj*$
zYCh7$M6rRcV!0fG7X_QRcm5nmt|LmelFiym0QJ%;C@05wM$?{$^JS>}NI^B0Ly@2$
zKE4DKX(zEE6fo%IaY^P;zMO^9rZ(i+nxI!B>0c=%+j37_zIY7Lt0Bf42jx{EFx{Ct
zX!ceXpPdPu;ogb)9wbwKIgUz!=n?dInIMa9h;;bIH1|4HBf+pS!}Ed{)NuKA{YjG0
z<M1SSwDA3^morWEYt?vT`qHSr$8K%N^+R>324wF<r6_wDJk`6Lv(k{11s;r4P0h1X
zTUI(yDBjCHIatRJ@#Xgw1NX~=mQR06W)6G?z2f4!o`m(%ATZXMW@nP!+$S_*eaqY0
z97`N;;DM<-@dU|!%Yp9?*Nq;F@~p0Fffh8T8Sb9Yt%anK{0-h^%nAB5wZ=|ORM1AQ
zg$|16H3&;VRg@G042%t^k2KaTGlv8h#`XOIDembNuGeGkHI-2l(j7*>N91Piexcz$
zs3I+^t>>YacR-{X#F|vo*^D7p;ZUk_DmNjQjk^c0S$M3jYda(UO?Y!!wG>-#th}~v
zkS@x~&1uu`%;E?8L6!Q|u{qd4uMY^RA}=HvvmjsD4vJ$a#Pidp9pW!L8hk?N#pS(|
zYQJ$7gwGtL?eL+R_C(_roc{pcVt0==i%$<Zoi@O+d}DBInSG$hubVH*BFjQQSo|Kt
ze!)9R%dxmjPi0JpVKH)Q%gtkd03RM5@<B4e*^yPh@9FdL*)`%fe2wE5(qB~V#GTD1
z9z<I#b5d-oC-mH>+Oo26<WAHO>lF!RMD8!rPW4iGqeJBkLfi(Xt#o5ytq-au<ZohX
z<pCeynuj<xu3Mc`)F{!4Ndy}iz1cu$wQ{l|G;=eed>n3->+`l`I?fd3P*NYb4>vii
zTjfPmNpy`PBCE~$O$Y0>a!kW2tE4xrn~7RytA;F6A}f>!2qr#SElF|T`<QNR8JIKq
zRE8Z_)}etG9pr1lXa@bReMJAz#Q=_lBy(gefol%E^4-5RJzV0J7>YANc3$At!Bz1U
zbEuL2*&3CqPSg&>J}LjucV%xzp32akEK^lfJHmW7Mb=O7nb!1{j?*$KgPJ@joaeYG
zCS%1KcFMRO_$%6@Qn8`#WSaZ&;#_+Qc>71nm!n64!G)kte*o$IGRLqi7*B6ip!aJS
zX}(QlJBDJgyLFK1XBj*%Fttt)FVTE%*j1}RPhC#5ZYK$4MJI({wT&URr1%8=4pK`Q
z{%NUtH4WvUJI?HyCH8S-vz$w%O1w8E#m}iuU+nG`D6U!EeAf^1L=-3>{vgn?dG4Ew
z@2x&JB4TcG30IPc;nY}=AoYKB|Cye$n__l3&_E=XQru(PCLm+(#|sl#s+{(c3n(UE
zi2Fp;L8JNuoj50FoLifwwYXk3#<hzz7Y(QIei*rbRl@sc@N2+hyJUWOr)@H!q+?it
zD5~?Ia5l$66IUB+{J@Q<+qX*=yZ=K5<V$RY<)b9B1_HV2ftb(Qd+YOAae+rFLsxk!
zZyno6;j{#X#W*o2d&QdSG6}XF$f9FUntyrm%a{PyrsBc1HRBdVwp`gKUJ&yDy!rK6
zh&2VHr`sUPG1M3Hws<B$o@Y=jo24}SimMA*V`q5ZO6f=zosvVY!UR%Sissyl=31E)
z{ccs2RFr}GDkleYjBxg*uN>caTzorkZL;VJw+J>Xw`)I}!C7a{mk2l)e|!1_IdNF6
zG@`;R8vCY3w#%WM%-{BnlX;ir{Y4%jWjg-#RVQ<Gl)uNz6|0C0AI)6`OE8C{IqM_-
zb+-@qn0$sVSxRQP{O}zhYwZ81@z{3z@cZ(IyAF>VGwyr#%M1h%fWY{&5*=1iOn^;9
z$Hlcdv0E6^iTj;ZPxJdY*9CaDFNA&h^c;C<)~d$bLL|St#BB{`@%qN}rLaR&yISlb
zb{oh|#3vMV&T~K}LzcHUrb-N1=Z#{znzCt0qn1I0T#q|HcTljD@w}W{CACnb@&{Dv
zIO;(xiCO+uRlrsSh1<fo$gezO3sMTNhAkIimv<)G(G*Y_bN1Pjmh&F0$i-N0#agq|
zrvOhQAf|8y?`)CSjnz$lo_;sIyr-<r>0X47;7!>$)Lp2`{Y4u+#RaVFu$f&85+8y%
z`+P<mMsQltmjYi`AQ~qKl??>3i7;|mCP>tRxJw$k+y@6vRpOydMtQd$I^C_hv+gM#
z)mA5Q<FRAAz^+tZxLG3?<Hr*UKB~NMd`PKX!2Wwc`)IujjJuHGyfwOfT-H93ea5>=
zZp`Yz;x$;a!Nr8kkMj=YL|4-H1zcrIPpn-%*Y($q?I?(1(&7_EuxE#tZXyO=UmKQV
z!V^dyLWo?E5mR3-m=f1euS}M3Jns|tKHhfKs`RBdkA`@JAxr^}7e4yfht<W$@5-%0
z7hEc;aEfgchtO0Jal_s+@$|ZgYVnWQbJLPR{Z5M26dof;@eANt$1E<s62Eq5D(~Kx
zKY(*<rz&k%5oe2bqJ9x<<5h68L6X}w>f1p`iBA?UH_P0IK;!tYO0Kn$@3{u)xY@d-
z#OW6TFhUYdjpSw^cecufUvhN41KD0S^bf-zH@;T_O_jTO7)C2KZZ3ilXK(XC8$pkX
zzX#vslSOtd+Vr1_Mbw(e2pkCbIW|#b#&a3P3-lj8fj~_L(OQM2$9F5U3gb(SUdY4I
zA9|G>3E0omvgC#tCA;LA;rBj5yFf$N<2Z~KB6eK3?zr;cVxMgUPE+29za4JoE}GsT
z*fo-mlASJ1wqso67jwJyqvHAtFRi%Cwi(N~4(I6sr%rNN%3Tfd%eX6o&6)HIm6H9W
zI_khtjrwCP5`lN2)OwuOReN;LaDg|S=UKsk<BTT5LRqp+dWHNPsMt=VmNp+`f)(ub
zfLrUYp{Ms3L6pm``hLy*mKIzX)Y}Ofdx3+I0pv4%py`wf-iobaW)Lr6@0AgFE8#Y-
z3EQ76mk_DTyLmb&Y{`awx1;y6vn;ES+gX1L|3o237aBFRAk{kYes|jK1}Beo{T99d
zo4#&Jf*o$j5}XK|3H^yHYuZM4Ucc`-iyn{jZ)d%P=XtbP9NXYw>KM^3V>Pl3nTlN$
zEVps4ak~zMkL7wRD&F?6Q`F(Zn-k*I6Zz~~ILAotAvWv!{)p)NOhb4lQPzxRKb@@D
z#~^PgKB5}arC02Qor|=l1A}@?Gt_kXN)$cfHBj|>t(<p*M>N-8E|D5Q=Q4BqLwOsS
z`x>iJA43FKJ_SMU;#KfwPd{{66GSb4R;f5o5kKa7=<lVHO=KQZ2O8FK)~K3(C8%(Z
zYdlY;?-G3#POgj=Gcun{D;noAr5P<T+k+J0q{fu4Y1dxZmsKMe@P?V(zFvL~Co}ii
zzg)HRng4)maE@y!zMN+{%yZ(?$<+qYQ~eS6TPH@Aha(#UG%~!G3^=xg7@PXY1rD#y
zFD4*yw6-2bbQKU}7;R#_j2`1<1)Q`br}>+*n82*o&qf^tj?~5Ls}CzM`@FYSpov_o
z#_AX|FBsQ^JF52~hfi^|$E}ya^-Gq#je?|062My%MeeN=J2}vx#JI^DoYj3q#~JV*
zdtO@ML|H?zga@B^`l?xn7|Bk_rAgsoEW<Cc@!jO+!E>STJp(q>WdTVFV%SwkE~$<V
z$$jt(*N~=KzUOgC=aJgB_y@1h53a!#Y69!71L?o$f_VZ;qjK||tkc}y%`v>$Rt=+z
zrRh&fmC;!X?ZF2WSbI^(?J+=P(K7JPPrf>f%(Lg8@;T(TrRZnwp&*Km52i%VQIFeP
zq58Lh>Zqaw`u%p5s=dDFtm%(xr4yZ&=Q*O&8`dpPwGuU)g9YYR0KK<J@i5ZP`6Yk*
z<FKpni7~fieO4o@AGO=NgZ}27iNWj7NNtM06Yd~VZu}CBK3iBIgDA0#|AN52AH&Ft
SGRwe%hBAK|nl%5J{C@z}-g;jE

literal 0
HcmV?d00001

diff --git a/README.md b/README.md
index e481c00..af97bc7 100644
--- a/README.md
+++ b/README.md
@@ -1,14 +1,17 @@
 # Deepfakes Detection
 
-<p align="center">
-  <img src="https://raw.githubusercontent.com/HanMoonSub/DeepGuard/main/Images/deepfake2.png" alt="DeepGuard Banner" width="800" height="400">
-</p>
-
 <p align="center">
   <img src="https://img.shields.io/github/license/HanMoonSub/DeepGuard?style=flat-square&color=555555&logo=github&logoColor=white" alt="License">
   <img src="https://img.shields.io/github/stars/HanMoonSub/DeepGuard?style=flat-square&color=FFD700&logo=github&logoColor=white" alt="Stars">
+  <img src="https://img.shields.io/github/downloads/HanMoonSub/DeepGuard/total?style=flat-square&color=brightgreen&logo=github&logoColor=white" alt="Downloads">
+  <img src="https://img.shields.io/github/last-commit/HanMoonSub/DeepGuard?style=flat-square&color=lightgrey&logo=github&logoColor=white" alt="Last Commit">
   <img src="https://img.shields.io/badge/Status-Active-brightgreen?style=flat-square" alt="Status">
   <img src="https://img.shields.io/badge/Release-v0.2.0-orange?style=flat-square&logo=github&logoColor=white" alt="Release">
+  <img src="https://img.shields.io/github/repo-size/HanMoonSub/DeepGuard?style=flat-square&color=blueviolet" alt="Repo Size">
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/HanMoonSub/DeepGuard/main/Images/deepfake2.png" alt="DeepGuard Banner" width="800" height="400">
 </p>
 
 <p align="center">
@@ -43,6 +46,8 @@
 - [📈 Model Evaluation](#-model-evaluation) - Benchmarking results
 - [💻 Model Usage](#-model-usage) - How to integrate DeepGuard models into your own Python code or via timm
 - [🔮 Predict Image & Video](#-predict-image--video) - Simple Inference examples for detecting deepfakes in image and video
+- [🎨 DeepFake AI Explainability](#-deepfake-ai-explainability) - Visualizing model focus using Grad-CAM and attention maps
+- [📊 Metrics and Evaluation for DeepFake XAI](#-metrics-and-evaluation-for-deepfake-xai) - Quantitative assessment of explanation reliability
 - [📬 Authors](#-authors)
 - [📝 Reference](#-reference)
 - [⚖️ License](#-license)
@@ -303,6 +308,41 @@ print(f"Deepfake Probability: {result:.4f}")
 
 ```
 
+## 🎨 DeepFake AI Explainability
+
+Deepfake detection requires high reliability and interpretability. DeepGuard integrates a robust Explainable AI (XAI) Toolkit to visualize the decision-making process of our hybrid models
+
+⭐ Validated on hybrid CNN-ViT architectures, specifically `MS-EffViT` and `MS-EffGCViT`.  
+⭐ Dual-Branch Analysis: `Low-Level Branch`(focus on local forgery region), `High-Level Branch`(focus on global forgery region)
+
+### Deepfake Detection: XAI Methods by Multi-Scale Branch
+
+| Branch | Method | 🎯 What it does |
+| :--- | :--- | :--- |
+| ![](https://img.shields.io/badge/Low_level-blue?style=flat-square) | **HiResCAM** | Like GradCAM but element-wise multiply the activations with the gradients; provably guaranteed faithfulness for certain models |
+| ![](https://img.shields.io/badge/Low_level-blue?style=flat-square) | **GradCAMElementWise** | Like GradCAM but element-wise multiply the activations with the gradients then apply a ReLU operation before summing |
+| ![](https://img.shields.io/badge/Low_level-blue?style=flat-square) | **LayerCAM** | Spatially weight the activations by positive gradients. Works better especially in lower layers |
+| --- | --- | --- | --- |
+| ![](https://img.shields.io/badge/High_level-red?style=flat-square) | **EigenGradCAM** | Like EigenCAM but with class discrimination: First principle component of Activations*Grad. Looks like GradCAM, but cleaner |
+| ![](https://img.shields.io/badge/High_level-red?style=flat-square) | **GradCAM++** |  Like GradCAM but uses second order gradients |
+| ![](https://img.shields.io/badge/High_level-red?style=flat-square) | **XGradCAM** | Like GradCAM but scale the gradients by the normalized activations |
+
+### Multi Scale Efficient Vision Transformer
+
+| Model | Branch-Level | Image | HiresCam | GradCamElementwise | LayerCam |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **⚡ ms-eff-vit-b0** | ![](https://img.shields.io/badge/Low_level_Branch-blue?style=flat-square) | <img src="Images/deepfake-xai/서양인_가짜_1.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_low_hirescam.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_low_gradcamelementwise.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_low_layercam.JPG" width="100"> |
+| **⚡ ms-eff-vit-b5** | ![](https://img.shields.io/badge/Low_level_Branch-blue?style=flat-square) | <img src="Images/deepfake-xai/서양인_가짜_1.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_low_hirescam.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_low_gradcamelementwise.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_low_layercam.JPG" width="100"> |
+
+| Model | Branch-Level | Image | EigenGradCam | GradCamPlusPlus | XGradCam |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **⚡ ms-eff-vit-b0** | ![](https://img.shields.io/badge/High_level_Branch-red?style=flat-square) | <img src="Images/deepfake-xai/서양인_가짜_1.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_high_eigengradcam.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_high_gradcamplusplus.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b0_high_xgradcam.JPG" width="100"> |
+| **🔥 ms-eff-vit-b5** | ![](https://img.shields.io/badge/High_level_Branch-red?style=flat-square) | <img src="Images/deepfake-xai/서양인_가짜_1.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_high_eigengradcam.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_high_gradcamplusplus.JPG" width="100"> | <img src="Images/deepfake-xai/ms_eff_vit_b5_high_xgradcam.JPG" width="100"> |
+
+
+
+## 📊 Metrics and Evaluation for DeepFake XAI 
+
 
 ## 📬 Authors
 
@@ -323,6 +363,7 @@ _**This project was developed as a Senior Graduation Project by the Department o
 6. [`deepfake-detection-project-v4`](https://github.com/ameencaslam/deepfake-detection-project-v4) - _Multiple Deep Learning Models by Ameen Caslam_
 7. [`Awesome-Deepfake-Detection`](https://github.com/Daisy-Zhang/Awesome-Deepfakes-Detection
 ) - _A curated list of tools, papers and code by Daisy Zhang_
+8. [`Pytorch-Grad-Cam`](https://github.com/jacobgil/pytorch-grad-cam) - _Advanced Visual Explanations for PyTorch Models_
 
 ## ⚖️ License 
 

From 4942477c8a4cce3824dadce7e94617c621104ef6 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Sat, 16 May 2026 18:54:07 +0900
Subject: [PATCH 36/48] =?UTF-8?q?=F0=9F=A7=B9=20Slim=20cache=20key=20to=20?=
 =?UTF-8?q?(model,=20dataset,=20type,=20branch)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/explain_svc.py                | 17 +++---
 explainability/explainer/base_explainer.py |  2 -
 explainability/explainer/cam_explainer.py  | 61 +++++++++++-----------
 3 files changed, 37 insertions(+), 43 deletions(-)

diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
index 6bd0e95..07118a5 100644
--- a/App/services/explain_svc.py
+++ b/App/services/explain_svc.py
@@ -26,29 +26,26 @@
 _explainer_cache: dict = {}
 
 def _get_or_create_explainer(model_name: str, dataset: str, explain_req: ExplainRequest):
-    cache_key = (model_name, dataset,
-                 explain_req.explainer_type, explain_req.branch_level,
-                 explain_req.category, explain_req.aug_smooth, explain_req.eigen_smooth)
+    cache_key = (model_name, dataset, explain_req.explainer_type, explain_req.branch_level)
     
     if cache_key not in _explainer_cache:
         _explainer_cache[cache_key] = EXPLAINER_REGISTRY[explain_req.explainer_type](
             model_name = model_name,
             dataset = dataset,
             branch_level = explain_req.branch_level,
-            category = explain_req.category,
-            aug_smooth = explain_req.aug_smooth,
-            eigen_smooth = explain_req.eigen_smooth,
         )
     return _explainer_cache[cache_key]
         
 def _run_visualization(explainer, image_path: str, explain_req: ExplainRequest) -> bytes:
     if explain_req.display_type == "heatmap":
-        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio, threshold=explain_req.threshold)
+        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio, threshold=explain_req.threshold,
+                                                 category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth)
     elif explain_req.display_type == "contour":
-        img = explainer.display_contour_on_image(image_path, threshold=explain_req.threshold)
+        img = explainer.display_contour_on_image(image_path, threshold=explain_req.threshold,
+                                                 category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth)
     else:  
-        img = explainer.display_bbox_on_image(image_path, threshold=explain_req.threshold)
- 
+        img = explainer.display_bbox_on_image(image_path, threshold=explain_req.threshold,
+                                              category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth) 
     img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
     _, buf = cv2.imencode(".png", img_bgr)
     return buf.tobytes()
diff --git a/explainability/explainer/base_explainer.py b/explainability/explainer/base_explainer.py
index 6054772..2897a1e 100644
--- a/explainability/explainer/base_explainer.py
+++ b/explainability/explainer/base_explainer.py
@@ -18,7 +18,6 @@ def __init__(
         dataset: str,  # celeb_df_v2 | ff++ | both
         margin_ratio: float = 0.2,
         conf_thres: float = 0.5,
-        category: int = 1, # 0: real | 1: fake
         branch_level: str = "both", # low | high | both
         l_stage_idx: int = -1, # gcvit low branch stage index (0~3)
         block_idx: int = -1 # vit, gcvit block index 
@@ -28,7 +27,6 @@ def __init__(
         self.device = "cuda:0" if torch.cuda.is_available() else 'cpu'
         self.margin_ratio = margin_ratio
         self.conf_thres = conf_thres
-        self.category = category
         self.model_name = model_name
         self.dataset = dataset
         self.branch_level = branch_level
diff --git a/explainability/explainer/cam_explainer.py b/explainability/explainer/cam_explainer.py
index 36ea75b..27e6f58 100644
--- a/explainability/explainer/cam_explainer.py
+++ b/explainability/explainer/cam_explainer.py
@@ -33,35 +33,36 @@ def _draw_label(img: np.ndarray, text: str, x: int, y: int, prob_ratio: float =
 
 
 class CAMExplainer(BaseExplainer, ABC):
-    def __init__(self, 
-                 aug_smooth: bool = False,
-                 eigen_smooth: bool = False,
-                 **kwargs):
+    def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        self.aug_smooth = aug_smooth
-        self.eigen_smooth = eigen_smooth
         self.cam = self._build_cam()
     
     @abstractmethod
     def _build_cam(self):
         ...
 
-    def _build_grayscale_cam(self, img_path: str):
+    def _build_grayscale_cam(self, img_path: str,
+                             category: int = 1,
+                             aug_smooth: bool = False,
+                             eigen_smooth: bool = False):
         face, tensor = self._get_transform(img_path)
         tensor = tensor.to(self.device)
-        targets = [BinaryClassifierOutputTarget(self.category)]
+        targets = [BinaryClassifierOutputTarget(category)]
         
         grayscale_cam = self.cam(
             input_tensor=tensor,
             targets=targets,
-            aug_smooth=self.aug_smooth,
-            eigen_smooth=self.eigen_smooth,
+            aug_smooth=aug_smooth,
+            eigen_smooth=eigen_smooth,
         )
         return grayscale_cam[0], tensor, face
 
-    def _prepare_cam(self, img_path: str):
+    def _prepare_cam(self, img_path: str,
+                     category: int = 1,
+                     aug_smooth: bool = False,
+                     eigen_smooth: bool = False):
         """Build grayscale CAM and recover it to original face resolution."""
-        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path)
+        grayscale_cam, tensor, face = self._build_grayscale_cam(img_path, category, aug_smooth, eigen_smooth)
         cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
         return cam_recovered, face
     
@@ -75,31 +76,29 @@ def _get_binary_mask(self, cam: np.ndarray, threshold: float | str):
             )
         return binary_mask
 
-    def display_heatmap_on_image(self, 
-                                 img_path: str, 
-                                 colormap: int = cv2.COLORMAP_JET, 
-                                 image_weight: float = 0.5,
-                                 threshold: float | str = "auto"
+    def display_heatmap_on_image(self, img_path: str, colormap: int = cv2.COLORMAP_JET, 
+                                 image_weight: float = 0.5, threshold: float | str = "auto",
+                                 category: int = 1,
+                                 aug_smooth: bool = False, eigen_smooth: bool = False,
                                  ):
-        cam_recovered, face = self._prepare_cam(img_path)
+        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
         
         binary_mask = self._get_binary_mask(cam_recovered, threshold)
         cam_recovered = np.where(binary_mask > 0, cam_recovered, 0.0)
                 
-        heatmap = show_cam_on_image(
+        # (H,W,3), range(0~255), np.uint8
+        return show_cam_on_image(
             np.float32(face) / 255.0,
             cam_recovered,
             use_rgb=True,
             colormap=colormap,
             image_weight=image_weight,
-        )
-        return heatmap  # (H,W,3), range(0~255), np.uint8
+        )  
     
-    def display_contour_on_image(self, 
-                                 img_path: str,
-                                 threshold: float | str = 'auto',
-                                 thickness: int = 2):
-        cam_recovered, face = self._prepare_cam(img_path)
+    def display_contour_on_image(self, img_path: str, threshold: float | str = 'auto', thickness: int = 2,
+                                 category: int = 1,
+                                 aug_smooth: bool = False, eigen_smooth: bool = False):
+        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
         
         binary_mask = self._get_binary_mask(cam_recovered, threshold)
             
@@ -107,11 +106,10 @@ def display_contour_on_image(self,
         annotated_img = cv2.drawContours(face.copy(), contours, -1, (255, 0, 0), thickness)
         return annotated_img
     
-    def display_bbox_on_image(self, 
-                              img_path: str,
-                              threshold: float | str = 'auto',
-                              thickness: int = 2):
-        cam_recovered, face = self._prepare_cam(img_path)
+    def display_bbox_on_image(self, img_path: str, threshold: float | str = 'auto', thickness: int = 2,
+                              category: int = 1,
+                              aug_smooth: bool = False, eigen_smooth: bool = False):
+        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
         
         binary_mask = self._get_binary_mask(cam_recovered, threshold)
         
@@ -127,6 +125,7 @@ def display_bbox_on_image(self,
             box_color = (r, g, 30)  # RGB
 
             cv2.rectangle(annotated_img, (x, y), (x + w, y + h), box_color, thickness)
+            
             _draw_label(
                 annotated_img,
                 text=f"{roi_prob * 100:.1f}%",

From 8043fc69e5b2ddafb5cf3ee235e99bef899f5dd5 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Tue, 19 May 2026 09:16:28 +0900
Subject: [PATCH 37/48] =?UTF-8?q?=E2=9C=A8=20Add=20validate=5Fexplainer=5F?=
 =?UTF-8?q?for=5Feigen=5Fsmooth?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/schemas/explain_schema.py | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
index b730a51..8dd1773 100644
--- a/App/schemas/explain_schema.py
+++ b/App/schemas/explain_schema.py
@@ -19,6 +19,8 @@
 
 _LOW_ALLOWED  = {"hirescam", "gradcamelementwise", "layercam"}
 _HIGH_ALLOWED = {"eigengradcam", "gradcamplusplus", "xgradcam"}
+_EIGEN_ALLOWED = {"gradcamelementwise", "layercam", "xgradcam"}
+
 
 class ExplainRequest(BaseModel):
     branch_level: Literal["low","high"] = Field("high", description="브랜치 레벨\nlow: 국소 위조 흔적 포착\nhigh: 전역적 위조 흔적 포착")
@@ -32,11 +34,21 @@ class ExplainRequest(BaseModel):
     
     @model_validator(mode="after")
     def validate_explainer_for_branch(self) -> "ExplainRequest":
-        allowed = {"low": _LOW_ALLOWED, "high": _HIGH_ALLOWED}
-        if self.explainer_type not in allowed[self.branch_level]:
+        branch_allowed = {"low": _LOW_ALLOWED, "high": _HIGH_ALLOWED}
+        if self.explainer_type not in branch_allowed[self.branch_level]:
             raise ValueError(
                 f"branch_level='{self.branch_level}'에서 허용된 기법: "
-                f"{sorted(allowed[self.branch_level])} "
+                f"{sorted(branch_allowed[self.branch_level])} "
+                f"(입력값: '{self.explainer_type}')"
+            )
+        return self
+
+    @model_validator(mode="after")
+    def validate_explainer_for_eigen_smooth(self) -> "ExplainRequest":
+        if self.eigen_smooth and self.explainer_type not in _EIGEN_ALLOWED:
+            raise ValueError(
+                f"eigen_smooth가 허용된 기법: "
+                f"{sorted(_EIGEN_ALLOWED)} "
                 f"(입력값: '{self.explainer_type}')"
             )
         return self
\ No newline at end of file

From cfee097eda3c398fb2e5cff02544c46e6916b667 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 09:46:22 +0900
Subject: [PATCH 38/48] =?UTF-8?q?=E2=9A=A1=EF=B8=8F=20Include=20explain=5F?=
 =?UTF-8?q?svc,=20image=5Fsvc,=20video=5Fsvc=20in=20celery=5Fapp?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/celery_app.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/App/celery_app.py b/App/celery_app.py
index df5fd33..5503e8b 100644
--- a/App/celery_app.py
+++ b/App/celery_app.py
@@ -15,7 +15,11 @@
     "deepguard",
     broker=REDIS_URL, # 메세지를 전달할 브로커
     backend=REDIS_URL, # 작업 결과를 저장할 백엔드
-    include=["services.inference_svc"] # 비동기 추론 작업 파일
+    # 비동기 추론 작업 파일
+    include=["services.inference_svc", # process_image_task, process_video_task 
+             "services.explain_svc", # process_explain_image_task
+             "services.image_svc", # cleanup_anonymous_image, cleanup_image_cam
+             "services.video_svc"] # cleanup_anonymous_video
 )
 
 # Celery Instance 상세 설정

From 667e93cf56d62a55a439a3a19e748e23f6967808 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 09:52:28 +0900
Subject: [PATCH 39/48] =?UTF-8?q?=E2=9C=A8=20Add=20get=5Fexplain=5Fimage?=
 =?UTF-8?q?=5Fresult?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py | 74 +++++++++++++++++++++++++++++++++++++++----
 1 file changed, 68 insertions(+), 6 deletions(-)

diff --git a/App/routes/explain.py b/App/routes/explain.py
index d15a71d..d6083d4 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -1,16 +1,20 @@
+import os
 import io
+from pydantic import Field
 from fastapi import APIRouter, status, Depends
 from fastapi.exceptions import HTTPException
-from utils.responses import PNGStreamingResponse
+from fastapi.responses import JSONResponse
 from sqlalchemy import Connection
 from db.database import context_get_conn
 from schemas.explain_schema import ExplainRequest
 from services import session_svc, explain_svc, image_svc
+from celery_app import celery_app
+from celery.result import AsyncResult
 
 router = APIRouter(prefix="/explain", tags=["explain"])
 
-@router.post("/image/{image_id}", status_code=status.HTTP_200_OK, 
-             response_class=PNGStreamingResponse , summary="딥페이크 이미지 위조 흔적 시각화(CAM)")
+@router.post("/image/{image_id}", status_code=status.HTTP_202_ACCEPTED, 
+             response_class=JSONResponse, summary="딥페이크 이미지 위조 흔적 시각화 비동기 접수")
 async def explain_image(
     image_id: int,
     explain_req: ExplainRequest,
@@ -18,14 +22,72 @@ async def explain_image(
     session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수 
 ):
     result = await image_svc.get_image_result(conn, image_id)
-    
+      
     if result.status != "SUCCESS":
         raise HTTPException(
             status_code = status.HTTP_400_BAD_REQUEST,
             detail = "이미지 위조 흔적 분석은 추론이 성공한 이미지만 가능합니다"
         )
+
+    image_path = "." + result.image_loc
+    
+    if not os.path.exists(image_path):
+        raise HTTPException(
+            status_code = status.HTTP_404_NOT_FOUND,
+            detail =  f"요청하신 이미지 파일을 찾을 수 없습니다. 삭제하였는지 다시 확인해주세요."
+        )
+        
+    if result.model_type == "pro" and explain_req.aug_smooth:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Pro 모델은 aug_smooth 기능을 지원하지 않습니다",
+        )
+        
+    # Celery Task 호출(Redis 브로커로 작업 전달)
+    task = explain_svc.process_explain_image_task.delay(
+                                user_email = session_user["email"],
+                                result_dict = result.model_dump(), # 직렬화 (Pydantic 불가)
+                                explain_req_dict = explain_req.model_dump()) # 직렬화 (Pydantic 불가)
+    return {
+        "message": "딥페이크 이미지 위조 흔적 시각화 접수 완료. 시각화 분석 시작 ...",
+        "task_id": task.id, 
+    }
+    
     
-    img_bytes = await explain_svc.explain_image(conn, result, explain_req)
-    return PNGStreamingResponse(io.BytesIO(img_bytes))
+@router.get("/image/result/{task_id}", status_code=status.HTTP_200_OK,
+            response_class=JSONResponse, summary="딥페이크 이미지 위조 흔적 시각화 결과 가져오기")
+async def get_explain_image_result(
+        task_id: str,
+        session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수
+        ):
+    
+    # celery_app에서 해당 task 가져오기
+    task = AsyncResult(task_id, app=celery_app)
+    
+    if task.state == "PENDING":
+        return {
+            "status": "PENDING", 
+            "message": "딥페이크 이미지 위조 흔적 시각화 분석 중 ..."
+            }
+
+    if task.state == "FAILURE":
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 
+            detail="딥페이크 이미지 위조 흔적 시각화 중 알 수 없는 오류가 발생하였습니다")
+    
+    # Celery Task 성공 시
+    result = task.result
+    
+    if result["status"] == "FAILED":
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 
+            detail=result["message"])
+
+    return {
+        "status": result["status"],
+        "message": result["message"],
+        "cam_loc": result["cam_loc"],   
+    }
     
+        
     
\ No newline at end of file

From 7efe4e134fca669ed5e4f07ff695e1c2483805d3 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 10:02:08 +0900
Subject: [PATCH 40/48] =?UTF-8?q?=E2=9C=A8=20Add=20cleanup=5Fimage=5Fcam?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/image_svc.py | 74 +++++++++++++++++++++++++++++----------
 1 file changed, 55 insertions(+), 19 deletions(-)

diff --git a/App/services/image_svc.py b/App/services/image_svc.py
index e094f6f..5dab7f3 100644
--- a/App/services/image_svc.py
+++ b/App/services/image_svc.py
@@ -1,6 +1,8 @@
 import os
 import asyncio
 import time
+import cv2
+import numpy as np
 import aiofiles as aio
 from dotenv import load_dotenv
 
@@ -14,34 +16,30 @@
 
 
 load_dotenv()
-UPLOAD_DIR = os.getenv("UPLOAD_DIR")
+UPLOAD_DIR = os.getenv("UPLOAD_DIR", "./static/uploads")
+EXPLAIN_UPLOAD_DIR = os.getenv("EXPLAIN_UPLOAD_DIR", "./static/explain")
 
-# 사용자 업로드 이미지 서버 내 저장 (회원/비회원 공통)
+os.makedirs(EXPLAIN_UPLOAD_DIR, exist_ok=True)
+os.makedirs(UPLOAD_DIR, exist_ok=True)
+
+# 사용자 업로드 이미지 파일 서버 내 저장 (회원/비회원 공통)
 async def upload_image(user_email: str | None, imagefile: UploadFile) -> str:
     try:
-        # 1. 사용자별 하위 디렉토리 결정
+        # 사용자별 하위 디렉토리 결정
         sub_dir = user_email if user_email else "anonymous"
         
         user_dir = os.path.join(UPLOAD_DIR, sub_dir)
 
-        # 3. 디렉토리 존재 확인 및 생성
-        if not os.path.exists(user_dir):
-            try:
-                os.makedirs(user_dir, exist_ok=True)
-            except OSError as e:
-                print(f"[Storage Error] 디렉토리 생성 실패: {e}")
-                raise HTTPException(
-                    status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
-                    detail="서버 내 저장 공간을 준비하지 못했습니다."
-                )
-
-        # 4. 파일명 중복 방지 (파일명_타임스탬프.확장자)
+        # 디렉토리 존재 확인 및 생성
+        os.makedirs(user_dir, exist_ok=True)
+    
+        # 파일명 중복 방지 (파일명_타임스탬프.확장자)
         filename_only, ext = os.path.splitext(imagefile.filename)
         
         upload_filename = f"{filename_only}_{int(time.time())}{ext}"
         upload_image_loc = os.path.join(user_dir, upload_filename)
 
-        # 5. 비동기 이미지 저장 (Chunk 단위 읽기)
+        # 비동기 이미지 저장 (Chunk 단위 읽기)
         try:
             async with aio.open(upload_image_loc, "wb") as outfile:
                 while content := await imagefile.read(1024 * 1024):
@@ -65,6 +63,25 @@ async def upload_image(user_email: str | None, imagefile: UploadFile) -> str:
             status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
             detail="이미지 업로드 과정에서 예상치 못한 오류가 발생했습니다.")
 
+# 딥페이크 이미지 위조 흔적 시각화 파일 서버 내 저장 (회원 전용)
+async def upload_image_cam(user_email: str, image_id: int, image: np.ndarray) -> str:
+    user_dir = os.path.join(EXPLAIN_UPLOAD_DIR, user_email)
+    os.makedirs(user_dir, exist_ok=True)
+    
+    cam_filename = f"{image_id}_{int(time.time())}.png"
+    cam_loc = os.path.join(user_dir, cam_filename)
+            
+    image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+    _, buf = cv2.imencode(".png", image_bgr)
+
+    try:
+        async with aio.open(cam_loc, "wb") as outfile:
+            await outfile.write(buf.tobytes())
+    except Exception as e:
+        raise e
+
+    return cam_loc[1:].replace("\\", "/")
+
 # 사용자 업로드 이미지 서버 내 삭제
 async def delete_image(image_loc: str):
     try:
@@ -76,8 +93,7 @@ async def delete_image(image_loc: str):
         else:
             print(f"File not found: {file_path}")
 
-    except Exception as e:
-        print(f"{e}")
+    except Exception:
         raise HTTPException(
             status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
             detail="알수없는 이유로 문제가 발생하였습니다."
@@ -185,7 +201,27 @@ async def _delete():
     finally:
         loop.close()
 
-
+@celery_app.task(name="cleanup_image_cam")
+def cleanup_image_cam(cam_loc: str):
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+    try:
+        async def _delete():
+            is_deleted = True
+            try:
+                await delete_image(cam_loc)
+            except Exception as e:
+                is_deleted = False
+                print(f"[Cleanup] 이미지 시각화 파일 삭제 실패 - cam_loc: {cam_loc}, error: {e}")
+            
+            if is_deleted:
+                print(f"[Cleanup] 이미지 시각화 파일 삭제 프로세스 완료 - cam_loc: {cam_loc}")
+        
+        loop.run_until_complete(_delete())
+    
+    finally:
+        loop.close()
+        
 # 이미지 DB 레코드 및 물리 파일 완전 삭제
 async def delete_image_db(conn: Connection, image_id: int):
     try:

From 8b89a071a598e619feeb778b2c1f8f414fb91028 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 10:03:03 +0900
Subject: [PATCH 41/48] =?UTF-8?q?=E2=9C=A8=20Add=20process=5Fexplain=5Fima?=
 =?UTF-8?q?ge=5Ftask?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/explain_svc.py | 127 ++++++++++++++++++++++++------------
 1 file changed, 86 insertions(+), 41 deletions(-)

diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
index 07118a5..2367bf1 100644
--- a/App/services/explain_svc.py
+++ b/App/services/explain_svc.py
@@ -1,5 +1,6 @@
 import cv2
 import asyncio
+import numpy as np
 from functools import partial
 from fastapi import status
 from fastapi.exceptions import HTTPException
@@ -8,63 +9,107 @@
 from schemas.image_schema import ImageData_indi
 from services import image_svc, inference_svc
 from explainability import (
+    CAMExplainer, 
     HiResCAMExplainer, GradCAMElementWiseExplainer, LayerCAMExplainer,
     EigenGradCAMExplainer, GradCAMPlusPlusExplainer, XGradCAMExplainer,
 )
+from celery_app import celery_app
 
+# --- CAM Explainer Registry ---
+# LOW branch: 국소 위조 흔적 포착
+# HIGH branch: 전역 위조 흔적 포착 
 EXPLAINER_REGISTRY = {
-    # LOW branch: 국소 위조 흔적 포착 
-    "hirescam":   HiResCAMExplainer,
-    "gradcamelementwise": GradCAMElementWiseExplainer,
-    "layercam":   LayerCAMExplainer,
-    # HIGH branch: 전역 위조 흔적 포착
-    "eigengradcam": EigenGradCAMExplainer,
-    "gradcamplusplus":   GradCAMPlusPlusExplainer,
-    "xgradcam":     XGradCAMExplainer,
+    "hirescam":   HiResCAMExplainer, #LOW branch
+    "gradcamelementwise": GradCAMElementWiseExplainer, #LOW branch
+    "layercam":   LayerCAMExplainer, #LOW branch
+    "eigengradcam": EigenGradCAMExplainer, # HIGH branch
+    "gradcamplusplus":   GradCAMPlusPlusExplainer, # HIGH branch
+    "xgradcam":     XGradCAMExplainer, # HIGH branch
 }
 
 _explainer_cache: dict = {}
 
-def _get_or_create_explainer(model_name: str, dataset: str, explain_req: ExplainRequest):
-    cache_key = (model_name, dataset, explain_req.explainer_type, explain_req.branch_level)
+# 캐시된 CAMExplainer 객체 반환하거나 새로 생성 
+def _get_or_create_explainer(model_name: str, dataset: str, explain_req_dict: dict):
+    cache_key = (model_name, dataset, explain_req_dict["explainer_type"], explain_req_dict["branch_level"])
     
     if cache_key not in _explainer_cache:
-        _explainer_cache[cache_key] = EXPLAINER_REGISTRY[explain_req.explainer_type](
+        _explainer_cache[cache_key] = EXPLAINER_REGISTRY[explain_req_dict["explainer_type"]](
             model_name = model_name,
             dataset = dataset,
-            branch_level = explain_req.branch_level,
+            branch_level = explain_req_dict["branch_level"],
         )
     return _explainer_cache[cache_key]
         
-def _run_visualization(explainer, image_path: str, explain_req: ExplainRequest) -> bytes:
-    if explain_req.display_type == "heatmap":
-        img = explainer.display_heatmap_on_image(image_path, image_weight=explain_req.overlay_ratio, threshold=explain_req.threshold,
-                                                 category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth)
-    elif explain_req.display_type == "contour":
-        img = explainer.display_contour_on_image(image_path, threshold=explain_req.threshold,
-                                                 category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth)
+# 시각화 이미지 생성 (heatmap, contour, bbox 선택)
+def _run_visualization(explainer: CAMExplainer, image_path: str, explain_req_dict: dict) -> np.ndarray:
+    if explain_req_dict["display_type"] == "heatmap":
+        return explainer.display_heatmap_on_image(image_path, image_weight=explain_req_dict["overlay_ratio"], threshold=explain_req_dict["threshold"],
+                                                 category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"])
+    elif explain_req_dict["display_type"] == "contour":
+        return explainer.display_contour_on_image(image_path, threshold=explain_req_dict["threshold"],
+                                                 category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"])
     else:  
-        img = explainer.display_bbox_on_image(image_path, threshold=explain_req.threshold,
-                                              category=explain_req.category, aug_smooth=explain_req.aug_smooth, eigen_smooth=explain_req.eigen_smooth) 
-    img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
-    _, buf = cv2.imencode(".png", img_bgr)
-    return buf.tobytes()
+        return explainer.display_bbox_on_image(image_path, threshold=explain_req_dict["threshold"],
+                                              category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"]) 
 
-async def explain_image(conn: Connection, result: ImageData_indi, explain_req: ExplainRequest):
-    
-    model_name, dataset = inference_svc.MODEL_CONFIG[result.version_type][result.model_type][result.domain_type]
-    
-    explainer = _get_or_create_explainer(model_name, dataset, explain_req)
-    image_path = "." + result.image_loc
-    
+# 딥페이크 이미지 위조 흔적 시각화 처리
+@celery_app.task(name="process_explain_image_task")
+def process_explain_image_task(user_email: str, 
+                               result_dict: dict,
+                               explain_req_dict: dict):
+    async def run_explain():
+        cam_loc = None
+        try:
+            # 추론에 사용된 모델 및 데이터셋 설정 로드
+            version_type = result_dict["version_type"]
+            model_type = result_dict["model_type"]
+            domain_type = result_dict["domain_type"]
+            
+            model_name, dataset = inference_svc.MODEL_CONFIG[version_type][model_type][domain_type]
+        
+            explainer = _get_or_create_explainer(model_name, dataset, explain_req_dict)
+            image_path = "." + result_dict["image_loc"]
+
+            # 시각화 이미지 생성 시작
+            try:
+                image = _run_visualization(explainer, image_path, explain_req_dict)
+            except Exception:
+                raise HTTPException(
+                    status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+                    detail="딥페이크 이미지 위조 흔적을 생성하는 중 오류가 발생하였습니다"
+                )
+
+            # 생성된 시각화 이미지 파일 저장 
+            try:
+                cam_loc = await image_svc.upload_image_cam(user_email, result_dict["image_id"], image)
+            except Exception:
+                raise HTTPException(
+                    status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+                    detail="딥페이크 이미지 위조 흔적 파일을 저장하는 중 오류가 발생했습니다."
+                )
+            
+            return {"status": "SUCCESS", 
+                    "message": "딥페이크 이미지 위조 흔적 시각화가 성공적으로 이루어졌습니다",
+                    "cam_loc": cam_loc}
+        
+        except HTTPException as e:
+            print(e.detail)
+            return {"status": "FAILED", "message": str(e.detail)}
+        
+        except Exception as e:
+            print(str(e))                    
+            return {"status": "FAILED", "message": str(e)}
+        
+        finally:
+            # 임시 저장된 시각화 이미지 파일 삭제 
+            if cam_loc:
+                image_svc.cleanup_image_cam.apply_async(args=[cam_loc], countdown=60) 
+
+    # 동기식 Celery 워커 내 비동기 이벤트 루프 구동
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
     try:
-        loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(
-            None, 
-            partial(_run_visualization, explainer, image_path, explain_req)
-        )
-    except Exception as e:
-        raise HTTPException(
-            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
-            detail="알수없는 이유로 문제가 발생하였습니다."
-            )
+        return loop.run_until_complete(run_explain())
+    finally:
+        loop.close()

From 1aa3e14b92651db641d790457a96da42f417af50 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 10:22:14 +0900
Subject: [PATCH 42/48] =?UTF-8?q?=E2=9A=A1=EF=B8=8F=20Add=20Celery=20Task?=
 =?UTF-8?q?=20Status(STARTED,=20RETRY)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/App/routes/explain.py b/App/routes/explain.py
index d6083d4..1ad8552 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -20,7 +20,8 @@ async def explain_image(
     explain_req: ExplainRequest,
     conn: Connection = Depends(context_get_conn),
     session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수 
-):
+):  
+    # 딥페이크 이미지 추론 결과 가져오기 
     result = await image_svc.get_image_result(conn, image_id)
       
     if result.status != "SUCCESS":
@@ -43,11 +44,11 @@ async def explain_image(
             detail="Pro 모델은 aug_smooth 기능을 지원하지 않습니다",
         )
         
-    # Celery Task 호출(Redis 브로커로 작업 전달)
+    # Celery Task 호출(Redis Broker 활용)
     task = explain_svc.process_explain_image_task.delay(
                                 user_email = session_user["email"],
-                                result_dict = result.model_dump(), # 직렬화 (Pydantic 불가)
-                                explain_req_dict = explain_req.model_dump()) # 직렬화 (Pydantic 불가)
+                                result_dict = result.model_dump(mode='json'),
+                                explain_req_dict = explain_req.model_dump())
     return {
         "message": "딥페이크 이미지 위조 흔적 시각화 접수 완료. 시각화 분석 시작 ...",
         "task_id": task.id, 
@@ -61,10 +62,10 @@ async def get_explain_image_result(
         session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수
         ):
     
-    # celery_app에서 해당 task 가져오기
+    # Redis Broker에서 Task ID에 해당하는 비동기 작업 상태 가져오기
     task = AsyncResult(task_id, app=celery_app)
     
-    if task.state == "PENDING":
+    if task.state in ("PENDING", "STARTED", "RETRY"):
         return {
             "status": "PENDING", 
             "message": "딥페이크 이미지 위조 흔적 시각화 분석 중 ..."
@@ -75,7 +76,7 @@ async def get_explain_image_result(
             status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, 
             detail="딥페이크 이미지 위조 흔적 시각화 중 알 수 없는 오류가 발생하였습니다")
     
-    # Celery Task 성공 시
+    # Celery Task 결과 가져오기
     result = task.result
     
     if result["status"] == "FAILED":

From bf2a7b002dfa9409185f9a5185a25e55d5afd8ff Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 10:24:34 +0900
Subject: [PATCH 43/48] =?UTF-8?q?=E2=9A=B0=EF=B8=8F=20Remove=20is=5Fdelete?=
 =?UTF-8?q?d=20flag=20in=20cleanup=5Fimage=5Fcam?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/services/image_svc.py | 13 +++----------
 1 file changed, 3 insertions(+), 10 deletions(-)

diff --git a/App/services/image_svc.py b/App/services/image_svc.py
index 5dab7f3..f83194e 100644
--- a/App/services/image_svc.py
+++ b/App/services/image_svc.py
@@ -1,15 +1,14 @@
 import os
-import asyncio
 import time
 import cv2
 import numpy as np
 import aiofiles as aio
+import asyncio
 from dotenv import load_dotenv
-
 from fastapi import UploadFile, status
 from fastapi.exceptions import HTTPException
 from sqlalchemy import text, Connection
-from sqlalchemy.exc import SQLAlchemyError, DBAPIError
+from sqlalchemy.exc import SQLAlchemyError
 from schemas.image_schema import UserHistory, UserHistory_indi, ImageData_indi
 from db.database import celery_db_conn
 from celery_app import celery_app
@@ -207,15 +206,11 @@ def cleanup_image_cam(cam_loc: str):
     asyncio.set_event_loop(loop)
     try:
         async def _delete():
-            is_deleted = True
             try:
                 await delete_image(cam_loc)
+                print(f"[Cleanup] 이미지 시각화 파일 삭제 완료 - cam_loc: {cam_loc}")
             except Exception as e:
-                is_deleted = False
                 print(f"[Cleanup] 이미지 시각화 파일 삭제 실패 - cam_loc: {cam_loc}, error: {e}")
-            
-            if is_deleted:
-                print(f"[Cleanup] 이미지 시각화 파일 삭제 프로세스 완료 - cam_loc: {cam_loc}")
         
         loop.run_until_complete(_delete())
     
@@ -233,9 +228,7 @@ async def delete_image_db(conn: Connection, image_id: int):
             
         await conn.commit()
 
-    
     except SQLAlchemyError as e:
-        print(e)
         await conn.rollback()
         raise HTTPException(status_code=status.HTTP_503_SERVICE_UNAVAILABLE, detail="요청하신 서비스가 잠시 내부적으로 문제가 발생하였습니다.")
 

From a64a4838267ceb87b09506b2cb3b38933ddd1eae Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Wed, 20 May 2026 10:47:52 +0900
Subject: [PATCH 44/48] =?UTF-8?q?=F0=9F=9A=9A=20Rename=20ExplainRequest=20?=
 =?UTF-8?q?to=20ExplainImageRequest?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/routes/explain.py         | 4 ++--
 App/schemas/explain_schema.py | 6 +++---
 App/services/explain_svc.py   | 2 --
 3 files changed, 5 insertions(+), 7 deletions(-)

diff --git a/App/routes/explain.py b/App/routes/explain.py
index 1ad8552..674edcb 100644
--- a/App/routes/explain.py
+++ b/App/routes/explain.py
@@ -6,7 +6,7 @@
 from fastapi.responses import JSONResponse
 from sqlalchemy import Connection
 from db.database import context_get_conn
-from schemas.explain_schema import ExplainRequest
+from schemas.explain_schema import ExplainImageRequest
 from services import session_svc, explain_svc, image_svc
 from celery_app import celery_app
 from celery.result import AsyncResult
@@ -17,7 +17,7 @@
              response_class=JSONResponse, summary="딥페이크 이미지 위조 흔적 시각화 비동기 접수")
 async def explain_image(
     image_id: int,
-    explain_req: ExplainRequest,
+    explain_req: ExplainImageRequest,
     conn: Connection = Depends(context_get_conn),
     session_user = Depends(session_svc.get_session_user_prt), # 로그인 필수 
 ):  
diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
index 8dd1773..3ce9729 100644
--- a/App/schemas/explain_schema.py
+++ b/App/schemas/explain_schema.py
@@ -22,7 +22,7 @@
 _EIGEN_ALLOWED = {"gradcamelementwise", "layercam", "xgradcam"}
 
 
-class ExplainRequest(BaseModel):
+class ExplainImageRequest(BaseModel):
     branch_level: Literal["low","high"] = Field("high", description="브랜치 레벨\nlow: 국소 위조 흔적 포착\nhigh: 전역적 위조 흔적 포착")
     explainer_type: str = Field("eigengradcam", description = ("선택 가능한 XAI 기법. low: [hirescam, gradcamelementwise, layercam], ""high: [eigengradcam, gradcamplusplus, xgradcam]"))
     display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
@@ -33,7 +33,7 @@ class ExplainRequest(BaseModel):
     eigen_smooth: bool = Field(False, description = "PCA 기반 노이즈 제거. 지배적인 패턴만 남김")
     
     @model_validator(mode="after")
-    def validate_explainer_for_branch(self) -> "ExplainRequest":
+    def validate_explainer_for_branch(self) -> "ExplainImageRequest":
         branch_allowed = {"low": _LOW_ALLOWED, "high": _HIGH_ALLOWED}
         if self.explainer_type not in branch_allowed[self.branch_level]:
             raise ValueError(
@@ -44,7 +44,7 @@ def validate_explainer_for_branch(self) -> "ExplainRequest":
         return self
 
     @model_validator(mode="after")
-    def validate_explainer_for_eigen_smooth(self) -> "ExplainRequest":
+    def validate_explainer_for_eigen_smooth(self) -> "ExplainImageRequest":
         if self.eigen_smooth and self.explainer_type not in _EIGEN_ALLOWED:
             raise ValueError(
                 f"eigen_smooth가 허용된 기법: "
diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
index 2367bf1..54953f4 100644
--- a/App/services/explain_svc.py
+++ b/App/services/explain_svc.py
@@ -5,8 +5,6 @@
 from fastapi import status
 from fastapi.exceptions import HTTPException
 from sqlalchemy import Connection
-from schemas.explain_schema import ExplainRequest
-from schemas.image_schema import ImageData_indi
 from services import image_svc, inference_svc
 from explainability import (
     CAMExplainer, 

From aabfc70780cf6b5ab3705da89917d3d0ac5546f6 Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 21 May 2026 10:36:04 +0900
Subject: [PATCH 45/48] =?UTF-8?q?=E2=9C=A8=20Add=20=5Fget=5Ftransform=5Ffr?=
 =?UTF-8?q?om=5Farray?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/explainer/base_explainer.py | 101 ++++++++++++++++-----
 1 file changed, 80 insertions(+), 21 deletions(-)

diff --git a/explainability/explainer/base_explainer.py b/explainability/explainer/base_explainer.py
index 2897a1e..5d46ce4 100644
--- a/explainability/explainer/base_explainer.py
+++ b/explainability/explainer/base_explainer.py
@@ -3,26 +3,22 @@
 import torch
 import numpy as np
 from pathlib import Path
-from typing import List, Optional, Tuple
+from typing import List, Optional, Tuple, Literal, Callable
 from preprocess.face_detector import FaceDetector2
 from deepguard.data import get_test_transforms
 from explainability.utils.reshape_transforms import reshape_transform_vit, reshape_transform_gcvit
 
-DATASETS = {"celeb_df_v2", "ff++", "kodf"}
-MODELS   = {"ms_eff_vit_b0", "ms_eff_vit_b5", "ms_eff_gcvit_b0", "ms_eff_gcvit_b5"}
-
 class BaseExplainer:
     def __init__(
         self,
-        model_name: str, # ms_eff_vit_b0 | ms_eff_vit_b5 | ms_eff_gcvit_b0 | ms_eff_gcvit_b5 
-        dataset: str,  # celeb_df_v2 | ff++ | both
+        model_name: Literal["ms_eff_vit_b0", "ms_eff_vit_b5", "ms_eff_gcvit_b0", "ms_eff_gcvit_b5"],
+        dataset: Literal["celeb_df_v2", "ff++", "kodf"], 
         margin_ratio: float = 0.2,
         conf_thres: float = 0.5,
-        branch_level: str = "both", # low | high | both
+        branch_level: Literal["low", "high"] = "high",
         l_stage_idx: int = -1, # gcvit low branch stage index (0~3)
-        block_idx: int = -1 # vit, gcvit block index 
+        block_idx: int = -1, # vit, gcvit block index
     ):  
-        self._validate_inputs(model_name, dataset, branch_level)
         
         self.device = "cuda:0" if torch.cuda.is_available() else 'cpu'
         self.margin_ratio = margin_ratio
@@ -48,19 +44,24 @@ def __init__(
         # Explainability setup
         self.reshape_fn = self._get_reshape_fn()
         self.target_layers = self._resolve_target_layers()
-    
-    def _validate_inputs(self, model_name: str, dataset: str, branch_level: str):
-        if model_name not in MODELS:
-            raise ValueError(f"Unsupported model: '{model_name}'. Must be one of {MODELS}")
-        if dataset not in DATASETS:
-            raise ValueError(f"Invalid dataset: '{dataset}'. Choose from {DATASETS}")
-        if branch_level not in ("low", "high", "both"):
-            raise ValueError(f"Invalid branch_level: '{branch_level}'. Expected 'low', 'high', or 'both'.")
-        
-    def _get_reshape_fn(self):
+
+    def _get_reshape_fn(self) -> Callable[[torch.Tensor], torch.Tensor]:
+        """모델 종류에 맞는 Feature map reshape 함수를 반환한다.
+
+        Returns:
+            Callable: 3D/4D 텐서를 GradCAM 연산에 적합한 형태로 변환하는 함수.
+        """
         return reshape_transform_vit if self.transformer_type == "vit" else reshape_transform_gcvit    
     
-    def _resolve_target_layers(self) -> list:
+    def _resolve_target_layers(self) -> list[torch.nn.Module]:
+        """시각화(XAI)의 대상이 될 Target Layer들을 선택한다.
+
+        Returns:
+            list[torch.nn.Module]: 분석 대상 정규화(Normalization) 레이어 리스트.
+
+        Raises:
+            ValueError: l_stage_idx나 block_idx가 모델 범위를 벗어날 경우 발생.
+        """
         layers = []
 
         if self.transformer_type == "gcvit":
@@ -95,16 +96,46 @@ def _resolve_target_layers(self) -> list:
 
     @staticmethod
     def _resolve_block_idx(block_idx: int, n_blocks: int, name: str) -> int:
+        """음수 인덱싱을 지원하며, 해당 블록의 유효한 인덱스를 계산 및 검증한다.
+
+        Args:
+            block_idx (int): 접근하려는 블록 인덱스 (음수 가능).
+            n_blocks (int): 해당 스테이지/모델의 총 블록 개수.
+            name (str): 에러 메시지 출력용 모듈 이름.
+
+        Returns:
+            int: 정제된 양수 형태의 블록 인덱스.
+
+        Raises:
+            ValueError: 인덱스가 유효 범위를 벗어난 경우 발생.
+        """
         idx = block_idx if block_idx >= 0 else n_blocks + block_idx
         if not (0 <= idx < n_blocks):
             raise ValueError(f"block_idx={block_idx} 범위 초과 ({name} block 수: {n_blocks})")
         return idx
     
     def _get_face_bbox(self, img: np.ndarray) -> List[float]:
+        """얼굴 탐지기(YOLO)를 통해 이미지에서 얼굴 바운딩 박스를 좌표를 추출한다.
+
+        Args:
+            img (np.ndarray): HWC 형태의 RGB 이미지 데이터.
+
+        Returns:
+            list[float]: [xmin, ymin, xmax, ymax, confidence] 형태의 좌표 리스트.
+        """
         result = self.face_detector.detect_batch([img], [1.0])
         return result[0]
     
     def _crop_face(self, img: np.ndarray, bbox: List[float]) -> np.ndarray:
+        """마진 비율을 고려하여 바운딩 박스 기준으로 얼굴 영역을 크롭한다.
+
+        Args:
+            img (np.ndarray): 전체 원본 이미지.
+            bbox (list[float]): 탐지된 [xmin, ymin, xmax, ymax] 얼굴 좌표.
+
+        Returns:
+            np.ndarray: 크롭된 얼굴 이미지 영역.
+        """
         xmin, ymin, xmax, ymax = bbox
         h, w = img.shape[:2]
         pw = int((xmax - xmin) * self.margin_ratio)
@@ -115,6 +146,14 @@ def _crop_face(self, img: np.ndarray, bbox: List[float]) -> np.ndarray:
         ]
         
     def _preprocess_img(self, img_path: str) -> np.ndarray:
+        """이미지 경로를 읽어 얼굴을 탐지하고 크롭하는 전과정을 수행한다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+
+        Returns:
+            np.ndarray: 전처리가 완료된 크롭된 얼굴 이미지 (RGB).
+        """
         img = cv2.imread(img_path)        
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
         detect_result = self._get_face_bbox(img)                
@@ -122,10 +161,30 @@ def _preprocess_img(self, img_path: str) -> np.ndarray:
         cropped = self._crop_face(img, bbox)
         return cropped
                   
-    def _get_transform(self, img_path: str):
+    def _get_transform(self, img_path: str) -> tuple[np.ndarray, torch.Tensor]:
+        """이미지 경로를 받아 크롭 이미지와 모델 입력용 텐서를 동시에 생성한다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+
+        Returns:
+            tuple[np.ndarray, torch.Tensor]: (크롭된 얼굴 이미지, (1, 3, H, W) 형태의 배치 텐서).
+        """
         face = self._preprocess_img(img_path)
         tensor = self.transforms(image=face)['image'].unsqueeze(0) # (1,3,H,W)
         return face, tensor
+    
+    def _get_transform_from_array(self, face: np.ndarray) -> tuple[np.ndarray, torch.Tensor]:
+        """이미 크롭된 얼굴 배열을 받아 모델 입력용 텐서로 변환한다.
+
+        Args:
+            face (np.ndarray): 이미 잘려진 얼굴 이미지 배열 (RGB).
+
+        Returns:
+            tuple[np.ndarray, torch.Tensor]: (원본 얼굴 이미지, (1, 3, H, W) 형태의 배치 텐서).
+        """
+        tensor = self.transforms(image=face)['image'].unsqueeze(0) # (1,3,H,W)
+        return face, tensor
             
     def __repr__(self):
         return (f"{self.__class__.__name__}("

From 3474cb82ed4054f2001360dd88b3803ef6b3d16b Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 21 May 2026 11:12:52 +0900
Subject: [PATCH 46/48] =?UTF-8?q?=F0=9F=94=A5=20Remove=20contour=20display?=
 =?UTF-8?q?=20type=20from=20image=20explainability?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 App/schemas/explain_schema.py | 2 +-
 App/services/explain_svc.py   | 3 ---
 2 files changed, 1 insertion(+), 4 deletions(-)

diff --git a/App/schemas/explain_schema.py b/App/schemas/explain_schema.py
index 3ce9729..37a51d8 100644
--- a/App/schemas/explain_schema.py
+++ b/App/schemas/explain_schema.py
@@ -25,7 +25,7 @@
 class ExplainImageRequest(BaseModel):
     branch_level: Literal["low","high"] = Field("high", description="브랜치 레벨\nlow: 국소 위조 흔적 포착\nhigh: 전역적 위조 흔적 포착")
     explainer_type: str = Field("eigengradcam", description = ("선택 가능한 XAI 기법. low: [hirescam, gradcamelementwise, layercam], ""high: [eigengradcam, gradcamplusplus, xgradcam]"))
-    display_type: Literal["heatmap", "contour", "bbox"] = Field("heatmap", description="시각화 형태. heatmap: 전체 분포, contour: 외곽선, bbox: 위조 의심 영역 사각형")
+    display_type: Literal["heatmap", "bbox"] = Field("heatmap", description="시각화 형태. heatmap: 전체 분포, bbox: 위조 의심 영역 사각형")
     category: Literal[0, 1] = Field(1, description="판단 클래스 인덱스 (0: Real / 1: Fake)")
     overlay_ratio: float = Field(0.5, ge=0.0, le=1.0, description = "Heatmap 투명도 (0: 히트맵만 강조, 1: 원본 이미지 위주)")
     threshold: float = Field(0.5, ge=0.5, le=1.0, description="contour/bbox 이진화 임계값 (0.0~1.0)")
diff --git a/App/services/explain_svc.py b/App/services/explain_svc.py
index 54953f4..83acc09 100644
--- a/App/services/explain_svc.py
+++ b/App/services/explain_svc.py
@@ -44,9 +44,6 @@ def _run_visualization(explainer: CAMExplainer, image_path: str, explain_req_dic
     if explain_req_dict["display_type"] == "heatmap":
         return explainer.display_heatmap_on_image(image_path, image_weight=explain_req_dict["overlay_ratio"], threshold=explain_req_dict["threshold"],
                                                  category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"])
-    elif explain_req_dict["display_type"] == "contour":
-        return explainer.display_contour_on_image(image_path, threshold=explain_req_dict["threshold"],
-                                                 category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"])
     else:  
         return explainer.display_bbox_on_image(image_path, threshold=explain_req_dict["threshold"],
                                               category=explain_req_dict["category"], aug_smooth=explain_req_dict["aug_smooth"], eigen_smooth=explain_req_dict["eigen_smooth"]) 

From 4b59f54ecaafec984845b542449349e3ceafa1ed Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 21 May 2026 11:19:11 +0900
Subject: [PATCH 47/48] =?UTF-8?q?=F0=9F=93=9D=20Update=20=5F=5Frepr=5F=5F?=
 =?UTF-8?q?=20with=20removing=20unnecessary=20argument?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/explainer/base_explainer.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/explainability/explainer/base_explainer.py b/explainability/explainer/base_explainer.py
index 5d46ce4..dc4fc62 100644
--- a/explainability/explainer/base_explainer.py
+++ b/explainability/explainer/base_explainer.py
@@ -191,4 +191,4 @@ def __repr__(self):
                 f"model={self.model_name}, dataset={self.dataset}, "
                 f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
                 f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, block_idx={self.block_idx},"
-                f"category={self.category},device={self.device})")
\ No newline at end of file
+                f"device={self.device})")
\ No newline at end of file

From d8fafd99aafc9f28cd5cbafb877adb48334a1b5b Mon Sep 17 00:00:00 2001
From: seoyunje <origin6165@naver.com>
Date: Thu, 21 May 2026 11:29:11 +0900
Subject: [PATCH 48/48] =?UTF-8?q?=E2=9C=A8=20Add=20display=5Fheatmap=5Ffro?=
 =?UTF-8?q?m=5Farray,=20display=5Fbbox=5Ffrom=5Farray?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 explainability/explainer/cam_explainer.py | 269 ++++++++++++++++++----
 1 file changed, 229 insertions(+), 40 deletions(-)

diff --git a/explainability/explainer/cam_explainer.py b/explainability/explainer/cam_explainer.py
index 27e6f58..f630055 100644
--- a/explainability/explainer/cam_explainer.py
+++ b/explainability/explainer/cam_explainer.py
@@ -1,3 +1,4 @@
+from typing import Union, Literal
 import torch
 import cv2
 import numpy as np
@@ -38,13 +39,33 @@ def __init__(self, **kwargs):
         self.cam = self._build_cam()
     
     @abstractmethod
-    def _build_cam(self):
-        ...
+    def _build_cam(self) -> any:
+        """하위 클래스에서 각 알고리즘(GradCAM, HiResCAM 등)에 맞는 CAM 객체를 생성한다.
+
+        Returns:
+            Any: pytorch-grad-cam 패키지의 CAM 기반 인스턴스.
+        """
+        pass
 
     def _build_grayscale_cam(self, img_path: str,
                              category: int = 1,
                              aug_smooth: bool = False,
-                             eigen_smooth: bool = False):
+                             eigen_smooth: bool = False) -> tuple[np.ndarray, torch.Tensor, np.ndarray]:
+        """이미지 경로를 받아 전처리를 수행하고 raw 흑백(Grayscale) CAM을 생성한다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+            category (int, optional): 타겟 클래스 인덱스 (위조: 1, 정상: 0). Defaults to 1.
+            aug_smooth (bool, optional): Test-time augmentation 가동 여부. Defaults to False.
+            eigen_smooth (bool, optional): PCA 기반 노이즈 제거 가동 여부. Defaults to False.
+
+        Returns:
+            tuple[np.ndarray, torch.Tensor, np.ndarray]: 
+                - grayscale_cam: (H, W) 형태의 0~1 사이로 정규화된 CAM 맵.
+                - tensor: (1, 3, H_model, W_model) 형태의 모델 입력용 GPU 텐서.
+                - face: 크롭된 원본 얼굴 이미지 배열 (RGB).
+        """
+        
         face, tensor = self._get_transform(img_path)
         tensor = tensor.to(self.device)
         targets = [BinaryClassifierOutputTarget(category)]
@@ -56,17 +77,92 @@ def _build_grayscale_cam(self, img_path: str,
             eigen_smooth=eigen_smooth,
         )
         return grayscale_cam[0], tensor, face
+    
+    def _build_grayscale_cam_from_array(self,
+                                        face: np.ndarray,
+                                        category: int = 1,
+                                        aug_smooth: bool = False,
+                                        eigen_smooth: bool = False) -> tuple[np.ndarray, torch.Tensor, np.ndarray]:
+        """이미 크롭된 얼굴 이미지 배열로부터 raw 흑백(Grayscale) CAM을 생성한다.
+
+        Args:
+            face (np.ndarray): 전처리(크롭)가 완료된 얼굴 이미지 배열 (RGB).
+            category (int, optional): 타겟 클래스 인덱스. Defaults to 1.
+            aug_smooth (bool, optional): Test-time augmentation 가동 여부. Defaults to False.
+            eigen_smooth (bool, optional): PCA 기반 노이즈 제거 가동 여부. Defaults to False.
+
+        Returns:
+            tuple[np.ndarray, torch.Tensor, np.ndarray]: 
+                - grayscale_cam: (H, W) 형태의 0~1 사이로 정규화된 CAM 맵.
+                - tensor: (1, 3, H_model, W_model) 형태의 모델 입력용 GPU 텐서.
+                - face: 입력받은 원본 얼굴 이미지 배열.
+        """
+        face, tensor = self._get_transform_from_array(face)
+        tensor = tensor.to(self.device)
+        targets = [BinaryClassifierOutputTarget(category)]
+        
+        grayscale_cam = self.cam(
+            input_tensor=tensor,
+            targets=targets,
+            aug_smooth=aug_smooth,
+            eigen_smooth=eigen_smooth,
+        )
+        return grayscale_cam[0], tensor, face
 
     def _prepare_cam(self, img_path: str,
                      category: int = 1,
                      aug_smooth: bool = False,
-                     eigen_smooth: bool = False):
-        """Build grayscale CAM and recover it to original face resolution."""
+                     eigen_smooth: bool = False) -> tuple[np.ndarray, np.ndarray]:
+        """흑백 CAM을 생성하고, 모델 입력 시 발생한 패딩 제거 및 원래 얼굴 해상도로 복원한다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+            category (int, optional): 타겟 클래스 인덱스. Defaults to 1.
+            aug_smooth (bool, optional): Test-time augmentation 가동 여부. Defaults to False.
+            eigen_smooth (bool, optional): PCA 기반 노이즈 제거 가동 여부. Defaults to False.
+
+        Returns:
+            tuple[np.ndarray, np.ndarray]: 
+                - cam_recovered: 원본 얼굴 크기로 복원된 CAM 맵 (H_face, W_face).
+                - face: 크롭된 원본 얼굴 이미지 배열 (RGB).
+        """
         grayscale_cam, tensor, face = self._build_grayscale_cam(img_path, category, aug_smooth, eigen_smooth)
         cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
         return cam_recovered, face
     
-    def _get_binary_mask(self, cam: np.ndarray, threshold: float | str):
+    def _prepare_cam_from_array(self, face: np.ndarray,
+                                category: int = 1,
+                                aug_smooth: bool = False,
+                                eigen_smooth: bool = False) -> tuple[np.ndarray, np.ndarray]:
+        """얼굴 배열로부터 흑백 CAM을 생성하고 원래 얼굴 해상도로 복원한다.
+
+        Args:
+            face (np.ndarray): 전처리(크롭)가 완료된 얼굴 이미지 배열 (RGB).
+            category (int, optional): 타겟 클래스 인덱스. Defaults to 1.
+            aug_smooth (bool, optional): Test-time augmentation 가동 여부. Defaults to False.
+            eigen_smooth (bool, optional): PCA 기반 노이즈 제거 가동 여부. Defaults to False.
+
+        Returns:
+            tuple[np.ndarray, np.ndarray]: 
+                - cam_recovered: 원본 얼굴 크기로 복원된 CAM 맵 (H_face, W_face).
+                - face: 입력받은 원본 얼굴 이미지 배열.
+        """
+        grayscale_cam, tensor, face = self._build_grayscale_cam_from_array(face, category, aug_smooth, eigen_smooth)
+        cam_recovered = remove_padding_and_resize(grayscale_cam, face.shape, tensor.shape)
+        return cam_recovered, face
+    
+    def _get_binary_mask(self, cam: np.ndarray, threshold: Union[float, Literal["auto"]]) -> np.ndarray:
+        """입력된 CAM 맵을 기반으로 위조 유력 부위를 이진화(Binary Mask)한다.
+
+        Args:
+            cam (np.ndarray): 0~1 사이 값을 가진 복원된 CAM 맵.
+            threshold (Union[float, 'auto']): 
+                - float (0.0~1.0): 상위 백분위수 기준으로 임계값 설정 (예: 0.7 이면 상위 30% 영역 선택).
+                - 'auto': Otsu 알고리즘을 사용해 최적의 임계값을 자동으로 계산.
+
+        Returns:
+            np.ndarray: 0 또는 255 값을 가지는 이진화 마스크 맵 (np.uint8).
+        """
         if threshold != 'auto':
             t = np.percentile(cam, threshold * 100)
             binary_mask = (cam > t).astype(np.uint8) * 255
@@ -76,42 +172,143 @@ def _get_binary_mask(self, cam: np.ndarray, threshold: float | str):
             )
         return binary_mask
 
-    def display_heatmap_on_image(self, img_path: str, colormap: int = cv2.COLORMAP_JET, 
-                                 image_weight: float = 0.5, threshold: float | str = "auto",
-                                 category: int = 1,
-                                 aug_smooth: bool = False, eigen_smooth: bool = False,
-                                 ):
-        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
+    def display_heatmap_on_image(self, img_path: str, **kwargs) -> np.ndarray:
+        """이미지 경로를 입력받아 원본 얼굴 위에 위조 흔적 히트맵을 중첩한 이미지를 생성한다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+            **kwargs:
+                category (int): 타겟 클래스 인덱스 (기본값: 1).
+                aug_smooth (bool): Test-time augmentation 가동 여부 (기본값: False).
+                eigen_smooth (bool): PCA 기반 노이즈 제거 가동 여부 (기본값: False).
+                threshold (float | str): 이진화 임계값 (기본값: "auto").
+                colormap (int): OpenCV 컬러맵 상수 (기본값: cv2.COLORMAP_JET).
+                image_weight (float): 원본 이미지 불투명도 가중치 (기본값: 0.5).
+
+        Returns:
+            np.ndarray: 히트맵이 합성된 얼굴 이미지 (RGB, np.uint8).
+        """
+        cam_recovered, face = self._prepare_cam(
+            img_path, 
+            category = kwargs.get("category", 1), 
+            aug_smooth = kwargs.get("aug_smooth", False), 
+            eigen_smooth = kwargs.get("eigen_smooth", False))
+        
+        binary_mask = self._get_binary_mask(cam_recovered, kwargs.get("threshold", "auto"))
+        cam_recovered = np.where(binary_mask > 0, cam_recovered, 0.0)
+                
+        return show_cam_on_image(
+            np.float32(face) / 255.0,
+            cam_recovered,
+            use_rgb = True,
+            colormap = kwargs.get("colormap", cv2.COLORMAP_JET),
+            image_weight = kwargs.get("image_weight", 0.5),
+        )
+        
+    def display_heatmap_from_array(self, face: np.ndarray, **kwargs) -> np.ndarray:
+        """얼굴 배열을 직접 입력받아 원본 얼굴 위에 위조 흔적 히트맵을 중첩한 이미지를 생성한다.
+
+        Args:
+            face (np.ndarray): 전처리(크롭)가 완료된 얼굴 이미지 배열 (RGB).
+            **kwargs:
+                category (int): 타겟 클래스 인덱스 (기본값: 1).
+                aug_smooth (bool): Test-time augmentation 가동 여부 (기본값: False).
+                eigen_smooth (bool): PCA 기반 노이즈 제거 가동 여부 (기본값: False).
+                threshold (float | str): 이진화 임계값 (기본값: "auto").
+                colormap (int): OpenCV 컬러맵 상수 (기본값: cv2.COLORMAP_JET).
+                image_weight (float): 원본 이미지 불투명도 가중치 (기본값: 0.5).
+
+        Returns:
+            np.ndarray: 히트맵이 합성된 얼굴 이미지 (RGB, np.uint8).
+        """
         
-        binary_mask = self._get_binary_mask(cam_recovered, threshold)
+        cam_recovered, face = self._prepare_cam_from_array(
+            face, 
+            category = kwargs.get("category", 1), 
+            aug_smooth = kwargs.get("aug_smooth", False), 
+            eigen_smooth = kwargs.get("eigen_smooth", False))
+        
+        binary_mask = self._get_binary_mask(cam_recovered, kwargs.get("threshold", "auto"))
         cam_recovered = np.where(binary_mask > 0, cam_recovered, 0.0)
                 
-        # (H,W,3), range(0~255), np.uint8
         return show_cam_on_image(
             np.float32(face) / 255.0,
             cam_recovered,
-            use_rgb=True,
-            colormap=colormap,
-            image_weight=image_weight,
+            use_rgb = True,
+            colormap = kwargs.get("colormap", cv2.COLORMAP_JET),
+            image_weight = kwargs.get("image_weight", 0.5),
         )  
     
-    def display_contour_on_image(self, img_path: str, threshold: float | str = 'auto', thickness: int = 2,
-                                 category: int = 1,
-                                 aug_smooth: bool = False, eigen_smooth: bool = False):
-        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
+    def display_bbox_on_image(self, img_path: str, **kwargs) -> np.ndarray:
+        """이미지 경로를 입력받아 위조 의심 부위를 찾아 사각형 바운딩 박스와 확률을 그린다.
+
+        Args:
+            img_path (str): 입력 이미지 파일 경로.
+            **kwargs:
+                category (int): 타겟 클래스 인덱스 (기본값: 1).
+                aug_smooth (bool): Test-time augmentation 가동 여부 (기본값: False).
+                eigen_smooth (bool): PCA 기반 노이즈 제거 가동 여부 (기본값: False).
+                threshold (float | str): 이진화 임계값 (기본값: "auto").
+                thickness (int): 사각형 선 굵기 (기본값: 2).
+
+        Returns:
+            np.ndarray: 바운딩 박스와 어노테이션이 추가된 원본 얼굴 이미지 (RGB).
+        """
+        cam_recovered, face = self._prepare_cam(
+            img_path, 
+            category = kwargs.get("category", 1), 
+            aug_smooth = kwargs.get("aug_smooth", False), 
+            eigen_smooth = kwargs.get("eigen_smooth", False)
+            )
+        
+        binary_mask = self._get_binary_mask(cam_recovered, kwargs.get("threshold", "auto"))
         
-        binary_mask = self._get_binary_mask(cam_recovered, threshold)
-            
         contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-        annotated_img = cv2.drawContours(face.copy(), contours, -1, (255, 0, 0), thickness)
+        
+        annotated_img = face.copy()
+        for c in contours:
+            x, y, w, h = cv2.boundingRect(c)
+            roi_prob = float(np.max(cam_recovered[y:y + h, x:x + w]))
+
+            r = int(min(255, 2 * roi_prob * 255))
+            g = int(min(255, 2 * (1 - roi_prob) * 255))
+            box_color = (r, g, 30)  # RGB
+
+            cv2.rectangle(annotated_img, (x, y), (x + w, y + h), box_color, kwargs.get("thickness", 1))
+            
+            _draw_label(
+                annotated_img,
+                text=f"{roi_prob * 100:.1f}%",
+                x=x,
+                y=max(y - 6, 14),
+                prob_ratio=roi_prob,
+            )
+
         return annotated_img
-    
-    def display_bbox_on_image(self, img_path: str, threshold: float | str = 'auto', thickness: int = 2,
-                              category: int = 1,
-                              aug_smooth: bool = False, eigen_smooth: bool = False):
-        cam_recovered, face = self._prepare_cam(img_path, category, aug_smooth, eigen_smooth)
+
+    def display_bbox_from_array(self, face: np.ndarray, **kwargs) -> np.ndarray:
+        """얼굴 배열을 직접 입력받아 위조 의심 부위를 찾아 사각형 바운딩 박스와 확률을 그린다.
+
+        Args:
+            face (np.ndarray): 전처리(크롭)가 완료된 얼굴 이미지 배열 (RGB).
+            **kwargs:
+                category (int): 타겟 클래스 인덱스 (기본값: 1).
+                aug_smooth (bool): Test-time augmentation 가동 여부 (기본값: False).
+                eigen_smooth (bool): PCA 기반 노이즈 제거 가동 여부 (기본값: False).
+                threshold (float | str): 이진화 임계값 (기본값: "auto").
+                thickness (int): 사각형 선 굵기 (기본값: 2).
+
+        Returns:
+            np.ndarray: 바운딩 박스와 어노테이션이 추가된 원본 얼굴 이미지 (RGB).
+        """
+        cam_recovered, face = self._prepare_cam_from_array(
+            face, 
+            category = kwargs.get("category", 1), 
+            aug_smooth = kwargs.get("aug_smooth", False), 
+            eigen_smooth = kwargs.get("eigen_smooth", False)
+            )
         
-        binary_mask = self._get_binary_mask(cam_recovered, threshold)
+        binary_mask = self._get_binary_mask(cam_recovered, kwargs.get("threshold", "auto"))
         
         contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
         
@@ -124,7 +321,7 @@ def display_bbox_on_image(self, img_path: str, threshold: float | str = 'auto',
             g = int(min(255, 2 * (1 - roi_prob) * 255))
             box_color = (r, g, 30)  # RGB
 
-            cv2.rectangle(annotated_img, (x, y), (x + w, y + h), box_color, thickness)
+            cv2.rectangle(annotated_img, (x, y), (x + w, y + h), box_color, kwargs.get("thickness", 1))
             
             _draw_label(
                 annotated_img,
@@ -134,12 +331,4 @@ def display_bbox_on_image(self, img_path: str, threshold: float | str = 'auto',
                 prob_ratio=roi_prob,
             )
 
-        return annotated_img
-
-    def __repr__(self):
-        return (f"{self.__class__.__name__}("
-                f"model={self.model_name}, dataset={self.dataset}, "
-                f"margin_ratio={self.margin_ratio}, conf_thres={self.conf_thres}, "
-                f"branch_level={self.branch_level}, l_stage_idx={self.l_stage_idx}, block_idx={self.block_idx},"
-                f"aug_smooth={self.aug_smooth}, eigen_smooth={self.eigen_smooth}, "
-                f"category={self.category},device={self.device})")
\ No newline at end of file
+        return annotated_img
\ No newline at end of file