🔍 Deep Dive: Detecting AI-Generated Faces

The goal of this project was to engineer a Convolutional Neural Network (CNN) model for image classification between datasets consisting of authentic human portraits and AI-generated faces (StyleGAN). Leveraging Python's frameworks for machine learning such as TensorFlow and OpenCV I developed a Jupyter Notebook pipeline for image preprocessing, data normalization, feature extraction, and model validation, achieving high-precision classification.

🖼️ View Research Poster

_{Click the image above to view the full PDF.}

📂 Project Structure

/DeepDive-Detecting-AI-Generated-Faces
├── data/               # Image datasets (Human/AI)
├── documentation/      # Research poster & icon files
├── models/             # Trained .keras models
├── notebooks/          # Data processing & training scripts
├── .env.example        # Configuration template
├── .gitignore          # Version control exclusions
├── README.md           # Project documentation
├── Tensorboard.bat     # Opens Tensorboard interface
├── Tensorflow.bat      # Opens Tensorflow interface
├── requirements.txt    # Python library dependencies
└── setup.bat           # Automated environment setup

🚀 Getting Started

1. Clone the repository using: git clone https://github.com/cdmanning/DeepDive-Detecting-AI-Generated-Faces.git

2. Ensure you have Python 3.12 installed and correctly added to your system PATH.

3. Run setup.bat to automatically install all required Python libraries and generate the necessary project folders.

4. Rename the .env.example file to .env and verify the DATADIR path points to your target dataset location.

5. Download the image sets and place them into the ./data/Human and ./data/AI directories:

   • Human Dataset: CelebA Dataset
   • AI-Generated Dataset: 1M Fake Faces Dataset

6. Run Tensorboard.bat to run the Jupyter Notebook environment and begin training the model.

⚖️ Licensing

• All source code in this repository is All Rights Reserved.
• The research poster and findings are licensed under CC BY-NC 4.0 Attribution-NonCommercial.

Developed as part of the 2023 UNG Annual Research Conference.