Transfer-Based Adversarial Attack Study on CIFAR-10

A reproducible study of transfer-based adversarial attacks using non-robust surrogate models to attack robust CIFAR-10 models from RobustBench.

Overview

This project generates adversarial examples on surrogate models (non-robust) and evaluates how well they transfer to robust target models. Results are compared against each target's reported AutoAttack robust accuracy to analyze transferability.

Models

Role	Model	Architecture	Source
Surrogate	Standard	WideResNet-28-10	RobustBench (non-robust baseline)
Surrogate	cifar10_vgg16_bn	VGG-16-BN	chenyaofo/pytorch-cifar-models
Surrogate	cifar10_resnet56	Resnet-56	chenyaofo/pytorch-cifar-models
Surrogate	cifar10_mobilenetv2_x1_4	MobileNetV2	chenyaofo/pytorch-cifar-models
Surrogate	cifar10_repvgg_a2	RepVGG	chenyaofo/pytorch-cifar-models
Target	Bartoldson2024Adversarial_WRN-94-16	WideResNet-94-16	RobustBench (73.71% robust acc)
Target	Ding2020MMA	WideResNet-28-4	RobustBench (41.44% robust acc)
Target	Bai2024MixedNUTS	ResNet-152 + WideResNet-70-16	RobustBench (69.71% robust acc)

Attacks (from torchattack)

Attack	Description
DI-FGSM	Diverse Input FGSM (input diversity)
TI-FGSM	Translation-Invariant FGSM
Admix	Admix input transformation
VMI-FGSM	Variance-tuned MI-FGSM

Setup

Prerequisites

• Python 3.10+ with PyTorch, in a conda environment.
• Minimum 16 GB NVIDIA GPU recommended.
• >=8 GB RAM recommended.

Install dependencies

pip install -r requirements.txt

All dependencies (torch, torchvision, robustbench, torchattack, etc.) should already be installed if you followed the project setup. Install suitable CUDA version of PyTorch for your device.

Usage

Quick test (100 samples)

python main.py --num-samples 100

Full run (1000 samples, default config)

python main.py

Full dataset (10,000 samples)

python main.py --num-samples 10000

Skip ensemble attacks

python main.py --no-ensemble

Custom config

python main.py --config configs/my_config.yaml

Force CPU

python main.py --device cpu

Output

Results are saved to results/:

File	Description
results.csv	Full results table
summary.md	Markdown-formatted summary table
attack_success_rates.png	Bar chart of attack success rates by attacks and target
best_transfer_vs_reported.png	Best transfer result vs reported AutoAttack for each target
config_snapshot.yaml	Copy of the config used for the run

Project Structure

transfer_attack_study/
├── README.md
├── requirements.txt
├── configs/
│   └── default.yaml           # Default experiment configuration
├── src/
│   ├── __init__.py
│   ├── data.py                # CIFAR-10 data loading
│   ├── models.py              # Model loading and normalization wrappers
│   ├── attacks.py             # Attack factory and adversarial generation
│   ├── evaluate.py            # Transfer evaluation metrics
│   ├── metadata.py            # Reported model accuracies
│   └── reporting.py           # CSV, markdown, and plot generation
├── main.py         # Main experiment runner
└── results/                   # Output directory (created at runtime)

Citations

• RobustBench: Croce et al., 2021
• torchattack: SpencerWoo
• chenyaofo/pytorch-cifar-models