Adversarial Attack Visualizer

A comprehensive GUI application for visualizing and understanding adversarial attacks on deep neural networks.

Attack Progression

Attack Surface Visualization

Features

Attack Methods

• FGSM (Fast Gradient Sign Method)
• PGD (Projected Gradient Descent)
• DeepFool (Minimal perturbation attack)
• C&W (Carlini & Wagner attack)

Visualizations

• Real-time attack progression tracking
• Side-by-side original vs adversarial image comparison
• Enhanced perturbation visualization (10x amplified)
• Top 5 prediction confidence analysis
• 3D attack surface mapping
• Gradient flow visualization
• Vulnerability heatmaps

Image Sources

• Pexels API integration for image search
• Unsplash fallback support
• Direct URL image loading
• Automatic fallback to dummy images (optional)

Installation

Prerequisites

• Python 3.8 or higher
• CUDA-compatible GPU (optional, but recommended)

Setup

git clone https://github.com/imjbassi/adversarial-visualizer.git
cd adversarial-attack-visualizer
pip install -r requirements.txt
cp .env.example .env

Edit .env and insert your Pexels API key.

Usage

Quick Start

python scripts/run_attack.py

Test Installation

python test_setup.py

Configuration

Environment Variables

PEXELS_API_KEY=your_pexels_api_key_here

Attack Parameters

• Epsilon: Perturbation magnitude (0.001 - 0.1)
• Iterations: Number of attack iterations (10 - 100)
• Attack Method: FGSM, PGD, DeepFool, or C&W

Project Structure

adversarial-attack-visualizer/
├── scripts/
│   └── run_attack.py
├── attacks/
│   ├── fgsm.py
│   ├── pgd.py
│   ├── deepfool.py
│   └── cw.py
├── utils/
│   ├── image_utils.py
│   └── visualization.py
├── requirements.txt
├── test_setup.py
├── .env.example
├── .gitignore
└── README.md

Examples

Basic Attack

1. Launch the application
2. Enter search term (e.g., "cat", "dog", "car")
3. Select attack method
4. Adjust parameters using sliders
5. Click "Search & Attack"

Advanced Visualizations

• 3D Attack Surface: Compare attack effectiveness across methods and parameters
• Gradient Flow: Visualize gradient magnitudes during attacks
• Vulnerability Heatmap: Identify most vulnerable image regions

Technical Details

Model

• Architecture: ResNet-18
• Dataset: ImageNet pretrained
• Input Size: 224x224 RGB
• Normalization: ImageNet standard (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Attack Implementations

• FGSM: Single-step gradient-based attack
• PGD: Multi-step projected gradient descent with momentum
• DeepFool: Iterative minimal perturbation method
• C&W: Optimization-based attack with L2 norm constraints

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/your-feature)
3. Commit your changes (git commit -m 'Add feature')
4. Push to your branch (git push origin feature/your-feature)
5. Open a Pull Request

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgments

• PyTorch team for the deep learning framework
• ImageNet dataset contributors
• Pexels and Unsplash for image APIs
• Researchers behind adversarial methods

Citation

@software{adversarial_attack_visualizer,
  title={Adversarial Attack Visualizer},
  author={Your Name},
  year={2025},
  url={https://github.com/yourusername/adversarial-attack-visualizer}
}

Troubleshooting

Common Issues

• CUDA out of memory: Reduce batch size or switch to CPU
• Missing dependencies: Run pip install -r requirements.txt
• API key errors: Check .env file
• Slow performance: Use GPU acceleration

Getting Help

• Open a GitHub issue
• Search existing issues
• Review documentation and logs

Roadmap

• Support for additional attacks (e.g., JSMA, BIM)
• Custom model uploads
• Batch attack support
• Export result capability
• Performance benchmarking