Subset Scanning Adversarial Detection Demo

A comprehensive demonstration of subset scanning for adversarial detection using the Adversarial Robustness Toolbox (ART). This project showcases how subset scanning can effectively identify adversarial examples hidden within clean data.

🎯 Overview

This demo implements the complete workflow for detecting adversarial examples using subset scanning:

1. FGSM Attack Generation: Create adversarial examples using Fast Gradient Sign Method
2. Mixed Dataset Creation: Hide adversarial samples within clean data (simulating real-world scenarios)
3. Subset Scanning Detection: Use ART's SubsetScanningDetector to identify the anomalous subgroup
4. Performance Evaluation: Comprehensive metrics and visualizations
5. Qualitative Analysis: Visual evidence of subtle perturbations and detection results

🖼️ Visual Showcase

Here are some key visualizations from the demo:

Original vs Adversarial Images

Shows the subtle differences between clean and adversarial images, with perturbation patterns and magnified views.

Detection Performance

Demonstrates how well the subset scanning detector separates clean from adversarial samples.

Perturbation Patterns

Reveals systematic patterns in adversarial attacks and identifies vulnerable image regions.

🚀 Quick Start

Prerequisites

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Run the Complete Showcase

# Train the model first (if not already trained)
python src/models/cnn_model.py

# Run the comprehensive showcase
python showcase_subset_scanning_demo.py

This will demonstrate the complete workflow and generate both quantitative and qualitative visualizations.

Interactive Visualization App

For a standalone interactive visualization experience:

# Run the interactive visualization app
python visualization_app.py

This app will:

• Load real data if available, or create realistic dummy data
• Display interactive visualizations one by one
• Save visualizations to files
• Provide detailed explanations for each visualization

Each visualization opens in a separate window - close it to proceed to the next one.

Screenshot Demo

For creating high-resolution screenshots for documentation or presentations:

# Create high-resolution screenshots
python screenshot_demo.py

This will generate publication-ready screenshots in the screenshots/ directory.

Visualization Guide

For detailed information about all visualization features, customization options, and troubleshooting, see the comprehensive Visualization Guide.

📊 What You'll See

The showcase demonstrates:

• FGSM Attack Effectiveness: Shows how adversarial examples can fool the model
• Hidden Adversarial Detection: Subset scanning finds adversarial samples even when they're mixed with clean data
• Performance Comparison: Raw pixels vs. feature-based detection
• Comprehensive Metrics: ROC AUC, precision, recall, F1-score
• Quantitative Visualizations: Score distributions, ROC curves, confusion matrices
• Qualitative Visualizations: Original vs adversarial images, perturbation patterns, detected anomalies (see screenshots above)
• Interactive Exploration: Each visualization opens in a separate window for detailed examination

🏗️ Project Structure

├── showcase_subset_scanning_demo.py  # Main showcase script
├── visualization_app.py              # Standalone interactive visualization app
├── screenshot_demo.py                # High-resolution screenshot generator
├── src/
│   ├── data/
│   │   ├── dataset.py              # MNIST data loading
│   │   └── subset_data.py          # Mixed dataset preparation
│   ├── models/
│   │   └── cnn_model.py            # CNN model for MNIST
│   ├── attacks/
│   │   └── fgsm_attack.py          # FGSM attack implementation
│   ├── scanning/
│   │   ├── art_subset_scanner.py   # ART subset scanning utilities
│   │   ├── subset_evaluation.py    # Evaluation and visualization
│   │   ├── simple_subset_scanning_demo.py  # Simple demo
│   │   └── run_subset_scanning_demo.py     # Comprehensive demo
│   ├── visualization/
│   │   ├── qualitative_analysis.py # Qualitative visualizations
│   │   └── __init__.py
│   └── utils/
│       ├── art_utils.py            # ART classifier utilities
│       └── data_utils.py           # Data preparation utilities
├── screenshots/                     # High-resolution visualization screenshots
├── interactive_visualizations/      # Interactive visualization outputs
├── VISUALIZATION_GUIDE.md          # Comprehensive visualization guide
└── requirements.txt

🔬 Key Features

1. FGSM Attack Integration

• Uses ART's FastGradientMethod for generating adversarial examples
• Configurable perturbation magnitude (ε)
• Attack effectiveness evaluation

2. Subset Scanning Detection

• Raw Pixel Detection: Flattened image data (784 dimensions for MNIST)
• Feature-Based Detection: Extracted features from CNN layers
• Mixed Dataset Support: Hidden adversarial samples within clean data

3. Comprehensive Evaluation

• ROC curves and AUC scores
• Precision-recall analysis
• Confusion matrices
• Score distribution visualization

4. Qualitative Visualization

• Original vs Adversarial: Side-by-side comparison showing subtle perturbations
• Perturbation Analysis: Statistical analysis of perturbation patterns
• Detected Anomalies: Visualization of images identified as anomalous
• Score Distribution: Analysis of detection scores across clean and adversarial samples

5. ART Integration

• Seamless integration with ART's SubsetScanningDetector
• PyTorch model compatibility
• Configurable detection parameters

📈 Example Results

When you run the showcase, you'll see output like:

SUBSET SCANNING ADVERSARIAL DETECTION SHOWCASE
================================================================================

PHASE 1: Loading Data and Model
----------------------------------------
✓ Loaded 1000 test samples
✓ Data shape: torch.Size([1000, 1, 28, 28])
✓ Model accuracy on clean data: 98.50%

PHASE 2: Generating FGSM Adversarial Examples
----------------------------------------
✓ FGSM attack completed with eps=0.3
✓ Attack success rate: 85.20%
✓ Adversarial accuracy: 14.80%

PHASE 3: Creating Mixed Dataset with Hidden Adversarial Subgroup
----------------------------------------
✓ Mixed dataset created:
  - Total samples: 1000
  - Clean samples: 900
  - Adversarial samples: 100
  - Contamination rate: 10.0%

PHASE 4: Running Subset Scanning Detection
----------------------------------------
✓ Raw pixel detection completed!
✓ Feature-based detection completed!

PHASE 5: Evaluation and Visualization
----------------------------------------
DETECTION PERFORMANCE COMPARISON
============================================================
Metric               Raw Pixels      Features        
------------------------------------------------------------
ROC AUC              0.892           0.945           
PR AUC               0.756           0.823           
F1 Score             0.723           0.789           
Precision            0.689           0.756           
Recall               0.760           0.823           

PHASE 6: Qualitative Visualization
----------------------------------------
✓ Original vs adversarial visualization saved
✓ Detection scores distribution saved
✓ Detected anomalies visualization saved
✓ Perturbation analysis saved

🎯 SUCCESS: Subset scanning successfully detected the hidden adversarial subgroup!

🖼️ Qualitative Visualizations

The showcase generates several qualitative visualizations that provide intuitive understanding:

Interactive Display

All visualizations are displayed interactively in separate windows, allowing you to:

• Zoom and pan to examine details
• Save screenshots using the plot window controls
• Explore different aspects of the visualizations
• Close windows to proceed to the next visualization

1. Original vs Adversarial Images

• Side-by-side comparison of clean and perturbed images
• Perturbation visualization showing the subtle differences
• Magnified perturbations to highlight imperceptible changes
• Perturbation statistics (L∞ and L2 norms)

See the Visual Showcase section above for an example of this visualization.

2. Detection Score Analysis

• Score distributions for clean vs adversarial samples
• Box plots and violin plots showing separation
• ROC curves demonstrating detection performance

See the Visual Showcase section above for an example of this analysis.

3. Detected Anomalies

• True positives: Adversarial images correctly detected
• False positives: Clean images incorrectly flagged
• Score rankings: Top detected anomalies
• Threshold analysis: Detection boundary visualization

4. Perturbation Analysis

• Statistical patterns across the dataset
• Average perturbation patterns showing systematic changes
• Perturbation variance highlighting consistent vs random changes
• Pixel-level analysis of where perturbations occur most

See the Visual Showcase section above for an example of this analysis.

🛠️ Usage Examples

Basic Demo

from src.scanning.simple_subset_scanning_demo import main
main()

Comprehensive Demo

from src.scanning.run_subset_scanning_demo import main
main()

Custom FGSM Attack

from src.attacks.fgsm_attack import generate_fgsm_attack
adv_data, adv_preds, success_rate = generate_fgsm_attack(
    model, test_data, test_labels, eps=0.3
)

Subset Scanning Detection

from art.defences.detector.evasion import SubsetScanningDetector
from src.utils.art_utils import create_art_classifier

classifier = create_art_classifier(model)
detector = SubsetScanningDetector(classifier, window_size=784)
scores, p_values, scan_stats = detector.detect(x_combined_flat)

Qualitative Visualization

from src.visualization.qualitative_analysis import create_comprehensive_qualitative_analysis

create_comprehensive_qualitative_analysis(
    clean_data, adversarial_data, true_labels, adv_predictions, 
    scores, y_true, output_dir='my_visualizations'
)

📚 Technical Details

Subset Scanning Algorithm

• Statistical Approach: Uses likelihood ratio tests to identify anomalous subgroups
• Unsupervised Detection: No labels required during detection
• Scalable: Efficient implementation for large datasets

ART Integration

• PyTorchClassifier: Wrapper for PyTorch models
• SubsetScanningDetector: Main detection class
• Configurable Parameters: Window size, scoring functions, etc.

Evaluation Metrics

• ROC AUC: Area under the receiver operating characteristic curve
• PR AUC: Area under the precision-recall curve
• F1 Score: Harmonic mean of precision and recall
• Confusion Matrix: Detailed classification results

Visualization Features

• Multi-panel layouts: Comprehensive analysis in single figures
• Color-coded results: Intuitive distinction between clean and adversarial
• Statistical overlays: Thresholds, confidence intervals, and distributions
• High-resolution output: Publication-ready visualizations

🤝 Contributing

This project demonstrates the integration of subset scanning with ART. Contributions are welcome!

📄 License

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

🙏 Acknowledgments

• ART Team: For the Adversarial Robustness Toolbox
• Subset Scanning Research: For the statistical foundation
• MNIST Dataset: For the standard benchmark dataset

Contact

For questions and support, please open an issue on GitHub or contact the maintainers.

---

Note: This is a demonstration project. For production use, please ensure proper testing and validation of all algorithms and implementations.