Model Scanning

This project contains tools for model security analysis, including serialization attack detection, membership inference attacks, and adversarial example generation.

Setup

1. Create and activate a Python virtual environment:

python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
venv\Scripts\activate  # Windows

cd ModelScanning

2. Install dependencies:

pip install -r requirements.txt

Project Structure

• serialisation/ - For detecting malicious model serialization

  • scanner.py - Model file scanner
  • inject.py - Pickle injection utilities
  • helper.ipynb - Helper notebook for model downloading and injection

• mia/ - Membership Inference Attack implementation

  • cifar10.py - MIA example on CIFAR10
  • utils.py - Shadow model utilities

• adversarial/ - Adversarial example generation and detection

  • adv_pgd.py - PGD attack implementation
  • imagenet_classes.txt - ImageNet class labels

Usage

Serialization Attack Detection

You can use the helper.ipynb file to download model and inject a malicious code.

python serialisation/scanner.py /path/to/model.pt

Membership Inference Attack

python mia/cifar10.py --target_epochs 12 --attack_epochs 6 --num_shadows 3

Adversarial Example Generation

cd adversarial
python adv_pgd.py

Requirements

See requirements.txt for full list of dependencies. Key requirements:

• PyTorch
• TensorFlow
• scikit-learn
• numpy
• matplotlib
• tqdm