Adversarial Attacks and Defences

Adversarial Attacks and Defences (AAD) is a Python framework for defending machine learning models from adversarial examples.

Required Libraries

• PyTorch (www.pytorch.org)
• ART (https://github.com/IBM/adversarial-robustness-toolbox)

Build And Install The Package

Install module as a package using setuptools

1. Use a virtual environment

   virtualenv --system-site-packages -p python3 ./venv
   source ./venv/bin/activate

2. Install you project from pip

   pip install --upgrade pip
   pip install -e .
   pip check
   pip freeze

3. Run the code demo from Jupyter Lab

   cd ./examples
   jupyter lab

4. Run the script from terminal

   python ./cmd/train.py -d MNIST -e 5 -vw

Code Structure

root
├─┬ aad
│ ├── attacks    # modules of adversarial attacks
│ ├── basemodels # modules of base classification models
│ ├── datasets   # data loader helper module
│ └── defences   # modules of adversarial defences
├── cmd          # scripts for terminal interface
├── data         # dataset
├── examples     # code examples
├── log          # logging files
├── save         # saved pre-trained models
├── tests        # unit tests

Run Script From Terminal

The terminal scripts are separated into 3 parts: train, attack and defence.

• To train a model:

  python ./cmd/train.py --help

• To attack a model:

  python ./cmd/attack.py --help

• To defend a model:

  python ./cmd/defend_ad.py --help

Examples

Examples are available under the ./examples/ folder.