Adversarial machine learning

This repository contains an exploration of adversarial attacks and defenses. Some simple gradient-driven attacks are implemented and tested. Moreover, a demonstration of adversarial training is given. It uses adversarial examples during training in order to robustify the model.

Everything here is implemented with PyTorch and Lightning. The dedicated ART library is employed in addition to that. It provides a unified NumPy-based API for adversarial ML that, under the hood, supports all major deep learning frameworks.

Notebooks

• Introduction

• Adversarial attacks (ART)

• Adversarial attacks (PyTorch)

• Adversarial attacks (Hugging Face)

• Adversarial training (Hugging Face)

Installation

pip install -e .

Training

python scripts/main.py fit --config config/std_train.yaml

python scripts/main.py fit --config config/adv_train.yaml