Adversarial Training Project
Adversarial training is a defensive technique in machine learning where models are trained not only on clean data, but also on adversarially perturbed examples. By exposing the model to attacks during training, it can build resilience against them. This project demonstrates implementations using MNIST, FGSM, and PGD attacks.
- Technologies Used: Python, PyTorch, Jupyter Notebook, NumPy, Matplotlib, torchvision (MNIST dataset)
- Goals: Implement adversarial training on MNIST, evaluate robustness against FGSM and PGD attacks, and compare clean vs adversarial accuracy.
- Challenges: Balancing model accuracy with robustness, ensuring stable training with adversarial examples, and tuning attack parameters (epsilon, step size, iterations).
