This was my first hands-on implementation of adversarial machine learning where I implemented the PGD (projected gradient descent) white-box attack from Madry et al. against LeNet on MNIST and ResNet18 on CIFAR-10, then systematically analyzed how attack hyperparameters, training augmentations, and regularization techniques affect a model's vulnerability. Finally, I defended both models using adversarial training, reducing the accuracy gap from ~45–65% down to just 2–13% with nearly no cost to clean accuracy. This project gave me hands-on experience with adversarial example crafting, robustness evaluation, and the practical tradeoffs between model utility and security.An implementation and analysis of PGD adversarial attacks on image classifiers. By iteratively perturbing inputs within a bounded ε-ball, the attack reliably fools undefended models. Training models on these attack-generated examples produces models robust to PGD with minimal clean accuracy sacrifice.

	LeNet - MNIST	Resnet18 - CIFAR10
Clean Accuracy	99.41%	88.31%
Adversarial Accuracy	54.78%	23.37%
Accuracy Difference	44.63%	64.94%

Adversarially Trained	LeNet - MNIST	Resnet18 - CIFAR10
Clean Accuracy	99.50%	86.75%
Adversarial Accuracy	97.43%	72.97%
Accuracy Difference	2.07%	13.78%