PGD Adversarial Training

A classical adversarial training method that uses projected gradient descent (PGD) to generate adversarial examples. Proposed by Madry et al., it is one of the benchmark methods for evaluating adversarial robustness.