A stab at robustness through MadryLab's library.
The aim is to explore https://github.com/MadryLab/ robustness package. FashionMnist was added to the dataloaders, to be adv-trained and compared to the std-trained model accuracy.
CINIC was used as well. For both sets, the following 3 metrics were computed
- Standard training of a model against the natural(original set) and then validated against an PGD adversary
- PGD-trained model validated against the natural and the adversarial dataset
- Standard model trained on Adversarial dataset, then validated aginst natural and adv dataset.