- Train a Resnet on CIFAR 10 using standard training and PGD training and Analyze the class-distribution of misclassification using untargeted attacks.
- Perform PGD training with targeted adversarial training using adversarial examples with all target classes except ground truth for each sample.
- Repeat targeted PGD training considering different norms of robust losses across classes instead of averaging.
- Analyze class-wise natural and robust accuracies in each scenario.
Results: Standard training accuracy = 0.999 (100 epochs), Standard test accuracy = 0.930, PGD_training accuracy (Adversial Images + Orginal Images) = 0.943 (200 epochs ), PGD Test accuracy on adverisl images = 0.454, PGD test accuracy on orginal images = 0.850,