The repo is the PyTorch-1.0 implementation for the adversarial training on MNIST/CIFAR-10. And I also reproduce part of the visualization results in [1].
Note: Not an official implementation.
| Objective Function | |
|---|---|
| Standard Training | Adversarial Training |
| |
|
where p in the table is usually 2 or inf.
The objective of standard and adversarial training is fundamentally different. In standard training, the classifier minimize the loss computed from the original training data, while in adversarial training, it trains with the worst-case around the original data.
In [1], the authors discover that the features learned by the robustness classifier are more human-perceivable. Related results are shown in mnist/cifar-10 folder.
Part of the codes in this repo are borrowed/modified from [2], [3], [4] and [5].
[1] D. Tsipras, S. Santurkar, L. Engstrom, A. Turner, A. Madry. Robustness May Be at Odds with Accuracy, https://arxiv.org/abs/1805.12152
[2] https://github.com/MadryLab/mnist_challenge
[3] https://github.com/MadryLab/cifar10_challenge
[4] https://github.com/xternalz/WideResNet-pytorch
[5] https://github.com/utkuozbulak/pytorch-cnn-visualizations
Yi-Lin Sung, [email protected]