Skip to content

PyTorch-1.0 implementation for the adversarial training on MNIST/CIFAR-10 and visualization on robustness classifier.

Notifications You must be signed in to change notification settings

ylsung/pytorch-adversarial-training

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

27 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Adversarial Training and Visualization

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.

Adversarial Training

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.

Visualization

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.

Implementation

Part of the codes in this repo are borrowed/modified from [2], [3], [4] and [5].

References:

[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

Contact

Yi-Lin Sung, corumlouis123@gmail.com

About

PyTorch-1.0 implementation for the adversarial training on MNIST/CIFAR-10 and visualization on robustness classifier.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published