Unified Distributional Robustness

Introduction

Pytorch implementation for the paper "A Unified Wasserstein Distributional Robustness Framework for Adversarial Training" (accepted to ICLR 2022). Link to the paper.

Citation

Please consider to cite our paper if you find any useful in this source code or our paper

@inproceedings{
bui2022a,
title={A Unified Wasserstein Distributional Robustness Framework for Adversarial Training},
author={Anh Tuan Bui and Trung Le and Quan Hung Tran and He Zhao and Dinh Phung},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=Dzpe9C1mpiv}
}

Structure

• Cifar10-WideResNet: an implementation with WideResNet architecture, which is based on AWP's implementation. To reproduce Table 4 in our paper, including: UDR-AWP-AT, AWP-AT.
• Cifar10-Resnet18: an implementation with Resnet18 architecture (supports WideResNet as well). To reproduce Table 3 in our paper, including baseline methods: PGD-AT, TRADES, MART and our methods UDR-PGD, UDR-TRADES and UDR-MART.
• Toy2D: an implementation to demonstrate the benefit of soft-ball projection.