SEAT (Self-Ensemble Adversarial Training)

This is the official code for ICLR'22 paper "Self-Ensemble Adversarial Training for Improved Robustness"

Prerequisites

• Python (3.7)
• Pytorch (1.5)
• Torchvision
• CUDA
• Numpy
• AutoAttack

Training and Testing

• Train ResNet-18 on CIFAR10:

  $ CUDA_VISIBLE_DEVICES={your GPU number} python3 seat.py 

• Train WRN-32-10 on CIFAR10

  $ CUDA_VISIBLE_DEVICES={your GPU number} python3 seat.py --arch 'WRN'

Then, it will automatically run all the robustness evaluation in our paper, including NAT, PGD20/100, MIM, CW, APGD_ce, APGD_dlr, APGD_t, FAB_t, Square and AutoAttack.

Citation

If you are interested in our work, please consider citing our paper:

@inproceedings{
wang2022selfensemble,
title={Self-ensemble Adversarial Training for Improved Robustness},
author={Hongjun Wang and Yisen Wang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=oU3aTsmeRQV}
}