Exploring Memorization in Adversarial Training

This repository contains the code for the following paper

Exploring Memorization in Adversarial Training (ICLR 2022)

Yinpeng Dong, Ke Xu, Xiao Yang, Tianyu Pang, Zhijie Deng, Hang Su, and Jun Zhu

Citation

If you find our methods useful, please consider citing:

@inproceedings{dong2022exploring,
  title={Exploring Memorization in Adversarial Training},
  author={Yinpeng Dong and Ke Xu and Xiao Yang and Tianyu Pang and Zhijie Deng and Hang Su and Jun Zhu},
  booktitle={International Conference on Learning Representations},
  year={2022}
}

Prerequisites

  • Python (3.6.8)
  • Pytorch (1.3.0)
  • torchvision (0.4.1)
  • numpy

Training with Random Labels

On CIFAR-10

For PGD-AT

python train.py --wd 0 --noise-type label_symmetric --noise-rate 1.0

For TRAEDS

python train.py --wd 0 --noise-type label_symmetric --noise-rate 1.0 --loss-type trades

Overcome Robust Overfitting

On CIFAR-10

For PGD-AT + TE

python train_te.py

For TRADES + TE

python train_te.py --loss-type trades

GitHub

View Github