This is the official code for our paper Efficient Robust Training via Backward Smoothing (aceepted by AAAI'2022) by Jinghui Chen (PSU), Yu Cheng (Microsoft), Zhe Gan (Microsoft), Quanquan Gu (UCLA), Jingjing Liu (Tsinghua University).
- Python (3.6.9)
- Pytorch (1.7.1)
- CUDA
- numpy
alpha
: step size for perturbationepsilon
: input space perturbation strengthgamma
: output space perturbation strengthbeta
: TRADES robust regularization parameter
- Train Backward Smoothing on CIFAR10 using Resnet-18:
$ python3 train_trades_backward.py --arch resnet --dataset cifar10 --beta 10.0 --gamma 1.0 --alpha 0.031 --epsilon 0.031
For technical details and full experimental results, please check the paper.
@inproceedings{chen2022efficient,
author = {Chen, Jinghui and Cheng, Yu and Gan, Zhe and Gu, Quanquan and Liu, Jingjing},
title = {Efficient robust training via backward smoothing},
booktitle = {AAAI},
year = {2022}
}