Code for ECML-PKDD 2023 "A3T: Accuracy Aware Adversarial Training"
python train_resnet.py # to test with ResNet18
python train_wideresnet.py # to test with WideResNet
python glue.py $dataset $adv_method
- The dataset parameter can take a value of one of the GLUE datasets.
- The adv_method parameter must be one of the "clean", "AT", and "A3T" as an adversarial attack method. Each dataset should first run with a 'clean' argument, then one of 'AT' or 'A3T' can be used.
If you use this code in your work, please cite the accompanying paper:
@article{altinisik2022a3t,
title={A3T: Accuracy Aware Adversarial Training},
author={Altinisik, Enes and Messaoud, Safa and Sencar, Husrev Taha and Chawla, Sanjay},
journal={arXiv preprint arXiv:2211.16316},
year={2022}
}
The vision part of the code is forked from the repository of YisenWang/MART.