Effective and Robust Adversarial Training Against Data and Label Corruptions

This is the official PyTorch repository for the implementation of ERAT.

Requirements

Python3

PyTorch (> 1.0)

Prepare datasets

1. Download datasets

python downloaddata.py

2. Generate perturbed training data The code for generating perturbed data can be accessible in Delusive-Adversary, DeepConfuse, Unlearnable-Examples.

Train

python Dual_main.py

If you find this code helpful for your research, please consider citing our paper:

@article{zhang2024effective,
  title={Effective and Robust Adversarial Training Against Data and Label Corruptions},
  author={Zhang, Peng-Fei and Huang, Zi and Xu, Xin-Shun and Bai, Guangdong},
  journal={IEEE Transactions on Multimedia},
  year={2024},
  publisher={IEEE}
}

Acknowledge

Some of our code and datasets are based on DivideMix.