This repository contains the PyTorch implementation of our ICLR 2023 paper titled "Is Adversarial Training Really a Silver Bullet for Mitigating Data Poisoning?".

python train_reference_model.py --robust_eps 4 --reference_path ./reference_model

You can control the robustness of the reference model by ajusting --robust_eps parameter. The reference model will be saved at --reference_path. This file can also be used to evaluate the attack performance, with poisoned data as input.

python get_centroid.py --centroid_path ./centroid

The centroid will be saved at --centroid_path

python poison_generate.py --eps 8 --recipe push

The poison budget can be controlled by adjusting --eps. You can select the poisoning method by setting --recipe push corresponding to EntF-Push or --recipe pull corresponding to EntF-Pull.

@inproceedings{wen2023is,
    title={Is Adversarial Training Really a Silver Bullet for Mitigating Data Poisoning?},
    author={Rui Wen and Zhengyu Zhao and Zhuoran Liu and Michael Backes and Tianhao Wang and Yang Zhang},
    booktitle={International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=zKvm1ETDOq}
}

If you are interested in our work, feel free to drop me an email at [email protected]

We would like to acknowledge the work of Fowl et al. for their excellent framework for generating poisons based on adversarial examples. Our code leverages their code for the gradient descent and poison saving parts.