README.md

Augmented Unlearnable Examples (AUE)
Augmented Adversarial Poisoning (AAP)

Code for NeurIPS 2024 paper "Efficient Availability Attacks against Supervised and Contrastive Learning Simultaneously".

Environment

The repository is tested on the following environment:

python=3.11.4
pytorch=2.0.1
torchvision=0.15.2
kornia=0.7.0
numpy=1.25.2

Example Scrtipts

In example.sh, we provide generation and evaluation scripts for AUE and AAP attacks on CIFAR-10 and CIFAR-100.

Generation

• For AUE attack, run the following command:

python aue.py --dataset <dataset> --backbone <model> --eps <eps> --strength <s> 

• For AAP attack, run the following command:

python aap.py --dataset <dataset> --backbone <model> --eps <eps>  --ref-strength <s1> --gen-strength <s2> \   
    [--untargeted] 

Activate the --untargeted flag for untargeted AAP.

Evaluation

When evaluating the attack performance of a pretrained poison, specify the flag --poison-path <.../poison.pt>. When training baseline models, delete the flag --poison-path.

• Supervised learning

python evaluation_sl.py --dataset <dataset> --backbone <model> --experiment <poison_name> \
    --poison-path  <.../poison.pt> 

• Contrastive learning

python evaluation_cl.py --dataset <dataset> --backbone <model> --experiment <poison_name> \
    --method [SimCLR/MoCo/BYOL/SimSiam] \
    --poison-path  <.../poison.pt> 

Citation

@inproceedings{WZG24,
    title={Efficient Availability Attacks against Supervised and Contrastive Learning Simultaneously},
    author={Yihan Wang and Yifan Zhu and Xiao-Shan Gao},
    booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
    year={2024},
    url={https://openreview.net/forum?id=FbUSCraXEB}
}