Combating-Adversaries-with-Anti-Adversaries

This is the official repor for the work: "Combating Adversaries with Anti-Adversaries"

This work is accepted to AAAI Conference on Artificial Intelligence (AAAI'22).

preprint: https://arxiv.org/pdf/2103.14347.pdf

Environment Installations:

First, you need to install the environment from the provided yml file by running:

conda env create -f anti_adv.yml

Then, activate the envionment by running:

conda activate anti_adv

Reproducing our Results:

To reproduce the results against AutoAttack (APGD, ADLR, FAB and Square attacks) for AWP and AWP+Anti-Adv, you need to download the weights of AWP model for CIFAR 10 from their official repo. Then, go to adv_weight_pert.py and change the WEIGHTS_PATH argument to the path were you saved the weights in.

Then, run the following command for AWP+Anti-Adv.

python main.py --checkpoint runs --experiment awp --k 2 --alpha 0.15

and run the following for AWP.

python main.py --checkpoint runs --experiment awp --k 0 --alpha 0.0

For other baselines, copy their model defintion in the experiments directory in a similar fashion to adv_weight_pert.py, and import their get_model function in utils.py line 166. Finally, add the experiment name to the opts.py in the EXP_CHOICES argument.

Wanna Use the Anti-Adversary Layer with Your Classifier ?

Simply use the wrapper in Anti_Adv.py. The wrapper takes the following arguments:

model: a PyTorch trained model

mean: a list or a numpy arry containing the mean used to normalize the data.

std: a list or a numpy arry containing the standard deviation used to normalize the data.

k: a scalar representing the number of iterations used to compte the anti-adversary.

alpha: a scalar representing the learning rate used to compute the anti-adversary.

Citation

If you find our work useful, please consider citing it as

@misc{alfarra2021combating,
      title={Combating Adversaries with Anti-Adversaries}, 
      author={Motasem Alfarra and Juan C. Pérez and Ali Thabet and Adel Bibi and Philip H. S. Torr and Bernard Ghanem},
      year={2021},
      eprint={2103.14347},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

This project is licensed under the terms of the MIT license.