This repository includes the official PyTorch implementation of our paper:

Adversarial Purification with Score-based Generative Models

Jongmin Yoon, Sung Ju Hwang, Juho Lee

the 38th International Conference for Machine Learning (ICML 2021)

ArXiv: https://arxiv.org/abs/2106.06041

We propose a method that gives adversarial robustness to a neural network model against (stochastic) adversarial attacks by using an Energy-based Model (EBM) trained with Denoising Score Matching (DSM), which is called Adversarial denosing purification (ADP).

Run the following command to install some necessary python packages to run our code.

pip install -r requirements.txt

To run the experiments with adp.py or adp_decision.py, enter the following command.

python main.py --config <config-file>

For example, we provide the example configuration file configs/cifar10_bpda_eot_sigma025_eot15.yml in the repository.

For adversarial attacks, the classifier PGD attack and BPDA+EOT attack are implemented in attacks/clf_pgd.py and attacks/bpda_strong.py, respectively. At the configuration file, setting the attack.attack_method into clf_pgd or bpda_strong will run these attacks, respectively. For defense, we implemented the main ADP algorithm and ADP after detecting adversarial examples (Appendix F.) in purification/adp.py and purification/adp_decision.py, respectively.

• For the configuration files, we use the pixel ranges [0, 255] for the perturbation scale attack.ptb and the one-step attack scale attack.alpha. And the main experiments are performed within the pixel range [0, 1] after being rescaled during execution.
• For training the EBM and classifier models, we primarily used the pre-existing methods such as NCSNv2 and WideResNet classifier. Here is the repository we used for training the WideResNet classifier. Nevertheless, other classifiers, such as the pre-trained adversarially robust classifier implemented in here can be used.

If you find our work useful for your research, please consider citing this.

@inproceedings{
yoon2021advpur,
title={Adversarial Purification with Score-based Generative Models},
author={Jongmin Yoon and Sung Ju Hwang and Juho Lee},
booktitle={Proceedings of The 38th International Conference on Machine Learning (ICML 2021)},
year={2021},
}

For further details, please contact [email protected].

MIT