🔍 ARES 2.0 (Adversarial Robustness Evaluation for Safety) is a Python library dedicated to adversarial machine learning research. It aims at benchmarking the adversarial robustness of image classification and object detection models, and introduces mechanisms to defend against adversarial attacks through robust training.
- Developed on Pytorch.
- Supports various attacks on classification models.
- Employs adversarial attacks on object detection models.
- Provides robust training for enhanced robustness and various trained checkpoints.
- Enables distributed training and testing.
-
Optional: Initialize a dedicated environment for ARES 2.0.
conda create -n ares python==3.10.9 conda activate ares -
Clone and set up ARES 2.0 via the following commands:
git clone https://github.com/thu-ml/ares2.0 cd ares2.0 pip install -r requirements.txt mim install mmengine==0.8.4 mim install mmcv==2.0.0 mim install mmdet==3.1.0 pip install -v -e .
- For robustness evaluation of image classification models against adversarial attacks, please refer to classification.
- For robustness evaluation of object detection models, please refer to detection.
- For methodologies on robust training, please refer to robust-training.
📚 Access detailed tutorials and API docs on strategies to attack classification models, object detection models, and robust training here.
If you derive value from ARES 2.0 in your endeavors, kindly cite our paper on adversarial robustness, which encompasses all models, attacks, and defenses incorporated in ARES 2.0:
@inproceedings{dong2020benchmarking,
title={Benchmarking Adversarial Robustness on Image Classification},
author={Dong, Yinpeng and Fu, Qi-An and Yang, Xiao and Pang, Tianyu and Su, Hang and Xiao, Zihao and Zhu, Jun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages={321--331},
year={2020}
}