Group-based Robustness

This is the official code repository for the NDSS'24 paper: "Group-based Robustness: A General Framework for Customized Robustness in the Real World" (https://arxiv.org/pdf/2306.16614.pdf).

We propose the following to better evaluate the threat of evasion attacks in realistic attack scenarios.

1. A new metric, group-based robustness, that empirically evaluates the threat more accurately.
2. Two new loss functions, MDMAX and MDMUL, that more efficiently estimate the threat.
3. Several attack strategies that more efficiently estimate the threat.
4. A defense approach that mitigates the threat.

Usage

You may find the instructions to reproduce our results in each directory corresponding to each dataset. GPUs are highly recommended for all experiments. We also highly recommend the use of virtual environments for all experiments.

Citation

@inproceedings{weiran2024GBR,
    title = {Group-based Robustness: A General Framework for Customized Robustness in the Real World},
    author = {Weiran Lin and Keane Lucas and Neo Eyal and Lujo Bauer and Michael K. Reiter and Mahmood Sharif},
    booktitle = {Proceedings of the Network and Distributed System Security Symposium (NDSS)},
    year = {2024}
}

Acknowledgement

Our code is heavily based on the following implementations:

• Autoattack (https://github.com/fra31/auto-attack)
• robustness package (https://github.com/MadryLab/robustness)
• T-PGD (https://github.com/Phantivia/T-PGD)
• Defense against Occlusion Attacks (DOA) (https://github.com/tongwu2020/phattacks)

We greatly appreciate authors of the above for their fantastic and innovative work.

We also reused some of the code from our previous CGD attack implementation (https://github.com/linweiran/CGD), which you are very welcome to check out!