CURE

Towards Universal Certified Robustness with Multi-Norm Training [Arxiv]
Enyi Jiang, David S. Cheung, Gagandeep Singh

Setup

Create and activate a conda environment

conda create --name CURE python=3.10.4
conda activate CURE

Install the requirements

conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Add the main directory to the PYTHONPATH (make sure you are in the top level directory)

export PYTHONPATH=$PWD:$PYTHONPATH

To download the TinyImageNet dataset navigate into the data directory and execute:

bash tinyimagenet_download.sh

Training

All training scripts with different baselines and CURE framework can be found in the scripts folder, which consists of experiments on MNIST, CIFAR-10, and TinyImagenet.

Evaluation

One can use the onnx file from training outputs and verify the model using alpha-beta-crown. The config files can be found in configs_abc folder.

Evaluation of Geometric Transformations

Please refer to the GitHub https://github.com/uiuc-arc/CGT.

Evaluation of Patch Attacks

Please refer to the GitHub https://github.com/Ping-C/certifiedpatchdefense.

Credits

Parts of the code in this repo is based on https://github.com/eth-sri/SABR.

Citation

Cite the paper/repo:

@article{jiang2024towards,
  title={Towards Universal Certified Robustness with Multi-Norm Training},
  author={Jiang, Enyi and Singh, Gagandeep},
  journal={arXiv preprint arXiv:2410.03000},
  year={2024}
}