DorPatch: Distributed and Occlusion-Robust Adversarial Patch to Evade Certifiable Defenses

By Chaoxiang He and Bin Benjamin Zhu

This repository contains the code for the paper "DorPatch: Distributed and Occlusion-Robust Adversarial Patch to Evade Certifiable Defenses", published at the Network and Distributed System Security Symposium (NDSS) 2024.

Files

├── README.md                        #this file 
├── requirement.txt                  #required package
| 
├── main.py                          #entrance to run the code
├── utils.py                         #utils for the code
├── attack.py                        #code for attacking methods (e.g., DorPatch)
| 
├── defenses
|   └── PatchCleanser.py             #code for the SOTA certifiable defense of PatchCleanser
|
├── pretrained_models
|   └──imagenet                      #model directory for imagenet
|
└── results                          #directory for generated results

$dataset_directory
    └──imagenet                      #data directory for imagenet

Instructions to Run the Code

1. Create the Python environment using conda from requirements.txt:

   $ conda create -n pytorch_dorpatch --file requirements.txt -c pytorch -c nvidia

   and switch to the environment by

   $ conda activate pytorch_dorpatch

2. Prepare the dataset and pre-trained model.

   • Download the corresponding dataset (e.g., download the validation set of ImageNet from here) and prepare the data to the directory $dataset_directory, where $dataset_directory is a variable referring to the path that will store all datasets (e.g., if $dataset_directory = './data', then the resulting validation set of ImageNet is prepared to the directory './data/imagenet/val').

   • Download the corresponding pretrained model (e.g., 'resnetv2_50x1_bit_distilled_cutout2_128_imagenet.pth' for ImageNet from here, provided by PatchCleanser) and put it into "pretrained_models/$dataset/", where$dataset is a variable referring to the directory name of the corresponding dataset (e.g., $dataset = 'imagenet' for the ImageNet model dataset to match the above directory name of the ImageNet dataset in$dataset_directory).

3. Run the code with $ python main.py --data_dir $dataset_directory

Citation

If you are using our code for research purposes, please cite our paper.

@inproceedings{dorpatch_ndss24,
  author    = {Chaoxiang He and Xiaojing Ma and Bin Benjamin Zhu and Yimiao Zeng and 
               Hanqing Hu and Xiaofan Bai and Hai Jin and Dongmei Zhang},
  title     = {{DorPatch}: Distributed and Occlusion-Robust Adversarial Patch
               to Evade Certifiable Defenses},
  booktitle = {Network and Distributed System Security Symposium, {NDSS}, 2024,
               San Diego, CA, USA, February 26 - March 1, 2024},
  publisher = {The Internet Society},
  year      = {2024},
  url       = {https://dx.doi.org/10.14722/ndss.2024.24920}
}

Paper's Link

Our paper can be downloaded from here.