BSR

This repository contains code to reproduce results from the paper:

Boosting Adversarial Transferability by Block Shuffle and Rotation (CVPR 2024)

Kunyu Wang, Xuanran He, Wenxuan Wang, Xiaosen Wang†

We also include the torch version code in the framework TransferAttack.

Requirements

• Python >= 3.6.5
• Keras==2.4.3
• numpy == 1.19.2
• Tensorflow == 2.5.0
• scikit-image==0.19.3
• opencv-python >= 3.4.2
• scipy ==1.6.3
• pandas == 1.2.4
• imageio >= 2.6.1
• tf-slim == 1.1.0
• tensorflow-addons==0.14.0

Qucik Start

Prepare the data and models

The data and pretrained models are provided by previous work. You should download them and place the data and pretrained models in dev_data/ and models/, respectively.

BSR

All the provided codes generate adversarial examples on inception_v3 model. If you want to attack other models, replace the model in graph and batch_grad function and load such models in main function.

Runing attack

Taking BSR attack for example, you can run this attack as following:

CUDA_VISIBLE_DEVICES=gpuid python mi_bsr.py

Evaluating the attack

The generated adversarial examples would be stored in directory ./outputs. Then run the file simple_eval.py to evaluate the success rate of each model used in the paper:

CUDA_VISIBLE_DEVICES=gpuid python simple_eval.py

Acknowledgments

Codes refers to Admix

Citation

If our paper or this code is useful for your research, please cite our paper:

@inproceedings{wang2024boosting,
     title={{Boosting Adversarial Transferability by Block Shuffle and Rotation}},
     author={Kunyu Wang and Xuanran He and Wenxuan Wang and Xiaosen Wang},
     booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
     year={2024}
}