Skip to content

Latest commit

 

History

History
50 lines (36 loc) · 2.18 KB

README.md

File metadata and controls

50 lines (36 loc) · 2.18 KB

Precision-Guided Adversarial Attack

The official repository for Precision-Guided Adversarial Attack (PG-Attack).

First-Place in the CVPR 2024 Workshop Challenge: Black-box Adversarial Attacks on Vision Foundation Models

Paper: PG-Attack: A Precision-Guided Adversarial Attack Framework Against Vision Foundation Models for Autonomous Driving (https://arxiv.org/abs/2407.13111)

Brief Introduction

Vision foundation models are widely used in autonomous driving, but they are vulnerable to adversarial attacks, risking vehicle safety. We propose PG-Attack, a framework combining Precision Mask Perturbation (PMP-Attack) and Deceptive Text Patch (DTP-Attack). PMP-Attack precisely targets regions to maximize impact, while DTP-Attack adds misleading text patches, confusing the model’s scene understanding. Our method effectively deceives advanced multi-modal models like GPT-4V, Qwen-VL, and imp-V1.

Prepare checkpoints

Please download the model file from the following link

Framework

Visualization

Citation

If you find our paper interesting or helpful to your research, please consider citing it, and feel free to contact [email protected] if you have any questions.

@article{fu2024pgattack,
  title={PG-Attack: A Precision-Guided Adversarial Attack Framework Against Vision Foundation Models for Autonomous Driving}, 
  author={Jiyuan Fu and Zhaoyu Chen and Kaixun Jiang and Haijing Guo and Shuyong Gao and Wenqiang Zhang},
  journal={arXiv preprint arXiv:2407.13111},
  year={2024}
}

@misc{fu2024cmi,
  title={Improving Adversarial Transferability of Vision-Language Pre-training Models through Collaborative Multimodal Interaction}, 
  author={Jiyuan Fu and Zhaoyu Chen and Kaixun Jiang and Haijing Guo and Jiafeng Wang and Shuyong Gao and Wenqiang Zhang},
  year={2024},
  eprint={2403.10883},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
}

License

The project is only free for academic research purposes but has no authorization for commerce.