Home

Awesome

Jailbreak Large Vision-Language Models Through Multi-Modal Linkage

Code for the paper Jailbreak Large Vision-Language Models Through Multi-Modal Linkage

Data

We uploaded the encrypted images here

Run before Download it as ./dataset

Attack Commands

python attack.py --dataset 'safebench' \
    --data-path 'dataset' \
    --save-dir 'save_dir' \
    --image-format 'images_wr'\

You can choose the encryption or attack methods by replacing the image-format parameter. Here are some options:

Main Results

Reference

If you find the code useful for your research, please consider citing

@article{wang2024jailbreak,
         title={Jailbreak Large Vision-Language Models Through Multi-Modal Linkage}, 
         author={Wang, Yu and Zhou, Xiaofei and Wang, Yichen and Zhang, Geyuan and He, Tianxing},
         journal={arXiv preprint arXiv:2412.00473},
         year={2024}
}