This repository contains code to reproduce results from the paper:
Attention! Your Vision Language Model Could Be Maliciously Manipulated (NeurIPS 2025)
Xiaosen Wang, Shaokang Wang, Zhijin Ge, Yuyang Luo, Shudong Zhang
To create a conda environment, use the following command:
conda env create -f environment.yml
The original prompt and target output are written in data.json. To run the attack, you can run the script as follows:
CUDA_VISIBLE_DEVICES=gpu_id python main.py
After finishing the attack, you can evaluate the adversarial attack as follows:
CUDA_VISIBLE_DEVICES=gpu_id python main.py --eval
For more parameters, you can run the script as follows:
python main.py --help
If our paper or this code is useful for your research, please cite our paper:
@inproceedings{wang2025attention,
title={{Attention! Your Vision Language Model Could Be Maliciously Manipulated}},
author={Xiaosen Wang and Shaokang Wang and Zhijin Ge and Yuyang Luo and Shudong Zhang},
journal={Advances in Neural Information Processing Systems (NeurIPS)},
year={2025}
}
















