VLMToolMisuse

This is the official code implementation for the paper Misusing Tools in Large Language Models With Visual Adversarial Examples.

Environment

You can create a conda enviroment to run the code following create_env.sh.

Description

Currently, we only support one Vision Language Model, the LLaMAAdapter Model. You need the original checkpoints of the LLaMA model downloaded, as specified here.

Modeling: model.py and LLaMAAdapterModel.py contains code to support tuning the LLaMAAdapter Model.

Training: train_adversarial_image.py contains the code to conduct the adversarial training and evaluation. run.sh is an example execution script with some pre-defined hyper-parameters.

Tool Use Templates: All our tool use templates are placed in data/attacks/Attack.py

Images: The original images are placed in data/images/. data/images_pp/ contains the preprocessed images that will actually be used. Note that preprocessing could be different for different Vision Language Models.

Citation

If you find our code helpful, please consider citing

@article{fu2023misusing,
  title={Misusing Tools in Large Language Models With Visual Adversarial Examples},
  author={Fu, Xiaohan and Wang, Zihan and Li, Shuheng and Gupta, Rajesh K and Mireshghallah, Niloofar and Berg-Kirkpatrick, Taylor and Fernandes, Earlence},
  journal={arXiv preprint arXiv:2310.03185},
  year={2023}
}