https://github.com/Haochen-Luo/CroPA

https://github.com/Haochen-Luo/CroPA

Last synced: 10 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/Haochen-Luo/CroPA
• Owner: Haochen-Luo
• Created: 2024-03-10T14:13:16.000Z (9 months ago)
• Default Branch: main
• Last Pushed: 2024-06-16T07:49:15.000Z (6 months ago)
• Last Synced: 2024-08-12T08:13:04.229Z (4 months ago)
• Language: Python
• Size: 15.4 MB
• Stars: 22
• Watchers: 1
• Forks: 2
• Open Issues: 3
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• Awesome-MLLM-Safety - Github - Luo/CroPA.svg?style=social&label=Star) (Attack)
• Awesome-LVLM-Attack - Github

README

        
# CroPA_ICLR24

This repository contains the code and data for the paper ["An Image Is Worth 1000 Lies: Transferability of Adversarial Images across Prompts on Vision-Language Models"](https://openreview.net/forum?id=nc5GgFAvtk) accepted by ICLR 2024.

## Overview

Overview of CroPA 

![Cropa Overview](Cropa.png)

## Requirements

The environment.yml file contains the necessary packages to run the code. You can create the environment using the following command:

```

conda env create -f environment.yml

```

We provided a customized version of the Huggingface's transformers library. The code is based on the transformers library version 4.33.2.

## Train

To obtain the perturbation, you can use the following command:

```

python main.py --model_name {VLM_NAME} --prompt_num {NUM_OF_PROMPTS} 

```

## Inference

We provide the pretrained perturbation in the model/ directory. To run the inference, you can use the following command:

```bash

python inference.py  --noise_dir /path/to/perturbation

```

Pretrained perturbation can be found here: https://1drv.ms/u/c/4f881fa19ba8dfee/ERYe-4sAhPVEqbujubnuWxUByYtM676mnW8FQzaBkxtF-w?e=bIeOCh

## Clarifications

- Open-Flamingo model used in the paper

    ```py

    from huggingface_hub import hf_hub_download import torch

    checkpoint_path = hf_hub_download("openflamingo/OpenFlamingo-9B-deprecated", "checkpoint.pt")

    ```

- Dataset used in the paper

    The dataset used in the paper is a subset of MSCOCO validation set. We provide the dataset in the data/ directory.

## Acknowledgement

We would like to thank the authors of the following repositories for their code: https://github.com/mlfoundations/open_flamingo/

## Citation

If you find this repository useful, please consider citing our paper:

```

@inproceedings{

luo2024an,

title={An Image Is Worth 1000 Lies: Transferability of Adversarial Images across Prompts on Vision-Language Models},

author={Haochen Luo and Jindong Gu and Fengyuan Liu and Philip Torr},

booktitle={The Twelfth International Conference on Learning Representations},

year={2024},

url={https://openreview.net/forum?id=nc5GgFAvtk}

}

```