ECSO

This repository contains the implementation of the paper:

ECSO: Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation
Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Aoxue Li, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, Yu Zhang
European Conference on Computer Vision (ECCV), 2024

Installation

1. Clone this repository and navigate to ECSO folder.

   git clone https://github.com/gyhdog99/ecso/
   cd ECSO-main

2. Install Package

    conda create -n ecso python=3.10 -y
    conda activate ecso
    pip install --upgrade pip  # enable PEP 660 support
    pip install -e .

Demo

We show the 4 core steps (i.e, 1. direct answer, 2. harm detect, 3. query-aware I2T caption, 4. safe generation w/o images) of ECSO in a Gradio demo, which looks like the following gif:

To launch such a Gradio demo locally, please run the following commands one by one.

Launch a controller

python -m llava.serve.controller --host 0.0.0.0 --port 10000

Launch a gradio web server

python -m llava.serve.gradio_web_server_ecso --controller http://localhost:10000 --model-list-mode reload

You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker

Launch a model worker

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path llava-v1.5-7b

Wait until the process finishes loading the model and you see "Uvicorn running on ...". Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.

Evaluation on Safety Benchmarks

Data/Model Preparation

Download VLSafe, MM-SafetyBench and COCO images

VLSafe

1. Generate direct/ECSO responses

   bash scripts/v1_5/eval_safe/gen_vlsafe.sh
   bash scripts/v1_5/eval_safe/gen_vlsafe_tell_ask.sh

2. Evaluation

   bash llava/eval/eval_vlsafe.sh

MM-SafetyBench

1. Generate direct/ECSO responses

   bash scripts/v1_5/eval_safe/gen_mmsafe.sh
   bash scripts/v1_5/eval_safe/gen_mmsafe_tell_ask.sh

2. Evaluation

   bash llava/eval/eval_mmsafe_loop.sh

Evaluating Utilities on MLLM benchmarks

Data/Model Preparation

Follow the guideline to download the evaluation data of MME, MMBench and MM-Vet.

MME

Generate direct/ECSO responses

bash scripts/v1_5/eval/mme.sh
bash scripts/v1_5/eval_safe/gen_mme_unsafe_ask.sh

MMBench

Generate direct/ECSO responses

bash scripts/v1_5/eval/mmbench.sh
bash scripts/v1_5/eval_safe/gen_mmbench_unsafe_ask.sh.sh

MM-Vet

Generate direct/ECSO responses

bash scripts/v1_5/eval/mmvet.sh
bash scripts/v1_5/eval_safe/gen_mm-vet_unsafe_ask.sh.sh

Acknowledgement

• LLaVA This repository is built upon LLaVA!

Citation

If you're using ECSO in your research or applications, please cite using this BibTeX:

@article{gou2024eyes,
  title={Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation},
  author={Gou, Yunhao and Chen, Kai and Liu, Zhili and Hong, Lanqing and Xu, Hang and Li, Zhenguo and Yeung, Dit-Yan and Kwok, James T and Zhang, Yu},
  journal={arXiv preprint arXiv:2403.09572},
  year={2024}
}