Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks
Paper • 2504.01308 • Published • 14
How to use from
SGLangUse Docker images
docker run --gpus all \
--shm-size 32g \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<secret>" \
--ipc=host \
lmsysorg/sglang:latest \
python3 -m sglang.launch_server \
--model-path "Jarvis1111/MiniGPT4-RobustVLGuard" \
--host 0.0.0.0 \
--port 30000# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "Jarvis1111/MiniGPT4-RobustVLGuard",
"prompt": "Once upon a time,",
"max_tokens": 512,
"temperature": 0.5
}'Quick Links
🚀 Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks
Welcome! This repository hosts the official implementation of our paper, "Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks."
Paper link: arxiv.org/abs/2504.01308
Project page:
🌟 What’s New?
We propose state-of-the-art solutions to enhance the robustness of Vision-Language Models (VLMs) against Gaussian noise and adversarial attacks. Key highlights include:
🎯 Robust-VLGuard: A pioneering multimodal safety dataset covering both aligned and misaligned image-text pair scenarios.
🛡️ DiffPure-VLM: A novel defense framework that leverages diffusion models to neutralize adversarial noise by transforming it into Gaussian-like noise, significantly improving VLM resilience.
✨ Key Contributions
- 🔍 Conducted a comprehensive vulnerability analysis revealing the sensitivity of mainstream VLMs to Gaussian noise.
- 📚 Developed Robust-VLGuard, a dataset designed to improve model robustness without compromising helpfulness or safety alignment.
- ⚙️ Introduced DiffPure-VLM, an effective pipeline for defending against complex optimization-based adversarial attacks.
- 📈 Demonstrated strong performance across multiple benchmarks, outperforming existing baseline methods.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Install from pip and serve model
# Install SGLang from pip: pip install sglang# Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Jarvis1111/MiniGPT4-RobustVLGuard" \ --host 0.0.0.0 \ --port 30000# Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Jarvis1111/MiniGPT4-RobustVLGuard", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'