Reproduce “BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunhan Zhao, Xiang Zheng, Lin Luo, Yige Li, Xingjun Ma, Yu-Gang Jiang
A black-box defense framework, BLUESUFFIX, was developed to protect Vision-Language Models (VLMs) from jailbreak attacks by combining diffusion-based image purification, LLM-based text purification, and a reinforcement learning-tuned cross-modal suffix generator. The system effectively reduced attack success rates to near-zero for various attack types on diverse VLMs, maintaining high performance on legitimate prompts.
Key insights
- Reinforcement learning, guided by an external judge and optimized for cross-modal VLM responses, can effectively generate suffixes that robustify VLM safety mechanisms.
- A combination of unimodal input purification with a proactively generated "blue-team" suffix offers superior defense against diverse jailbreak attacks, including complex cross-modal ones.
- The generated suffixes empirically enhance the VLM's utilization of cross-modal information for safer outputs, as validated by mutual information analysis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.