Summary of Bluesuffix: Reinforced Blue Teaming For Vision-language Models Against Jailbreak Attacks, by Yunhan Zhao et al.
BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks
by Yunhan Zhao, Xiang Zheng, Lin Luo, Yige Li, Xingjun Ma, Yu-Gang Jiang
First submitted to arxiv on: 28 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a novel defense method, BlueSuffix, to protect Visual Language Models (VLMs) from jailbreak attacks. Existing methods are either unimodal or bimodal, enhancing specific modules or realigning text-image representations. However, these methods fail to fully exploit cross-modal information or compromise model performance on benign inputs. BlueSuffix addresses these limitations by incorporating three key components: visual and textual purifiers, and a blue-team suffix generator using reinforcement fine-tuning. The method is evaluated on four VLMs (LLaVA, MiniGPT-4, InstructionBLIP, and Gemini) and four safety benchmarks (Harmful Instruction, AdvBench, MM-SafetyBench, and RedTeam-2K), outperforming baseline defenses by a significant margin. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper is about protecting computer models that understand both pictures and words from being hacked. These models are useful for things like image search or chatbots, but hackers can make them do bad things if they want to. The researchers came up with a new way to keep these models safe called BlueSuffix. It works by cleaning out any bad images or text messages and making the model stronger at recognizing good ones. They tested it on four different models and it worked better than other methods. This is important because it could help make sure that AI technology is used responsibly. |
Keywords
» Artificial intelligence » Fine tuning » Gemini
