Summary of Autodefense: Multi-agent Llm Defense Against Jailbreak Attacks, by Yifan Zeng et al.
AutoDefense: Multi-Agent LLM Defense against Jailbreak Attacks
by Yifan Zeng, Yiran Wu, Xiao Zhang, Huazheng Wang, Qingyun Wu
First submitted to arxiv on: 2 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL); Cryptography and Security (cs.CR)
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 proposed AutoDefense framework is a multi-agent defense mechanism designed to filter out harmful responses from large language models (LLMs) and prevent jailbreak attacks. This framework assigns different roles to LLM agents, allowing them to work together collaboratively to complete the defense task. The response-filtering mechanism is robust against various attack prompts and can be used to defend different victim models. AutoDefense also enables integration with other defense components as tools, enhancing the overall instruction-following of LLMs. Experimental results demonstrate that the framework can effectively defend against different jailbreak attacks while maintaining normal user request performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) are powerful tools, but they’re vulnerable to “jailbreak” attacks that make them generate harmful content. To fix this problem, researchers created a new system called AutoDefense. It’s like a team of agents working together to keep the bad guys out. Each agent has a special job, and they all work together to make sure only good responses come out. This system is really good at stopping jailbreak attacks, even against super powerful models. For example, it was able to reduce the success rate of an attack on GPT-3.5 from 55% to just 8%. |
Keywords
* Artificial intelligence * Gpt
