用连续参数平滑离散优化，提升大模型越狱攻击成功率
Functional Homotopy: Smoothing Discrete Optimization via Continuous Parameters for LLM Jailbreak Attacks
- 通过构建由易到难的优化问题序列，实现对离散输入空间的平滑优化
- 在Llama-2和Llama-3上实现20%-30%的越狱成功率提升
- 适合研究大模型安全与对抗攻击的学者参考
优化方法广泛用于深度学习中识别并缓解模型的不当响应。尽管基于梯度的方法在图像模型中表现良好，但其在语言模型中的应用受限于输入空间的离散性。本文提出一种新型优化方法——功能同伦（functional homotopy），利用模型训练与输入生成之间的函数对偶性，构建一系列由易到难的优化问题，通过借鉴经典同伦方法原理进行迭代求解。该方法应用于大语言模型（LLMs）的越狱攻击生成，在绕过Llama-2和Llama-3等主流开源安全模型时，成功率达现有方法的20%-30%提升。
原文摘要 · Abstract (English)
Optimization methods are widely employed in deep learning to identify and mitigate undesired model responses. While gradient-based techniques have proven effective for image models, their application to language models is hindered by the discrete nature of the input space. This study introduces a novel optimization approach, termed the \emph{functional homotopy} method, which leverages the functional duality between model training and input generation. By constructing a series of easy-to-hard optimization problems, we iteratively solve these problems using principles derived from established homotopy methods. We apply this approach to jailbreak attack synthesis for large language models (LLMs), achieving a $20\%-30\%$ improvement in success rate over existing methods in circumventing established safe open-source models such as Llama-2 and Llama-3.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。