arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2510.22014cs.CLcs.AI
收藏

面向大型语言模型中对抗性后缀的可迁移性理解

Toward Understanding the Transferability of Adversarial Suffixes in Large Language Models

Ludwig-Maximilians-Universität München(慕尼黑莱布尼茨-马克斯大学);Munich Center for Machine Learning (MCML)(慕尼黑机器学习中心);Carnegie Mellon University(卡内基梅隆大学);University of Maryland(马里兰大学)

Sarah Ball, Niki Hasrati, Alexander Robey, Avi Schwarzschild, Frauke Kreuter, Zico Kolter, Andrej Risteski

AI总结:

该研究针对大型语言模型的对抗性后缀可迁移性,分析出三个与迁移成功高度相关的统计属性,其成果可用于提升越狱攻击的实际效果。

摘要:

基于离散优化的大型语言模型越狱攻击旨在生成简短、无意义的后缀,将其附加到输入提示中时会引发违禁内容。值得注意的是,这些后缀通常具有可迁移性——在未针对其优化的提示和模型上也能成功。然而,尽管可迁移性令人惊讶且在经验上已得到充分证实,但该领域缺乏对何时以及为何发生迁移的严格分析。为填补这一空白,我们确定了三个在众多实验设置中与迁移成功高度相关的统计属性:(1)无后缀的提示激活模型内部拒绝方向的程度;(2)后缀促使偏离该方向的强度;(3)这些在与拒绝正交方向上的偏移量大小。另一方面,我们发现提示语义相似性与迁移成功仅呈弱相关。这些发现促成了对可迁移性更细致的理解,我们在干预实验中利用这一点展示了我们的统计分析如何转化为攻击成功率的实际提升。

英文摘要:

Discrete optimization-based jailbreaking attacks on large language models aim to generate short, nonsensical suffixes that, when appended onto input prompts, elicit disallowed content. Notably, these suffixes are often transferable -- succeeding on prompts and models for which they were never optimized. And yet, despite the fact that transferability is surprising and empirically well-established, the field lacks a rigorous analysis of when and why transfer occurs. To fill this gap, we identify three statistical properties that strongly correlate with transfer success across numerous experimental settings: (1) how much a prompt without a suffix activates a model's internal refusal direction, (2) how strongly a suffix induces a push away from this direction, and (3) how large these shifts are in directions orthogonal to refusal. On the other hand, we find that prompt semantic similarity only weakly correlates with transfer success. These findings lead to a more fine-grained understanding of transferability, which we use in interventional experiments to showcase how our statistical analysis can translate into practical improvements in attack success.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑