arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2502.09755cs.CRcs.LG
收藏

越狱攻击初始化作为合规方向的提取器

Jailbreak Attack Initializations as Extractors of Compliance Directions

Department of Computer Science, Technion - Israel Institute of Technology(技术学院计算机科学系);Department of Data and Decision Science, Technion - Israel Institute of Technology(技术学院数据与决策科学系);School of Electrical and Computer Engineering Engineering, Ben-Gurion University of the Negev(内盖夫本· Gurion大学电气与计算机工程学院)

Amit Levi, Rom Himelstein, Yaniv Nemcovsky, Avi Mendelson, Chaim Baskin

AI总结:

本文发现基于梯度的越狱攻击初始化会收敛到抑制拒绝的单一合规方向,并据此提出CRI框架,通过沿合规方向投影未见提示来提高攻击成功率并降低计算开销。

摘要:

安全对齐的LLM对提示的响应要么是合规要么是拒绝,每种响应对应模型激活空间中的不同方向。最近的研究表明,通过从其他提示进行自我迁移来初始化攻击可以显著提升其性能。然而,这些初始化的潜在机制仍不清楚,并且攻击使用任意或手动选择的初始化。本文表明,每个基于梯度的越狱攻击及其后续初始化逐渐收敛到一个抑制拒绝的单一合规方向,从而能够实现从拒绝到合规的高效转换。基于这一见解,我们提出了CRI,一个旨在将未见提示进一步投影到合规方向的初始化框架。我们在多种攻击、模型和数据集上展示了我们的方法,实现了更高的攻击成功率(ASR)并降低了计算开销,突显了安全对齐LLM的脆弱性。参考实现可在以下网址获取:https://amit1221levi.github.io/CRI-Jailbreak-Init-LLMs-evaluation

英文摘要:

Safety-aligned LLMs respond to prompts with either compliance or refusal, each corresponding to distinct directions in the model's activation space. Recent works show that initializing attacks via self-transfer from other prompts significantly enhances their performance. However, the underlying mechanisms of these initializations remain unclear, and attacks utilize arbitrary or hand-picked initializations. This work presents that each gradient-based jailbreak attack and subsequent initialization gradually converge to a single compliance direction that suppresses refusal, thereby enabling an efficient transition from refusal to compliance. Based on this insight, we propose CRI, an initialization framework that aims to project unseen prompts further along compliance directions. We demonstrate our approach on multiple attacks, models, and datasets, achieving an increased attack success rate (ASR) and reduced computational overhead, highlighting the fragility of safety-aligned LLMs. A reference implementation is available at: https://amit1221levi.github.io/CRI-Jailbreak-Init-LLMs-evaluation.

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。

↑