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 2603.24543cs.CRcs.CL
收藏

分析转向向量的安全隐患

Analysing the Safety Pitfalls of Steering Vectors

Technical University of Munich(慕尼黑技术大学);Munich Center for Machine Learning(慕尼黑机器学习中心)

Yuxiao Li, Alina Fastowski, Efstratios Zaradoukas, Bardh Prenkaj, Gjergji Kasneci

AI总结:

本文研究转向向量在对抗攻击中的影响,发现其能显著改变攻击成功率,揭示可控性与安全性的权衡。

摘要:

激活转向已成为一种强大的工具,无需权重更新即可塑造大语言模型(LLM)的行为。尽管其固有的脆弱性和不可靠性已得到充分记录,但其安全性影响仍被忽视。在本工作中,我们采用统一的评估协议,对使用对比激活添加(CAA)获得的转向向量进行了系统安全性审计。通过JailbreakBench基准测试,我们证明转向向量一致地影响对抗攻击的成功率,简单模板攻击下效果更显著。在不同LLM家族和大小中,向特定方向转向可使攻击成功率(ASR)大幅增加(最高57%)或减少(最高50%),这归因于转向向量与拒绝行为潜在方向的重叠。因此,我们提供了可追溯的解释。共同发现揭示了LLM中此前未观察到的安全差距起源,突显了可控性与安全性的权衡。

英文摘要:

Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety.

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。

↑