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 2410.01438cs.LG
收藏

大矛盾对决:越狱与隐蔽性在视觉语言模型中如何博弈?

The Great Contradiction Showdown: How Jailbreak and Stealth Wrestle in Vision-Language Models?

Institute of Information Science (IIS), Academia Sinica(中央研究院资讯科学研究所);National Taiwan University(国立台湾大学);National Yang Ming Chiao Tung University(国立阳明交通大学)

Ching-Chia Kao, Chia-Mu Yu, Chun-Shien Lu, Chu-Song Chen

AI总结:

本文提出信息论框架揭示视觉语言模型中越狱攻击有效性与隐蔽性的权衡,基于Fano不等式证明二者内在联系,并提出高效检测算法提升模型鲁棒性。

摘要:

视觉语言模型(VLMs)在多种任务上取得了显著性能,但仍易受破坏安全性和可靠性的越狱攻击。本文提出了一个信息论框架,用于理解这些攻击的有效性与隐蔽性之间的基本权衡。基于Fano不等式,我们证明了攻击者的成功概率与生成提示的隐蔽性存在内在联系。在此基础上,我们提出了一种高效算法来检测非隐蔽的越狱攻击,显著提升了模型鲁棒性。实验结果突出了强攻击与其可检测性之间的张力,为对抗策略和防御机制提供了见解。

英文摘要:

Vision-Language Models (VLMs) have achieved remarkable performance on a variety of tasks, yet they remain vulnerable to jailbreak attacks that compromise safety and reliability. In this paper, we provide an information-theoretic framework for understanding the fundamental trade-off between the effectiveness of these attacks and their stealthiness. Drawing on Fano's inequality, we demonstrate how an attacker's success probability is intrinsically linked to the stealthiness of generated prompts. Building on this, we propose an efficient algorithm for detecting non-stealthy jailbreak attacks, offering significant improvements in model robustness. Experimental results highlight the tension between strong attacks and their detectability, providing insights into both adversarial strategies and defense mechanisms.

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。

↑