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 2602.21496cs.AI
收藏

超越拒绝:探求代理自我校正对语义敏感信息的极限

Beyond Refusal: Probing the Limits of Agentic Self-Correction for Semantic Sensitive Information

Department of Computer Science, Virginia Tech(弗吉尼亚理工大学计算机科学系);School of IT and Engineering, ADA University(ADA大学信息科技与工程学院)

Umid Suleymanov, Zaur Rajabov, Emil Mirzazada, Murat Kantarcioglu

AI总结:

本文提出SemSIEdit框架,通过代理代理迭代批评和重写敏感信息,减少泄露并平衡隐私与效用,揭示规模依赖的安全分歧和推理悖论。

摘要:

尽管针对结构化敏感个人信息的防御措施已经成熟,但大语言模型(LLMs)带来了新的威胁:语义敏感信息(SemSI),其中模型推断敏感身份属性、生成损害声誉的内容或产生可能错误的信息。LLMs在不破坏实用性的情况下自我调节这些复杂且依赖上下文的敏感信息泄露的能力仍然是一个开放的科学问题。为此,我们引入SemSIEdit,一个推理时间框架,其中一个“编辑者”代理会迭代地批评和重写敏感片段,以保持叙述流畅,而不是简单地拒绝回答。我们的分析揭示了一个隐私-效用帕累托前沿,其中这种代理重写在所有三个SemSI类别中减少了泄露量34.6%,同时导致了9.8%的边际效用损失。我们还发现了一个规模依赖的安全分歧:大型推理模型(例如GPT-5)通过构造性扩展(增加细微差别)实现安全,而容量受限的模型则退化为破坏性截断(删除文本)。最后,我们识别了一个推理悖论:推理时间推理会增加基础风险,因为模型能够进行更深入的敏感推断,但同时使防御机制能够执行安全重写。

英文摘要:

While defenses for structured PII are mature, Large Language Models (LLMs) pose a new threat: Semantic Sensitive Information (SemSI), where models infer sensitive identity attributes, generate reputation-harmful content, or hallucinate potentially wrong information. The capacity of LLMs to self-regulate these complex, context-dependent sensitive information leaks without destroying utility remains an open scientific question. To address this, we introduce SemSIEdit, an inference-time framework where an agentic "Editor" iteratively critiques and rewrites sensitive spans to preserve narrative flow rather than simply refusing to answer. Our analysis reveals a Privacy-Utility Pareto Frontier, where this agentic rewriting reduces leakage by 34.6% across all three SemSI categories while incurring a marginal utility loss of 9.8%. We also uncover a Scale-Dependent Safety Divergence: large reasoning models (e.g., GPT-5) achieve safety through constructive expansion (adding nuance), whereas capacity-constrained models revert to destructive truncation (deleting text). Finally, we identify a Reasoning Paradox: while inference-time reasoning increases baseline risk by enabling the model to make deeper sensitive inferences, it simultaneously empowers the defense to execute safe rewrites.

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。

↑