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 2501.18532cs.CLcs.LG
收藏

面向大语言模型对齐的差分隐私引导方法

Differentially Private Steering for Large Language Model Alignment

Technical University of Darmstadt(达姆施塔特工业大学);Max Planck Institute for Intelligent Systems(马克斯·普朗克智能系统研究所);University of Copenhagen(哥本哈根大学)

Anmol Goel, Yaxi Hu, Iryna Gurevych, Amartya Sanyal

AI总结:

本研究针对私有数据集对齐大语言模型时的隐私泄露问题,提出带差分隐私保障的PSA激活编辑算法,在多基准多模型实验中实现隐私保护与极小性能损失的平衡,还首次提出对应成员推理攻击方法验证隐私效果。

摘要:

使大语言模型(LLM)与人类价值观对齐、避免产生不良行为(如幻觉)已变得愈发重要。近期,通过激活编辑引导LLM呈现预期行为的方法逐渐兴起,可在推理阶段有效缓解有害生成内容。激活编辑通过保留正向示例(如实情内容)的信息、最小化负向示例(如幻觉内容)的信息,来修改LLM的表征。当这些示例来自私有数据集时,对齐后的LLM可能会泄露这些私有样本中包含的隐私信息。\n本研究首次开展了利用私有数据集对齐LLM行为的相关研究,提出了面向LLM对齐的私有引导(Private Steering for LLM Alignment, PSA)算法,可在差分隐私(DP)保障下编辑LLM的激活值。我们使用不同规模(0.5B至7B)、不同模型家族(LlaMa、Qwen、Mistral和Gemma)的开源LLM,在七个不同基准上开展了大量实验。结果表明,PSA在为LLM对齐提供DP保障的同时,仅带来极小的性能损失,涵盖对齐指标、开放式文本生成质量以及通用推理能力等方面。我们还首次提出了成员推理攻击(Membership Inference Attack, MIA)方法,用于评估和审计基于激活编辑的LLM引导问题的经验隐私水平。实验验证了理论保障的有效性,表明与现有多种非隐私技术相比,PSA算法的隐私保障效果更优。

英文摘要:

Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the Private Steering for LLM Alignment (PSA) algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with open-source LLMs of different sizes (0.5B to 7B) and model families (LlaMa, Qwen, Mistral and Gemma). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, open-ended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our experiments support the theoretical guarantees by showing improved guarantees for our PSA algorithm compared to several existing non-private techniques.

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。

↑