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 2605.29979cs.CRcs.LG
收藏

大型语言模型的推理系统指纹识别

Fingerprinting Inference Systems of Large Language Models

BIFOLD & TU Berlin(BIFOLD与柏林技术大学)

Anna Wimbauer, Jonas Möller, Erik Imgrund, Konrad Rieck

AI总结:

本文提出一种通过分析LLM的提示-响应行为来识别推理系统组件(如推理引擎、注意力后端和硬件平台)的指纹方法,并论证了防御该指纹识别的根本困难性。

摘要:

LLM的行为不仅仅取决于模型本身。推理系统的组件,如推理引擎、注意力后端和硬件平台,微妙地影响输入的处理方式。这些组件在实现上存在差异,因此在运行相同模型时,不同系统之间会产生微小的数值偏差。虽然先前的工作已经建立了这种偏差的理论存在性,但其安全影响尚未被探索。在本文中,我们表明这些偏差是特定组件的特征,并传播到可观察的文本输出中,从而将推理系统暴露给任何能够查询模型的方。基于这一观察,我们引入了一种指纹识别方法,通过分析LLM的提示-响应行为来识别推理系统的组件。我们的实证评估表明,即使在LLM以非零温度运行时,推理引擎、注意力后端和底层硬件平台也能被可靠地识别。我们证明,防止指纹识别从根本上来说是困难的,因为它需要消除硬件和软件堆栈之间的数值差异。因此,我们提出了部分缓解措施并讨论了它们的影响。

英文摘要:

The behavior of LLMs does not depend solely on the model itself. Components of the inference system, such as the inference engine, attention backend, and hardware platform, subtly influence how inputs are processed. These components differ in their implementations and thereby induce small numerical deviations across systems when running the same model. While prior work has established the theoretical existence of such deviations, their security implications have remained unexplored. In this paper, we show that these deviations are characteristic of specific components and propagate to observable textual outputs, exposing the inference system to any party that can query the model. Building on this observation, we introduce a fingerprinting method that analyzes the prompt-response behavior of LLMs to identify components of the inference system. Our empirical evaluation demonstrates that the inference engine, attention backend, and underlying hardware platform can be identified reliably, even when the LLM is operated at non-zero temperature. We show that preventing fingerprinting is fundamentally hard, as it would require eliminating numerical differences between hardware and software stacks. We therefore propose partial mitigations and discuss their impact.

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。

↑