arXiv 2606.050042026-06-03cs.CRcs.AI

SharedRequest: 面向大型语言模型的隐私保护模型无关推理

SharedRequest: Privacy-Preserving Model-Agnostic Inference for Large Language Models

National University of Singapore (Chongqing) Research Institute（新加坡国立大学（重庆）研究院）；Chongqing Key Laboratory of Trusted Perception and Interaction Technology for Intelligent and Connected Vehicles（重庆智能网联车辆可信感知与交互技术重点实验室）；National University of Singapore（新加坡国立大学）；Hebrew University of Jerusalem（耶路撒冷希伯来大学）；State Key Laboratory of Intelligent Vehicle Safety Technology, Chongqing, China（重庆智能车辆安全技术国家重点实验室）；CHONGQING CHANGAN AUTOMOBILE Co., Ltd（重庆长安汽车有限公司）

Peihua Mai, Xuanrong Gao, Youlong Ding, Xianglong Du, Wei Liu, Yan Pang

备注：: accepted by ACL 2026 (main)

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

提出一种模型无关的隐私保护推理框架SharedRequest，通过批量级别混淆和语义分组实现高效隐私保护，相比差分隐私基线效用提升20%以上，查询成本降低5倍。

摘要：

随着ChatGPT等公共大型语言模型（LLMs）的广泛部署，保护用户提示隐私已成为一个日益关键的问题。现有的隐私保护推理方法要么牺牲效用，要么牺牲效率，并且通常需要特定于模型的修改，限制了其兼容性。在本文中，我们提出了SharedRequest，一个模型无关的隐私保护LLM推理框架，它将隐私保护重新定义为批量级别而非单个提示级别。关键思想是通过将原始提示与噪声变体混合来混淆敏感信息，同时将语义等效的指令分组，以在大量查询批次中分摊推理成本，对LLM响应质量影响最小。该设计独立于LLM架构，无需访问模型参数或进行架构修改。实验结果表明，与先前的差分隐私基线相比，SharedRequest实现了超过20%的效用提升，并且其共享提示机制相比非批量推理将查询成本降低了5倍。

英文摘要：

With the widespread deployment of public large language models (LLMs) such as ChatGPT, protecting user prompt privacy has become an increasingly critical issue. Existing privacy-preserving inference methods sacrifice either utility or efficiency, and often require model-specific modifications that limit their compatibility. In this paper, we propose SharedRequest, a model-agnostic framework for privacy-preserving LLM inference that reformulates privacy protection at the batch level rather than the individual-prompt level. The key idea is to obscure sensitive information by mixing original prompts with noisy variants, while grouping semantically equivalent instructions to amortize the inference cost over a large batch of queries with minimal impact on LLM response quality. This design is independent of the LLM architecture, requiring no access to model parameters or architectural modification. Empirical results demonstrate that SharedRequest achieves over $20\%$ higher utility compared to prior differential privacy baselines, and its shared-prompt mechanism reduces query cost by up to $5\times$ compared to non-batched inference.

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。