文章背景与核心概要
检索增强生成(RAG)虽然是增强大语言模型(LLM)能力的强大技术,但也带来了显著的隐私风险。现有研究多集中于防止对数据的未授权访问,而本文则关注了在RAG过程中向外部第三方生成器泄露敏感信息这一常被忽视的风险。
为此,作者引入了敏感实体别名生成器(SEAG),这是一个旨在保护用户数据的隐私保护框架。SEAG利用轻量级模型识别敏感实体、生成安全别名,并创建实体替换表。该表在将用户查询和检索到的文档发送给外部LLM处理之前,对其敏感信息进行屏蔽。实验结果表明,该框架在保持高实用性的同时,模型在提供正确响应方面的准确率超过了80%,并有效地隐匿了敏感数据。
Privacy-Preserving RAG by Concealing Sensitive Information from External LLMs
arXiv: 2608.12675
Submitted: August 13, 2026
Subjects: Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR)
Summary
Retrieval-Augmented Generation (RAG) is a powerful technique for enhancing Large Language Models (LLMs); however, it introduces significant privacy risks. While existing research focuses on preventing unauthorized access to data, this paper addresses the often-overlooked risk of exposing sensitive information to external third-party generators during the RAG process.
检索增强生成(RAG)是增强大语言模型(LLM)能力的强大技术;然而,它也带来了显著的隐私风险。现有研究多集中于防止对数据的未授权访问,而本文则关注了在RAG过程中向外部第三方生成器泄露敏感信息这一常被忽视的风险。
The authors introduce the Sensitive Entity Alias Generator (SEAG), a privacy-preserving framework designed to protect user data. SEAG utilizes a lightweight model to identify sensitive entities, generate secure aliases, and create an entity replacement table. This table masks sensitive information in both user queries and retrieved documents before they are processed by an external LLM. Experimental results demonstrate that the framework maintains high utility, with models achieving over 80% accuracy in providing correct responses while effectively concealing sensitive data.
作者引入了敏感实体别名生成器(SEAG),这是一个旨在保护用户数据的隐私保护框架。SEAG利用轻量级模型识别敏感实体、生成安全别名,并创建实体替换表。该表在将用户查询和检索到的文档发送给外部LLM处理之前,对其敏感信息进行屏蔽。实验结果表明,该框架在保持高实用性的同时,模型在提供正确响应方面的准确率超过了80%,并有效地隐匿了敏感数据。
Key Features of the SEAG Framework
- Entity Masking: Automatically identifies and replaces sensitive entities with aliases.
- Privacy-First Architecture: Ensures that confidential information is never disclosed to third-party providers.
- Performance: Validated across multiple models, including Qwen-3, LLaMA-3.2, and Phi-4, showing robust performance in both response accuracy and entity concealment.
SEAG 框架的核心特性
- 实体屏蔽: 自动识别敏感实体并将其替换为别名。
- 隐私优先架构: 确保机密信息绝不会向第三方服务提供商泄露。
- 性能表现: 在包括 Qwen-3、LLaMA-3.2 和 Phi-4 在内的多个模型上进行了验证,在响应准确率和实体隐匿方面均表现出强大的性能。
Authors
- Saleh Almohaimeed
- Saad Almohaimeed
- Mousa Jari
- Fahad Alotaibi
- Khalid A. Alobaid
作者
- Saleh Almohaimeed
- Saad Almohaimeed
- Mousa Jari
- Fahad Alotaibi
- Khalid A. Alobaid
Access & Resources

访问与资源
Citation
If you use this work, please cite it via the arXiv DOI.
引用
如果您使用了本工作,请通过 arXiv DOI 进行引用。