首页|期刊导航|网络与信息安全学报|大语言模型隐私度量与保护技术研究

大语言模型隐私度量与保护技术研究OA

Survey of privacy measurement and protection techniques for large language model

中文摘要英文摘要

大语言模型(large language model,LLM)的快速发展极大地推动了人工智能技术的迭代升级,并在各行各业得到广泛应用.然而,LLM庞大的参数规模和基于海量网络语料的数据训练方式,使其在全生命周期内面临诸多隐私泄露风险.为此,以LLM全生命周期的"隐私风险分析—隐私风险度量—隐私保护技术"为主线,首先,系统剖析LLM从数据准备、模型训练到推理与交互阶段面临的数据提取、成员推断、内容注入、属性推断、提示窃取、越狱6类隐私威胁;其次,聚焦6类隐私威胁的量化评估方法,即隐私度量技术,科学评估LLM面临的隐私威胁;最后,详细介绍贯穿LLM全生命周期的隐私保护技术体系,涵盖差分隐私、联邦学习和安全多方计算等关键技术.该研究成果可为LLM的设计、训练及应用提供完善的隐私安全理论与实践参考,助力LLM在强大功能与用户隐私保护之间实现有效平衡.

The rapid development of large language models(LLM)has been greatly facilitated by the iterative up-grading of artificial intelligence technologies,and they have been widely applied across various industries.However,due to their massive parameter scales and data training approaches based on vast amounts of Web corpora,numerous privacy leakage risks have been faced by LLM throughout their entire lifecycle.To address this issue,a systematic analysis was conducted along the main line of"privacy risk analysis-privacy risk measurement-privacy protection technology"throughout the LLM lifecycle.Firstly,Six types of privacy threats,including data extraction,member-ship inference,content injection,attribute inference,prompt stealing,and jailbreak attacks were systematically ana-lyzed during the stages of data preparation,model training,and inference interaction.Secondly,The quantitative evaluation methods for these six types of privacy threats,namely privacy measurement techniques,were focused on,so as to scientifically assess the privacy threats faced by LLM.Finally,A detailed introduction was also given to the privacy protection technology system throughout the entire LLM lifecycle,covering key technologies such as differ-ential privacy,federated learning,and secure multi-party computation.The research findings presented in this paper can be provided as comprehensive theoretical and practical references for privacy and security in the design,training,and application of LLM,and an effective balance can be facilitated between the powerful capabilities of LLM and user privacy protection.

陈天宇;狄永正;史红周

中国科学院计算技术研究所,北京 100190中国科学院计算技术研究所,北京 100190中国科学院计算技术研究所,北京 100190

信息技术与安全科学

大语言模型隐私风险隐私度量隐私保护技术

large language modelprivacy riskprivacy measurementprivacy protection technology

《网络与信息安全学报》 2026 (4)

115-133,19

10.11959/j.issn.2096-109x.AQ25252

评论