信息网络安全

信息网络安全 ›› 2025, Vol. 25 ›› Issue (7): 1138-1152.doi: 10.3969/j.issn.1671-1122.2025.07.012

• 理论研究 • 上一篇    下一篇

多智能体系统安全防护技术研究综述

汪正阳1, 刘晓露1(), 沈卓炜2, 韦梦立1   

  1. 1.东南大学数学学院，南京 211189
    2.东南大学网络空间安全学院，南京 211189
  • 收稿日期:2025-05-19 出版日期:2025-07-10 发布日期:2025-08-07
  • 通讯作者: 刘晓露 E-mail:101300566@seu.edu.cn
  • 作者简介:汪正阳（2004—），男，安徽，本科，主要研究方向为多智能体系统、数据安全|刘晓露（1992—），女，四川，副教授，博士，主要研究方向为多智能体协调控制|沈卓炜（1974—），男，江苏，副教授，博士，CCF会员，主要研究方向为高可信分布式软件架构、工业互联网、车联网及其安全|韦梦立（1993—），男，江苏，博士研究生，主要研究方向为分布式学习与隐私保护
  • 基金资助:
    国家自然科学基金(622303210)

Review of Security Protection Technologies for Multi-Agent Systems

WANG Zhengyang1, LIU Xiaolu1(), SHEN Zhuowei2, WEI Mengli1   

  1. 1. School of Mathematics, Southeast University, Nanjing 211189, China
    2. School of Cyber Science and Engineering, Southeast University, Nanjing 211189, China
  • Received:2025-05-19 Online:2025-07-10 Published:2025-08-07
  • Contact: LIU Xiaolu E-mail:101300566@seu.edu.cn

RichHTML

28

PDF (PC)

61

摘要：

文章聚焦于多智能体系统的安全防护技术，从系统面临的威胁视角展开全面且深入的探讨。首先，基于多智能体系统具备开放性、异构性、自治性、协同性、动态适应性和涌现性6个主要特性，探讨其内生安全风险。从攻击目标、攻击方式和攻击者属性3个维度，对系统的安全风险进行分类，给出相关攻击方式。其次，概述安全威胁识别方法，指出威胁建模方法的局限性。在安全防御技术方面，梳理加密与认证、入侵检测与响应、信誉管理、容错设计、安全策略与审计等领域面临的挑战和研究进展。再次，探讨大模型直接调用智能体可能引发的跨域攻击威胁，分析视觉、音频攻击手段被大模型利用后可能导致的损害，并从打断攻击链的角度分析可能的防御措施。从次，阐述安全架构的演进方向，介绍弹性安全架构和内部工作逻辑。最后，对国内外研究现状进行总结，并从理论、技术和学科融合创新方面给出后续研究建议。

关键词: 多智能体系统, 安全防护, 加密算法, 隐私防护, 跨域攻击

Abstract:

This study focused on security protection technologies for multi-agent systems, conducting a comprehensive and in-depth exploration from the perspective of threats faced by the system. Firstly, based on the six major characteristics of multi-agent systems,including openness, heterogeneity, autonomy, collaboration, dynamic adaptability, and emergence, the intrinsic security risks were discussed. The security risks were categorized from three dimensions: attack targets, attack methods, and attacker attributes, with relevant attack methods provided. Secondly, A summary of threat identification methods was given, highlighting the limitations of threat modeling approaches. In terms of security defense technologies, challenges and research progress in areas such as encryption and authentication, intrusion detection and response, reputation management, fault tolerance design, and security policies and audits were reviewed. Thirdly, The potential cross-domain attack threats caused by large models directly invoking agents were explored, analyzing the damage that visual and audio attack methods could cause when exploited by large models, and proposing possible defensive measures from the perspective of disrupting the attack chain. Finally, The evolution direction of security architecture was elaborated, introducing elastic security architecture and its internal working logic. Finally, the current research status was summarized, and suggestions for future research are provided from theoretical, technological, and interdisciplinary innovation perspectives.

Key words: multi-agent system, security protection, encryption algorithm, privacy protection, cross-domain attack

中图分类号: