基于TD3的恶意节点检测与鲁棒联邦聚合算法  

Malicious Node Detection and Robust Federated Aggregation Algorithm Based on TD3

作　　者：孙凡 文红[1,2,3] 侯文静 王永丰[1,2,3] 姚瑞祥 严地宝 SUN Fan;WEN Hong;HOU Wenjing;WANG Yongfeng;YAO Ruixiang;YAN Dibao(School of Aeronautics and Astronautics,UESTC,Chengdu Sichuan 611731,China;Aircraft Swarm Intelligent Sensing and Cooperative Control Key Laboratory of Sichuan Province,UESTC,Chengdu Sichuan 611731,China;Sichuan Provincial Engineering Research Center of Communication Technology for Intelligent IoT,UESTC,Chengdu Sichuan 611731,China)

机构地区：[1]电子科技大学航空航天学院,四川成都611731 [2]电子科技大学飞行器集群智能感知与协同控制四川省重点实验室,四川成都611731 [3]电子科技大学四川省智慧物联通信技术工程研究中心,四川成都611731

出　　处：《通信技术》2024年第8期845-849,共5页Communications Technology

基　　金：国家自然科学基金(U23B2021,61901089)。

摘　　要：针对联邦学习中的数据安全与隐私问题,提出了一种基于双延迟深度确定性策略梯度(Twin Delayed Deep Deterministic policy gradient,TD3)的恶意节点检测与鲁棒联邦聚合算法,旨在提高系统在面对恶意攻击时的鲁棒性。以标签翻转攻击和梯度上升攻击为例,展示了该算法对于恶意节点的有效识别与抵御能力,并对比分析了所提方法与传统的聚合算法,如FedAvg、Krum、MKrum等。试验结果表明,相较于传统算法,所提算法不仅能有效防御恶意攻击,同时能保持高效的学习效率和模型准确性,显著提升了模型的整体安全防御能力。To address the data security and privacy issues in federated learning,this paper proposes a malicious node detection and robust federated aggregation algorithm based on TD3(Twin Delayed Deep Deterministic) policy gradient,which aims to enhancing the system robustness against malicious attacks.By using the label flipping attack and gradient ascent attack as examples,this paper demonstrates the algorithm's ability to effectively identify and defend against malicious nodes,and compares and analyzes the proposed method with conventional aggregation algorithms,such as FedAvg,Krum,MKrum,and so on.Experimental results indicate that compared with the conventional algorithm,the proposed algorithm not only can effectively defend against malicious attacks,but also can maintain high learning efficiency and model accuracy,which significantly improves the overall security defense ability of the model.

关 键 词：联邦学习 恶意节点检测 鲁棒联邦聚合 TD3算法 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]