一种子图感知的横向图联邦鲁棒聚合算法  

A Subgraph-aware Robust Aggregation Algorithm for GNN-based Horizontal Federated Learning

作　　者：黄瑞辉 曾昭延 郭磊 HUANG Ruihui;ZENG Zhaoyan;GUO Lei(China Southern Airlines Co.,Ltd,Guangzhou,China,510405)

机构地区：[1]中国南方航空股份有限公司,广州510405

出　　处：《福建电脑》2025年第4期1-8,共8页Journal of Fujian Computer

基　　金：《联邦学习在航空公司客户洞察的应用研究》(No.KIA120003)项目资助。

摘　　要：为了防御横向图联邦学习中的投毒攻击,本文提出一种新的子图感知鲁棒聚合算法。算法包含恶意用户识别和个性化聚合两个模块。恶意用户识别模块考虑用户梯度异常性和图结构感知异常性,迭代地识别出所有恶意用户;个性化聚合模块对恶意用户进行梯度修正,并且对良性用户进行个性化聚合,从而优化各个本地模型的性能。本文的实验结果表明,与Krum、Median、TrimmedMean等方法相比,本文提出的防御算法具有预测精度更高、收敛速度更快等优势。In order to defend against poisoning attacks in horizontal graph federated learning,this paper proposes a new subgraph aware robust aggregation algorithm.The algorithm consists of two modules:malicious user identification and personalized aggregation.The malicious user identification module considers user gradient anomalies and graph structure perception anomalies,iteratively identifying all malicious users;The personalized aggregation module performs gradient correction on malicious users and personalized aggregation on benign users to optimize the performance of various local models.The experimental results of this article show that compared with Krum,Median,TrimedMean and other methods,the defense algorithm proposed in this article has advantages such as higher prediction accuracy and faster convergence speed.

关 键 词：联邦学习 图神经网络 图联邦学习 投毒攻击 鲁棒性 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]