计算机工程2025，Vol.51Issue(2)：170-178,9.DOI:10.19678/j.issn.1000-3428.0068389
基于EMD最优匹配的分层联邦学习算法
Hierarchical Federated Learning Algorithm Based on EMD Optimal Matching
摘要
Abstract
Federated learning allows multiple clients to cooperatively train a high-performance global model without sharing private data.In a horizontal federated learning environment involving cross-silo scenarios,the statistical heterogeneity in the distribution of local client data degrades the performance of the global model.To improve the global model performance of federated learning,prevent sacrificing client privacy,and reduce computing costs,a new hybrid federated learning method,FedAvg-Match,is proposed in this paper.The basic idea is to improve model quality for clients by improving the federated learning method.Aiming at the data heterogeneity characterized by an unbalanced label distribution,a client group aggregation algorithm is designed under a hierarchical federated learning framework,to reduce the impact of client data heterogeneity on model performance.A client-matching algorithm,Dynamic Programming(DP)-ClientMatch,is designed to solve the problem of optimal client matching,whereby optimal client group matching is determined according to the client data distribution using Earth Mover's Distance(EMD).Experimental results across three datasets,MNIST,Fashion-MNIST and CIFAR-10,showed that compared with other federated learning algorithms,the proposed FedAvg-Match algorithm can significantly improve the performance of the global model in federated learning for image classification tasks.In federated learning scenarios with high statistical heterogeneity,the accuracy of the global model testing can be improved by at least more than 10 percentage points.关键词
联邦学习/非独立同分布数据/最优匹配/EMD最优匹配/模型质量Key words
federated learning/non-Independent Identically Distribution(non-IID)data/optimal matching/EMD optimal matching/model quality分类
信息技术与安全科学引用本文复制引用
吴小红,李佩,顾永跟,陶杰..基于EMD最优匹配的分层联邦学习算法[J].计算机工程,2025,51(2):170-178,9.基金项目
国家自然科学基金青年科学基金项目(61906066,2022ZD2002). （61906066,2022ZD2002）
