异构环境下基于历史模型的动态联邦学习方法

打开文本图片集

关键词：联邦学习；数据异构；梯度校正；历史模型；聚类采样

中图分类号：TP181 文献标志码：A 文章编号：1001-3695（2026）04-028-1202-06

doi：10.19734/j. issn.1001-3695.2025.07.0286

History-aware dynamic federated learning method for data heterogeneous environments

Luo Haowen a，b，c ，Sheng Zhiwei a，b.e† （aSchoolofCyberseurity（XinGuIndustrialCllge），bAdancedCrytogahSystemSecurityKeyLboratoryofSicuanProce， c.SUGON IndustrialControlandSecurityCenter，ChengduUniversityofInformation TechnologyChengdu 61o25，China）

Abstract：Federatedlearningenables globalmachinelearningmodel trainingwhilepreservinglocaldataprivacy.However，it suffers from low model accuracyand slow convergence speeddue to data heterogeneity.To addressthese isues，this paper proposed adynamic federated learning method，named FedHD.FedHDdynamicallycorretedtheupdate directionof client models basedonhistoricallocalmodelsand globalmodel information.Thisapproach improvedtheglobal model’saccuracy whileensuring smothnessandrobustness duringtheconvergenceprocess.Inadition，clustering basedon model similarity optimizedclientsampling，asliding window mechanismincorporated historicalmodeldata，enhancingtheusabilityandrobustnessof the sampling processandacceleratingglobal modelconvergence.The studyevaluatedthe method under various heterogeneous datascenarios.Theresults showthatFedHDoutperforms otheralgorithms interms of modelaccuracyandconvergence speedon MNISTand CIFAR-10，while demonstrating beterrobustnessThese findings indicate that the proposed method performswell in heterogeneous data environments.

Key words：federated learning；data heterogeneity；gradientcorrction；historical models；clustered samplint

0 引言

随着信息化和数字化进程的加速，各行业在发展过程中积累了大量数据，涉及医疗、金融、教育、交通等多个领域[12]然而，由于不同来源的数据在进行共享的同时面临隐私泄露、数据滥用和安全风险等挑战[3，4]。（剩余15404字）