面向空天地一体化网络的低时延高性能联邦学习方法研究

打开文本图片集

中图分类号：TN929 文献标志码：A文章编号：1003-3114（2025）06-1180-09开放科学（资源服务）标识码（OSID）：

Abstract：Space-Air-Ground IntegratedNetwork （SAGIN）constituteacrtical development paradigmforfuture networks.Training inteligentmodelsusingdatafromtisnetworkcanenablediversesmartservices.However，giventhecalengesofdatailosandprivacy concerns，FederatedLeaing（FL）hasmergedasapromisingsolutio.Neverthele，deploingFLalgorithsinesourcecostred SAGINenvironmentstillecounterschallngessuchasprolongedrainingtimesandsuboptialleangperformance.TispaperinvestigateseficientFLmetodsuderNon-IdependentandIdenticallistrbuted（Non-ID）datasenariosinSAGNandroposs thre-tierFLachiteturetilordforAGSpecificalyitstablsstwo-lvelmodelaggregationmechansms：oteeevic andunmannederialvehicles，ndnoterbetweenunmadarialvehicesandsateltes.Furthermre，itucidatesthifueceof deviceassociatioategisodeleorasedotseisihts，eapproachropsdinspaperjitlyotiesdviceociationandbandwidthalocation strategiesbyleveragingthediminishing marginalgainmehanismandconvexoptimizationtechniques， therebyehagthngcyofLAdioallosidegliidceiofatelits，dami tionstrategyisdsignedtoiterativelyadjustthenumberofparticipatingdevicesandresourceallcationoutcomes，ensuringtielyompletionofmodelrainingwithinconstrainedtimeframes.Simulationresultsdemonstratethattheproposedmethodsignificantlyreduces trainingdelaywhile increasing the number of participating devices and maintaining learning performance.

Keywords：SAGIN;FL;device association;resource allocation

0引言

1 国内外研究现状

近年来，随着移动通信网络与物联网（InternetofThings，IoT）的持续发展，移动设备数量呈指数级增长，其产生的海量数据可以用于机器学习（Machine Learning，ML）任务，从而提供智能化服务[1]传统集中式ML方法需要将设备本地数据上传到中心服务器，不仅会占用大量带宽，还涉及隐私安全问题，因此，FL的概念被提出[2]。（剩余13504字）