基于合作博弈的联邦学习聚合算法

打开文本图片集

关键词：联邦学习；异构数据；合作博弈；个性化联邦学习；TMC-Shapley值；SFedMeta中图分类号：TN919-34；TP311.5 文献标识码：A 文章编号：1004-373X（2026）09-0162-10

Cooperative game based aggregation algorithm for federated learning

LiuYingl²，LiYong1，³，WenMing²，HeZhenzhen1 （1.CollegeofComputerScienceand Technology，XinjiangNormal University，Urumqi 83oo54，China; 2.ScientificResearchDepartment，XinjiangElectronicResearch InstituteCo.，Ltd.，Urumqi83ooo1，China; 3.KeyLe

Abstract：Acooperativegamebased federated learning aggregationalgorithmisproposed toadress the model training chalengesduetoclientdataheterogeneityandpotentiallymaliciousclientsinfederatedlearning.Themethodintegrates personalizedfederatedmeta-learningwithcooperativegameShapleyvalueoptimizationstrategy，whichaimstoimprove the performanceandrobustnessoftheglobalmodelandreducethecommunicationandcomputationcosts.Firstly，bycolecting clientsoftlabels，themaximumentropyjudgment methodisusedtoselecttheclientswithhighcontributiontotheglobalmodel. Secondly，afastestimationstrategybasedontheTMC-Shapleyvalueisdesignedtoeficientlyestimatethmarginalcontribution ofclientsbyfinitetimessampling，soastoavoidexponentialcomputationalcomplexity.Finall，weightedagregationis performedbasedonclient Shapleyvaluesanddatadistributioncharacteritics.Experimentsshow thattheproposed method performswellontherealdatasetclasificationtask，significantlyimprovestheacuracyandreducesthecomputationalcostin comparisonwith thebaselinemethod.Itsadvantagesaremoreprominentinthescenariowithalargenumberofclientsand significant data heterogeneity.

Keywords：federated learning;heterogeneous data;cooperativegame;personalized federatedlearning;TMC-Shapleyvalue; SFedMeta

0 引言

随着人工智能物联网（AIoT）的快速发展，联邦学习（FederatedLearning，FL）[-2作为一种新兴的分布式机器学习范式，已被广泛应用于安全关键领域，如自动驾驶、商业监控、软件缺陷预测等。（剩余19968字）