基于局部差分隐私的动态聚类个性化联邦学习

打开文本图片集

关键词：个性化联邦学习；差分隐私；DBSCAN聚类；非独立同分布数据；隐私预算中图分类号：TP309 文献标志码：A 文章编号：1001-3695（2025）10-031-3137-07doi：10. 19734/j.issn. 1001-3695.2025.02.0050

Dynamic clustering-based personalized federated learning based on local differential privacy

Zhao Jinye，LiXiaohui†，Jia Xu （SchoolofElectronics&InformationEnginering，LiaoningUniversityofTechnology，JinzhouLiaoning121o，China）

Abstract：Thispaper proposedalocaldifferential privacypersonalized federatedlearningalgorithmintegratedwithDBSCAN clustering （DBLDP-PFL）to addressmodelaccuracydegradationcaused bysingle global models innon-independent and identicallydistributedfederatedlearningscenarios.Themethodfirstlyintroducedalocal diferentialprivacymechanismtoprotect single-clientdataprivacybyinjectingnoiseduringclientraining.Itthenemployedaprobabilisticselectionstrategybasedon clientprivacybudgetstoscrenparticipatingtraining clients.Finall，thealgorithmdesignedadynamicclientclusteringalgorithmwithinthefederatedlearningframework，enablingiterativerefinementoftheglobalmodelthroughcollaborativelearning among clients with similardata distributions.Experimentalresults demonstratethat the proposed algorithmoutperforms NbAFL， LDPKmeans，and HDP-EFL by achieving accuracy improvements of 10.09% 4.38% and 3.33% ，respectively on the MNIST dataset while preserving data privacy.

Keywords：personalizedfederated learning；differential privacy；DBSCANclustering；non-independentandidenticaly distributed data；privacy budget

0 引言

为了满足日益增长的机器学习需求，企业或机构通过共享数据来增强机器学习模型的性能。（剩余18640字）