通信学报2025，Vol.46Issue(z1)：92-101,10.DOI:10.11959/j.issn.1000-436x.2025235
基于原型聚类机制的个性化联邦学习方法
Personalized federated learning method based on prototype clustering mechanism
摘要
Abstract
Existing studies mostly adapt knowledge distillation or multi-task training for personalized federated learning,but these methods typically require additional distillation steps or high communication overhead,affecting overall model performance.To address this challenge,a personalized federated learning method FedPC based on prototype clustering was proposed.By introducing a client clustering mechanism,FedPC grouped clients with similar data distributions into clusters based on prototypes,thereby reducing the impact of data distribution differences on model performance.To bet-ter adapt the personalized needs of local models of participants,the client model was decoupled into a feature extractor and a personalized classifier.At the same time,an adaptive weighted aggregation strategy and a joint loss function were used to co-optimize the training processes of clients and the server,achieving better model performance.Experimental re-sults on three commonly used datasets,Cifar10,Cifar100,and FMNIST,show that FedPC outperforms traditional feder-ated learning methods in terms of model accuracy,verifying its effectiveness in handling data heterogeneity issues.关键词
个性化联邦学习/数据隐私/数据异质性/原型Key words
personalized federated learning/data privacy/data heterogeneity/prototype分类
信息技术与安全科学引用本文复制引用
刘海军,王浩龙,刘雅辉,马洪亮..基于原型聚类机制的个性化联邦学习方法[J].通信学报,2025,46(z1):92-101,10.基金项目
兵团重大科技基金资助项目(No.2023AA001) （No.2023AA001）
八师石河子市财政科技计划基金资助项目(No.2024GY08) （No.2024GY08）
兵团指导性科技计划基金资助项目(No.2023ZD045) （No.2023ZD045）
兵团重点领域科技攻关基金资助项目(No.2024AB080) （No.2024AB080）
兵团科技创新人才计划基金资助项目(No.2023CB005,No.2023ZD066,No.2022CB002-08)Bingtuan Major Science and Technology Project(No.2023AA001),Shihezi Financial Science and Technology Project(No.2024GY08),Bingtuan Science and Technology Program(No.2023ZD045),Bingtuan Key Areas Science and Technology Research Project(No.2024AB080),Bingtuan Science and Technology Innovation Talent Program(No.2023CB005,No.2023ZD066,No.2022CB002-08) （No.2023CB005,No.2023ZD066,No.2022CB002-08）
