基于改进CFSFDP算法的聚类联邦学习  

Clustering Federated Learning Based On Improved CFSFDP Algorithm

作　　者：彭璐璐 张淑芬 陈海田 徐超 PENG Lu-lu;ZHANG Shu-fen;CHEN Hai-tian;XU Chao(College of Science,North China University of technology,Tangshan Hebei 063210,China;Hebei Key Laboratory of data science and application,Tangshan Hebei 063210,China;Tangshan Key Laboratory of Big Data Security and Intelligent Computing,Tangshan Hebei 063210,China)

机构地区：[1]华北理工大学理学院,河北唐山063210 [2]河北省数据科学与应用重点实验室,河北唐山063210 [3]唐山市大数据安全与智能计算重点实验室,河北唐山063210

出　　处：《华北理工大学学报（自然科学版）》2024年第3期112-121,共10页Journal of North China University of Science and Technology：Natural Science Edition

基　　金：国家自然科学基金项目(U20A20179):基于Sketch的网络行为测量关键技术与系统。

摘　　要：联邦学习是一种基于多个用户协作训练全局模型的分布式训练范式,旨在解决由数据孤岛造成的数据碎片化问题。然而,在真实环境中,参与联邦学习的各个节点的数据分布通常是不平衡的,这会导致联邦学习模型的精确率下降。针对上述问题,提出了基于改进CFSFDP算法的聚类联邦学习方法——FL_CDP。此方法对CFSFDP算法中的局部密度定义进行了优化,并通过二阶差分实现自动选择聚类中心。通过评估客户端模型的推理相似度,对客户端节点进行了聚类,进而缓解了联邦学习中因节点数据分布不均所带来的模型准确度下降问题。在MNIST和FashionMNIST数据集上的实验结果证明,相较于传统的联邦学习算法,基于改进CFSFDP算法的聚类联邦学习算法在模型准确率方面有着显著提升。Federated learning is a distributed training paradigm based on multi user collaborative training of global models,aimed at solving the problem of data fragmentation caused by data silos.However,in the real environment,the data distribution of each node participating in federated learning is usually uneven,which can lead to a decrease in the accuracy of the federated learning model.In response to the above issues,a clustering federated learning method based on the improved CFSFDP algorithm,FL_CDP,was proposed.This method optimizes the definition of local density in the CFSFDP algorithm and automatically selects clustering centers through second-order difference.By evaluating the inference similarity of the client model,the client nodes were clustered,thereby alleviating the problem of model accuracy degradation caused by uneven distribution of node data in federated learning.The experimental results on the MNIST and Fashion MNIST datasets demonstrate that compared to traditional federated learning algorithms,the clustering federated learning algorithm based on the improved CFSFDP algorithm has a significant improvement in model accuracy.

关 键 词：联邦学习 隐私保护 不平衡数据 二阶差分 

分 类 号：TQ301[化学工程]