平衡信息与动态更新的原型表示联邦学习  

Balance Information and Dynamic Update in Prototype-Based Federated Learning

作　　者：徐炜钦 肖婷 王喆[1] XU Weiqin;XIAO Ting;WANG Zhe(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237

出　　处：《华东理工大学学报（自然科学版）》2024年第6期905-912,共8页Journal of East China University of Science and Technology

基　　金：国家自然科学基金(62076094);上海市科技计划项目‘联邦框架下跨域/跨任务增量学习方法研究’(21511100800)。

摘　　要：联邦学习(FL)是一种分布式机器学习方法,旨在通过训练模型而不共享客户之间的原始数据来解决隐私问题。然而,跨客户端数据的异构性会阻碍FL中的优化收敛性和泛化性能。为了解决这个问题,本文提出了平衡信息与动态更新的联邦原型学习(BD-FedProto)框架,它由两个组件组成:原型调度的动态聚合(DA)和对比原型聚合(CPA)。前者动态地调整局部学习和全局学习之间的比例,以平衡局部知识和全局知识的有效性;后者利用缺失的类作为负样本,通过统一的原型集群来学习未知的分布。在CIFAR-10和MNIST数据集上的实验结果表明,BD-FedProto能有效提高FL的分类性能和稳定性。Federated Learning(FL)is a distributed machine learning approach that aims to solve privacy issues by training models without sharing the original data among clients.However,the heterogeneity of client data across FL can hinder the convergence and generalization performance of optimization.To address this issue,this paper proposes a balanced information and dynamic updating federated prototype learning(BD-FedProto)framework consisting of two components:dynamic aggregation(DA)of prototype scheduling and contrastive prototype aggregation(CPA).The former dynamically adjusts the proportion between local learning and global learning to balance the effectiveness of local knowledge and global knowledge.The latter utilizes missing classes as negative samples by learning unknown distributions through a unified prototype clustering.The experimental results on the CIFAR-10 and MNIST datasets show that BDFedProto is effective in improving the classification performance and stability of FL.

关 键 词：联邦学习 特征空间聚合 原型表示 对比学习 数据异构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]