基于局部性原理的高效软聚类联邦学习  

Efficient soft clustering federated learning based on locality principle

作　　者：许航 范艳芳[1] 蔡英[1] XU Hang;FAN Yanfang;CAI Ying(Computer School,Beijing Information Science&Technology University,Beijing 102206,China)

机构地区：[1]北京信息科技大学计算机学院,北京102206

出　　处：《北京信息科技大学学报（自然科学版）》2024年第4期88-95,共8页Journal of Beijing Information Science and Technology University

基　　金：北京市自然科学基金项目(L192023);北京信息科技大学“勤信人才”培育计划项目(QXTCP C202111)。

摘　　要：软聚类常用于解决多任务联邦学习(federated learning,FL)场景下存在非独立同分布(non-independent and identically distributed,non-IID)数据时的模型精度下降问题。然而,使用软聚类需上传和下载更多的模型参数。为了应对这一挑战,提出了基于局部性原理的联邦学习(federated learning with principle of locality,FedPol)算法。采用近端局部更新机制以确保客户端的本地更新在一定范围内波动;利用客户端本地数据分布的局部特性,整合历史数据分布信息至模型训练过程,加速了模型收敛,减少了需要传输的参数量。仿真实验证明,针对non-IID数据,FedPol算法可以在保持模型精度的前提下比其他算法减少约10%的迭代轮次,有效降低了通信成本。Soft clustering is often used to solve the problem of model accuracy degradation in multi-task federated learning(FL)scenarios in the presence of non-independent and identically distributed(non-IID)data.However,using soft clustering requires uploading and downloading more model parameters.To address this challenge,federated learning with principle of locality(FedPol)algorithm was proposed.The proximal local update mechanism was adopted to ensure that the local update of the client fluctuates within a certain range.By utilizing the local characteristics of the local data distribution of the client,the historical data distribution information was integrated into the model training process,which accelerated the model convergence and reduced the number of parameters that needed to be transmitted.Simulation experiments show that for non-IID data,the FedPol algorithm can reduce the number of iteration rounds by about 10%compared with other algorithms under the premise of maintaining the accuracy of the model,which effectively reduces the communication cost.

关 键 词：联邦学习 聚类算法 局部性原理 近端局部更新 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]