一种自适应的网格化联邦学习客户端调度算法  

Adaptive gridding client schedule algorithm for federated learning

作　　者：吴家皋[1,2] 蒋宇栋 刘林峰 WU Jiagao;JIANG Yudong;LIU Linfeng(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;Jiangsu Key Laboratory of Big Data Security&Intelligent Processing,Nanjing 210023,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210023 [2]江苏省大数据安全与智能处理重点实验室,江苏南京210023

出　　处：《南京邮电大学学报(自然科学版)》2025年第1期79-89,共11页Journal of Nanjing University of Posts and Telecommunications:Natural Science Edition

基　　金：国家自然科学基金(62272237,61872191)资助项目。

摘　　要：针对联邦学习(Federated Learning,FL)系统异构性而导致的训练性能下降问题,提出了一种自适应的网格化联邦学习客户端调度算法。首先,全面考虑FL的异构性特点,将3种异构性分别定义为3个独立的维度,包括训练速度、数据量和数据分布维度,提出了一种新的FL客户端三维网格模型,并将所有客户端分配到该模型中相应的单元格内,以对其进行分类管理。在此基础上,为了克服传统启发式算法的不足,提出了一种基于多臂老虎机的网格化客户端调度算法,该算法能自适应地选择模型精度较低的单元格中的客户端子集参与每轮的FL训练,以改善客户端选择的公平性。仿真实验表明,与几种相关的最新FL算法相比,所提出的算法能显著提高模型精度,同时减少训练时间,从而验证了其有效性。Given the degradation of training performance caused by the heterogeneity of federated learn⁃ing(FL)systems,an adaptive gridding client schedule algorithm for FL is proposed.First,three types of heterogeneity of FL are defined as three independent dimensions,including the dimensions of training speed,data quantity and data distribution.Then,a novel 3D grid model for FL clients is proposed,and all clients in FL are assigned to corresponding cells in the model for classification and management.Sec⁃ond,to overcome the shortcomings of traditional heuristic algorithms,a multi-armed bandit-based grid⁃ding client scheduling algorithm is proposed.It can adaptively select a subset of clients in the cell with lower model accuracy to participate in FL training each round,thus the fairness of the client selection is improved.The simulation experiments show that the proposed algorithm can significantly increase the model accuracy while reducing the training time,compared with several related state-of-the-art FL algo⁃rithms,and thus the effectiveness is verified.

关 键 词：联邦学习 异构性 三维网格 客户端选择 多臂老虎机 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]