边缘计算网络中联邦学习算法设计与优化  

Design and Optimization of Federated Learning Algorithms in Edge Computing Networks

作　　者：陈柱[1] 莫俊彬 陈青霞 叶绍雄 郭俊滨 CHEN Zhu;MO Junbin;CHEN Qingxia;YE Shaoxiong;GUO Junbin(China United Network Communications Co.,Ltd.,Qingyuan Branch,Qingyuan 511500,China)

机构地区：[1]中国联合网络通信有限公司清远市分公司,广东清远511500

出　　处：《移动通信》2024年第12期122-128,共7页Mobile Communications

摘　　要：针对海量物联网终端训练所面临的网络资源瓶颈和数据安全问题,提出边缘计算网络中联邦学习算法设计与优化。首先,设计端边云三层联邦学习架构实现高效的联邦学习,设备端负责特征提取,边缘端和云端分别负责模型参数训练和整合;然后,引入模型压缩技术构建通信模型和能耗模型并对联邦学习问题进行建模,最小化云端模型训练的时延和能耗;最后,设计一种粒子群的方法获得原始问题的最优解,实现算法快速收敛,在一定程度上降低联邦学习算法的能耗和时延。仿真表明:相较于其他算法,该算法可以适应大规模的物联网场景,在保证算法收敛速率的同时,还在一定程度上降低了能耗和时延。Aiming at the bottleneck of network resources and data security problems faced by massive IoT terminals'training,this paper proposes the design and optimization of federated learning algorithms in edge computing networks.Firstly,an end-edge-cloud three-layer federated learning architecture is designed to achieve efficient federated learning.The device side is responsible for feature extraction,while the edge and cloud sides are responsible for model parameter training and integration,respectively.Then,model compression techniques are introduced to construct communication and energy consumption models and to model the federated learning problem,minimizing the latency and energy consumption of cloud-based model training.Finally,a particle swarm optimization method is designed to obtain the optimal solution of the original problem,achieve fast convergence of the algorithm,and to some extent reduce the energy consumption and latency of federated learning algorithms.Simulation results show that compared to other algorithms,the proposed algorithm can be adapted to large-scale IoT scenarios with the ensurance of the convergence rate of the algorithm,while reducing energy consumption and latency to a certain extent.

关 键 词：边缘计算网络 联邦学习 通信模型 能耗模型 

分 类 号：TN929.5[电子电信—通信与信息系统]