摘要
针对5G网络下的联邦学习架构及关键技术展开研究,通过5G网络的帮助来提升移动终端收集的小样本数据对于训练全局模型的重要意义。从具有不同本地数据集的终端可以加速模型训练和增强模型泛化能力的理论分析入手,详细阐述了如何利用5G系统优势,实现在通信资源约束下选择具有典型特征的终端成员,从而达到联邦学习效果最大化的目的。基于3GPP 5G系统现有架构,提出了支持联邦学习的5G架构以及典型解决方案流程,最后给出了仿真结果,证明了5G网络对于联邦学习具有良好增益。
This paper focuses on the federated learning architecture and key technologies based on the 5G network.5G network can improve the small sample data collected by mobile terminals,which is of great significance for training the global model.Starting from the theoretical analysis that terminals with different local datasets can accelerate model training and enhance model generalization ability,this paper expounds on how to use the advantages of the 5G system to select terminal members with typical characteristics subject to the communication resource constraint,so as to maximize the performance of federated learning.Based on the existing architecture of the 3GPP 5G system,the 5G architecture supporting federated learning and the typical solution flow are proposed.Finally,the simulation results are given,which prove that the 5G network has a good gain for federated learning.
作者
许阳
陈景然
XU Yang;CHEN Jingran(Beijing OPPO Telecommunication Co.,Ltd.,Beijing 100048,China)
出处
《移动通信》
2023年第1期24-28,共5页
Mobile Communications
基金
国家重点研发计划“宽带通信和新型网络”重点专项“6G全场景按需服务关键技术”项目(BZ0300320)。
关键词
5G
核心网
联邦学习
模型训练
5G
core network
federate learning
model training
