«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年02期

页码:

98-104

栏目:

移动与物联网络

出版日期:

2024-02-10

文章信息/Info

Title:

Dynamic Federated Learning Optimization Method in Edge Scenarios

文章编号:

1673-629X(2024)02-0098-07

作者:

王志良¹ ; 何 刚¹; 2* ; 俞文心¹ ; 许 康¹ ; 文 军¹ ; 刘 畅¹

1. 西南科技大学 计算机科学与技术学院,四川 绵阳 621010;

2. 国家卫生健康委核技术医学转化实验室(绵阳市中心医院),四川 绵阳 621010

Author(s):

WANG Zhi-liang1 ; HE Gang1; 2* ; YU Wen-xin1 ;  XU Kang1 ; WEN Jun1 ; LIU Chang1

1. School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,China;

2. National Health Commission Laboratory for Translational Medicine of Nuclear Technology ( Mianyang Central Hospital) ,Mianyang 621010,China

关键词:

边缘计算; 机器学习; 联邦学习; 服务动态缩放; 数据不均衡; 设备异构

Keywords:

edge computing; machine learning; federated learning; service dynamic scaling; data imbalance; device heterogeneity

分类号:

TP181

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 02. 015

摘要:

边缘计算( Edge Computing)是一种新的计算方式,通过在网络边缘提供计算服务,与传统的云计算模式相比,具有高可信度和低延迟等特点。 联邦学习( FL)作为一种分布式机器学习方法,

尽管具备保护隐私和数据安全的特性,却仍然面临设备异构和数据不均衡等问题,导致出现部分参与者( 边缘端) 训练时间长、训练效率低下等问题。 为了解决上述问题,该文提出了一种名

为 FlexFL 的动态联邦学习优化算法。 该算法引入了两层联邦学习策略,通过在同一参与者部署多个联邦学习训练服务和一个联邦学习聚合服务,将本地数据集平均分配给各个联邦学习训

练服务,并每回合激活一定数量的训练服务。 未激活的服务将休眠,不占用计算资源,并将资源平均分配给激活的服务,以加快训练速度。 该算法能够平衡参与者设备异构和数据不均衡性带

来的训练时间差异,从而提高整体训练效率。 在 MINST 数据集和 CIFAR 数据集上与 FedAvg 联邦学习算法进行了对比实验,结果显示,FlexFL 算法在减少时间消耗的同时,不降低模型性能。

Abstract:

Edge computing is a new computing paradigm that provides computational services at the network edge. Compared totraditional cloud computing, edge computing offers advantages such as high reliability and low latency. However, federated learning( FL) ,a distributed machine learning method,still faces challenges related to device heterogeneity and data imbal-ance,leading to issueslike prolonged training time and low training efficiency for certain participants ( edge devices) . To address these challenges,we propose adynamic federated learning optimization algorithm called FlexFL. The algorithm introduces a two - tier federated learning strategy bydeploying multiple federated learning training services and a federated learning aggregation service on the same edge device. It evenlypartitions the local dataset among the federated learning training services and activates a certain number of training services per round.Inactive services go into a dormant state, freeing up computing resources and redistributing them evenly among the active services toaccelerate training. The algorithm balances the discrepancies in training time caused by device heterogeneity and data imbalance,therebyimproving overall training efficiency. Experimental comparisons between the FlexFL algorithm and the FedAvg federated learningalgorithm were conducted on the MINST dataset and CIFAR dataset,and the results demonstrate that FlexFL reduces time consumptionwithout compromising model performance.

常用功能

更新日期/Last Update: 2024-02-10