资源异构场景下事件触发联邦学习机制研究  

Research on Event-Triggered Federated Learning Mechanism in Heterogeneous Resource Scenarios

作　　者：舒龙 SHU Long(College of Physics and Electronic Engineering,Shanxi University,Taiyuan,030006,China)

机构地区：[1]山西大学物理电子工程学院,太原030006

出　　处：《网络新媒体技术》2025年第1期9-15,共7页Network New Media Technology

基　　金：山西省重点研发计划资助项目(编号:202202020101004)。

摘　　要：联邦学习中中心服务器与各客户端之间频繁地进行模型参数交换,产生巨大的通信开销,如何减小通信开销成为了联邦学习的研究热点。首先,通过引入事件触发的通信机制降低客户端与中心服务器之间的通信频率,从而降低联邦学习的通信开销;考虑到客户端设备之间的资源异质性,根据各个设备的通信能力、计算能力、数据质量设置一个与资源异质性相关的阈值,提出基于资源异质性的事件触发联邦平均算法(FedAvg-RHB)。其次,为了提高全局模型精度,采用全局模型重用策略。最后,在MNIST数据集上对所提算法及其对比算法进行python仿真,通过对实验结果的对比分析,验证了所提算法是一种降低联邦学习通信开销的有效算法。Model parameters are frequently exchanged between the central server and each client in federated learning,which generates huge communication cost,and how to reduce the communication cost has become a research hotspot in federated learning.Firstly,the event-triggered communication mechanism is introduced to reduce the communication frequency between the client and the central server,thereby reducing the communication cost of federated learning.Considering the resource heterogeneity between client devices,a threshold related to resource heterogeneity was set according to the communication ability,computing power and data quality of each device,and a Resource Heterogeneity-based event-triggering strategy in the Federated Average algorithm(FedAvg-RHB)was proposed.Secondly,in order to improve the accuracy of the global model,the global model reuse strategy was adopted.Finally,python simulation is carried out on the proposed algorithm and its comparison algorithms on the MNIST dataset,and the experimental results are compared and analyzed,which proves that the proposed algorithm is an effective algorithm to reduce the communication cost of federated learning.

关 键 词：通信开销 联邦学习 事件触发 资源异质性 模型重用 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]