边缘联邦学习的客户端选择机制

Client selection mechanism of federated learning in edge computing

针对传统边缘联邦学习(FL)由于客户端资源异质性导致联邦学习模型性能低下等问题,提出面向边缘计算的联邦学习客户端选择机制。该机制综合考虑了客户端的计算资源、通信资源以及数据资源,在联邦学习每轮给定的时间阈值内,使得边缘服务器能够选取尽可能多的客户端数量的同时避免资源不足的客户端,保证参与到联邦学习过程中的客户端的质量,在一定程度上降低了联邦学习的训练成本。该联邦学习客户端选择机制在MNIST和CIFAR-10数据集上与现有的联邦学习客户端选择算法——联邦平均算法(FedCS)和基于多标准的联邦学习客户端选择算法(FedMCCS)进行了对比模拟实验,实验结果表明当所提方法达到FedCS和FedMCCS的最终精度时:在MNIST数据集上时间消耗分别减少了79.55%和72.73%,且最终精度分别提升了2.0%和1.8%;在CIFAR-10数据集上时间消耗分别减少了70.83%和70.83%,且最终精度分别提升了23.6%和27.8%。实验结果验证了提出的客户端选择算法能够有效提升联邦学习的效率。

Aiming at the problem of low performance of traditional edge federated learning model due to heterogeneity of client resources,a client selection mechanism for edge computing was proposed.It comprehensively considered the computing resources,communication resources and data resources of clients.In each time threshold within each round of federated learning,the edge server could select as many clients as possible while avoiding clients with insufficient resources,thus ensuring the quality of clients participating in the process of federated learning and reducing the training cost of federated learning to some extent.Compared with the existing federated learning client selection algorithms FedCS and FedMCCS on MNIST and CIRAR-10 datasets,the simulation results show that when the proposed method achieves the final accuracy of FedCS and FedMCCS:the time consumption on MNIST was reduced by 79.55%and 72.73%,respectively,and the final accuracy was improved by 2.0%and 1.8%,respectively;the time consumption on CIFAR-10 was reduced by 70.83%and 70.83%,respectively,and the final accuracy was improved by 23.6%and 27.8%,respectively.Experimental results show that the proposed client selection algorithm can effectively improve the efficiency of federated learning.