基于代理选举的高效异构联邦学习方法

Efficient and heterogeneous federated learning based on agent election

物联网多样性终端设备在计算、存储、通信方面的异构性导致联邦学习效率不足。针对上述联邦训练过程中面临的问题,基于代理选举思路,提出了一种高效联邦学习算法。设计了基于马氏距离的代理节点选举策略,将设备的计算能力与闲置时长作为选举因素,选举性价比高的设备作为代理节点,充分发挥设备计算能力。进一步设计了基于代理节点的新型云边端联邦学习架构,提升了异构设备之间的联邦学习效率。基于MNIST和CIFAR-10公开数据集与智能家居设备真实数据的实验表明,该联邦学习方法的效率提高了22%。

The heterogeneity of diverse end devices in terms of computation,storage,and communication leads to insufficient accuracy and efficiency in federated learning.To address the issues faced in the aforementioned federated training process,this paper presented an efficient federated learning algorithm based on the idea of device agent election.To select agent nodes from diverse devices,it designed a device agent node election strategy based on Mahalanobis distance by considering the devices’computational capabilities and idle time as election factors to fully leverage their computing power.Furthermore,it proposed a novel cloud-edge-end federated learning architecture using the agent node to improve the efficiency of federated learning between heterogeneous devices.Experimental results based on the MNIST and CIFAR-10 public datasets and the practical smart home datasets demonstrate that the proposed efficient federated learning algorithm achieves average improvement about 22%in learning efficiency.