An efficient federated learning method based on the proportional-integral-derivative mechanism for embedded devices
TL;DR
An efficient federated learning method based on the PID mechanism (FL-PID) to adaptively adjust clients’ local training rounds and aggregation weights, which effectively solves the client data and computational heterogeneity problems in FL.
Abstract
For federated learning (FL) deployed on embedded devices, clients (embedded devices) face severe data skew and diverse hardware capabilities. This leads to both data and computational heterogeneity, and the level of heterogeneity varies from device to device. To address the above challenges, this paper proposes an efficient federated learning method based on the PID mechanism (FL-PID). Firstly, we construct a federated learning testbed with multiple embedded devices, and assign training samples with different quantities and heterogeneity levels to each device. Secondly, we propose to apply the PID mechanism to adaptively adjust clients’ local training rounds and aggregation weights, which effectively solves the client data and computational heterogeneity problems in FL. Finally, the proposed approach is validated on the MNIST, Fashion-MNIST and Synthetic datasets. Experimental results show that in comparison with existing approaches, our method achieves faster convergence, and reaches an accuracy of 93.54%, 74.62% and 72.06% on MNIST, Fashion-MNIST and Synthetic datasets respectively.
