Federated Learning for Intelligent Resources Allocation in Internet of Things

Abstract

By using federated learning (FL), multiple Internet-of-Things (IoT) devices can construct a shared learning model without sending raw data to a centralized server. While FL has come a long way, it still has a ways to go. Issues such as heterogeneous user equipment (UEs) and data that is not independently and uniformly distributed are still obstacles. Facilitating a numerous UEs to participate in the learning in each cycle poses a possible problem of the huge communication budget. A weighted adjoining factor is presented to the localized gradient descent, generalizing the present FedAvg to solve these concerns. At the start of each global round, the proposed FL method randomly selects a fraction of the UEs to perform stochastic gradient descent in parallel. Then, we utilize the suggested FL method in cellular IoT to reduce either total power usage or execution duration of FL, in which a straightforward but effective path-following method is constructed for its explanations. At last, obtained simulations on poorly balanced data are presented to show that the presented FL algorithm is superior to FedAvg in terms of performance with respect to fast convergence. Moreover, they show that the suggested algorithm needs significantly less time and energy to train than the FL algorithm does when users contribute heavily to the learning process. These findings provide strong support for the suggested FL algorithm as a potential paradigm change for training mobile IoT networks with limited bandwidth.