Power allocation and communication resource scheduling for federated learning in wireless IoT networks

Abstract

Federated learning (FL) allows devices to train a machine learning model collaboratively without compromising data privacy. In wireless networks, FL presents challenges due to limited resources and the unstable nature of transmission channels that can cause delays and errors that compromise the consistency of global model updates. Furthermore, efficient allocation of communication resources is crucial in Internet of Things (IoT) environments, where devices often have limited energy capacity. This work introduces a novel FL algorithm called DFed-wOptDP, designed for wireless networks within the IoT framework. This algorithm incorporates a device selection mechanism that evaluates the quality of device data distribution and connection quality with the aggregate server. By optimizing the power allocation for each device, DFed-wOptDP minimizes overall energy consumption while enhancing the success rate of transmissions. The simulation results demonstrate that DFed-wOptDP effectively operates with low transmission power while preserving the accuracy of the global model compared to other algorithms.