Federated Learning Game in IoT Edge Computing

Abstract

Edge Computing provides an effective solution for relieving IoT devices from the burden of handling Machine Learning (ML) tasks. Further, given the limited storage capacity of these devices, they can only accommodate a restricted amount of data for training, resulting in higher error rates for ML predictions. To address this limitation, IoT devices can leverage Edge Computing and collaborate in the learning process through a designated peer acting as an Edge device. However, the transmission of offloaded tasks over a wireless access network poses challenges in terms of time and energy consumption. Consequently, although collaborative learning can diminish the variance of the learned model, it introduces a communication cost, dependent on the chosen Edge device. In light of these considerations, this paper introduces a coalition formation game that proposes a distributed Federated Learning approach, where devices autonomously and efficiently select the most suitable Edge device, aiming to minimize both their learning error and communication cost.

Keywords

• Edge computing
• game theory
• federated learning
• linear regression
• IoT