Energy-efficient client selection in federated learning with heterogeneous data on edge

Due to large scale deployment of machine learning applications, a vast amount of data is increasingly generated from mobile and edge devices. Federated Learning (FL) has recently attracted a lot of attention from both industry and academy to explore the potential of such data. It is a distributed optimisation paradigm where a central server coordinates learning from heterogeneous data distributed across a wide range of clients. Typical participating clients in FL are energy-restricted mobile devices, and thus energy efficiency is a key challenge. One approach to reduce energy cost is to choose only a small number of suitable clients to finish training tasks. However, the current approach of the random selection method tends to require more participants than needed. Therefore, in this paper, we propose FedNorm, a client selection framework that finds the clients that provide significant information in each round of FL training. Furthermore, based on FedNorm, we further propose a more energy-efficiency variant that requires only the client selection to be conducted every certain round. With extensive experiments in PyTorch implementation and FEMNIST-based datasets, the evaluation results demonstrate that the proposed algorithms outperforms existing client selection methods in FL in various heterogeneous data distribution properties, and reduces energy cost by decreasing the number of participating clients.