Novel Approach for Curbing Unfair Energy Consumption and Biased Model in Federated Edge Learning

Researchers and practitioners have recently shown interest in deploying federated learning for enhanced privacy preservation in wireless edge networks. In such settings, resource-constrained user equipment (UE) often experiences unfair energy consumption and performance degradation of machine learning models due to data heterogeneity and constrained computation and communication resources. Several approaches have been proposed in the literature to reduce energy consumption, including scheduling a subset of UEs to undertake learning tasks based on their energy budgets. However, these approaches are inherently unfair as the frequently selected UEs rapidly deplete their energy and are rendered inaccessible. Furthermore, the server may be unable to capture the incongruent data distribution, resulting in a biased model. In this paper, we propose a novel approach that addresses those challenges, considering the historical participation of the UEs to ensure that all the training data of the UEs are incorporated into the global model. Specifically, using Jain’s fairness index, we formulate the overall optimization problem, decompose it into two sub-problems, and iteratively solve the sub-problems. Towards this end, we partition the optimization variables into two-blocks; one on the server-side and another on the UEs’ side. The server-side algorithm aims to balance energy usage and learning performance, while the client-side algorithm seeks to optimize CPU frequency and transmit power. Extensive experiments using two realistic datasets, FEMNIST and CIFAR-10, indicate that the proposed algorithms minimize overall energy while curbing unfair energy consumption between the UEs, accelerating convergence rates, and significantly enhancing local accuracy for all UEs.