A Novel Joint Dataset and Computation Management Scheme for Energy-Efficient Federated Learning in Mobile Edge Computing
Jingyeom Kim, Doyeon Kim, Joohyung Lee, Jungyeon Hwang
Abstract
In this letter, a novel joint dataset and computation management (DCM) scheme for energy-efficient federated learning (FL) in mobile edge computing (MEC) is proposed. For this purpose, with respect to the amount of dataset and computation resources, we rigorously formulated analytical models for i) learning efficiency, which considers the estimated global accuracy tendency according to the amount of dataset and service latency, and ii) the overall energy consumption of FL participants, including local training and model parameter transmission. To consider the trade-off between these two factors in the FL procedure with MEC, a theoretical framework for the DCM problem that jointly optimizes the amount of dataset and the computation resources used for local training over multiple FL clients was designed. Additionally, the extensive simulation-based performance evaluations validate the superior performance of the proposed DCM; compared to the various benchmarks in terms of the proposed cost function and test accuracy on the MNIST dataset with independent identically distributed (IID) / non-IID settings.