An Energy-Aware Multi-Criteria Federated Learning Model for Edge Computing

Abstract

The successful convergence of Internet of Things (IoT) technology and distributed machine learning have leveraged to realise the concept of Federated Learning (FL) with the collaborative efforts of a large number of low-powered and smallsized edge nodes. In Wireless Networks (WN), an energy-effcient transmission is a fundamental challenge since the energy resource of edge nodes is restricted. In this paper, we propose an Energyaware Multi-Criteria Federated Learning (EaMC-FL) model for edge computing. The proposed model enables to collaboratively train a shared global model by aggregating locally trained models in selected representative edge nodes (workers). The involved workers are initially partitioned into a number of clusters with respect to the similarity of their local model parameters. At each training round a small set of representative workers is selected on the based of multi-criteria evaluation that scores each node representativeness (importance) by taking into account the trade-off among the node local model performance, consumed energy and battery lifetime. We have demonstrated through experimental results the proposed EaMC-FL model is capable of reducing the energy consumed by the edge nodes by lowering the transmitted data.

Distributed and Adaptive Edge-based AI Models for Sensor Networks (DAISeN)

“…However, concerning the latter category of approaches, it is vital to find out more efficient FL schemes other than FedAvg, which converge with the same speed as FedAvg and apply to any FL applications [82]. For example, the studies in [83] and [84] have explored an approach that applies clustering optimization to bring efficiency and robustness in FL's communication: only the most representative updates are uploaded to the central server for reducing network communication costs.…”