Energy-efficient Federated Learning for Data Analytics in Fog Networks

Abstract

Federated Learning(FL) is a collaborative and distributed machine learning technique that enables training over many clients without sharing the client’s data. The advent of a massive number of low-powered Internet of Things (IoT) devices and local fog devices with sufficient computational power have made it possible to take advantage of this distributed framework in real-life scenarios. However, the standard IoT-enabled fog framework suffers from significant energy expense due to the intercommunication between the computing devices. The existing state-of-the-art strategies have proposed altering the core architecture to reduce energy expenses that work only under ideal conditions on independent and identically distributed (IID) data. Nevertheless, the vast deployment of low-cost sensor devices in use cases like Smart Agriculture makes it impossible for such ideal conditions to prevail in real life. Motivated by the above-mentioned challenges, in this thesis, an energy-efficient fog framework for smart irrigation is proposed to mitigate these issues. The proposed algorithm utilizes data sampling and optimal resource provisioning methodologies to maximize resource utilization, which results in a significant energy reduction in the framework. Besides that, the local gateway devices of the proposed fog framework serve as functional units based on redundant data filtering, outlier removal, and lossy data aggregation to minimize data transmission. The analysis of this proposed model is done by training on data from agricultural field sensors using a data simulator to predict irrigation requirements. From the simulation results, it is observed that the proposed algorithm reduces the total energy consumption by 51.5% and 15.2% compared with Split Learning(SL) and standard FL, respectively, while achieving the prediction accuracy of 91.1%.