Implementing collaborative decentralized machine learning for Internet of Things (IoT)

Abstract

Federated Edge Learning (FEL) is a novel technique for collaborative machine learning through distributed edge devices that offers improved data privacy over traditional machine learning training that is centralized in nature. In FEL, a global model is downloaded on edge devices where training is done locally. Updated local models are then aggregated back to a central server and used to update the global model. Through this, data remains on-device. However, FEL still faces several key challenges as a nascent field, and existing implementations incur high communication overhead between edge devices and the central server. In order to overcome some of these challenges, the blockchain-empowered federated edge learning (BFEL) framework was proposed by Kang et al. [4] In this paper, the FEL aspect of the BFEL is evaluated and explored. Firstly, a study of hyperparameters in both IID and non-IID environments is done in order to identify any relationships particular to FEL. Following which, several gradient compression schemes are implemented and evaluated in the BFEL setting to explore their viability in reducing communication overhead cost with little degradation in model performance.