System Partitioning with Virtualization for Federated and Distributed Machine Learning on Critical IoT Edge Systems

Abstract

Machine learning transforms the fledgling IoT landscape by making meaningful business decisions utilizing data from a vast number of sensors. However, the scale of connected devices puts a toll on system networks. Federated and distributed learning systems have been introduced to offload the network stress into edge and fog nodes. However, this approach presents a new challenge in integrating and deploying machine learning algorithms into existing systems. Due to the complex nature of machine learning algorithms and the associated data interaction paradigms, most traditional edge node systems today require a total system re-architecture to incorporate machine learning on the edge. This paper presents a novel virtualization-based system partition approach to system design that enables the execution of machine learning algorithms on edge nodes without modifications to existing software and hardware in a system. In addition to easing the development process, this approach also prevents inadvertent introduction errors by virtue of complete memory isolation of the learning systems on the same hardware.