EDDL: A Distributed Deep Learning System for Resource-limited Edge Computing Environment

Abstract

This paper investigates the problem of performing distributed deep learning (DDL) to train machine learning (ML) models at the edge with resource-constrained embedded devices. Existing solutions mostly focus on data center environments, where powerful serverclass machines are interconnected with ultra-high-speed Ethernet, and are not suitable for edge environments where much less powerful computing devices and networks are used. Due to the resource constraint on computing devices and the network connecting them, there are three main challenges for performing edge-based DDL: (1) susceptibility to struggling workers, (2) difficulty of scaling up to a large training cluster, and (3) frequent changes in training device availability and capability. To address these challenges, we design and implement EDDL, an edge-based DDL system, with ARM-based ODROID-XU4 and Raspberry Pi 3 Model B boards. We evaluate the prototype EDDL system by performing edge-based mobile malware detection and classification on a large Android APK dataset. The evaluation results show that EDDL can efficiently train deep learning models with consumer-grade embedded devices and wireless networks while incurring small overhead.