Abstract
The advance in machine learning (ML) has greatly bolstered many real-time applications of Internet-of-Things (IoT). However, IoT devices can barely afford complex ML models to meet requirements of latency, accuracy, and energy efficiency. A remedy to these issues is to leverage edge computing for quicker responses and richer computational resources in the proximity of end users. The goal of this dissertation is therefore to investigate the feasibility and problems when applying ML to edge computing, from deploying ML models at IoT devices, to edge servers, and to a hierarchy. In the first work, we explore the advantage of applying ML at IoT devices to solve a dilemma between saving energy for wearable IoT devices (powered by small batteries) and reliable data transmission of body sensor networks (BSN) for healthcare applications. We propose a cross-layer adaptive MAC scheduling scheme driven by user behaviors which are detected by a trained ML model at the IoT gateway of BSN. We build a BSN testbed which demonstrates that the ML-based adaptive MAC scheduling scheme ensures reliable data transmission and energy efficiency simultaneously. In the second work, we study the problem of mobile users (MUs) in edge computing who offload inference tasks to deep ML models running in nearby edge servers. Due to the user mobility, we need to both migrate the offloaded services—Docker containers—to new edge servers for a lower end-to-end (E2E) delay, and hand over wireless communication links to new base stations. Thus, we propose a coordinated migration-handover mechanism to jointly achieve low E2E delay and minimal service interruption. We implement an edge computing testbed with our proposed mechanism to evaluate the performance, and demonstrate that the mechanism achieves 30%-40% lower service downtime and 13%-22% lower E2E delay as compared to other baseline mechanisms. In the final work, we address the trade-off between high accuracy and low detection delay in anomaly detection, where the former is typically achieved by offloading anomaly detection tasks to a complex ML model in the cloud, and the latter by running an adequate ML model at IoT devices. To this end, we propose a contextual-bandit anomaly detection scheme for IoT data in distributed hierarchical edge computing (HEC) systems. This scheme deploys a series of anomaly detection ML models in a HEC system from the bottom layer to the top, with increasing complexity, and adaptively selects one of the ML models using a contextual-bandit policy neural network. We have implemented our proposed approach on a real HEC testbed, and evaluated it using real IoT datasets. The evaluation results show that our proposed approach achieves the best accuracy-delay trade-off in comparison with several other baseline schemes.