TOWARDS IN-NETWORK IMAGE CLASSIFICATION FOR LATENCY-CRITICAL IOT APPLICATIONS

Abstract

As the demand for latency-critical applications (intelligent transport systems, medical imaging, surveillance, AR/VR) continues to increase, tasks such as computer vision, which are imperative to these applications, require expedited processing to keep up with this latency requirement. However, the distant cloud servers and low-powered IoT devices could not offer that latency, which then force the realm of edge computing. It exists as an intersection between offering computational power at a considerably lower latency access point. Edge servers have less computing resources compared to the distant cloud, which initiated a plethora of works on distributing load among IoT devices and edge servers. This thesis introduces a novel in-network computing paradigm by leveraging programmable routers sit between IoT devices and edge servers. Specifically, we propose a novel in-network framework, NetPixel for learning-based operations, e.g., image classification, to support the latency demand of the aforementioned IoT applications. The framework incorporates two ML and DL-based classifiers: decision tree and convolutional neural network. While the former can interpret the model better, the latter is the standard image classifier. We implement a prototype of NetPixel over programmable software switch Bmv2 and perform an extensive evaluation using four standard datasets. The results reveal that NetPixel offers a competitive performance compared to traditional server-based solutions while can process network traffic at high speed (Tbps) to support latency-critical applications.