Implementing a Traffic Classifier on Embedded SoC Systems with Deep Learning Processor Unit

This paper presents the design and implementation of a traffic classifier on a Zynq UltraScale MPSoC platform using a Deep Learning Processor Unit (DPU). The system classifies flows from applications such as YouTube, Netflix, Teams, Roblox, E-Mail, and File Transfer. A Fully Connected Neural Network (FCNN) was trained and deployed using Vitis Artificial Intelligence (AI) tools, achieving an average accuracy of 80.5% on hardware. The classifier operates at 700 flows per second, demonstrating the feasibility of efficient AI inference in resourceconstrained environments.