A Deployable Platform for Real-World Network Traffic Classification Edge Devices

Network Traffic Classification (NTC) is fundamental to network management, service differentiation, and performance optimization. However, most existing studies rely on offline datasets or controlled simulations, limiting their applicability to real operational environments. This paper presents a deployable end-to-end NTC platform designed for continuous monitoring of live network traffic. The system operates as an edge-side passive probe that captures Ethernet traffic non-intrusively using TShark and extracts$\mathbf{2 8}$-dimensional statistical flow features within sliding observation windows. To support real-time classification, we introduce a Per-Feature Normalized Mean Square Error (PFNMSE) metric that normalizes feature-wise deviations between live samples and a pre-built application feature library. Each incoming flow is categorized by identifying the library entry with the smallest PF-NMSE score. Experiments on real network sessions demonstrate that the proposed PF-NMSE-based classifier accurately recognizes application types and streaming resolutions under varying bandwidth conditions, achieving an average accuracy of 97%. The platform can be directly deployed in operational networks and serves as a practical solution for continuous traffic monitoring and application-level behavior analysis.