Abstract:

This paper addresses the critical challenge of traffic-aware routing in modern Software-Defined Networking (SDN) environments, which are increasingly defined by the dual pressures of pervasive traffic encryption and the demand for real-time, adaptive network control. The widespread adoption of protocols like TLS 1.3 and QUIC, particularly with privacy-enhancing features such as Encrypted Client Hello (ECH), has rendered traditional visibility tools obsolete, hindering intelligent network management. Concurrently, controller-centric machine learning approaches introduce significant performance bottlenecks and lack the transparency required for operator trust. To overcome these limitations, this work introduces the Lightweight, Explainable, and Programmable (LEAP) framework. LEAP presents a novel, synergistic architecture that integrates a Deep Reinforcement Learning (DRL) agent in the control plane for adaptive, high-level routing policy generation; a highly efficient, lightweight Gradient Boosting Decision Tree (GBDT) classifier, compressed via knowledge distillation and deployed in P4- programmable data planes for line-rate traffic identification; and a dedicated Explainable AI (XAI) module to provide human- interpretable justifications for both classification and routing decisions. Through extensive emulation in a realistic network environment using modern, encrypted traffic datasets, the LEAP framework demonstrates significant improvements in network throughput and end-to-end latency compared to state-of-the-art baselines, establishing a new paradigm for efficient, transparent, and autonomous network management.