SMART TREE NET: HYBRID AND ENTROPY-DRIVEN MACHINE LEARNING FOR HIGH-PERFORMANCE SDN PACKET CLASSIFICATION

Abstract

 Software-Defined Networking (SDN) environments increasingly encounter encrypted data flows and evolving attack behaviors, complicating the task of reliable packet identification. This study introduces SmartTreeNet, a hybrid machine learning framework developed to enhance packet classification performance in SDN-based architectures. The framework combines Random Forest, LightGBM, and a shallow decision-tree meta-layer to balance accuracy, throughput, and processing efficiency. An entropy-based feature refinement strategy is further incorporated to eliminate redundant attributes, thereby lowering computational overhead. The models were trained using publicly available SDN traffic datasets and evaluated on Mininet-emulated network flows to ensure reproducible evaluation under near-real conditions. In baseline experiments, SmartTreeNet achieved 99.75% classification accuracy with throughput beyond two million inferences per second, while an entropy-optimized variant sustained 99.44% accuracy and reduced controller CPU load by 3.62%. Comparative tests against standalone classifiers (Random Forest, LightGBM, and a 1D CNN) demonstrated that SmartTreeNet delivers a superior accuracy–latency trade-off, maintaining extremely low latency and high scalability without sacrificing accuracy. These results show that the proposed entropy-guided hybrid approach provides real-time SDN operation capability and improved controller efficiency, making it suitable for adaptive defense against encrypted and malicious network traffic in modern programmable networks.