Intelligent Data-Planes for Network Traffic Management

Abstract

Network traffic management is increasingly complex due to the exponential growth of connected devices and bandwidth-intensive applications. As the volume of network traffic continues to rise, managing this data flow efficiently is a key challenge for modern communication networks. Traditional management techniques, which rely on static, protocol-driven methods, struggle to keep pace with the dynamic demands of today’s networks. Emerging technologies like software-defined networking (SDN) and machine learning (ML) offer promising solutions by enabling more intelligent, real-time network management. However, in traditional ML-assisted SDN architectures, ML models are deployed in the control-plane, which relies on receiving relevant traffic information from the data-plane for analysis. This design poses a key limitation, as frequent control-plane/data-plane communication introduces latency that can delay time-sensitive services such as congestion mitigation and intrusion response. This thesis addresses this challenge by proposing a novel approach: introducing ML directly into the data-plane to enable real-time, autonomous decision-making, thus reducing the delays associated with traditional SDN architectures. The primary objective of this research is to design and implement an intelligent data-plane with built-in ML inference, enabling real-time, local decisions and reducing reliance on the control-plane. First, we develop a quantization-aware ML toolbox that facilitates the training of ML models while simplifying their storage and execution within resource-limited data-planes. This approach ensures that quantized model inference can be effectively implemented in the data-plane, satisfying its operational and computational constraints. Second, to enable multi-phase decision-making within the data-plane, we design a confidence-based intrusion detection system that detects malicious flows at both early and later phases by leveraging the confidence level from early detection. Third, to support concurrent management tasks, we develop a novel in-network multi-task learning framework that performs simultaneous inference for multiple tasks in the data-plane. This approach is both resource-efficient and more accurate than single-task models by sharing feature representations among related tasks. Additionally, we enhance scalability by supporting distributed deployment, where different layers of a multi-task model can be offloaded across multiple switches. Finally, we address the challenge that offline trained models often struggle to adapt to dynamic network environments, where changing traffic patterns can degrade performance. We design an unsupervised drift detection mechanism in the data-plane that monitors distributional changes in traffic and triggers model updates when drift is detected. In addition, we present an in-network drift-aware traffic classification framework that not only classifies known traffic accurately but also identifies drifting samples that deviate from all known classes.