Adaptive In-Network Traffic Classifier: Bridging the Gap for Improved QoS by Minimizing Misclassification

In-network traffic classification presents an innovative approach to developing early-stage and accurate traffic classification solutions. However, despite its initial accuracy, the one-size-fits-all Machine Learning (ML) model becomes obsolete as traffic patterns evolve. This evolution in traffic patterns inevitably leads to misclassification, resulting in the erroneous mapping of traffic flows to Quality of Service (QoS) classes. Consequently, misclassification may lead to service quality violations, imposing penalties on Infrastructure Providers (InPs). The impact, however, is not solely tied to misclassification rates, as a multi-path network with paths of varying capacities can redirect traffic from low data rate classes to high data rate paths and vice versa, thereby influencing the overall outcome. Therefore, precisely quantifying the impact of misclassification on network performance, i.e., QoS, is paramount. This research aims to investigate and address the effects of traffic misclassification on Service Level Agreement (SLA) violations within multi-class, multi-path networks. To achieve this, we propose a novel framework to quantify SLA violations caused by misclassification, an economic model to assess its impact on provider profitability, and adaptive ML techniques to enhance traffic classification accuracy continually. The evaluation results reveal that the optimal path allocation for various traffic classes determines the targeted revenue. Meanwhile, the adaptivity of the classifier maintains prediction accuracy, ensuring the integrity of SLA through precise QoS class assignments. Hence, implementing an adaptive traffic classifier can mitigate penalties and sustain profitability. This work provides valuable insights for network operators, enabling effective misclassification management, resource optimization, and the maintenance of SLA integrity.