Towards In-Network Drift-Aware Traffic Classification

Abstract

Recent advances in intelligent data-plane designs have enabled efficient inference of machine learning models on programmable switches using P4. However, existing solutions focus primarily on classifying traffic that belongs to previously observed classes, i.e., in-distribution traffic, and often overlook the impact of concept drift, i.e., the emergence of out-of-distribution traffic. In this paper, we propose an approach that not only accurately classifies in-distribution traffic but also detects drifting samples that deviate from known classes. Our approach leverages a triplet network to learn an encoder that maps traffic input features to a latent space where representations of the same class form compact clusters. Drifting samples are identified based on their distance to class centroids in the latent space. Experimental results on two use cases demonstrate that our method achieves superior drifting sample detection performance compared to confidence-based schemes, while maintaining comparable in-distribution classification accuracy.