Adapting at Line-Speed: A Demonstration of In-Network Concept Drift Detection

We present a fully in-network system for detecting and adapting to concept drift in real-time traffic classification. Implemented on a P4-programmable switch, our architecture integrates an Isolation Forest for one-class classification, a multiclass classifier for packet labeling, and a lightweight, bitwise EWMA-based drift detector, all within the data plane. When significant traffic shifts are detected, the control plane triggers automated model retraining and redeployment, ensuring continued accuracy at line rate. This demo highlights the feasibility and effectiveness of adaptive, resource-aware Machine Learning in high-speed networks.