ADAPTIVE REAL-TIME NETWORK INTRUSION DETECTION SYSTEM

Abstract

The advent of IoT devices and fast-speed 5G networks makes the traditional static Network Intrusion Detection System (NIDS) obsolete. The static model, which is trained on a static dataset, is unable to cope with concept drift, or the changing statistical patterns of the network and polymorphic attacks. In this paper, we propose and test an adaptive real-time NIDS based on online deep learning approaches to counter model degradation when working in dynamic settings. We utilize a Hoeffding Adaptive Tree (HAT) integrated with an Adaptive Windowing (ADWIN) concept drift detector. In the prequential simulation of the continuous UNSW-NB15 data stream with 250,000 sequential network packets, we demonstrate that while the static Hoeffding Tree baseline degrades to 96.50% accuracy during critical concept drift events, the adaptive model successfully adapts and maintains 100% accuracy. Furthermore, empirical resource profiling proves the model's suitability for Edge AI deployment, requiring only 198.08 MB of peak RAM with an average inference latency of 0.37 milliseconds per packet. By combining incremental learning techniques with stringent statistical drift analysis, a highly efficient, edge-native security solution is achieved.