← Back to VOLUME 14, ISSUE 11, NOVEMBER 2026
This work is licensed under a Creative Commons Attribution 4.0 International License.
Explainable Network Intrusion Detection Using Two-Stage Tree Ensembles, Novelty-Aware Thresholding and Risk-Based Alert Prioritisation
Shirin Anna Viju*, S. Anitha Jebamani, Swetha A, A. Sujithra Raj, Saranya S
π 9 viewsπ₯ 6 downloads
Abstract: Machine-learning intrusion detectors are frequently evaluated in ways hiding their weakness on attack types absent from training. This paper presents an explainable network intrusion detection system evaluated on the official NSL-KDD split, whose test set contains 17 attack types unseen in training. A two-stage tree-ensemble model detects and categorises attacks, an Isolation Forest flags anomalous flows, and each alert carries a calibrated probability, a SHAP explanation and a rule-based severity score; the pipeline is served through FastAPI and Streamlit. Leave-attack-type-out validation on training data alone selected a LightGBM and Extra Trees ensemble with a 0.02 decision threshold. On the held-out test set it reaches 91.9% accuracy, 93.4% recall and 92.7% unseen-type recall, against 77.9%, 63.3% and 31.1% for the original LightGBM, at a 10.2% false-positive rate instead of 2.9%. The gain comes mainly from the threshold (ROC-AUC 0.963 vs. 0.964); calibration lowers the expected calibration error from 0.214 to 0.050.
Keywords: Intrusion detection, LightGBM, Extra Trees, Isolation Forest, unknown attacks, threshold selection, probability calibration, explainable AI, SHAP, alert prioritisation, NSL-KDD
Keywords: Intrusion detection, LightGBM, Extra Trees, Isolation Forest, unknown attacks, threshold selection, probability calibration, explainable AI, SHAP, alert prioritisation, NSL-KDD
How to Cite:
[1] Shirin Anna Viju*, S. Anitha Jebamani, Swetha A, A. Sujithra Raj, Saranya S, βExplainable Network Intrusion Detection Using Two-Stage Tree Ensembles, Novelty-Aware Thresholding and Risk-Based Alert Prioritisation,β International Journal of Innovative Research in Electrical, Electronics, Instrumentation and Control Engineering (IJIREEICE), DOI: 10.17148/IJIREEICE.2026.141103
