Explainable AI for Intrusion Detection on Live Multi-Service Honeypot Telemetry

Final benchmark results for all six model architectures (Random Forest, XGBoost, LightGBM, MLP, LSTM, Transformer) evaluated on the official UNSW-NB15 train/test split and on CIC-IDS2017. Reports F1, ROC-AUC, MCC, and balanced accuracy. The `Verification` column indicates whether each row was independently verified (three-seed mean) or author-verified on the corrected split.