Development of an Adaptive and Explainable IIoT Intrusion Detection System Using Novel Ensemble Techniques

This project presents an intrusion detection system for IIoT networks using the X-IIoTID dataset. The system applies three novel approaches to improve classification accuracy across 19 attack and normal classes. The first novelty introduces an attention-weighted ensemble combined with SHAP for explainable predictions. The second novelty adds concept drift detection using ADWIN to support continual learning and model updates. The third novelty uses hierarchical feature clustering with stacked generalization, where specialist tree models handle feature groups and a meta-learner combines their outputs with isotonic calibration.

The web-based application allows users to register, login, perform classification on network traffic features, and view results with confidence scores. Preprocessing includes feature selection via mutual information and exact balancing to 5000 samples per class. Experiments show high performance with accuracy above 99% on the test set. The Flask-based system provides an easy interface for model interaction while maintaining strong detection capabilities.

Keywords: IIoT intrusion detection, X-IIoTID dataset, stacked generalization, hierarchical clustering, attention-weighted ensemble, concept drift detection, continual learning, SHAP explainability, mutual information selection, ensemble calibration.