A multi-level intrusion detection system for industrial IoT using bowerbird courtship-inspired feature selection and hybrid data balancing

Industrial Internet of Things (IIoT) deployments demand intrusion-detection systems (IDSs) that are both accurate and lightweight. We propose a two-tier IDS in which a Logistic Regression (LR) model running at the edge classifies the bulk of traffic, while about 16% of low-confidence flows are forwarded to a cloud-hosted 1-D Convolutional Neural Network (CNN) for deeper inspection. A Bowerbird Courtship-Inspired Feature Selection (BBFS) algorithm reduces the input space at both tiers-using a recall-weighted F1 wrapper for LR and a Fisher-score & correlation-penalty filter for the CNN-while a hybrid SMOTE+ENN+LOF strategy mitigates class imbalance. On the WUSTL-IIoT data set, the edge model with balancing plus BBFS attains 99.70% accuracy and 99.71% recall, and the cloud CNN reaches approximately 100.00% accuracy with 99.99% precision and a 99.98% F1-score. Feature reduction lowers LR inference from 52 ms to 38 ms and CNN inference from 27.86 s to 18.28 s on a Raspberry Pi 4 B, demonstrating the framework’s suitability for real-time, resource-constrained IIoT security.