Implementing an Attention-Guided Quantum-Inspired Extreme Learning Machine with Grey Wolf Optimization for Rare Command-Injection Detection in a Factory-Automation IIoT Network: A Simulation-Based Evaluation

Abstract

Background. Industrial Internet of Things (IIoT) traffic in factory-automation networks is overwhelmingly benign, so the attack class that matters most operationally, a rare command-injection intrusion that can alter a controller's behaviour, is also the hardest for a classifier to learn: it supplies the fewest training examples against a backdrop of common, lower-stakes anomalies and routine telemetry noise. Shahiduzzaman et al. proposed an attention-guided, quantum-inspired extreme learning machine tuned by a grey wolf optimizer for exactly this rare-attack setting. Methods. This study implements that method end to end: a quantum-inspired qubit-angle initialisation of an extreme learning machine's (ELM) input layer, a per-feature attention vector that rescales inputs before the hidden layer, and a grey wolf optimizer (GWO) that jointly tunes the attention vector, hidden-layer width, and ridge-regularisation strength to maximise a G-mean/rare-class-F1 blend evaluated by three-fold cross-validation. The implementation is evaluated on a synthetic, severely imbalanced ten-feature factory-automation traffic stream (90% benign, 6%/2.5% two common attack classes, 1.5% rare command-injection) against two baselines, a plain randomly-initialised ELM and a Random Forest, over five independent seeds. Results. The proposed method improved rare-class F1 from 0.353 (plain ELM) to 0.574 and G-mean from 0.314 to 0.454, each a roughly 60% relative gain over the un-tuned, non-attentive ELM it extends, and it matched or exceeded Random Forest on every aggregate metric (macro-F1 0.635 versus 0.616; rare-class F1 0.574 versus 0.562) while trading a few points of rare-class precision (0.896 versus 0.954) for substantially more consistent recall. The grey wolf optimizer consistently converged within the allotted 15 iterations, and its tuned attention vector assigned its two highest weights, on average over five seeds, to the two features the command-injection signal was actually planted on. Conclusion. In this synthetic but severely imbalanced setting, implementing Shahiduzzaman et al.'s attention-guided, quantum-inspired, GWO-tuned ELM delivered a substantial, consistent improvement over the plain ELM it is built from and performance competitive with a considerably more expensive Random Forest ensemble, at the cost of materially higher run-to-run variance in rare-class recall that any deploying operator should plan around.