Implementing an Attention-Guided Quantum-Inspired Extreme Learning Machine with Grey Wolf Optimization for Rare False-Data-Injection Detection in a Water-Treatment SCADA Network: A Simulation-Based Evaluation

Abstract

Background. A supervisory control and data acquisition (SCADA) network for a water-treatment plant carries sensor and control traffic that is overwhelmingly routine, so a rare false-data-injection attack, in which falsified sensor readings or forged control commands are fed into the system to mask an unsafe physical state, is both the most consequential and the hardest attack class for a learning-based detector to find: it is rare by construction and must be told apart from far more common, lower-stakes network anomalies. Shahiduzzaman et al. proposed an attention-guided, quantum-inspired extreme learning machine tuned by a grey wolf optimizer (GWO) for exactly this class of rare-attack detection problem. Methods. This study implements that method as a SCADA-traffic classifier: a quantum-inspired qubit-angle initialisation of an extreme learning machine's (ELM) input layer, a per-feature attention vector rescaling the ten SCADA traffic features used here, and a grey wolf optimizer that jointly tunes the attention vector, hidden-layer width, and ridge strength against a three-fold cross-validated G-mean/rare-class-F1 objective. It is evaluated on a synthetic, severely imbalanced water-treatment SCADA traffic stream (91% benign, 5%/2.7% two common anomaly classes, 1.3% rare false-data injection) against a plain, untuned ELM and a Random Forest, over five independent seeds. Results. The proposed method improved rare-class F1 from 0.365 (plain ELM) to 0.637 and G-mean from 0.379 to 0.534, each a 65-75% relative gain over the un-tuned, non-attentive ELM it extends, and it exceeded Random Forest on every aggregate metric (macro-F1 0.693 versus 0.660; G-mean 0.534 versus 0.480; rare-class F1 0.637 versus 0.575). The optimizer's tuned attention vector assigned its two highest average weights, across five seeds, to value-deviation-z and setpoint-delta, the two features the false-data-injection signal was planted on. Conclusion. In this synthetic but severely imbalanced SCADA setting, implementing Shahiduzzaman et al.'s attention-guided, quantum-inspired, GWO-tuned ELM delivered the strongest result of the two IIoT scenarios evaluated in this study series, outperforming both the plain ELM it extends and a considerably more expensive Random Forest ensemble on every rare-class and balanced-accuracy metric reported.