Anomaly Detection in Industrial Control Systems Using Bayesian Optimization-Enhanced Gated Recurrent Neural Networks

Abstract

Industrial Control Systems (ICS) are foundational to critical infrastructure, yet their increasing interconnection with information technology has exposed them to sophisticated cyber threats. Traditional anomaly detection strategies struggle to model the complex temporal dynamics of multivariate process data, while deep learning approaches often rely on manually tuned hyperparameters that compromise generalizability. This study proposes a Bayesian Optimization-Enhanced Gated Recurrent Unit (BO-GRU) model for anomaly detection in ICS environments. Using the Secure Water Treatment (SWaT) dataset, the model integrates GRU-based temporal learning with automated hyperparameter tuning via Bayesian Optimization. Experimental results demonstrate that the optimized model outperforms the baseline GRU across Accuracy 0.994, Precision 0.955, Recall 0.9954, F1-score 0.975, and Receiver Operating Characteristic – Area Under Curve (ROC-AUC) 1.000. Comparative analysis with recent ICS anomaly detection studies confirms the superior robustness, efficiency, and adaptability of the proposed BO-GRU framework. The findings highlight the potential of automated optimization techniques in enhancing deep learning-based ICS security mechanisms.