Securing Industrial Water Networks: Event-Level Anomaly Detection in SCADA Telemetry

Abstract

The increased interdependence between industrial control systems and digital infrastructures has made water treatment facilities more suscepti-ble to cyber-physical attack, which has resulted in a need to have robust and smart anomaly detection systems. In this study, MRTAF-Net (Multi-Resolution Temporal-Attention Fusion Network) is a modern hybrid deep-learning framework aimed at facilitating the safe and de-pendable functionality of industrial water networks due to the detection of anomalies at the event levels of SCADA telemetry. The proposed model uses a Multi-Resolution Temporal Convolutional Network (MR-TCN) to extract hierarchical temporal features, a Channel Squeeze-Excitation (SE) to induce changes in channel weights in adaptive mode, and a Multi-Head Self-Attention (MHSA) to identify long-range tem-poral features and contextual sensor interactions. The proposed model is based on the SWaT data of the iTrust laboratory. In order to further em-power the model to jointly leverage the dynamic temporal information and the static statistical information, some statistical descriptors are in-tegrated with these sequential features by means of a learnable gated fu-sion module. Numerous experimental assessments show that MRTAF-Net has a very high level of performance with 99.83 accuracy, 99.84 F1-score and an AUC of 99.99, which is significantly better than popu-lar baseline models (Bi-LSTM, XGBoost, CatBoost and TabNet). MRTAF-Net has created an effective and scalable base to protect indus-trial water infrastructure against changing cyber-physical attacks through a combination of interpretability, robustness and multi-scale time rea-soning.