Abstract
Supervisory Control and Data Acquisition (SCADA) systems form the backbone of modern critical national infrastructure, managing operations across smart power grids, water treatment facilities, and transportation networks. The convergence of Operational Technology (OT) and Information Technology (IT) has dramatically expanded the attack surface of these environments, exposing industrial processes to sophisticated cyber-physical threats such as false data injection and stealthy actuator tampering. While deep learning models offer substantial promise for anomaly detection in multivariate SCADA telemetry, their deployment is severely hindered by the operational necessity of centralizing sensitive, proprietary, and geographically dispersed telemetry data, which raises significant operational security and bandwidth concerns. In this study, we propose a novel, privacy-preserving anomaly detection framework for SCADA systems leveraging federated learning over multivariate time-series data. The framework introduces a local Attention-augmented Temporal Convolutional Network (Attn-TCN) Autoencoder architecture coupled with an adaptive aggregation mechanism, FedSCADA, engineered to mitigate non-Independent and Identically Distributed (non-IID) data heterogeneity across distributed edge sub-stations. Comprehensive empirical evaluations on the Secure Water Treatment (SWaT) and Water Distribution (WADI) benchmark datasets demonstrate that our decentralized approach achieves an F1-score of 0.941 and 0.887, respectively, closely matching the performance of centralized detection models while reducing uplink communication overhead by up to 68.4%. Our findings validate the viability of federated learning for robust, real-time industrial anomaly detection without compromising process confidentiality.
