Deep Learning Architectures for Anomaly Detection in SCADA Networks: A Comparative Analysis of LSTM, CNN, and Hybrid LSTM-Autoencoder Models
Abstract
Supervisory Control and Data Acquisition (SCADA) systems govern the physical processes underpinning water treatment, electric power, and industrial manufacturing, yet their deterministic legacy protocols were designed without security in mind, leaving them exposed to stealthy cyber-physical anomalies that manipulate sensor readings and actuator commands while mimicking legitimate operation. This paper presents a comparative empirical study of three deep learning architectures for anomaly detection in SCADA network traffic: a Long Short-Term Memory (LSTM) sequence model, a Convolutional Neural Network (CNN), and a hybrid LSTMAutoencoder. We evaluate the models on SWaT- and BATADAL-style industrial control system data augmented with synthetic Modbus/TCP and DNP3 attacks spanning command injection, replay, and false data injection. Features are engineered from parsed timestamps, function codes, register-access patterns, and payload sizes, then windowed and normalized to expose temporal and structural regularities. The hybrid LSTM-Autoencoder achieves the strongest performance (F1 approximately 0.95), followed by the LSTM (0.93) and the CNN (0.87), with all models sustaining sub-second detection latency suitable for near-real-time deployment. Analysis shows that the LSTM captures long-term temporal dependencies and gradual false-data drift, that the CNN excels at packet-structure and function-code irregularities, and the hybrid model unifies both while yielding interpretable reconstruction-error anomaly scores. External validation on unseendomain traffic preserves F1 above 0.92. The results confirm that behavioural modelling of highly deterministic legacy SCADA processes is an especially effective foundation for deep learningbased intrusion detection.