Real-Time Detection of Cyber-Physical Attacks on Water Distribution Systems Using Deep Learning

In this paper we explore the potential of deep learning for real-time detection of cyber-physical attacks on water distribution systems. Deep learning methods represent the state-of-the-art of artificial intelligence, and they have been successfully employed for solving classification and regression problems in many fields of engineering. We focus on AutoEncoders (AE), which are multi-layered artificial neural networks trained to reconstruct their own inputs. During the training process, the weights of the AE are tuned to minimize the reconstruction error computed for all training patterns. This entails that when used on new input data, anomalous patterns containing outliers are coarsely reproduced by the AE—with a corresponding higher reconstruction error. When employed for attack detection on water distribution systems, the AE are trained to map the status of the network by using the readings of the deployed sensors as both inputs and outputs of the model. If the training data are representative of normal operations, the AE return high reconstruction errors when fed with sensor readings recorded during the course of an attack. The detection performance is evaluated on a test dataset featuring several classes of plausible attacks—simulated with EPANET—and quantified with two evaluation criteria accounting for detection accuracy, false alarms and the time needed to identify a threat. We also carry out a sensitivity analysis to determine the effect of data availability and model complexity on AE detection performance.