Towards Securing Water Systems Communications from Adversarial Attacks using AI Assurance

Water systems are increasingly susceptible to cyberattacks due to their reliance on networked communications for monitoring and control. This paper introduces an AI-assured approach to detect anomalies in water distribution systems communication layers using Long Short-Term Memory models, which are trained, tuned, and validated through four different scenarios targeting disruptions in communication patterns. Using the Secure Water Treatment and anomalies and malicious acts datasets, and by simulating both operational and malicious scenarios, Long Short-Term Memory model is trained on normal data and tested on both regular (idle) and attack situations. The LSTM model employs a sliding time-window technique to capture sequential dependencies in communication data; it identifies anomalies by comparing predicted and actual sequence—significant deviations that signal potential cyber threats. Our scenarios demonstrate an anomaly detection accuracy of 94%, with a false positive rate of 2.5%, indicating the model’s capacity to identify both overt and subtle attacks. Further, we evaluate the intentionality of those outliers, whether they are cyber threats or merely a human operator error. The anomalies and malicious acts dataset scenario successfully detects multiple intentional sabotage scenarios, further validating the model’s potential in real-world cyber-physical security settings. The findings suggest that integrating AI-assured, LSTM-based anomaly detection into water systems can enhance security, enabling early threat detection and reducing the overall risk to critical infrastructures communications.