Critical infrastructure, such as water distribution networks (WDNs), relies on large volumes of sensor data to automate control processes and ensure reliable operation. However, data-driven system modeling and control is vulnerable to cyber-attacks that can compromise system safety.

To mitigate this risk, attack detection methods increasingly rely on deep learning models that analyze streams of sensor and control data to identify anomalous behavior. In this work, we propose a physics-augmented autoencoder framework for detecting cyber-attacks in WDNs.

Our approach incorporates domain knowledge of system dynamics, physical constraints, and control logic to enhance anomaly detection performance. We (1) introduce an improved autoencoder architecture benchmarked against state-of-the-art methods, and (2) use attack-aware training with simulated attack scenarios to increase model robustness.

Experimental results demonstrate that physics-augmented modeling substantially improves the resilience of critical water infrastructure against cyber threats.

K. Petrovic, B. Stojanovic, O. Saukh, Physics-Augmented Autoencoder-Based Cyber-Attack Detection for Critical Water Infrastructure, Proceedings of the 15th International Conference on the Internet of Things (2025) 43-51.