Physics-constrained generative adversarial networks for false data injection attack detection in smart water dispatching systems

Abstract

To address the False Data Injection Attack, which circumvents traditional safeguards by simulating natural hydrological fluctuations, this paper proposes a Physically Aware Dual Discriminator Generative Adversarial Network (PA-DGAN) for high-precision anomaly detection in water resource management systems. Unlike purely data-driven models, PA-DGAN incorporates the Saint-Venant equations as an embedded physical constraint layer. The generator employs a spatiotemporal graph convolutional network to capture the topological correlations between hydrological stations. Furthermore, a dual discriminator architecture is designed to simultaneously evaluate the consistency of data distribution and adherence to hydraulic physics laws, effectively enabling the model to learn the underlying physical manifold of river dynamics. Extensive experiments were conducted using real hydrological data from the SWaT international benchmark. The results demonstrate the superiority of the PA-DGAN algorithm. This research provides a robust and physically interpretable security framework for intelligent collaborative scheduling in the Yellow River Basin.