Integrating Physics into Digital Twins for Anomaly Detection

Modern urban infrastructure relies heavily on sensor networks to monitor pedestrian flow, yet these systems are vulnerable to False Data Injection (FDI) attacks. These attacks are often "stealthy," meaning they are designed to remain within the statistical bounds of normal operation, making them invisible to standard threshold-based detection methods. This paper proposes a solution by embedding physical constraints—specifically the conservation laws governing pedestrian movement—into a digital twin model. By forcing the AI to reconcile sensor data with the underlying physical reality of how crowds move, the system can identify discrepancies that indicate malicious tampering rather than natural variance.

Providing Statistical Rigor with Conformal Guarantees

To move beyond heuristic detection, the authors utilize conformal prediction, a framework that provides rigorous statistical guarantees on error rates. By applying this to the digital twin's output, the system generates prediction intervals for expected sensor readings. If the real-time sensor data falls outside these dynamically calculated intervals, the system flags a potential security breach. This approach allows for a quantifiable trade-off between sensitivity (detecting attacks) and specificity (avoiding false alarms), providing operators with a clear confidence level in the integrity of the urban flow data. The method is validated through extensive testing, demonstrating that it can effectively isolate stealthy attacks that would otherwise bypass traditional anomaly detection pipelines.