Physics-constrained digital twins to detect stealthy false data injection in urban pedestrian counts (arXiv:2609.17635v1)
This paper formalises stealthy false data injection (FDI) attacks on city-scale pedestrian sensing and proposes a physics-constrained digital twin that estimates directed flows on a street graph, assimilates counts via a learned graph-localised gain, and is trained with a flow-conservation residual coupling metered and unmetered segments. Detection combines the innovation and the residual with alarm thresholds set by adaptive conformal calibration; on six years of Melbourne data the authors report an attack margin of 0.54 against a single compromised device and 0.19 when one-third of devices are compromised, while replacing the street graph by a distance graph reduces the margin to 0.09, indicating the defense derives from the conservation law rather than locality.