Researchers have presented a method for protecting urban systems from data attacks

A group of researchers developed a system of digital twins capable of detecting hidden attacks of the "false data injection" type in urban pedestrian flow monitoring networks. The method relies on physical laws of flow conservation, which allows detecting anomalies even when attackers try to mimic legitimate sensor readings.

The solution is based on an algorithm of adaptive conformal calibration, which automatically sets alarm thresholds. Unlike traditional systems that rely on manual tuning, this approach uses graph localization to evaluate directed flows. The model is trained on residual values of the conservation law, linking measured and unmeasured segments of urban infrastructure.

The effectiveness of the method was tested on Melbourne data for 6 years. When one device was compromised, the "attack margin" (resistance to distortions) was 0.54. Even when 33% of all sensors in the network were captured, where coverage is only 1.18% of all pedestrian segments, the system maintained a margin of 0.19. For comparison, using standard distance graphs reduces this indicator to 0.09.

The research results confirm that integrating physical constraints into the architecture of AI models significantly increases their resistance to targeted manipulations. The work is published under the CC BY-SA 4.0 license and offers a new security standard for critical urban infrastructure, where sensor data directly affects management decision-making.