Trustworthiness evaluation of multi-sensor situation recognition in transit surveillance scenarios

Physical Security Information Management (PSIM) systems are a recent introduction in the surveillance of critical infrastructures, like those used for mass-transit. In those systems, different sensors are integrated as separate event detection devices, each of them generating independent alarms. In order to lower the rate of false alarms and provide greater situation awareness for surveillance operators, we have developed a framework-namely DETECT-for correlating information coming from multiple heterogeneous sensors. DETECT uses detection models based on (extended) Event Trees in order to generate higher level warnings when a known threat scenario is being detected. In this paper we extend DETECT by adopting probabilistic models for the evaluation of threat detection trustworthiness on reference scenarios. The approach also allows for a quantitative evaluation of model sensitivity to sensor faults. The results of a case-study in the transit system domain demonstrate the increase of trust one could expect when using scenarios characterized in a probabilistic way for the threat detection instead of single-sensor alarms. Furthermore, we show how a model analysis can serve at design time to support decisions about the type and redundancy of detectors.