How to Detect the Location and Time of a Covert Chemical Attack: A Bayesian Approach

In this thesis, we develop a Bayesian updating model that estimates the location and time of a chemical attack using inputs from chemical sensors and Atmospheric Threat and Dispersion (ATD) models. In bridging the
critical gap between raw sensor data and threat evaluation and prediction, the model will help authorities perform better hazard prediction and damage control.
The model is evaluated with respect to settings representing real-world operations. Factors that affect the model’s capability to accurately estimate the location and time of an attack are (i) the specific layout of the deployed
sensors relative to the attack location; (ii) the number of false positive signals; and (iii) the number of false negative errors. An experimental design is used to evaluate the model against the factors identified. The dominant factor is the
Expected Number of Correct Signals (ENCS), which depends on the specific layout of the deployed sensors relative to the attack location. From analyzing the effect of sensitivity (absence of false negative errors) and specificity
(absence of false positive errors) of the sensors deployed, we conclude that it is more worthwhile to invest in sensitivity than specificity. We also obtain insights on sensor coverage.