Statistical Algorithms for Threat Detection via Sensor Networks

George Washington University, Washington DC

Investigators

Abstract

Threat detection (TD) means an assessment of the presence of harmful agents, biological, chemical, or nuclear. The footprint (or signatures) of such agents could be qualitative and verbal, and/or quantitative. These signatures are generated by sensors which collectively form a networked system. The architecture of such systems could be series, parallel, or hierarchical. A key feature of such signatures is that they tend to be imprecise, incomplete, and unreliable. Furthermore, the sensors could be co-operative or adversarial, the latter due to sabotage and psychological ploys. The principal investigator and his colleagues propose to articulate the mathematical underpinnings of the TD scenario, in order to integrate signatures from a multitude of sensors in a principled way. The goal is to express the presence of threats in terms of numerical probabilities. This tantamounts to integrating signatures which are filtered via distributed network structures, and are contaminated by imprecision, camouflage, and parleying. As a research topic in probability and statistics, the matter of integrating contaminated and camouflaged signatures is new. Both Bayesian and classical methods, as well as a cunning combination of the two, will be invoked. The crux of the work will entail developing meaningful likelihood functions that capture the essence of the physical and psychological issues. Current practice in intelligence and national security is to express threat in verbal and qualitative terms like possible, probable, likely, etc. Such expressions are not actionable. This research will place the threat detection scenario in a probabilistic framework so that decisive actions to mitigate threats can be taken. The work will have broader impacts in civilian applications such as oil exploration, weather prediction, medical diagnosis, and socio-cultural modeling.