Simulation-based Transfer-learning for Explosive Alarm Resolution Sensors (STARS)

Security at transportation checkpoints is a key strategic aspect of the DHS mission of preventing future attacks against the U.S. and its allies. The ability to freely move assets within, into, and out of the nation is paramount to national security. While DHS currently employs highly-effective explosive alarm resolution sensors (AR), these sensors require special training to operate and produce a tremendous cognitive load on the user. Explosive devices come in various shapes, sizes, and concealment. Therefore, the proposed work will use machine learning to create a sensor suite and associated algorithms for threat detection and classification as a function of these variations. Specifically, this proposal will create a framework for using simulation-based transfer learning to create "context-aware" AR sensor suites. The framework will use a low-fidelity simulation environment to simulate output data for notional AR sensors. These data can be used to fully explore the performance of potential AR sensor integrations across explosive variations. These sensor suites could include integrations for combining visual and IR camera outputs to detect glass or combining thermal and LIDAR geometry data to detect fingerprints. This simulation framework will also provide a training ground for context-aware algorithms. Transfer learning techniques will be developed for training algorithms in simulation and then transferring these algorithms to real-world sensor outputs. This proposal will demonstrate a framework for simulation-based transfer learning for AR sensor suites. This framework is inherently applicable to commercial applications, as AR is used to ensure the resilience of the U.S. economy.