Main Subcommittee Category: Emerging Concepts and Innovative Technologies

How would you label your submission?: Industry

Abstract:

In the current global threat environment, homeland security depends on domain and situational awareness. The evolving threat of illegal smuggling and entry along the U.S. southern border requires efficient threat classification and resource allocation. The Department of Homeland Security (DHS) Science and Technology (S&T) and DHS Customs and Border Protection (CBP) Air and Marine Operations Center (AMOC) for National Air Domain Security are tasked with monitoring and interdicting smuggling in the air and maritime domains. Such a complex task requires sorting through large volumes of data to make timely and accurate decisions. This paper describes how leveraging advanced machine learning techniques can support this task.

The proposed approach involves developing predictive threat models (PTM), where multiple machine learning algorithms such as multilayer perceptron (MLP) classification, adaptive boosting, and artificial neural networks (ANN) are tested and evaluated. The top performing model is selected and compared to a hybrid stratified sampling approach, where the data is split into distinct groups before being used to train each of the models on the same classification and deep learning methods used before. By tailoring models using the hybrid approach and selecting the most applicable for each unique record, the new predictions outperform those of the single-model methodologies. By deploying the hybrid stratified sampling models using modern machine learning operations (MLOps) best practices such as automated pipelines, continuous integration/continuous deployment (CI/CD), and model performance evaluation, we can streamline the delivery of these models. This ensures our DHS operators can leverage the highest performing real-time predictive analytics to make informed decisions quickly and effectively in the face of evolving threats. Additionally, standalone time-series autoregressive models are continuously trained on live data and instantaneously produce accurate forecasted predictions, equipping our DHS operators with the ability to accurately monitor the future positionings of specific targets of interest on demand.

Keywords: ADAPTIVE;DATA;DECISION;EVALUATION;INTEGRATION;MACHINE LEARNING;NETWORKS;OPERATIONAL ENVIRONMENT;REAL-TIME;THREAT MODELING

Additional Keywords not in the list above: Illegal Smuggling, Interdiction, Prediction, Multilayer Perception, Artificial Neural Networks, Test & Evaluation, Hybrid, Predictive Analytics

Menu

Karen Messer