Capitalizing on technology advancements, Science and Technology Directorate (S&T) research and development is driving real results in utilizing artificial intelligence (AI) to augment visibility and help better identify emerging threats on land, air, sea and the cyber realm.
Paul Revere determined which route to ride using lantern signals – one lamp signal if the British took the land route, two if they went by water. During the Civil War, aeronautics experts started experimenting with hot air balloons to provide situational awareness. Today, our modern systems which help monitor and secure our borders must rely on more sophisticated methods to alert to danger at the kind of speed Artificial Intelligence (AI) can deliver.
S&T has several research and development efforts underway to lead this charge. These advancements are already achieving success in identifying and flagging suspicious activity on land, air and sea. Others in development harbor the promise of augmenting human-based sensory abilities to enable the Department of Homeland Security (DHS), its component agencies and first responders on the front lines to avert danger and respond to threats.
One of S&T’s greatest successes in this arena so far is Kestrel, a cloud-based analytics system that augments existing DHS systems by leveraging AI. Kestrel identifies and integrates sensor data and applies AI and machine learning analytics and predictive threat modeling to allow operators to evaluate all air and maritime tracks and make more timely decisions. In fiscal year 2023, these analytics drove a 500% greater rate of suspect activity detection by U.S. Customs and Border Protection (CBP) Air and Marine Operations Center (AMOC).
“There are illicit things and people coming in, and we need these kinds of capabilities to augment our operators to be able to identify them. The best way to do that is through automation,” said S&T’s Maritime Safety and Security Program Manager. “What we are doing is called spatiotemporal analytics. We are looking for activity patterns in certain places and at certain times which is enriched with other contextual data. It’s, basically, how is something moving in space and time.”
Kestrel research and development began in 2018 to help AMOC officers monitor massive volumes of national air and maritime sensor data. At any given time, there are between 180,000 and 200,000 individual air and maritime tracks, and the watch team monitors these tracks over different points in time. CBP expressed a need for technology to securely evaluate this volume of sensor data and generate real-time analytic reporting of suspect activities.
So, S&T designed and built an entire custom platform and analytic applications to answer this need. The challenge: continuously process five to ten terabytes of streaming data daily and create an analytic capacity to ensure only ten seconds or less elapse between the time data is received and the time KestreI delivers a suspicious activity report to AMOC’s watch team. The platform operates by performing spatiotemporal analytics, looking for certain patterns in how something is moving in space over time. It is designed to be multi-domain, monitoring aviation and maritime activities, but could extend to land and even cyber domains with the right data sources, to track movement across IP addresses as network traffic traverses geolocations.
The approach taken by the team was to address as many analytics as possible by applying statistical and heuristic-based analytics, which consume far less compute power than large language models and other sophisticated AI tools, and don’t introduce the same sort of risks of “hallucinations” – inaccurate or nonsensical outputs. In conducting this type of analysis, lessons are learned about the quality of the data, so when it comes time to introduce that data into the larger AI models, knowing more about the input data will provide more confidence in the resulting outputs. And when it comes to identifying a moving threat within 10 seconds, accuracy and reliability are critical.
Upon detection of suspect activity, Kestrel creates a structured message, which is published and displayed directly onto a watch officer’s existing screen. It was important that the output didn’t require an additional screen, with the understanding that watch officers are already inundated with information. The objective was to provide more precision to aid the monitoring process.
“We’re actually augmenting AMOC’s system and personnel with Kestrel,” the S&T program manager said. “[Law enforcement has] had a couple of significant busts where $81 million of narcotics were seized.”
Kestrel’s transition to CBP’s AMOC is scheduled to be complete by the end of fiscal year 2025. Initially developed and tested in AWS GovCloud, the system recently received approval from the CBP Chief Information Officer to “go live” within CBP’s Commercial Amazon Cloud East (CACE) enterprise platform.
This article is the third in a feature article series dedicated to S&T’s AI/ML R&D efforts. Additional information can also be found the Artificial Intelligence and Autonomous Systems webpage. For related media inquiries, contact STMedia@hq.dhs.gov.