Facial recognition technology has to account for many real-world variables that can throw off accurate matching, underscoring the need for rigorous testing and continued development, government researchers told the Connect:ID conference in Washington on Monday.
Biometric technology is changing very quickly “but technologies are getting a lot better,” expanding where the tech can be used, Arun Vemury, program manager at the Department of Homeland Security’s Science and Technology Directorate, said on a panel discussion, emphasizing a “focus on collaboration between government and industry.”
Vemury noted that there is often a rift “between expectations and where technology really is.”
“Industry can lack or have unrealistic performance estimates,” he said.
DHS S&T’s 2018 Biometric Technology Rally, which brought together subject matter experts and private-industry technology to test emerging biometric systems with the help of volunteers, included 363 diverse subjects tested in groups of 15 over a five-day period. The technology was rated on efficiency for the traveler (not more than an average 20 seconds per person), effectiveness for security screeners, and overall satisfaction.
Vemury said “reliable acquisition of biometric images is a primary driver of error,” and stressed to the biometric industry audience that “cameras matter.” Performance of a single face algorithm varied widely across different face cameras. “Some were very good but others struggled,” he added.
Biometric systems also must be tested with a variety of population demographics.
National Institute of Standards and Technology biometrics evaluator Patrick Grother noted some of the key hurdles in facial recognition technology, including aging faces and poor quality photos, such as those taken with surveillance cameras or in low-light conditions. Identical twins can also fool the facial recognition system.
“Algorithms tend to fail on difficult images,” Grother said.
Notable successes of facial recognition technology include the identification of accused mass shooter Jarrod Warren Ramos, who refused to ID himself after being arrested for the June 2018 massacre at the Capital Gazette in Annapolis, Md. Police officials were unable to get fingerprints from Ramos, so they took his photo and compared it to driver’s license and passport databases.
Lars Ericson, Intelligence Advanced Research Projects Activity (IARPA) program manager at the Office of the Director of National Intelligence, also cited constrained photos — like a passport mug — versus unconstrained pics, such as surveillance images, as a key current limitation in biometric technology.
IARPA’s Janus program focuses on improving the ability of facial recognition tools to work with “in the wild” photos by evolving from two-dimensional image matching to multi-view model-based matching.
Ericson stressed that evaluators must “incorporate non-face distractors,” like inanimate objects, into a data set and “use the right test protocol for your use case.”