Officials from the Intelligence Advanced Research Projects Activity, or IARPA, say they hope the organization’s new Face Recognition Prize Challenge will help them move the needle on biometric security.
They expect to start judging soon a range of new technologies submitted for consideration, with a winner to be announced in October.
“Face recognition is a critical technology for national security and the intelligence community,” said Chris Boehnen, a senior program manager at IARPA, part of the Office of the Director of National Intelligence. “Many of the adversaries we face don’t wear uniforms, and that means we have to have some other way to recognize who people are.”
Sixteen teams submitted software solutions intended to make face recognition faster and more accurate. IARPA will test the various algorithms against a standard data set to see how they perform.
While industry and military leaders have lately put a heavy emphasis on facial recognition as an identification tool, challenges remain. A chief problem lies in the arena of variability: If the pose, the lighting or the facial expression changes, it can be hard for even the best software to match a live subject to a file photo.
“When you get your driver’s license you look straight at the camera, in good lighting,” Boehnen said. “But if you have one face in heavy shadow and another in direct light, if you have one face frowning and another smiling, all those changes make it harder for a computer to compare two faces.”
The military has pursued several avenues recently in the search for better face-recognition solutions.
In 2016 the United States Special Operations Command put out a call for small businesses with strong R&D capabilities to tackle advanced tactical facial recognition at a distance, as called for in SOCOM’s Science & Technology mission statement.
The Defense Forensic Science Center in late 2016 went to industry for face-recognition solutions that could operate independent of gender and ethnicity. At about the same time the U.S. Army Research Laboratory Sensors and Electron Devices Directorate published early findings on the possibility of using thermal infrared band scanning to provide face-matching capabilities.
While issues around accuracy have long plagued facial recognition efforts, military and intelligence planners are eager to see the capability enhanced. Unlike other biometric scans, a facial match can be done discreetly and from a distance, a potential tactical advantage.
“With fingerprint and iris, you tend to need a close-range, fully cooperative subject,” Boehnen said. Facial recognition “is not the most accurate biometric, but it is the easiest one to capture.”
Even with so many diverse efforts in play around the subject, prize challenges can fill a special niche. Because biometrics developers may be reluctant to take part in conventional RFPs that require them to transfer their intellectual property rights to the government, challenges can bring ideas to the table that planners otherwise might not see. Challenges also tend to attract smaller players, those outside the usual round of major contractors. “I can let everyone talk, from the garage tinkerer to the large company,” Boehnen said.
While it’s too early to say what will come of this challenge, organizers say they expect the strongest proposals will likely focus on “deep learning,” a computer science discipline that enables machines to process vast volumes of data at high speed.
Thanks to such techniques, “even something a grad student could whip up in two or three months today performs as well as something that a team of the best engineers in the world could have produced five years ago,” Boehnen said.
In facial recognition a computer can be tasked to consider anywhere from six to 150 different layers of data. More layers mean greater accuracy, but at some point the process can get sluggish, without really adding any new information of value. The winning algorithm likely will strike that balance between speed and volume.
Deep learning could be a key to help identify that sweet spot. “That has been the turbocharger in face recognition,” Boehnen said. “All available evidence says deep learning has increased performance of face recognition potentially by several orders of magnitude.”