California cops hope to expand facial recognition, “eCrime” head says
“It doesn’t require a front face shot,” Robert Morgester, the senior assistant attorney general and the head of the state’s eCrime Unit, said on Wednesday. “The software has become so robust that you can do a side shot. If mugshots are in use—each region has their own database of mugshots—and agents can query pictures of a known suspect.”
That was just one of a handful of surveillance technologies that Morgester ran through—he explored the law enforcement benefits of not only facial recognition, but also mobile DNA testing, license plate readers, and drones. He faced off with an attorney from the Electronic Frontier Foundation, Jennifer Lynch, who eloquently articulated counter-arguments and raised privacy concerns.
The two of them spoke during a forum on Wednesday called “Protecting Our Communities, Respecting Our Liberties,” which was organized by the California Attorney General’s Office.
While facial recognition is hardly perfect (after all, it famously missed catching the Boston Marathon bombing suspects), state law enforcement will likely begin using the technology against social media networks like Facebook, Morgester said, citing 2011 research from Carnegie Mellon University. Anil Jain, a professor of computer science and facial recognition expert at Michigan State University, told Ars last year that "great advances" in the field in recent years have made mugshot database searches possible.
"However, the performance of face recognition systems depend on the ‘quality’ (pose, illumination, resolution, expression) of the acquired image," he said by e-mail in November 2014.
Jain pointed to a 2014 study by the National Institute for Standards and Technology that found that facial recognition was 95.9 percent accurate while searching a mugshot database of 1.6 million people.
But he warned that when taking an unknown image of questionable quality or one that is not taken under ideal conditions, recognition capability can drop to as low as 60 percent.