Better Face-Recognition Software

Continued from page 2

Still, Gross says, "you can already see the path building." Until recently, the video-surveillance industry still mostly relied on analog cameras, requiring cable to be set up for long distances to connect those cameras to monitoring equipment. Now, "the industry is switching to IP-based cameras, with which you can pretty easily tap into already existing Ethernet networks," Gross says. "So you have wireless cameras and cameras using POE [Power over Ethernet technology allows IP telephones, wireless LAN Access Points, and other appliances to receive power as well as data over existing LAN cabling] where you don't need a separate power plug. You can buy commercial solutions that are essentially a TiVo for these cameras, with motion sensors built in so they only record when there's motion happening. With digital storage, you can keep the data indefinitely and enhance it in ways that you can't with analog images. So all these things are coming together."

In principle, therefore, as face-recognition software continues its rapid advance, it will likely be possible to search for specific faces across a network of webcams. Accordingly, Gross's recent work at Carnegie Mellon, in conjunction with colleagues at the Data Privacy Lab there, has been the development of algorithms to protect individuals' privacy while under video surveillance. The usual methods that thwart human recognition of an individual's features on video--for example, those pixelated fields sometimes covering faces and body parts on reality-TV shows--already won't fool much face-recognition software. Completely blacking out each face in a video clip would do the job, but this would be of limited use if law-enforcement agencies wanted to follow up evidence of suspicious behavior once they had a court warrant. The function of the privacy-preserving algorithms that Gross is helping to create, he explains, is to automatically take the average values of individuals' faces and, from those, synthesize new facial images, then superimpose those new images over the originals. "It may seem like the opposite technology," Gross says, "but actually, it's just the other side of face recognition."