Fred Benenson got an email from Facebook recently, informing him that it had detected his face in a number of photos and asking if he would like to be tagged in them. When he clicked through the link, he was amused to find eleven black and white headshots of his mother from her college days. The social network’s DeepFace algorithms detect and identify faces using a neural network with more than 120 million parameters, an ever-improving machine learning process the company says is closely approaching human-level performance. Facebook says DeepFace can determine if two photographs are of the same person 97.25 percent of the time, but there are still plenty of images that can trip up their most powerful machines, especially when it’s a family affair.
"We all have genetic traits that are translated into how we look. If a person closely resembles one of their parents, a photo of the young mother or father could correlate and confuse the algorithm," says David Tunnell, the chief technology officer at NXT-ID, which specializes in three-dimensional facial recognition. In Facebook's case, confusing two family members was a bug, but one that points toward other potential uses for facial recognition, as the process also works in reverse. "We have done research that shows you can use facial recognition to identify a person’s ethnicity, region of origin, and family affiliation," Tunnell says. With good data from my parents, siblings, or cousins, he says it might be possible to identify me even if the system had no actual images of me to work from.
Fred and his mother Donna
Is that creepy, or cool? Mostly it depends on the context. "I did find it oddly compelling," Benenson says in an email to The Verge. "Here was a real life example of a false positive (when an algorithm incorrectly predicts an observation as being something its not — in this case it thought my mom was me) with personal implications." Benenson, who is a data scientist at Kickstarter, says he generally isn’t too troubled by the idea that photographs of our faces are constantly being scanned and analyzed for personal identification. "In general I think people tend to overstate the nefarious things that Facebook is doing with this kind of technology, but it certainly opens the door to larger and more difficult questions."
When it comes to more serious applications, for example the way the Chicago Police are now using facial recognition, Benenson sees the potential for trouble. "What about the cases where this algorithm isn't used for fun photo tagging?" he asked rhetorically. "What if another false positive leads to someone being implicated for something they didn't do? Facebook is a publicly traded company that uses petabytes of our personal data as their business model — data that we offer to them, but at what cost?"
Facebook's DeepFace system analyzing images of Sylvester Stallone
Tunnell says that the ability to identify people based on family members, even as they age, is frightening, but that the possibility to use this capability for good is equally exciting. "I did a study with my own daughter, where I scanned her entire head every day over the course of several years. What we learned is that, given the right data set, these technologies can identify someone two, four, six years down the road. So that could be critical in helping to locate someone who hasn’t been seen by their own relatives, and who has through substantial physical changes."
In a 2012 US Senate hearing on privacy, Al Franken asked Facebook point blank if they would ever sell their database of facial identities to other companies. Facebook’s Rob Sherman said there was no telling how the company might use the data in the future. Users can always remove the ability to be tagged in a photo, but not the ability to be scanned. "It’s valuable to think hard about the fairness of the tradeoffs we're making with our personal data," says Benenson. "This isn't a calculus we've had enough time to fully understand from a societal perspective. It may be years before we actually know the consequences of feeding so much personal content to such a powerful company."
For the time being, however, Fred and his mother Donna simply found the misidentification to be touching, especially since Fred works in data science. "I can't get over the fact that, for the first time in 30 years that I'm told that I look like Fred, it happens by way of an algorithm," she wrote in an email to The Verge. "What could be more wild or fitting!"