… And Always Wear The Master Face

Close Your Eyes, Can’t Happen Here

Facial recognition programs offer a mixed bag of consequences, from the ease of unlocking a phone by looking at it, to arrests of innocent people because they were flagged by an algorithm which incorrectly ID’d them.   When used for convenience it can be very handy, assuming your device can recognize you even first thing in the morning, and that you don’t have an evil twin out there somewhere.  Unlike in the movies,  when used by officials to track people’s movements or to try to quickly identify a suspect on the run in a city, the algorithms behind the scenes are more likely to flag a false positives than they are to find a specific person; not least of which because there are a lot of people to scan and only one is actually the correct person.

There were training biases discovered which make false positives far more likely among those gifted with a fair amount of melanin, and several lawsuits spawned from that flaw but now there is an even more worrisome issue which researchers in Israel have discovered.   They generated a very large amount of virtual faces using StyleGAN and tested them against Dlib, FaceNet, and SphereFace to see which could generate at least some false positives.  They fed those faces into an evolutionary algorithm so they could generate new faces even better at generating false positives until finally they ended up with nine “master faces”, seven male and two female.

These master faces are capable of successfully impersonating over 40% of the population, without any additional information or data of the person they are identifying.   That is not great news for those who unlock their devices or homes via facial recognition, nor for those who believe automated facial recognition is the future of law enforcement.