Photo “disguise” fools facial recognition.

The University of Chicago’s SAND Labs has released Fawkes, a system for personal privacy protection, with a BSD 3-Clause license hosted on GitHub, and a paper (pdf) to be published next month at the USENIX Security 2020 online conference. Fawkes is an algorithm and software tool running on a local computer that allows users to disguise photos and limit their use for facial recognition.

Fawkes processes photos and makes subtle changes to their pixel levels that the human eye cannot recognize. The researchers call this method “image cloaking”. Users can use photos in disguise, share them on social networks and send them to friends. But when someone collects these public face photos to build a facial recognition model, the camouflaged photos will create a highly distorted face. This camouflage effect is difficult to detect and does not show errors in model training. If the face recognition system tries to use this model to identify you, it will fail.