We can age as little as five years before face-recognition algorithms begin to struggle to identify us as the same person. This means systems that rely on facial recognition may need to get new images of users periodically or risk being unable to verify who they are.
Face recognition is now used at border crossings, by police to watch for known offenders in public and even to unlock our phones. But there has been little research on how these systems cope with faces changing over time. at the Norwegian University of Science and Technology and his colleagues designed a test to assess it.
The team used open-source alternatives to face-recognition tools used by police forces and smartphone makers, which don鈥檛 reveal how their own algorithms work, and also used AI-generated images of 50,000 humans aged synthetically to various degrees, due to the scarcity of real data sets collected over long time frames.
Advertisement
Grimmer says the accuracy of face-recognition systems dropped continuously over time from the point the reference image was taken, rather than suddenly. Within five years, AI algorithms usually identified people, but after that, they began to struggle. Gaps beyond 20 years proved very difficult for software to handle.
He says that people will age differently depending on a range of lifestyle factors, and that we also age more or less quickly at different points of our lives, so determining how often face recognition needs to update photos of users across the board to maintain security and accuracy is hard.
鈥淏abies will change within two months, so you could probably capture a new photo every month and it would still fail. Generally, even until the age of 20 years, the face still kind of changes,鈥� he says. 鈥淭hen you have a period from say 20 to 60 years where the identity is kind of stable and there will only be some texture changes, like wrinkles. And then the other extreme would be the seniors, where it again gets a bit problematic because the head shape changes again, and you have more pronounced wrinkles.鈥�
The experiment also revealed that the ageing algorithms used to create older faces from reference images were more effective when the target was aged between 20 and 40 years, and didn鈥檛 do as well with children and older people. This is a 鈥渢ypical sign of biased image generation models鈥�, says the team, because they tend to be created in environments filled with software developers in their 20s, 30s and 40s.
Reference:
