Some people really do have doppelgangers and that could be a biometric problem
If everyone really does have an unrelated twin somewhere on the planet, that is bad news for facial recognition-based security system.
It is not a situation keeping biometrics and AI practitioners awake at night, but neither is it an entirely academic question. A physical “copy” is no different than a divulged password. Actual twins are enough of a concern.
A large Barcelona-based team of biology and IT researchers – plus one from biometric firm Herta Security – found 16 unrelated look-alikes around the world using algorithmic analysis. Nine of that set were considered “‘ultra’ look-alikes.”
Writing in the journal Cell Reports, the researchers say their look-alike dataset came from a unique photographic collection shot by Canadian photographer François Brunelle. Brunelle has for years looked for unrelated people who share uncannily similar faces. He has found 32 pairs.
The researchers fed the faces of the 32 pairs to three facial recognition algorithms: Custom-Net, a custom deep convolutional neural net created by Herta; MatConvNet and Microsoft’s Oxford Project face API. All three found 16 “objectively similar” and, of them, 9 were counted as ultra-similar.
(Thirteen pairs were of European ancestry. One couple each were Hispanic, East Asian and Central South Asian.)
It appears that the grouping of 16 objectively similar (and including the nine ultra look-alikes) share more than face biometrics, too.
Researchers found that compared to people not judged to be similar by all three algorithms, the 16 pairs of lookalikes “share a more comprehensive physical and probably behavioral, phenotypes.” They did not delve too deeply into the implications of this discovery, but it is at least plausible that other physical traits, like personality and gait, could match.
AI | biometrics | biometrics research | facial recognition | twins