Wrong Person May Be Identified 20 Percent of the Time With Facial Recognition Software
As of June of this year, enough states had merged their databases of driver's license photos with facial recognition photos to subject 120 million Americans to a permanent police line-up. Some states exercise a little discretion in whom they allow to run searches on those databases, while Ohio permits roughly 30,000 cops, court personnel, and random hangers-on to go trawling for familiar faces. Either way, though, it's a crapshoot for those of us whose mugs have been digitized and stored. As it turns out, facial recognition software may kick back a bad identification about one time in five—at least, that's what official parameters allow.
According to documents the Electronic Privacy Information Center extracted from the FBI with the legal jaws of life known as a FOIA lawsuit:
NGI shall return the correct candidate a minimum of 85% of the time, when it exists in the searched repository, as a result of facial recognition search in support of photo investigation services.
NGI shall return an incorrect candidate a maximum of 20% of the time, as a result of facial recognition search in support of photo investigation services.
By contrast iris searches are supposed to return the correct candidate a minimum of 98 percent of the time, and the incorrect candidate a maximum of 10 percent of the time. Fingerprints are supposed to return a minimum of 99 percent correct candidates and a maximum of two percent incorrect.
Those interesting comparisons come at the end of the very dry Next Generation Identification System Requirements Document (PDF), dated October 1, 2010, in the "Accuracy" section. While technology is always a work in progress, and results may have improved since then, that false-positive rate, as EPIC notes, "is much greater than expected" for a system that's already being deployed and used to match your lovely driver's license photo to a blurry security cam shot of some jerk who knocked over a liquor store.
