Page 44 - GCN, June/July 2018

BIOMETRICS
Starting in 2017, NIST told developers they could submit algorithms to be test- ed at any time. The testing was limited, however, to one-to-one facial recogni- tion technologies, which match one in- put face to one known face. In February, NIST switched to testing one-to-many applications, where a single input image is run against a database to find a match.
NIST tests the algorithms against six categories of images: visa photos, mug shots, photos in the wild, webcam imag- es, selfies and child exploitation photos. Each category provides different chal- lenges for facial recognition algorithms. NIST runs the algorithms millions of times to reveal how often the facial rec- ognition programs are right and how of-
ten they’re wrong.
It uses sequestered image databases
that haven’t been seen by research- ers, which means an algorithm can- not be tweaked repeatedly until it runs smoothly against one database. “That is really the gold standard for doing testing because it reduces the opportunities to cheat,” Grother said.
NIST’s tests are at least partially re- sponsible for the improvement in facial recognition technology in the past cou- ple of decades, said Brian Martin, a se- nior director of research and technology at IDEMIA, a provider of trusted identi- ties. “It makes everyone compete to make a better face recognition technology,” Martin said. “I think that if it were not for
the NIST tests, face recognition technol- ogy probably would not be at the state where it is today.”
Going through the FRVT process is im- portant for another simple reason: There aren’t many alternatives. “It’s the only independent testing of these commercial technologies,” Martin said.
NIST is now studying the accuracy of facial recognition technology across dif- ferent demographics — such as gender and race — to identify bias. NIST usu- ally looks at observational data, but for the demographic testing, it has set up an experiment that will reveal differences from one demographic to the next. The results of the experiment will be released later this year, Grother said. •
Man plus machine makes for most accurate facial recognition
BY SUSAN MILLER
Facial recognition technology has been improving, leading to speculation about when algorithms will be accurate enough for widespread use. That could be soon if artificial intelligence is paired with human experts, according to researchers from the National Institute of Standards and Technology and three universities.
Those researchers conducted a study to determine who performs best: algorithms, professional facial examiners or people untrained in facial recognition. The results indicate that using a professional identifier in tandem with artificial intelligence produces the most accurate identifications.
The study involved 184 people, 87 of whom were trained professional facial examiners. Thirteen were “super recognizers,” a term implying exceptional natural ability, and the control group consisted of 53 fingerprint examiners and 31 undergraduate students, none of whom had training in facial comparisons.
The participants (humans and four of the latest algorithms) were shown 20 pairs of faces that featured images designed to be challenging because of limited control of illumination, expression and appearance.
Participants (including the algorithms) rated the likelihood of each pair being the same person on a seven-point scale.
Some of the results were predictable: The trained professionals did significantly better than the untrained control groups, and the algorithms performed on par with the experts. It was also not surprising that the fused responses of the experts were more accurate than those of any individual expert because the fusing of scores decreased variability.
The most accurate identifications, however, came by synthesizing single forensic facial examiners’ results with those of the best-performing algorithm. Those results suggest “that humans and machines have different strengths and weaknesses that can be exploited/mitigated by cross-fusion,” the study states.
The researchers concluded that more testing of humans and algorithms — including a broader set of facial recognition tasks that features low-quality images, video and faces from diverse demographics — would provide a roadmap for highly accurate face identification. •
44 GCN JUNE/JULY 2018 • GCN.COM