NIST finds biometric age estimation effective in first benchmark, coming soon

The U.S. National Institute of Standards and Technology presented a preview of its assessment of facial age estimation with selfie biometrics at the Global Age Assurance Standards Summit, which revealed promising initial results for assurance that an individual is over 18. The evaluation also suggests that with effective thresholds in place, biometric age estimation is potentially as effective as checking a date of birth field on an identity document.

Six algorithms were submitted to NIST’s Face Analysis Technology Evaluation (FATE) Age Estimation & Verification track. Yoti is among them, and presented an overview of the presentation in a blog post. Other biometrics providers submitting algorithms were ROC, Dermalog, Incode, Neurotechnology and Unissey.

NIST’s evaluation is based on the “Challenge 25” principle, Biometric Standards and Testing Lead Patrick Grother explained to online and in-person attendees. The database used did not include images of people below 14 years of age, and NIST does not have any plans at this time to test age estimation for younger ages.

The report is expected to be released in the next few weeks, and Grother notes that other reports from the organization are revised roughly a monthly basis. NIST has had indications that more developers will submit algorithms in the future.

The discussion touched on the distribution of the images in the dataset, in terms of factors like demographics and image quality, and the differences between the kids of images NIST uses and those captured in a supermarket setting. Grother also explained in response to a question from Yoti CEO Robin Tombs that NIST’s evaluation of facial age estimation but not other modalities is a practical matter, as that is the modality NIST has a large volume of samples for. He left unsaid that transmitting images is much more common online than speech samples.

Collecting facial images of children is a problem in general, Grother explains, as it takes on ethical and data privacy challenges, but also costs, which are generally not balanced by the same level of interest as there is in data collected from adults. AI-generated data may or may not be synthesizing the right thing, and could deliver very different algorithmic results, even if it looks to the naked eye like it is creating real images.

Grother says that NIST’s evaluation could lead to comments on part 3 of the ISO/IEC WD 27566 standard in development, which deals with benchmarking.

NIST’s mean error calculations, included in the report, indicate whether each given algorithm tends to over or under-estimate the ages of certain demographic groups of people, and some variability was observed, Grother says.

Improvement should be expected though, and while Grother declines to make predictions, he notes that facial recognition improved by two orders of magnitude, from 30 percent error rates to 0.3 percent. Age estimation, however, is a different kind of challenge, he cautions.

Article Topics

algorithms  |  biometric age estimation  |  biometric testing  |  DERMALOG  |  Face Analysis Technology Evaluation (FATE)  |  face biometrics  |  facial age estimation (FAE)  |  Global Age Assurance Standards Summit  |  Incode  |  ISO 27566-3  |  ISO standards  |  Neurotechnology  |  NIST  |  Rank One Computing  |  selfie biometrics  |  Unissey  |  Yoti