Facial age verification, age estimation coming of age across sectors

A recent webinar hosted as part of the European Association for Biometrics’ (EAB) Lunch Talk series gets opinions on the state of facial age verification and estimation (AVE) from Dr. Ana Calarasanu and Dr. Thomas Petit of French biometric authentication firm Unissey.

Age verification is increasingly finding use cases in sectors from retail to online gaming. Calarasanu says the standard age verification toolkit may contain document ID checks, credit card checks and behavioral analysis. But as life moves online, algorithmic face biometrics analysis is emerging as a valuable tool that can reduce friction for users while providing additional levels of security and accessibility.

“Privacy-wise, we are exposing less information than on a document ID,” Calarasanu says. And technologies such as liveness detection ensure face biometrics tools offer a level of security needed to satisfy rigorous standards. But she notes the need for precision in differentiating between age verification and age estimation. Estimation tools, she says, “approximate the age or an age interval of a subject,” versus the more binary approach found in age verification processes that aim to confirm a user is over a certain target age.

“This distinction is important because there are application use cases where age estimation is really what we are looking to solve, but als from a technical point of view, there are differences, and the metrics and the performance will not necessarily be the same.” Age estimation can add a layer of security in use cases such as online banking or ID access management, or in “lighter use cases” such as dating apps or scooter rentals.

Age verification, meanwhile, is needed for cases such as gaming, porn or age-restricted purchases.

Technical aspects differ

Petit runs through various machine learning approaches that can be considered when building an age verification or estimation model, and how the specifics of a particular use case might weigh on which one is appropriate to use. He also runs through how evaluation for age estimation differs from age verification evaluations – in the case of estimation, a combination of Mean Absolute Error (the mean difference between actual and estimated age, so lower is better) and Cumulative Score (the proportion of predictions for which the error rate is below a threshold, so higher is better).

Petit walks listeners through the “Challenge T” concept, in which T is the legal age plus seven, meaning that in cases where the legal age is 18, those who look under 25 will be asked for ID. (So, in that case, Challenge 25.) He also gives a brief tour of National Institute for Standards and Technology (NIST) datasets for age verification and estimation, NIST evaluation metrics (the first benchmark for which came out this year), ethnicity bias and how facial age verification compares against human performances – as well as how the algorithms are improving over time.

In conclusion, Unissey offers four takeaways. First, AEV is not infallible; Petit compares it to a doorman or a bartender, who is usually reliable but still might miss the occasional identification. That said, the technology already offers better results than human oversight. However, like humans, it has its biases, particularly in gender and ethnicity, which must be accounted for. Finally, each use case will bring its own particulars, which should guide the selection of a facial age verification or age estimation product.

Petit says all signs suggest facial age verification and facial age estimation are on their way to eventual widespread adoption, and expects R&D on the technology to be “active” in the coming years. “The industry is getting more and more mature, and I would say it has already reached a stage where it can be a reliable solution for many applications.”

Article Topics

age estimation  |  age verification  |  biometrics  |  biometrics research  |  EAB  |  face biometrics  |  facial analysis  |  Unissey