Facial age verification, puberty age estimation in various sectors

A recent webinar organized as part of the European Association of Biometrics (EAB) Lunch Talk series gathered views on the state of facial age verification and estimation (AVE) from Dr. Ana Calarasanu and Dr. Thomas Petit of French biometric authentication company Unissey.

Age verification is increasingly being used in sectors ranging from retail to online gaming. Calarasanu says a standard age verification toolkit may include document verification, credit card verification and behavioral analysis. However, as life continues to move online, algorithmic facial biometric analysis is becoming a valuable tool that can reduce user friction while providing an additional level of security and accessibility.

“For privacy reasons, we reveal less information than with an ID document,” Calarasanu says. Technologies such as activity detection ensure that facial biometric tools provide the level of security needed to meet rigorous standards. However, it notes the need for precision in the distinction between age verification and age estimation. Estimation tools, she says, “determine the approximate age or age range of a subject,” compared to the more binary approach used in age verification processes, which aim to confirm that a user is over a certain target age.

“This distinction is important because there are application use cases where what we really want to solve is age estimation, but also from a technical perspective there are differences and the metrics and performance won’t 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.

Meanwhile, age verification is necessary in cases such as games, pornography or age-restricted purchases.

The technical aspects vary

Petit discusses the different machine learning approaches that can be considered when building a verification or age estimation model, and how the specifics of a particular use case may impact which one is appropriate. It also talks about the difference between evaluation for age estimation and verification evaluation – in the case of estimation, a combination of average absolute error (the average difference between the actual and estimated age, so the lower the better) and the cumulative result (the percentage of predictions for which error rate is lower than the threshold, so higher is better).

Petit walks listeners through the concept of the “T Challenge”, where T is the required age plus seven, meaning that in cases where the legal age is 18, people appearing to be under 25 will be asked to show ID. (In this case, challenge 25.) Also tours National Institute of Standards and Technology (NIST) short datasets for age verification and estimation, NIST assessment metrics (first benchmark released this year), ethnicity bias, and age verification comparisons faces with human results, and how algorithms improve over time.

In summary, Unissey offers four takeaways. First, AEV is not infallible; Petit compares him to a doorman or bartender who can usually be relied upon, but who may nevertheless sometimes lack identification. That said, the technology already offers better results than human supervision. However, like humans, it has its biases, especially regarding gender and ethnicity, that must be taken into account. Finally, each use case will have its own details that should help you choose a product for facial age verification or estimation.

Petit says all indications are that facial age verification and estimation are on track for eventual widespread adoption, and he expects research and development on the technology to be “active” in the coming years. “The industry is becoming more mature and I would say it has reached a stage where it can be a reliable solution for many applications.”

Article topics

age assessment | age verification | biometrics | biometric tests | EAB | facial biometrics | face analysis | Unissey