Why do Yoti facial age estimation results published by NIST differ to those reported by Yoti in its white papers

profile picture Rachael Trotman 4 min read
Yoti's Facial Age Estimation results versus the NIST Age Estimation evaluation report

In September 2023, we submitted our facial age estimation model to the US National Institute of Standards and Technology (NIST), as part of a public testing process. This is the first time since 2014 that NIST has evaluated facial age estimation algorithms. NIST age estimation reports are likely to become a globally trusted performance guide for vendor models.

NIST assessed vendor Facial Age Estimation models using 4 data test sets at certain image sizes:

NIST image sizes used in the evaluation of Yoti's Facial Age Estimation

NIST provides some example images:

Fig. 5. The figure gives simulated samples of application type image used in the evaluation. Image source: Authors. Fig. 4. Examples of mugshot images used in the evaluation. Image source: NIST Special Database 32: Multiple Encounter Deceased Subjects (MEDS).

NIST note in their report that age estimation accuracy “will depend on the quality of the images” and the type of facial images captured.

For 6 years, Yoti have trained our model on primarily selfies of people looking into a mobile phone camera (or a laptop camera) because this is the obvious way customers can capture (live) their facial image to be age estimated. We capture these facial images at 720 x 800 pixels, with the face closely cropped to maximise the facial detail, because we have learned that we can attain higher age estimation accuracy for businesses by using this image size. 

We believe our training and testing on mobile phone images with closely cropped faces at 720 x 800 image size are key reasons why Yoti published MAEs (and FPRs) are lower (more accurate) for the Yoti model than the performance data published by NIST their 4 different test data sets.

Table displaying the differences in performance between the NIST evaluation results and Yoti's own testing results of Yoti's Facial Age Estimation.

NIST selected FPR objectives of 10%, 5% and 1% in their report as a way to benchmark their evaluation. As can be seen from the table above, NIST publish that Yoti’s age estimation model is more accurate on higher image size ‘Mugshot’ faces than lower image size ‘Application’ faces.  Consequently, the age thresholds required to meet FPRs of 10%, 5% and 1% are lower for Mugshot images than those needed using Application images. The age thresholds required to meet these FPR %s are lower still when the Yoti model is estimating age from mobile phone captured, higher image size, facial images. 

NIST used over 11 million facial images (with verified age) to test vendors. Some readers may wonder why NIST did not also test vendors with a test set of mobile phone camera facial images given, this is how most images will be captured for online age estimation.

The reality is that it is very challenging to capture, with consent, a database of millions of mobile phone facial images with ground truth date of birth evidence from individuals representative of many countries across the world.

Yoti is fortunate to have a very large set of consented and anonymised facial images, verified to government issued age data, from Yoti app users. By separating out ~120,000 of these images as diverse test data across each year of age, from the many millions of images used to train our algorithm, we have confidence in the accuracy figures we publish in our white paper (based primarily on mobile phone facial images at 720 x 800 pixels).

As part of our document authenticity in our identity verification service we compare the age estimation result of the selfie with the real age from their document, which also helps us test the accuracy of the model.

Finally Yoti’s facial age estimation model was first tested for accuracy, and positively certified, in November 2020 by ACCS, a UK accredited testing agency. Our age estimation model is used by some of the largest online brands, including Meta and OnlyFans, both of whom have publicly stated that it works very well.

Keep reading

Under the hood of facial age estimation

Since the UK Online Safety Act came into force on Friday 25th July, there has been a lot of discussion about how effective age checks are.  Facial age estimation has been approved by Ofcom as a high assurance method for online age checks. It’s a quick, private and effective way to confirm if someone is above or below an age threshold.  Each facial age estimation is quick and simple – taking around one second. This has sparked conversation when talking to businesses, press, regulators and governments – one of the most striking comments we’ve heard when showing them the

3 min read
An image of a boy sat at home, using his mobile phone

Yoti supports platforms as UK’s Online Safety Act comes into force

Today marks a major moment for the internet in the UK: the country’s new Online Safety Act officially comes into force, bringing in a new chapter of accountability for digital platforms and service providers. From social media and gaming, to porn and dating, we’re pleased to be supporting platforms with highly effective, robust and privacy-preserving age assurance solutions. The legislation requires platforms to protect children from harmful or inappropriate content, prevent underage access, and build safer digital environments. Yoti’s age assurance technology is helping platforms to meet these new standards whilst protecting user privacy and anonymity. We’ve also been

4 min read

Facial Age Estimation white paper

Making it faster and safer to prove your age Our age estimation technology accurately estimates a person’s age by looking at their face. We built it to give everyone a secure and private way of proving how old they are in different everyday scenarios: from age checking on social platforms and online stores, to supermarket self-checkouts, bars and clubs. This privacy-friendly approach to age verification doesn’t require any personal details or documents, and all information is instantly deleted once someone receives their estimated age – nothing is ever viewed by a human.   Key takeaways

2 min read