Effective facial age estimation: a privacy-preserving approach to age assurance

profile picture Amba Karsondas 8 min read

For platforms to deliver age-appropriate content, they need to know the age of their users. Age estimation technology can provide an inclusive and accessible solution. It’s possible to estimate a person’s age from a number of features, including their voice, face, palm or fingerprints.

Some age estimation methods are very accurate. Others collect little or even no personal data. But very few can do both.

When done to a high standard, effective facial age estimation can offer a high level of privacy and a high level of assurance.

 

Age estimation in the real world

Consider the following situation: a person who is clearly underage tries to buy alcohol from a supermarket. 

When asked to prove their age, they present an identity document stating they’re over 18. If the cashier thinks the person looks under 18, they would reject the document. They would trust their instincts that the document may be fake, tampered with or belongs to someone else.

This ‘common-sense’ approach can be thought of as human age estimation.

 

Human age estimation

Through our experiences, humans have developed an ability to assess someone’s age. Over time we learn “that’s what people of a particular age look like”. We do this by observing visual cues such as wrinkles and hair colour.

But there are lots of reasons that can make it difficult for humans to correctly estimate another person’s age. We’re quite easily able to work out the ages of those closely aged to us. However, when we go outside of this range, our accuracy decreases. This accuracy also tends to fall as we ourselves get older.

It should be noted that various factors can affect the ageing process, such as:

  • nutrition and quality of diet
  • exposure to sun/wind
  • use of narcotics
  • cosmetic surgery
  • physical labour
  • stress
  • lack of sleep

Retail staff often need to estimate the ages of customers who are buying age-restricted goods. But when viewing lots of faces in succession, a person’s judgement tends to be influenced by the faces they’ve just seen.

Some regulators have imposed safety buffers to allow for these challenges. For instance, the UK has implemented the Challenge 25 policy. If a person buying age-restricted goods looks under 25, they must present a physical ID to prove their age.

 

Facial age estimation

Facial age estimation has automated the age estimation process. It allows technology to estimate a person’s age. It is powered by an algorithm that’s learned to estimate age in the same way humans do – by looking at faces. The technology detects a live human face, analyses the pixels in the image and gives an age estimate.

Facial age estimation is not facial recognition; it estimates someone’s age without identifying them. It does so by converting the pixels of a facial image into numbers and comparing the pattern of numbers to patterns associated with known ages. It does not scan their face against a database or recognise anyone.

An illustration detailing how Yoti's facial age estimation technology works. There are three steps. They are as follows: firstly, a face is detected in an image and reduced to pixels. Each pixel is assigned a number that the AI can understand. The second step shows the numbers being computed by a neural network that has been trained to recognise age by looking at millions of images of faces. The final step shows the technology finding a pattern in the numbers and producing an age.

As soon as an age has been estimated, the facial image is immediately and permanently deleted.

 

Safety buffers

In certain sectors, it is illegal for someone under a specified age to access particular goods, services or experiences.

In these cases, companies can include a safety buffer when using facial age estimation. These account for potential errors and provide a necessary margin of safety. The size of this buffer can depend on the level of accuracy required by the business or on any regulatory requirements.

For example, if a business needs to know if someone is 18 or over, they might apply the Challenge 25 policy and set a seven-year buffer. This means that the technology must determine the user’s age to be at least 25. 

 

What is effective facial age estimation?

Facial age estimation can give a high level of assurance when done effectively. Here are some of the steps required:

An infographic detailing the stages of effective facial age estimation. This is presented as three layers. The first layer is facial age estimation with liveness detection only. The description reads "check the person behind the camera is a 'live' person and not a photo or video." The second layer adds independent testing. The description reads "have a credible third party assess the technology for data compliance, bias and accuracy". The third layer adds injection attack detection. It reads "add robust anti-spoofing technology at the point an image is being taken to detect injection attacks.

Liveness detection

Businesses need to be sure that the individual completing the age estimation is a real person. The technology must be able to detect whether the face shown is a live person, and not a photo, mask or video. Without this, a child could attempt to use a photo of their parent to pass the age check.

There are two types of liveness checks – active liveness and passive liveness. Active liveness checks require users to perform specific actions or follow some instructions. The user may have to nod their head or say certain words. Though effective, this may be difficult, particularly for users with disabilities.

Language, audio input and noisy environments can also be a problem for some users. Instructions could be misunderstood by those reading them in a second language. Or some people may be self-conscious if completing the liveness check in public. 

Passive liveness works from a facial image. This creates a smooth, inclusive and more accessible user experience. The technology does not recognise any faces and does not save any images. It simply checks that the face in the image belongs to a real, live person.

NIST provides a framework for testing performance levels of liveness. A good benchmark for liveness is NIST Level 2. This involves testing against specialist attacks such as latex face masks or 3D printers.

 

Independent testing

With any technology, people can have questions about how it works and if it’s safe to use. Therefore, independent testing is a crucial part of effective facial age estimation. External accreditation is necessary for three main factors:

 

Data handling

Users need to be sure that their privacy is protected. They need to know that the technology does not recognise anyone, is not storing the image of their face, and that their data is being handled safely. 

Facial age estimation providers need to deliver clear guidance on how the technology works. They must also be transparent about what happens with a user’s data.

 

Bias

Effective facial age estimation should be tested for bias. At a minimum, it will be assessed for its accuracy by gender, skin tone and each year of age. This is to ensure that the technology works fairly for all people. Where there are inconsistencies, efforts should be made to improve the algorithm.

 

Accuracy

The technology should also be assessed for accuracy. Alongside testing the algorithm, testing should be done in a live environment. For example, when used at a gambling machine in a bar, the technology needs to be able to adapt to different lighting.

The technology’s performance should be measured in terms of true positives, and false positives, success rates and completion times. 

  • If the model correctly estimates an 18-year-old to be under 25, this is a true positive.
  • If the model incorrectly estimates an 18-year-old to be over 25, this is a false positive.

The aim is for a high true positive rate and a low false positive rate. This means that the right people can efficiently access certain age-restricted goods, products or experiences. And those who are outside of the correct age bracket are protected from accessing them.

It is good practice for age estimation providers to be transparent about these rates so that both businesses and individuals know how accurate the technology is.

 

Injection attacks

An injection attack involves injecting an image or video into the system. This is so the technology estimates the age of the face from the injected image, rather than the one in front of the camera

This is a particular issue for online facial age estimation where the image is taken with a laptop or mobile phone camera. With free software and limited technical ability, it’s fairly easy to swap an image for one that’s designed to pass the checks.

An effective facial age estimation system will be able to detect injection attacks. This is achieved with integrated anti-spoofing technology which can detect attacks at the point the image is taken.

 

So what does this all mean?

Facial age estimation can be an inclusive and proportionate method of age assurance. But the necessary controls must be in place to ensure that the technology delivers a high level of assurance and high privacy.

To find out more about how effective facial age estimation may be useful for your organisation, please get in touch.

Keep reading

An image of a person holding smart phone in their left hand. On the phone's display is a Yoti Digital ID showing their verified credentials. In the person's right hand is their physical UK driving licence. The person is holding the two forms of ID next to each other.

From physical to digital: how we transform your ID

There can be hidden beauty in things taking on new forms. A caterpillar becoming a butterfly, water turning into wine, a physical ID transforming into a Digital ID. If you’ve ever wondered how we do it, wonder no more. As we continue to shape the future of digital identity, we’ll continue sharing our methods with you. After all, the Digital ID app is created for you, so it’s only right we let you in on the magic.   What is a Digital ID? A Digital ID is your ID on your phone and your data in your hands. It gives

5 min read
An image of a person's smartphone. The person is using Yoti's facial age estimation technology. Next to this image, the text reads: "externally audited, over 600 million age checks, used by leading brands". Underneath this list are the logos of Instagram, Aldi, SuperAwesome and Only Fans.

Age assurance: What makes for the maturity of a technology or an industry?

Last week, the Australian government decided against the eSafety Commissioner’s recommendation of a pilot before requiring adult sites to verify the ages of visitors. They said this was due to concerns about the privacy of people’s data and the maturity of age assurance technology. So what exactly would constitute a mature technology? Is it something comparable to NASA’s Technology Readiness level? This suggests the technology needs to have gone through a thorough process of research and prototyping, before testing in a live environment and then ultimately rolling it out. If that’s the case, we can say that Yoti and other

7 min read
An illustration of a padlock with the Yoti logo sitting at the centre. Alongside this are three smaller icons showing that the app is free (represented by a crossed out pound sign), cannot be hacked (represented by a crossed out pickaxe) or seen by any third parties (represented by a crossed out eye)

How we keep your Digital ID private and secure

We’re committed to making the digital world safer for everyone. Yoti was created as we wanted to give every person a secure way of proving their age or identity. It’s quite literally why we exist. So it only made sense that we’ve built our Digital ID app with privacy and security at its core. We’re building technology that makes it easier and safer for you to go about your business, but that doesn’t mean we have to know your business. Just as the right to an identity is a fundamental human right, so is the right to privacy. And we’ll

6 min read