FaceAssure: DEVELOPING A ROBUST, PRIVACY-FIRST, FACE-BASED AGE ESTIMATION TECHNOLOGY
Main outcomes
- Privately received the UK regulator’s prestigious Level 2 certification and joins a handful of global companies to deploy their age assurance product across many industries.
- FaceAssure is a privacy-preserving age estimation technology with liveness and quality checks in place.
- Lakera’s development platform was used during development to achieve leading accuracy (mean age error of 1.33 years).
- Lakera has enabled Privately to reduce product iteration times from 1–2 weeks to 1–2 days.
- Privately has the ML infrastructure that allows them to continue building out their products at high speed and reliability.
About Us
Privately is a Swiss company born in 2014. We help platforms deliver age appropriate and safe online experiences to children. We democratize age estimation for developers, and automate regulatory compliance for companies.
Our on-device AI technologies determine age, well-being and online safety markers of the child through text, image, voice and behavior analysis. Our technologies are privacy-preserving, and allow real-time support and interventions to be integrated within the user experience.
The technology can be integrated within our clients’ apps, games or devices running privately within a smart-device environment through our Software Development Kit (SDK) or via secure on-browser implementation.
What is Age Estimation?
Age Estimation technologies deduce the age range of a person based on the analyses of biometric features, such as facial patterns or voice patterns. Though relatively new, the automated detection of age promises to revolutionize many industries and make the internet a safer place for minors.
Historically the purchase of restricted goods and services has been subject to age verification in the following retail spaces:
- Alcohol, cigarettes, including E-Cigarettes and vaping products
- Lottery tickets and scratch cards
- Dangerous Weapons, e.g., crossbows, knives.
We now observe a regulatory push for age assurance in online worlds as well. This is on account of rapid spreads of online advertisement, social media, adult sites, online video games, augmented & virtual realities such as metaverses. As a response, new regulation in the UK and US mandates Age Appropriate redesign for all web services including retail. Sellers and advertisers must ensure the user’s age beyond self-declaration.
Traditionally, age estimation is undertaken by natural persons, in a wide range of settings, all of the time. These natural persons make a judgment, based on a person’s appearance, whether they are old enough to purchase age-restricted goods, content or services. However, this approach notably fails to respond to important challenges:
- Humans estimations of age can vary very significantly from real ages
- How to prevent underage access to restricted goods / content online?
- How to verify many adults who do not have or don’t want to share identity documents online?
- How to handle data processing liabilities GDPR / biometric data for underage users?
As such, AI-Assisted, On-Device age estimation technologies prove to be a vital solution for ensuring age appropriate experience. Privately has set out to build such a system, named FaceAssure.
FaceAssure analyzes patterns on faces in order to estimate the ages. It does so with complex statistical rules derived via Deep Learning methods. On top of the core age estimation routine, FaceAssure contains other useful modules for liveness & quality check and user guidance for optimal outcomes.
Our system runs on the edge: it runs either fully as a mobile SDK, or as an on-browser process: no biometric data ever leaves the user’s device. This has been noted by ICO, which audited us and published their report on their website. They noted:
The Zero Data Principle, which appears to fundamentally underpin Privately’s approach to development and commercialisation, is an excellent approach which helps to ensure that Privately’s final product is as privacy conscious as possible, and seriously reduces the risk of a data subject’s rights being breached, or any aspect of data protection legislation being infringed by the operation of this service.
Making FaceAssure Robust
While our results showed that we were in the right direction, we needed better performance to reliably deploy FaceAssure in age-restricted environments. In particular, we needed to improve the robustness of our system against variations in the presented images during operational use on the edge, namely:
- Skin tones and Gender
- Directional luminosity, and variations on contrast and brightness
- Image compression, noise, blur
For this need, we started to use the MLTest of Lakera. MLTest is a tool to identify critical performance vulnerabilities, and is composed of model testing, data testing, and a dashboard. Installation was quite straightforward, just with a pip install and a docker pull command.
MLTest has helped us improve our system, in particular by:
- Identifying problematic inputs and edge cases.
- Identifying efficient data collection strategies.
- Identifying the best training- and test-time augmentation strategies to increase robustness to real-world perturbations.
As a result, MLTest has effectively increased our development speed since it took care of many elements of stress testing, and helped us identify the corner cases with examples. In fact, our product iteration cycles (model training, testing, and deployment) got faster from 1–2 weeks to 1–2 days.
Thanks to the indicators we gathered from their dashboard, which ran locally on our premises, we were able to pick the right training-time and inference-time augmentations to improve our overall system.
Genuineness check
While our team worked relentlessly on a cycle of additional data collection, optimizing model performances, real life testing, robustness testing (through MLTest) and back to iterative optimization, we also found ourselves pitted against minors who are trying to pass off as adults and believe it or not, some adults trying to pass off as children. As such, we decided to add spoof detection as an integral part of an industry-grade age estimation/verification system.
We were also helped by our partners at IDIAP labs to test our systems against advanced spoofing attempts — including latex masks.
As a response, we have developed active and passive liveness checks, covering more than 95% of the spoof attempts.
Performance numbers
With iterative developments on age estimation models, we were able to bolster our system performance to the following levels:
- Robustness against random perturbations: MLTest helped us notice that a majority of our test scores would be impacted by perturbations that naturally occur in real-life usage. Thanks to MLTest’s indicators, we were able to put the appropriate measures and managed to minimize this impact down to 10% and are much better prepared for subsequent developments.
- Mean Age Error: In May 2022, ACCS, a UKAS accredited conformity assessment body, noted a mean age error of 1.33 in their tests, a 61% reduction of the margin of our error. As a result, FaceAssure is now a Challenge-25/EAL-2 certified age estimation technology.
- The system is also robust against biases on account of different skin tones and is therefore very well adapted to large-scale market deployment.
Future Outlook
Our technology has a wide variety of applications, from age gating to self-sovereign identity, to compliance automation and age-appropriate advertisement. Fueled by our successful certification, we’re already running pilot studies and large-scale deployments with our clients on these use cases. We will increase our focus on growing our developer community as well.
We’re looking forward to certifying VoiceAssure, our voice-based age estimation product!