Removing bias within facial verification AI with more diverse data

The image shows Unissey logo
Success Story


Type of service

Services icon from humans in the loop with man collecting data on purple globe background


Illustrative image of a person with video annotation



Services circle icon from humans in the loop with woman at boardi on purple globe background
Annotations completed


Model validation circle icon from humans in the loop girl celebrating with robot on purple globe background
Seconds worth of video footage

The client

Unissey is an innovative startup based in Paris, providing solutions in the field of identity confirmation by facial biometrics, allowing users to access daily services via facial recognition in a matter of seconds. Their solution is based on liveness detection and facial recognition algorithms to ensure the true identity of the user while preventing identity theft, based on a photo of their personal ID and a one second live video selfie. 

The solution is based on liveness detection algorithms, to ensure that the user is not a fraudster using a stolen photo, a 3D mask, or a video on another device. This is coupled with facial comparison algorithms to ensure the person is the same one as on the photo ID. With a team of more than 20 experts in the field, they have created an intuitive, secure, and accessible experience to access the digital world. Unissey focuses on critical issues such as the fight against biometric discrimination, data protection, and the fluidity of the customer experience.

The challenge

The issue of conscious or unconscious bias is often brought up when considering facial biometric algorithms. One major reason for this is the lack of diversity in the datasets used for training such algorithms. The focus here was to get a more representative dataset, through greater diversity within end-users physical characteristics from one point. A second key component is diversity within the environment of the collected data – that is to say: avoiding controlled environments, and representing real life conditions displaying similar characteristics to the ones met by the end customers and all the challenges it encompasses. These include acquisition devices, variety of luminosity, environment, and others. 

The main challenges revolved around:
1) outsourcing the task to a partner who could deal responsibly with personal and sensitive data
2) to create the most diverse and representative dataset possible – gathering participants from a wide range of gender, age, religion and geography.

Image in the blog article Removing bias within facial verification AI with more diverse data

Image of Unissey’s solution in action

The solution

A real differentiator with HITL is that whatever field you may be working in, you can be confident collaborating on your project as they will know how to adapt.

Image of Sophie Finet, Product manager

Sophie Finet,
Product manager

Unissey recognised Humans in the Loop as the right partner for 3 main reasons:

1) the expertise in data acquisition on a global scale thanks to their international outreach
According to Sophie, “one of Humans in the Loop’s strengths was to have data samples from all around the world, which was crucial for the development of our solution – our need for a diverse dataset offering a broad spectrum of faces and video acquisition conditions was met, with each obstacle paired with a quick and thorough solution – from finding sources, to managing data protection requirements”

2) the annotation expertise demonstrated and a strong ethical and committed approach to the end application
The flexible approach to adapt to Unissey’s in-house annotation platform allowed Unissey to work more efficiently by reviewing annotations and validating the output quality. The smooth, responsive communication maintained throughout ensured success to a global process.

3) The ability to comply fully with GDPR regulations along with a transparent, responsible work flow 
“We found ourselves on the same page on this subject – we saw a real commitment to the fight against bias which resonated with the one we lead against biometric bias. HITL did a great job explaining the purpose of the collection and its boundaries, as such we were confident that explicit and freely given consent had been obtained during the data collection process”

The result

A partnership formed in the span of a year thanks to the ability to continuously generate data and provide a high standard of consistent annotation output. The initial scope targeted predominantly the collection of data from various types of devices (mobile and desktop) along with a variety of backgrounds and environments (indoor, outdoor). This later expanded to include data collected from various geographic regions, age categories, and gender.

Humans in the Loop assembled collectors on the ground and was able to acquire selfies shot in real-life conditions, through a process that ensured that each user was informed about the intended use of their personal data and provided their informed consent. In addition, Humans in the Loop used Unissey’s in-house proprietary annotation platform in order to provide an additional layer of quality control according to pre-established rejection criteria. For instance, selfie videos that were made in poor lighting conditions (e.g. too dark), or with excessive movements were automatically set aside. 

Unissey and Humans in the Loop have since worked together on three separate projects, achieving a milestone of over 26,000 seconds of video footage and 22,400 different annotations done, with data collected from over five continents, pushing forward the mission of building unbiased and representative facial biometric systems for all.

Interested in collecting diverse data for your AI solution? Get in touch with us and we would be happy to have a call.