IDnow and MAMMOth: On the way to an unbiased AI-powered face verification
IDnow is a leading identity verification platform provider in Europe with a vision to make the connected world a safer place. At IDnow, innovation means more than just creating something new or novel. The company believes that innovation is finding solutions to problems that do not exist yet but are sure to impact the future. That is why, since its creation, IDnow has invested in research and artificial intelligence (AI) technologies, by recruiting a team of senior researchers and R&D engineers and by participating in collaborative research projects.
IDnow leverages AI in its proprietary algorithms to automate three critical tasks: 1) capturing and parsing identity documents, 2) ensuring document compliance and authenticity, and 3) verifying the document holder’s identity, including a face verification step.
MAMMOth toolkit, a solution to identify and mitigate AI-biases
While AI has seen increased adoption in recent years, concerns have arisen regarding its sensitivity to bias, which may lead to unfair data processing and discrimination against underrepresented populations. Preliminary research conducted by IDnow has revealed some demographic biases in the facial verification algorithms, primarily linked to the underrepresentation of several population groups in the available data. Although a manual analysis performed after the automated check can rectify the AI verdict, addressing bias issues in AI-powered services is a priority for the company: IDnow is committed to ensuring that customers and end users have access to reliable, trustworthy, and fair tools.
To tackle this challenge, IDnow has been working with 12 European partners within the MAMMOth project, funded by the European Research Executive Agency, for about a year. This project seeks to comprehensively study existing biases and implement corrective solutions. The goal is to offer a toolkit for engineers, developers, and data scientists working with AI to identify and mitigate biases in datasets and algorithm outputs. As a partner in this project, IDnow coordinates the implementation, testing, and evaluation of the proposed solutions for three use cases: assessment of loan applications, academic evaluation and visibility of academic work, and face verification applied to identity verification. The company’s research work will focus on the face verification use case, to implement methods to mitigate the biases that have been found in its algorithms.
Promote the voice of under-represented groups
In an effort to provide a tool that meets the needs of those affected by these biases, the project has the support of three associations working to promote the voice of under-represented groups: IASIS, DAF, and DDG. Their involvement in MAMMOth, thanks to the co-creation approach taken by the project, will allow IDnow to be as close as possible to the needs and expectations of under-represented groups during the technical developments.
The sociologist expert UNIBO and these three associations have collected input from underrepresented groups, especially women and ethnic minorities. Questionnaires circulated among these groups, represented by DAF, DDG and IASIS, revealed their apprehensions about being discriminated against by identity verification algorithms, emphasizing the necessity of ensuring algorithm reliability before use. They also advocated for human oversight of algorithm decisions.
During the co-creation workshops organized by DDG and DAF, participants expressed concerns about possible data leakages – thus highlighting the need to ensure that privacy and security-related issues are carefully addressed when these systems are developed – and unfair AI decision-making. Nonetheless, some participants highlighted that human input could also be biased; using AI could reduce this risk.
Once again, these studies underscore the need for trustworthy, unbiased facial verification algorithms. This is the challenge that IDnow and MAMMOth partners aim to overcome during the remaining two years of the MAMMOth project.