IDnow reduces bias in facial recognition with EU-funded MAMMOth project

Identity verification provider IDnow reports significant progress in reducing algorithmic bias in facial recognition systems, following its participation in the EU-funded MAMMOth (Multi-Attribute, Multimodal Bias Mitigation in AI Systems) project.

The company notes the widely-cited 2018 “Gender Shades” study from MIT Media Lab prompted the search for more inclusive data and better model calibration. Biometric testing by the National Institute of Standards and Technology (NIST) has found that the majority of facial recognition algorithms are more likely to misidentify people with darker skin, women and the elderly, though the most accurate algorithms show very low differentials in the Institute’s latest testing.

As part of MAMMOth, IDnow focused on identifying and mitigating bias in its own facial recognition algorithms. One key challenge was the variation in skin tone representation caused by ID photo color adjustments, which can distort comparisons between selfies and official documents. To address this, IDnow applied a style transfer technique to diversify its training data, improving model resilience and reducing bias toward darker skin tones.

Other tools developed by IDnow address bias at different levels, such as biometric matching algorithms.

The results were notable: verification accuracy increased by 8 percent, even while using only 25 percent of the original training data volume. The accuracy gap between lighter and darker skin tones was cut by more than 50 percent. The enhanced AI model was integrated into IDnow’s identity verification platform in March 2025 and has been in active use since.

“Research projects like MAMMOth are crucial for closing the gap between scientific innovation and practical application,” says Montaser Awal, director of AI and ML at IDnow. “By collaborating with leading experts, we were able to further develop our technology in a targeted manner and make it more equitable.”

IDnow plans to adopt the MAI-BIAS open-source toolkit developed during the project to evaluate fairness in future AI models. This will allow the company to document biometric bias mitigation efforts and ensure consistent standards across markets.

“Addressing bias not only strengthens fairness and trust, but also makes our systems more robust and adoptable,” Awal added. “This will raise trust in our models and show that they work equally reliably for different user groups across different markets.”

The nuances of AI bias and discrimination

The MAMMOth project, supported by Horizon Europe, brought together leading academic and industry partners to tackle fairness in artificial intelligence across multiple modalities. The European research project ran for 36 months, concluding this month, and set out to tackle gender, race and other biases in AI. It is part of a broader European Union injunction that prohibits discrimination in EU law.

As AI becomes more ubiquitous in various domains including health, education, justice, personal security, work and so on, MAMMOth sought to identify a list of characteristics that are not protected under the law, but which have been shown to lead to bias in AI systems. Such biases have been studied and documented in leading academic journals, which MAMMOth references.

For example, these characteristics could be school grades and other personal details collected about the offender, like living situation, in a justice system context. In healthcare, attributes shown to be potentially connected to bias in AI systems include disability, age, native language and dialect, sexuality and socioeconomic status, among others.

Since AI trains on available data, biases could become further entrenched. For example, it is well known that psychology studies disproportionately draw their participants from Western, Educated, Industrialized, Rich and Democratic (WEIRD) societies. This reliance may mean the findings do not accurately reflect the responses of individuals from diverse cultural backgrounds.

More on the characteristics can be found on the MAMMOth website here.

Article Topics

biometric bias  |  biometrics  |  biometrics research  |  demographic fairness  |  facial recognition  |  IDnow  |  MAMMOth (Multi-Attribute  |  Multimodal Bias Mitigation in AI Systems)