European identity verification provider IDnow has reported significant progress in mitigating algorithmic bias within its facial recognition systems. This achievement stems from IDnow’s participation in the EU-funded MAMMOth (Multi-Attribute, Multimodal Bias Mitigation in AI Systems) project.
IDnow’s research targeted a specific and common source of bias: variations in ID photo color adjustments, which can distort comparisons with live selfies and create inaccuracies for users with darker skin tones. Rather than undertaking the slow and expensive process of new data collection, IDnow’s team applied a style transfer technique. This algorithmically diversified its existing training data, making the model more resilient to these color variations.
The results were a clear success: the accuracy gap between lighter and darker skin tones was cut by more than 50 percent. This enhanced, fairer AI model has already been in production on IDnow’s platform since March 2025. This finding is significant for the industry, as it demonstrates that algorithmic data augmentation can be a more efficient and targeted method for solving bias than new data collection alone.
This work is also a commercial necessity. As an EU-based firm, IDnow is building a public, defensible compliance record ahead of regulations like the EU AI Act and eIDAS 2.0. The company plans to adopt the MAI-BIAS open-source toolkit, also developed by the project, to evaluate and document fairness in all future AI models, marketing trust and fairness as a core product feature.
—
By the ID Tech Editorial Team
