Bias Mitigation in Face Recognition via Demographic-based Supervised Contrastive Learning

2026-08-13

AI Summary

This work addresses the significant demographic bias exhibited by existing face recognition systems across groups defined by gender, age, or race, particularly the pronounced unfairness in tail distributions under low false match rate regimes. To mitigate this issue, the authors propose Demographic-aware Supervised Contrastive Learning (DeSCon), a novel approach that incorporates demographic information into the supervised contrastive learning framework—an integration not previously explored for fairness enhancement. DeSCon employs customized training batches and a demographic-guided sampling strategy for positive and negative pairs to explicitly shape the distribution of non-match scores in the tail regions. Crucially, it departs from conventional reliance on data balancing. Experimental results demonstrate that DeSCon substantially improves group fairness on both datasets with demographic annotations and standard benchmarks, while maintaining state-of-the-art face verification performance.

Abstract

Face recognition systems have been shown to be biased toward certain demographic groups by exhibiting different error rates across gender, age, or ethnicity. Though the imbalance of the training data with respect to these demographics is one cause of this bias, training on artificially balanced groups does not completely mitigate the problem. For deployment, face recognition typically works at operating points allowing very low false match rates and, hence, on the tail of the non-match score distribution. While class balancing can improve the means of these distributions, the aim of our approach is to improve fairness by addressing the behavior in the tail. Particularly, we propose the Demographic-based Supervised Contrastive loss (DeSCon) for face recognition, which relies on a well-designed composition of training batches and demographic-aware pair selection. Our experimental evaluation on both demographically-labeled datasets and standard verification benchmarks shows that DeSCon can improve fairness beyond balancing training datasets while maintaining competitive verification performance. Source code is available upon request.

Problem

• bias mitigation
• face recognition
• demographic fairness
• tail distribution
• error rate disparity

Innovation

• Supervised Contrastive Learning
• Demographic Bias Mitigation
• Face Recognition Fairness
• Tail Distribution Optimization
• Demographic-aware Pair Selection