Title: Enhancing Fairness and Accuracy in Facial Recognition via Adaptive Calibration

Abstract:

This paper presents Adaptive Calibration (AC), an innovative calibration framework for facial recognition systems. AC transforms the cosine similarity derived from normalized embeddings into accurately calibrated probability scores. A core issue in facial recognition is the inherent misalignment of cosine similarity, where identical distances yield varying match probabilities depending on the specific region of the embedding space. Our method addresses this discrepancy by integrating local context into the calibration process.

Crucially, AC achieves superior performance and enhanced fairness without the need for demographic metadata. Across various standard benchmarks and pretrained models, our approach consistently outperforms existing techniques in both accuracy and fairness metrics. By offering continuous, region-specific calibration, AC avoids the common pitfall of "leveling down"—a scenario where efforts to improve fairness inadvertently degrade performance for certain groups. Consequently, AC offers a practical pathway toward equitable facial recognition systems that boost overall performance while eliminating the requirement for demographic group annotations.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC