Ocenjevanje in zmanjševanje demografske pristranskosti v biometričnih sistemih

Face recognition systems are reaching superhuman levels of accuracy but remain uneven across demographic groups. This thesis tackles demographic bias with a data-centric method: identity-preserving face swapping to build large synthetic datasets where each identity is equally represented across gender and ethnicity. Using a U-Net with FiLM conditioning, we generate and validate a two million-image dataset (based on BUPT-BalancedFace) via failure analysis, face-image quality metrics, and embedding-space comparisons. Fine-tuning ArcFace and AdaFace on this data reduces inter-group error disparities per standardized metrics, at some cost to verification accuracy. Versus fine-tuning on real balanced data, synthetic augmentation yields lower absolute accuracy but measurable fairness gains without extensive demographically labeled data—highlighting both the promise and limits of synthetic data for bias mitigation.