Empirical Investigation of Algorithmic Bias in Facial Recognition Systems Applied to Nigerian Population Data

Facial recognition systems trained predominantly on datasets skewed toward non-African facial features demonstrate documented performance disparities across demographic groups, raising equity and civil liberties concerns when deployed in Nigeria's security and financial sectors. This study empirically investigates algorithmic bias in facial recognition systems using Nigerian population data. A diverse test dataset of 8,000 facial images was assembled, comprising 4,000 images collected with informed consent from volunteers across six geopolitical zones, and 4,000 images from publicly available Nigerian social media datasets, stratified by gender, age group, and skin tone (measured using the Fitzpatrick scale). Three commercially available and two open-source facial recognition systems were evaluated on face detection, verification (1:1 matching), and identification (1:N matching) tasks. Statistical analysis of performance across demographic subgroups used AUC-ROC, false match rate, and false non-match rate. Results confirm significant performance disparities: mean false non-match rate for dark skin tones (Fitzpatrick V-VI) is 4.3 times higher than for medium skin tones across all tested systems. Gender disparities are also present, with female subjects showing 2.1 times higher false non-match rates than male subjects on four of five systems. The study provides empirical evidence of facial recognition bias in the Nigerian context and recommends mandatory demographic bias auditing prior to deployment of facial recognition systems in any Nigerian public safety or financial application.

Keywords: facial recognition, algorithmic bias, Nigeria, demographic disparities, AI fairness