Assessing demographic bias in face detection models

Details

Supervisors

Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée: analyse de données

Abstract

enThis work investigates fairness in face detection models by analyzing representation biases across demographic groups, specifically sex and ethnicity. We propose a novel test set construction methodology that balances both group representation and detection difficulty to improve the reliability of fairness measurements. Using a face detection model, we adopt Equal Opportunity as the fairness metric. Our experiments evaluate the influence of training data composition and test set construction strategies on model performance and fairness. The dataset subset, annotated with demographic information, will be publicly available to encourage further research. Limitations such as annotation subjectivity, dataset simplification, and reliance on a single model implementation are discussed. This study provides a framework for more robust fairness evaluation in face detection and highlights the importance of carefully designed test sets.