Master's Thesis : Towards fairness in face recognition systems

Abstract

Nowadays, state-of-the-art algorithms for face recognition achieve great results, even over human performances on most known testing datasets. But these algoriithms tend to be biased as the training and testing dataset are usually over represented by people sharing common facial features and color skin. Recent studies show that results on balanced dataset or representing the world population distribution tend to give lower performances and discrepancy between groups of people with different skin colors than the over-represented one. In this work, we define three metrics to evaluate this discrepancy and present three methods to reduce this discrepancy and improve results on balanced testing datasets.