An Approach for Reducing Racial Bias in Facial Monitoring Systems

Abstract

This thesis provides a new approach to reduce racial bias issues and inaccuracies caused by unbalanced benchmark datasets in facial detection systems. It is well known that these unbalanced benchmark datasets significantly over-represent white individuals. It is also understood that a deep learning model performance is based on the data used for training. With these two conjectures, previous research has shown how inaccuracies across different racial groups can be hidden by tracking a single class's labels, faces, and performance. New balanced benchmark datasets have been developed however they lack the variability seen in transitional benchmark sets. Additionally, manually annotating and retraining these models is both computational and financially expensive. Therefore, this research proposed a financially inexpensive way to reduce racial bias within pre-trained facial monitoring systems using semisupervised self-supervised learning.