Towards Equitable Biometrics: Addressing Demographic Bias in Facial Recognition Systems

Muhsin A., Akande H. F., Musa S., Iliyas M. U.

Abstract: This research presents a comprehensive approach towards enhancing the inclusivity and accuracy of facial recognition technology by addressing its inherent bias towards certain skin tones. The motivation for the study lies in the increasing need for improved biometric identification tools due to heightened security concerns in Kaduna City and escalating instances of academic dishonesty, particularly student impersonation. This issue is amplified in our context, considering the predominantly indigenous African student population, which faces the brunt of facial recognition bias. Our work begins with a literature review of four principal face detection methodologies, underscoring the under-addressed gap in improving facial recognition accuracy for darker skin tones. To address this gap, we collected and utilised a balanced dataset of 15, 000 images from a diverse demographic to train a convolutional neural network model. The ensuing facial recognition software, borne out of this training, demonstrated superior performance compared to the default model. It exhibited improved accuracy across a range of conditions and demographics, effectively managing access control. Our results indicate that our model successfully mitigates demographic bias, pushing the boundaries of facial recognition technology towards enhanced equity. However, additional testing and continuous model refinement are critical to further cement these advancements. This research underscores the need for strategic efforts to address and overcome inherent bias in facial recognition systems, laying the groundwork for future research towards achieving truly equitable and inclusive biometric identification systems.

Keywords: Facial recognition systems, Biometrics, Demographic bias