Semester

Spring

Date of Graduation

2025

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Jeremy Dawson

Committee Co-Chair

Nima Karimian

Committee Member

Nima Karimian

Committee Member

Prashnna Gyawali

Abstract

Facial recognition technology is utilized in many facets of life. As the use has become more widespread these systems have improved in reliability and performance approaching the level of human accuracy. With these improvements the problem of bias still remains as a persistent problem. Efforts have been made to minimize the bias prevalent in the systems via studies into various demographic factors, creating training datasets that have a more uniform distribution of subjects, and other methods. As facial recognition is one of the most utilized forms of biometric recognition it is vital to analyze potential causes of bias to help mitigate impacts on the performance of the systems. Examining results based on facial phenotypes is a unique feature analysis that has not been the focus of studies. As models typically grab the facial features associated with facial phenotypes they could add bias to a system.

This work consists of creating a dataset of facial phenotype annotation values based on facial images. The selected annotations are based on features that are examined by forensic experts for facial identification. A variety of publicly available facial recognition matchers are used to produce match score results that are then used for the purpose of creating receiver operating characteristic curves and the associated area under the curve, detection error tradeoff curves and the associated equal error rates. Other analysis is based on the 99th percentile non-mated score differential performance to analyze trends in imposter images that had the highest similarity to gallery images. The results indicate that the facial measurements of certain phenotypes have an impact on match scores and vary based on demographic groups.

Download

Share

COinS