Statistical Methods for Testing Equity of False Non Match Rates Across Multiple Demographic Groups.

Biometric recognition is used for a variety of applications including authentication, identity proofing, and border security. One recent focus of research and development has been methods to ensure fairness across demographic groups and metrics to evaluate fairness. However, there has been little work in this area incorporating statistical variation. This is important because differences among groups can be found by chance when no difference is present or may be due to an actual difference in system performance. We extend previous work to consider when individuals are members of one or more demographics (age, gender, race). Our methodology is meant to be more comprehendable by a non-technical audience and uses a robust bootstrap approach for estimation of variation in false non-match rates. After presenting our methodology, we present a simulation study and we apply our approach to MORPH-II data.