Quantifying the Extent to Which Race and Gender Features Determine Identity in Commercial Face Recognition Algorithms

This study was sponsored by the U.S. Department of Homeland Security and conducted at the Maryland Test Facility as part of ongoing evaluations of biometric performance across demographics groups. Using data gathered from the 2018 Biometric Technology Rally, we show that commercial face, but not iris, recognition algorithms use features associated with race and gender to establish individual identity. Here, we propose a first-of-its-kind method to quantify the extent to which different biometric algorithms exhibit this effect. We discuss the implications of these findings in the context of equitable performance of face recognition in verification and identification use cases.