2019/12/16 by Jacqueline G. Cavazos, P. Jonathon Phillips, Cavazos, Jacqueline G. +5 · 3 citations
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Face and Expression Recognition #Face recognition and analysis #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.1912.07398
openalex publication_date 2019/12/16 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Previous generations of face recognition algorithms differ in accuracy for\nimages of different races (race bias). Here, we present the possible underlying\nfactors (data-driven and scenario modeling) and methodological considerations\nfor assessing race bias in algorithms. We discuss data driven factors (e.g.,\nimage quality, image population statistics, and algorithm architecture), and\nscenario modeling factors that consider the role of the "user" of the algorithm\n(e.g., threshold decisions and demographic constraints). To illustrate how\nthese issues apply, we present data from four face recognition algorithms (a\nprevious-generation algorithm and three deep convolutional neural networks,\nDCNNs) for East Asian and Caucasian faces. First, dataset difficulty affected\nboth overall recognition accuracy and race bias, such that race bias increased\nwith item difficulty. Second, for all four algorithms, the degree of bias\nvaried depending on the identification decision threshold. To achieve equal\nfalse accept rates (FARs), East Asian faces required higher identification\nthresholds than Caucasian faces, for all algorithms. Third, demographic\nconstraints on the formulation of the distributions used in the test, impacted\nestimates of algorithm accuracy. We conclude that race bias needs to be\nmeasured for individual applications and we provide a checklist for measuring\nthis bias in face recognition algorithms.\n