Jerry L. Tipton

John J. Howard, Eli J. Laird, Yevgeniy B. Sirotin et al.

The development of face recognition algorithms by academic and commercial organizations is growing rapidly due to the onset of deep learning and the widespread availability of training data. Though tests of face recognition algorithm performance indicate yearly performance gains, error rates for many of these systems differ based on the demographic composition of the test set. These "demographic differentials" in algorithm performance can contribute to unequal or unfair outcomes for certain groups of people, raising concerns with increased worldwide adoption of face recognition systems. Consequently, regulatory bodies in both the United States and Europe have proposed new rules requiring audits of biometric systems for "discriminatory impacts" (European Union Artificial Intelligence Act) and "fairness" (U.S. Federal Trade Commission). However, no standard for measuring fairness in biometric systems yet exists. This paper characterizes two proposed measures of face recognition algorithm fairness (fairness measures) from scientists in the U.S. and Europe. We find that both proposed methods are challenging to interpret when applied to disaggregated face recognition error rates as they are commonly experienced in practice. To address this, we propose a set of interpretability criteria, termed the Functional Fairness Measure Criteria (FFMC), that outlines a set of properties desirable in a face recognition algorithm fairness measure. We further develop a new fairness measure, the Gini Aggregation Rate for Biometric Equitability (GARBE), and show how, in conjunction with the Pareto optimization, this measure can be used to select among alternative algorithms based on the accuracy/fairness trade-space. Finally, we have open-sourced our dataset of machine-readable, demographically disaggregated error rates. We believe this is currently the largest open-source dataset of its kind.

John J. Howard, Yevgeniy B. Sirotin, Jerry L. Tipton et al.

With increasing adoption of face recognition systems, it is important to ensure adequate performance of these technologies across demographic groups. Recently, phenotypes such as skin-tone, have been proposed as superior alternatives to traditional race categories when exploring performance differentials. However, there is little consensus regarding how to appropriately measure skin-tone in evaluations of biometric performance or in AI more broadly. In this study, we explore the relationship between face-area-lightness-measures (FALMs) estimated from images and ground-truth skin readings collected using a device designed to measure human skin. FALMs estimated from different images of the same individual varied significantly relative to ground-truth FALM. This variation was only reduced by greater control of acquisition (camera, background, and environment). Next, we compare ground-truth FALM to Fitzpatrick Skin Types (FST) categories obtained using the standard, in-person, medical survey and show FST is poorly predictive of skin-tone. Finally, we show how noisy estimation of FALM leads to errors selecting explanatory factors for demographic differentials. These results demonstrate that measures of skin-tone for biometric performance evaluations must come from objective, characterized, and controlled sources. Further, despite this being a currently practiced approach, estimating FST categories and FALMs from uncontrolled imagery does not provide an appropriate measure of skin-tone.

John J. Howard, Yevgeniy B. Sirotin, Jerry L. Tipton et al.

Human face features can be used to determine individual identity as well as demographic information like gender and race. However, the extent to which black-box commercial face recognition algorithms (CFRAs) use gender and race features to determine identity is poorly understood despite increasing deployments by government and industry. In this study, we quantified the degree to which gender and race features influenced face recognition similarity scores between different people, i.e. non-mated scores. We ran this study using five different CFRAs and a sample of 333 diverse test subjects. As a control, we compared the behavior of these non-mated distributions to a commercial iris recognition algorithm (CIRA). Confirming prior work, all CFRAs produced higher similarity scores for people of the same gender and race, an effect known as "broad homogeneity". No such effect was observed for the CIRA. Next, we applied principal components analysis (PCA) to similarity score matrices. We show that some principal components (PCs) of CFRAs cluster people by gender and race, but the majority do not. Demographic clustering in the PCs accounted for only 10 % of the total CFRA score variance. No clustering was observed for the CIRA. This demonstrates that, although CFRAs use some gender and race features to establish identity, most features utilized by current CFRAs are unrelated to gender and race, similar to the iris texture patterns utilized by the CIRA. Finally, reconstruction of similarity score matrices using only PCs that showed no demographic clustering reduced broad homogeneity effects, but also decreased the separation between mated and non-mated scores. This suggests it's possible for CFRAs to operate on features unrelated to gender and race, albeit with somewhat lower recognition accuracy, but that this is not the current commercial practice.