Demographic Effects in Facial Recognition and Their Dependence on Image Acquisition: An Evaluation of Eleven Commercial Systems

By Cynthia M. Cook, John J. Howard, Yevgeniy B. Sirotin, Jerry L. Tipton, A. Vemury. 2019

125 citations recorded by the source corpus.

We examined the effect of demographic factors on the performance of 11 commercial face biometric systems tested as part of the 2018 U.S. Department of Homeland Security, Science and Technology Directorate biometric technology rally. Each system that participated in this evaluation was tasked with acquiring face images from a diverse population of 363 subjects in a controlled environment. Biometric performance was assessed using a systematic, repeatable test process measuring both efficiency (transaction times), and accuracy (mated similarity scores) using a leading commercial algorithm. Prior works have documented differences in biometric algorithm performance across demographic categories and proposed that skin phenotypes offer a superior explanation for these differences. To test this concept, we developed an automatic method for measuring relative facial skin reflectance using subjects’ enrollment images and quantified the effect of this metric and other demographic covariates on performance using linear modeling. Both the efficiency and accuracy of the tested acquisition systems were significantly affected by multiple demographic covariates, including skin reflectance, gender, age, eyewear, and height. Skin reflectance had the strongest net linear effect on performance. Linear modeling showed that lower (darker) skin reflectance was associated with lower efficiency (higher transaction times) and accuracy (lower mated similarity scores). Skin reflectance was also a statistically better predictor of these effects than self-identified race labels. Unlike other covariates, the degree to which skin reflectance altered accuracy varied between systems. We show that the size of this skin reflectance effect was inversely related to the overall accuracy of the system such that the effect was almost negligible for the system with the highest overall accuracy. These results suggest that, in evaluations of biometric accuracy, the magnitude of measured demographic effects depends on image acquisition.

Subfields: Cognitive Neuroscience, Computer Vision and Pattern Recognition, Evaluation & Benchmarks, AI Ethics, Fairness & Society, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

Brian Dolhansky

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

Albert Gordo

In Signal Processing

Citations in the corpus

Towards Measuring Fairness in AI: The Casual Conversations Dataset (2021, background): “In following, [22] examined eleven commercial face biometrics systems for the effect of demographic factors and found out that both speed and the accuracy of these systems were inﬂuenced by demographic factors.”