Demographic differentials

When a biometric works less well for some groups than others, based on traits like skin tone, age, or sex.

For example, a face check might wrongly reject genuine customers from one group more often than from another, even though the same system serves everyone.

Measured variation in error rates across demographic cohorts such as skin tone, age, and sex. A cohort is a group of people who share a trait, and error rates are the false acceptance and false rejection rates.

Testers look for differentials in both kinds of error:

• False rejection differences, where genuine people in some groups are turned away more often.
• False acceptance differences, where impostors are wrongly accepted more often for some groups.

Measuring them takes a large, labeled dataset that covers each group well. Independent evaluations, such as the National Institute of Standards and Technology (NIST) demographic effects study, publish results by algorithm and by group.

Demographic differentials sit next to terms that are easy to mix up:

• NIST FRVT is the benchmark program that measures and reports them.
• False rejection rate is one of the error rates that can differ across groups.
• Biometrics is the broader field the differentials apply to.

For more, see Facial Recognition: How to Train Models and Keep Improving Over Time.