Why Aggregate Accuracy is Inadequate for Evaluating Fairness in Law Enforcement Facial Recognition Systems

Aggregate accuracy masks demographic disparities: Systems can report high overall accuracy while simultaneously producing vastly different error rates across demographic groups—meaning a facial recognition system might work well on average but fail disproportionately for specific populations.

Similar accuracy can hide different fairness problems: Two facial recognition systems with identical overall accuracy scores can have completely different fairness profiles, with one potentially causing more false arrests in certain communities while another misses suspects from other groups at higher rates.

High stakes demand different metrics: In law enforcement, where facial recognition errors can lead to wrongful suspicion or missed identification, relying solely on aggregate accuracy is inadequate—the paper emphasizes that false positive rates (wrongly identifying someone) and false negative rates (missing a suspect) must be examined separately across demographic groups.

Current evaluation practices are insufficient: The paper argues that accuracy-centric evaluation approaches fail to capture real-world operational risks in high-stakes applications, necessitating "fairness-aware evaluation approaches and model-agnostic auditing strategies" that can assess systems after they're deployed.

Comprehensive evaluation frameworks are essential: Moving beyond a single accuracy metric requires adopting more comprehensive evaluation frameworks that examine subgroup-level performance, enabling stakeholders to understand how facial recognition systems actually perform across different populations before and after deployment.