Overview
ISO/IEC 5152:2024 specifies statistical methodologies to estimate biometric false match rates (FMRs) when only small non-mated sample sets are available. The standard applies extreme value theory (EVT) to extrapolate the tail of similarity or likelihood score distributions so evaluators can produce an extrapolated FMR and confidence interval even when observed false matches are rare or absent. The document covers methodology, data recording, reporting, and metrics for estimated biometric performance, while excluding one-to-many false positive identification rate estimation and false accept rate estimation for verification transactions.
Key Topics
- Extreme value statistical models: The standard introduces two EVT-based approaches - the generalized extreme value (rGEV) model (for r largest order statistics) and the generalized Pareto (GP) distribution (for tail modeling). These enable reliable extrapolation of score distributions beyond the empirical range.
- Estimation design and confidence: Sample design and choice of thresholds are guided by the target FMR and desired confidence interval. The standard explains trade-offs between sample size, confidence level and extrapolation uncertainty (for example, industry rules like the “rule of 30” illustrate sample requirements when using purely empirical methods).
- Model fitness and diagnostics: Evaluators are required to assess model fit using diagnostic plots (e.g., Q–Q plots) and model selection procedures to validate extrapolation results.
- Stratified analysis and demographic factors: Where subpopulations (kinship, demographics, health, occupation, etc.) materially affect extreme scores, the standard recommends stratified evaluation and reporting of sub-dataset extrapolations.
- Record keeping and reporting: Procedures for recording comparison scores, reporting one-to-one performance, and communicating extrapolated FMRs with confidence intervals are defined to ensure reproducibility and transparency.
Applications
This standard is practical for:
- Technology evaluations of biometric verification algorithms when collecting very large non-mated datasets is impractical.
- Scenario and operational evaluations where comparison scores are available but false match events are rare.
- Risk assessments and procurement specifications that require quantified estimates of rare false match behavior and associated confidence bounds.
Benefits include reduced data collection burden, well-founded rare-event probability estimates, and standardized reporting to facilitate comparison between systems.
Related Standards
- ISO/IEC 19795-1:2021 - Principles and framework for biometric performance testing and reporting. This is a normative reference for test design and reporting principles.
- ISO/IEC 2382-37 - Biometrics vocabulary and definitions used throughout biometric standards.
Keywords: biometric, false match rate, FMR, extreme value theory, generalized Pareto, generalized extreme value, extrapolation, small sample estimation, model diagnostics, reporting.