False rejection and false match are a threshold tradeoff, not independent metrics

What it is

The biometric threshold tradeoff arises because a similarity score must cross a boundary for acceptance, and moving that boundary usually changes false match and false non-match rates together. Relaxing it admits more genuine samples and more similar impostor samples; tightening it does the opposite. The two measures must be interpreted as a pair at one threshold and deployment condition.

Why it happens

Genuine and impostor score distributions overlap, so one threshold turns continuous uncertainty into accept or reject. A team can misleadingly select the most attractive false-match and false-non-match numbers from different thresholds. Averages can also hide costs concentrated by device or population. Quality gates, retries, and fusion across attempts change transaction-level error beyond a single-frame algorithm result.

Studying it

Collect genuine and impostor comparisons in a predefined target scenario, plot detection-error tradeoffs across candidate thresholds, and report both errors, confidence intervals, and strata at the selected operating point. Separate one comparison from the transaction after retries, including capture failure and fallback. Choose the threshold using harm from false acceptance, harm from rejection, and available remedy—not an abstract maximum-accuracy score.

Where it stops holding

The tradeoff does not mean every improvement merely exchanges errors: a better sensor, feature, or capture instruction can improve both. Presentation attacks produce distributions unlike zero-effort impostors and require separate testing; multiple modalities and risk signals can use more complex decisions. Rates from different datasets, base rates, or operating conditions are not directly comparable.

Applying it

• Record the fixed threshold, quality rule, retry count, false-match rate, and false-non-match rate for each risk scenario rather than “accuracy” alone.
• Select the operating point from the consequences of takeover and legitimate lockout, with explicit ownership of risk approval.
• Version thresholds and recalibrate after model, sensor, or environment changes instead of retaining stale historical metrics.
• Show paired measures, stratified differences, and final transaction outcomes together; prohibit marketing numbers spliced across thresholds.

Related

• Same group: O3.03.1 Probabilistic matching · O3.03.3 Modality exclusion · O3.03.7 Presentation attacks
• Adjacent: O3.18 Security–usability tradeoffs
• Search terms: biometric decision threshold · false match false non-match · DET curve