Biometric Performance Metrics¶
Definition¶
Standardized metrics for measuring the accuracy of biometric systems. These metrics are essential for comparing models, setting thresholds, and meeting certification requirements.
Face Recognition Metrics¶
| Metric | Full Name | What It Measures | Ideal |
|---|---|---|---|
| FAR / FMR | False Accept / Match Rate | % of impostors incorrectly accepted | Lower is better |
| FRR / FNMR | False Reject / Non-Match Rate | % of genuine users incorrectly rejected | Lower is better |
| EER | Equal Error Rate | Point where FAR = FRR | Lower is better |
| TAR@FAR | True Accept Rate at fixed FAR | Acceptance rate at security operating point | Higher is better |
| d' | Decidability | Separation between genuine and impostor score distributions | Higher is better |
ROC and DET Curves¶
- ROC (Receiver Operating Characteristic): Plots TAR vs FAR — area under curve (AUC) measures overall performance
- DET (Detection Error Tradeoff): Plots FRR vs FAR on normal deviate scale — standard in biometrics (ISO 19795)
Liveness / PAD Metrics (ISO 30107-3)¶
| Metric | Full Name | What It Measures | Ideal |
|---|---|---|---|
| APCER | Attack Presentation Classification Error Rate | % of attacks that fool the system | Lower is better |
| BPCER | Bona Fide Presentation Classification Error Rate | % of real users incorrectly rejected as spoof | Lower is better |
| ACER | Average Classification Error Rate | (APCER + BPCER) / 2 | Lower is better |
APCER Calculation¶
APCER is calculated per PAI species (per attack type):
APCER_PAI = (Number of attack presentations classified as bona fide) / (Total attack presentations of that PAI species)
The overall APCER is typically the maximum APCER across all PAI species (worst-case attack type).
End-to-End eKYC Metrics¶
| Metric | What It Measures | Typical Target |
|---|---|---|
| STP Rate | Straight-Through Processing — % auto-approved | > 80% |
| First-attempt success | Users passing on first try | > 75% |
| Completion rate | Users who finish the full flow | > 85% |
| Average verification time | End-to-end duration | < 60 seconds |
| Manual review rate | % requiring human review | < 15% |
Key Takeaways¶
Summary
- FAR and FRR are the fundamental recognition metrics — always in tension (threshold tradeoff)
- EER is the single-number model quality summary — lower is better
- APCER, BPCER, ACER are the ISO 30107-3 standard for liveness/PAD evaluation
- APCER should be reported per PAI species — overall APCER uses the worst-case species
- End-to-end metrics (STP rate, completion rate) matter as much as model accuracy in production