False Accept Rate and False Reject Rate
False Accept Rate is how often a biometric system matches two different people; False Reject Rate is how often it fails to match the same person twice. Both move with the matching threshold, in opposite directions, so a system is tuned to the cost of each error rather than to one correct number.
How does the matching threshold trade FAR against FRR?
A face matcher outputs a similarity score, not a decision. The threshold is where the operator draws the line between match and no match. Raising it means fewer strangers are accepted and more legitimate users are turned away; lowering it does the reverse. No threshold reduces both errors at once, which is why a system cannot be described by one number.
| Threshold | False accepts | False rejects | Suited to |
|---|---|---|---|
| Loose | More | Fewer | Convenience unlocking, where a retry costs nothing |
| Balanced | Moderate | Moderate | Staff access control with a fallback credential |
| Strict | Fewer | More | High-security areas, where a wrong entry is unacceptable |
Why is a single accuracy percentage not meaningful?
An accuracy figure quoted without its threshold, dataset, gallery size, image quality and demographic breakdown cannot be reproduced or compared. The same algorithm can be described as far more or far less accurate by changing any one of those, and none of them are visible in the headline number.
- Threshold: the setting that decides the trade, and the easiest one to omit.
- Dataset: mugshot-quality portraits and CCTV frames produce different results from the same system.
- Gallery size: a 1:N error rate that holds at a thousand enrolments may not hold at a million.
- Demographics: error rates vary across demographic groups, which an aggregate figure hides.
- Capture conditions: pose, lighting and occlusion change results more than most algorithm differences do.
Ayonix publishes no accuracy percentage for its own products for exactly this reason, and recommends a pilot on the buyer's own cameras as the only figure that describes their deployment.
What other error metrics appear in biometric reports?
- FMR and FNMR: false match and non-match rates, measured per comparison rather than per transaction, which is what NIST reports.
- FPIR and FNIR: the identification equivalents, false positive and false negative identification rate for a 1:N search.
- Equal Error Rate: the point where FAR and FRR are equal, useful for comparing algorithms and useless as an operating point.
- Failure to Enrol: how often a person cannot be registered at all, which no matching metric captures.
Reports from NIST use FMR and FNMR because they measure comparisons, which is the algorithm's behaviour; a deployment measures transactions, which include retries, capture failures and operator decisions. Both are legitimate, and quoting one as the other is the most common way a benchmark figure stops describing the system a buyer will run. NIST does not endorse vendors and issues no certification, so a report is evidence about an algorithm, not a recommendation.
Frequently asked questions
- What do FAR and FRR mean in biometrics?
- False Accept Rate is how often a system matches two different people. False Reject Rate is how often it fails to match the same person. They trade against each other through the matching threshold: tightening the threshold lowers false accepts and raises false rejects, so there is no single correct setting independent of what an error costs.
- What is a good FAR for access control?
- It depends on what a wrong entry costs and what a rejection costs. A staff door with a card fallback tolerates a looser setting than a server room. The useful question is not which number is good but which error the site can absorb, because the threshold moves both and cannot lower both.
- What is the Equal Error Rate?
- The Equal Error Rate is the threshold at which false accepts and false rejects occur equally often. It is a convenient single number for comparing algorithms in a paper, and a poor operating point for a real system, because almost no deployment values the two errors equally.
- How do FAR and FRR relate to FMR and FNMR?
- FMR and FNMR measure individual comparisons; FAR and FRR measure transactions, which may involve several comparisons, retries and a capture step. NIST reports use FMR and FNMR because they characterise the algorithm. A deployment experiences FAR and FRR, which is why the two sets of numbers do not transfer directly.
- Why does Ayonix not publish an accuracy figure?
- Because a percentage without its threshold, dataset, gallery size and demographic breakdown cannot be reproduced, and a buyer cannot act on it. Ayonix points to independent evaluations for algorithm comparison and to a pilot on the buyer's own cameras for the figure that describes their site.
- Do error rates differ between demographic groups?
- Independent evaluations have repeatedly found that they do, and that the size of the difference varies substantially between algorithms. An aggregate accuracy figure averages this away, which is why demographic breakdowns are reported separately and why they belong in any procurement review.