1:1 vs 1:N face recognition: verification and identification explained

This page is part of the face recognition section. It gives the two tasks their own names and error rates; the face verification metrics page goes deeper on the 1:1 side, and the hub's page on FMR, FNMR, FPIR and FNIR defines the four terms formally.

01What is 1:1 face recognition?

Verification is a single comparison. A person claims an identity, by presenting a passport, typing a username or unlocking their own phone, and the system compares the live probe face with the one template stored for that identity. The output is a similarity score, and a threshold turns the score into accept or reject. Two errors are possible: a false non-match, where the genuine person is rejected, and a false match, where a different person is accepted. NIST's 1:1 page defines the rates: "FNMR is the proportion of mated comparisons below a threshold set to achieve the false match rate (FMR) specified. FMR is the proportion of impostor comparisons at or above that threshold" (NIST FRTE 1:1).

The word "mated" matters. A mated comparison is one where the two faces truly belong to the same person; an impostor comparison is one where they do not. Because verification is one comparison, the number of people enrolled in the system does not enter the error rate: a bank with a million customers and a phone with one owner face the same two rates per comparison.

02What is 1:N face recognition?

Identification is a search. The probe is compared against every one of N enrolled templates, the scores are sorted, and the system returns the candidates above a threshold, or a fixed number of top candidates, or both. The question is no longer "is this the claimed person?" but "who, if anyone, in this gallery is this?". NIST's 1:N page defines its rates: "FNIR is the proportion of mated searches failing to return the mate above threshold. FPIR is the proportion of non-mated searches producing one or more candidates above threshold" (NIST FRTE 1:N).

A mated search is one where the person is in the gallery; a non-mated search is one where they are not. The second case is where identification differs most from verification. A non-mated search produces a false positive if any of the N comparisons clears the threshold, so the chance of a false alarm grows with the gallery, and a threshold that gives a tolerable FMR on one comparison can give an intolerable FPIR across a million. The 1:N page records that in its tables "each person is enrolled with one image only" and that the gallery size is printed in each column header, because the number is meaningless without it.

03How does NIST report 1:1 results?

The FRTE 1:1 track runs continuously. Developers may submit "no more frequently than four calendar months", and the participation table, last updated 6 November 2026, counts 1,501 algorithms from 461 developers since 2017. The main table reports FNMR for each developer's most accurate algorithm on several datasets, with the algorithms ordered by their rank on the visa and border dataset "calculated using FNMR @ FMR = 10 -6": one false match in a million impostor comparisons. The threshold is chosen per algorithm to hit that FMR, and the FNMR is read off at it. A second table summarises demographic variation in FMR, with the threshold "fixed for each algorithm to give FMR if 0.0003 overall".

So a 1:1 result is always a pair: an FNMR and the FMR at which it was measured. Quoting the FNMR alone would let an algorithm look accurate by accepting everyone, and the NIST FRTE page explains how the track's datasets and thresholds are chosen so that no algorithm can do that.

04How does NIST report 1:N results?

The FRTE 1:N track, whose participation table (last updated 6 October 2026) counts 713 algorithms from 224 developers since 2018, publishes two kinds of table. The identification table reports FNIR "computed at the threshold that limits the False Positive Identification Rate (FPIR) to 0.003", with "the threshold set for each algorithm and each column separately". The page explains the choice: thresholding "supports use of face recognition in making mostly automated decisions e.g. for access into aircraft", where a false alarm costs an interruption and the system must bound how often it happens.

The investigation table drops the threshold. It reports FNIR "for the case where the threshold is set to zero and the algorithm returns a fixed number (50) of candidates", and FNIR becomes "the proportion of mated searches for which the algorithm does not place the correct candidate at rank 1". This mode "supports investigate uses where it is assumed and necessary that a human will be used to review the candidates", so a false positive is not counted as an error at all: the reviewer is expected to reject the wrong candidates. The two tables measure the same algorithms and disagree on what an error is, which is why a 1:N figure must say which mode it came from.

Source: NIST FRTE 1:1 and 1:N pages, read 9 November 2026.

05Why the two tasks are not one task with a loop

A 1:1 algorithm can implement a 1:N search by running N comparisons and sorting the scores, and NIST's 1:1 page says so: "if the 1:1 algorithm were used to implement 1:N search (via N-comparisons and a sort operation), the demographic effects noted here would be germane to the 1:N application". But the page adds that "some operational 1:N algorithms do employ 1:1 algorithms, and results for those algorithms must be measured in separate 1:N tests". Search systems use indexing, candidate pruning and score normalisation across the gallery that a single comparison never sees, and the false positive behaviour of a search cannot be read off a verification table.

The demographic findings follow the same split. NISTIR 8280, the FRTE demographic effects report of December 2019, uses FMR and FNMR for its 1:1 analysis and FPIR and FNIR for 1:N, and found false positive differentials that "often vary by factors of 10 to beyond 100 times" across demographic groups in 1:1 comparisons while false negatives varied by smaller factors (NISTIR 8280). In a gallery search, a group with a higher per-comparison false match rate meets that rate N times over. The hub's page on demographic differentials covers the report's findings in full.

06Where these rates meet the rest of the pipeline

Both branches start from the same front end, described on the face recognition system components page: detect the face, align it, and compute a template. ISO/IEC 19795-1:2021, the biometric performance testing standard, "establishes general principles for testing the performance of biometrics systems in terms of error rates and throughput rates", and excludes from its scope "people deliberately trying to subvert the intended operation of the biometric system (e.g. by presentation attacks)". Whether the face presented was a live person is measured separately, with the error rates described on the APCER, BPCER and ACER page, and whether a video of that face was manipulated is a third question with a third vocabulary, where test design is set by the deepfake detection protocol rather than by a fixed false match rate.

The habit of reporting an error rate with the condition that fixes it is the common thread. NIST fixes FMR or FPIR and reports the miss rate; a deepfake benchmark fixes the test set and reports AUC; a human study fixes the chance level, as the human detection studies table shows, and reports accuracy. In each case the second number is what makes the first one comparable.