Face recognition definition and 25 terms, taken from NIST and the ISO standards

This page is the vocabulary entry of the face recognition section. The four error rates page goes deeper on FMR, FNMR, FPIR and FNIR, and the face verification metrics explainer shows how they are plotted.

01What is face recognition?

NISTIR 8280 separates face recognition from the other things a face algorithm can do. Estimation algorithms output a continuous quantity such as age; classification algorithms output a category such as an emotional state and are "built with inherent knowledge of the classes they aim to produce". Face recognition algorithms "have no built-in notion of a particular person"; they "include a face detector followed by a feature extraction algorithm that converts one or more images of a person into a vector of values that relate to the identity of the person" (NISTIR 8280, section 3).

Recognition is then "a differential operator": two feature vectors are compared and the algorithm "emit[s] a similarity score", a "vendor-defined numeric value expressing how similar the parent faces are", which "is compared to a threshold value to decide whether two samples are from, or represent, the same person or not". Everything in the table follows from those four steps.

02Where do the definitions come from?

Three kinds of source are used, and the table says which for every row. NIST's FRTE pages define the four headline error rates in one sentence each. NISTIR 8280 defines the pipeline, the score types and the search terms in its section 3, and states that "The ISO/IEC 19795-1 performance testing and reporting standard requires different metrics to be reported for identification and verification implementations". ISO/IEC 19795-1:2021, Edition 2 of May 2021, "specifies performance metrics for biometric systems" and sets principles for "testing the performance of biometrics systems in terms of error rates and throughput rates" (ISO catalogue); its full text is paid, so no definition here is quoted from it directly.

The presentation attack terms come from the ISO/IEC 30107 series. ISO/IEC 30107-3:2023 covers "principles and methods for the performance assessment of presentation attack detection (PAD) mechanisms", and the attacks it considers "take place at the biometric capture device during presentation" (ISO catalogue). The wording of the five PAD definitions is taken from the ISO/IEC JTC 1/SC 37 working group slides that quote Parts 1 and 3 (Busch, 2024), since those texts are also paid; the APCER and BPCER page uses the same source.

03The 25 terms

Each row gives the term, a one-sentence definition in the source's words where quoted, and the source. Terms 1 to 20 are recognition and testing terms; 21 to 25 are presentation attack terms.

04Why do verification and identification need separate terms?

Because the errors are counted on different events. In verification one comparison happens, so FMR and FNMR are proportions of impostor and mated comparisons. In identification one search produces many comparisons and a candidate list, so FPIR and FNIR are proportions of non-mated and mated searches. NISTIR 8280 notes that identification algorithms may run "an exhaustive search of all N enrollments and a sort operation" or fast-search techniques that skip many comparisons. The 1:1 versus 1:N recognition page works through the arithmetic.

The FRTE 1:N table fixes its threshold so that FPIR is 0.003 and reports FNIR at that point; the 1:1 table fixes FMR and reports FNMR. NISTIR 8280 explains the insistence on fixed thresholds: "the vast majority of biometric systems are configured with a fixed threshold against which all comparisons are made", so reporting false negatives at a fixed false positive rate can hide excursions in the false positive rate between groups.

05Which terms carry over to deepfake detection?

Fewer than the shared pipeline suggests. A deepfake detector such as UCF outputs one score for one image with no second face, so it has no genuine or impostor scores, no gallery and no FMR; it is judged by AUC, accuracy or equal error rate over real and fake samples. The terms that do carry over are the presentation attack ones: a swapped face shown to a camera is a presentation attack with a PAI, and a liveness check is a PAD mechanism scored by APCER and BPCER. A swap such as FaceShifter also borrows from the recognition side, since its identity encoder is a pretrained recognition model and its identity loss is a similarity score.

The three routes by which a fake reaches a system are set out on the deepfake attack vectors page; only the first of them is a presentation attack in the ISO/IEC 30107 sense.

06How NIST and ISO fit together

NIST's evaluations apply the standard's requirement to report verification and identification separately; the FRTE 1:1 page lists 1,501 algorithms submitted between 2017 and 2026. ISO/IEC 19795-1 excludes attacks from its scope: "Not within the scope of this document is the measurement of error and throughput rates for people deliberately trying to subvert the intended operation of the biometric system (e.g. by presentation attacks)". Those belong to ISO/IEC 30107-3. The NIST FRTE explainer describes the reports where terms 15 to 19 are published.