Liveness detection in biometrics: face, fingerprint, iris and voice

This page belongs to the deepfake detection hub and widens the face-only view of liveness and injection attacks to the other biometrics. Face has its own data page, four face anti-spoofing datasets compared, and voice has its own, so this one concentrates on fingerprint and iris.

01What does liveness detection mean across biometrics?

ISO/IEC 30107-1:2023 defines presentation attack detection as "automated discrimination between bona-fide presentations and biometric presentation attacks". The definitions speak of a "biometric characteristic" and name no modality, and the list of attack instruments includes "lifeless biometric characteristics" and "altered biometric characteristics". Liveness detection is one family of methods inside that frame.

The companion part, ISO/IEC 30107-3:2023, "establishes principles and methods for performance assessment of PAD mechanisms" and covers the reporting of test results. The page on ISO/IEC 30107-3 explained walks through its terms.

02Which error rates carry across modalities?

Two do. APCER is the share of attack presentations wrongly accepted as genuine, and BPCER is the share of genuine presentations wrongly rejected. The LivDet review maps its older fingerprint terms onto them: "FerrFake" is APCER and "FerrLive" is BPCER. The same pair is used in face papers, and the LivDet-Iris organisers use them for iris. The APCER and BPCER page defines both with the attack-type condition the standard attaches.

Speech is the exception. The ASVspoof challenges report the tandem detection cost function and the equal error rate, because a speaker verification system sits in the loop, as the voice liveness detection page explains.

03How has fingerprint liveness improved?

The LivDet fingerprint competition has run every two years since 2009. The 2009 to 2021 review counts 4 participants and 4 algorithms in 2009 and 13 participants and 23 algorithms in 2021. Each edition uses a different set of scanners, materials and challenges, so the review warns that comparing editions is "non-trivial".

Averaged over the three best algorithms of each edition, BPCER fell from 22.03 percent in 2009 to 1.1 percent in 2021, and APCER fell from 26.97 percent in 2011 to 4.75 in 2019. Then 2021 broke the trend: mean APCER was 27.02 percent. The review explains the jump partly by a new fabrication method, ScreenSpoof, and hypothesises that the 2021 training set holding no glue-like material left glue spoofs in the test set new to every system. Its conclusion is that "the representativeness of the training set has a serious influence on their accuracy".

The review also reports that on the integrated competitions of 2019 and 2021, where liveness runs inside a matching system, algorithms reached an average accuracy of around 90 percent, meaning a typical system "would miss 10% of both presentation attacks and legitimate presentations".

04How hard is iris liveness?

Iris results show the same sensitivity to what the test set contains. The LivDet-Iris series began in 2013, and its 2020 paper tabulates the best performance of each edition.

In 2020 the best performing entry reported a weighted average APCER of 59.10 percent over all five attack types, and the organisers say the paper "serves as the latest evaluation of iris PAD on a large spectrum of presentation attack instruments" (arXiv 2009.00749). The 2023 edition added GAN-generated iris images and compared algorithms with human accuracy. Its best weighted result was an average classification error rate of 37.31 percent, and the paper concludes that "iris PAD is still a challenging problem" (arXiv 2310.04541).

Notice the pattern from 2015 on: BPCER stays low while APCER rises as the attack list grows. A system tuned to reject little that is genuine can still accept many attacks it has not seen.

05Why can scores not be compared across modalities?

Each modality has its own attack species, capture devices and test data. A face system is attacked with prints and replays; a fingerprint system with moulded materials and, since 2021, screen-based spoofs; an iris system with printed eyes, lenses and cadaver eyes. An APCER of 5 percent on one tells you nothing about another.

The LivDet-Iris 2020 paper also notes that its summary rate, ACER, "has been deprecated in ISO/IEC 30107-3:2017" for industry-related evaluations, so the two component rates are the figures to read. Fingerprint and iris results are measured at the sensor, and the full-system tests of NIST FRTE and the face verification metrics measure a different layer: whether the right person is matched, not whether a person is present.