A liveness check is only as good as the test evidence behind it, and most buyers accept a single headline number, such as "certified" or "99% accurate", that says nothing about which attacks were tested or where the threshold was set. A liveness check, or presentation attack detection (PAD), decides whether the biometric sample in front of the sensor comes from a live person who is present, rather than a photo, screen replay, mask or other artefact. This guide explains the ISO/IEC 30107-3 metrics (APCER, BPCER and IAPAR), how to read a PAD test report, how to set thresholds for your journey, why injection attacks need separate evidence, and what to measure in production. It is for identity architects, fraud and authentication teams and anyone evaluating face, fingerprint or iris liveness. The wider threat picture, including deepfakes and remote proofing standards, is in the Deepfake-Resistant Identity Verification Guide; biometric basics are in the Biometrics Guide.
Key takeaways
- Two error rates, always together. APCER is the share of attacks wrongly accepted as genuine; BPCER is the share of genuine users wrongly rejected as attacks. One without the other, or without the threshold, is meaningless.
- Results are per attack type. NIST's FATE PAD evaluation of 82 algorithms found accuracy "varies widely across algorithms, use cases ... and PA types", so an average hides the attack that gets through.
- IAPAR measures the whole system. NIST SP 800-63B-4 says deployment testing should show an impostor attack presentation accept rate below 0.07, tested to Clause 13 of ISO/IEC 30107-3, and makes PAD mandatory for facial recognition.
- Injection is a different test. ISO/IEC 30107 covers attacks at the sensor; injected video from virtual cameras or emulators is covered by CEN/TS 18099, and NIST's PAD evaluation left it out of scope.
- Measure in production too. Track rejection rates, retry patterns, injection signals and confirmed fraud by journey, because lab results describe the attacks the lab chose.
What PAD and liveness actually decide
ISO/IEC 30107 uses the term presentation attack: a presentation to the biometric capture subsystem intended to interfere with the system's operation. The artefact used is the presentation attack instrument (PAI), and a family of artefacts made the same way, such as printed photos or silicone masks, is a PAI species. PAD is the automated decision that a presentation is an attack. "Liveness" is the common market name for the same function, whether it is passive (a single image or short video, no user action) or active (the user turns their head, follows a dot or reads digits).
There are two goals an attacker can pursue, and NIST's FATE PAD report evaluated them separately:
- Impersonation: being accepted as someone else, for example opening or taking over an account. This is the main concern in authentication and onboarding.
- Evasion: avoiding recognition, for example to slip past a watchlist or create a second account. This matters for duplicate detection and screening.
NIST found impersonation detection generally performed better than evasion detection. If your use case includes deduplication, ask for evasion results specifically.
The ISO/IEC 30107-3 metrics
ISO/IEC 30107-3 is the testing and reporting part of the series, now in its 2023 second edition. It defines how to measure PAD; it does not prescribe how PAD should work. The metrics fall into two groups: those that judge the PAD subsystem on its own, and those that judge the full biometric system including comparison.
| Metric | What it measures | Plain meaning | Who cares most |
|---|---|---|---|
| APCER (attack presentation classification error rate) | Proportion of attack presentations of one PAI species classified as bona fide | Attacks that PAD let through | Fraud and security teams |
| BPCER (bona fide presentation classification error rate) | Proportion of genuine presentations classified as attacks | Real users wrongly blocked by PAD | Product, customer experience, accessibility |
| APNRR and BPNRR (non-response rates) | Presentations the system failed to process at all, for attacks and genuine users | "Try again" outcomes; hide errors if excluded | Evaluators reading the fine print |
| IAPAR (impostor attack presentation accept rate) | Proportion of impostor attack presentations accepted by the full system, PAD plus comparison | Attacks that succeeded end to end | Regulators and assurance frameworks |
Three details change how you read these numbers:
- APCER is per species. The standard computes APCER for each PAI species, and the figure that matters is the worst one. A vendor reporting "APCER 0.5%" averaged over ten species can still accept a specific mask type most of the time.
- Failures to process must be counted. NIST's FATE PAD method treats a failed processing as a detected attack, scored +1, so failures cannot flatter the results. Check how a lab report handled them.
- APCER and BPCER trade off. Both come from one threshold on a score. NIST reported two operating points: APCER at BPCER = 0.01 (convenience first: one genuine user in a hundred rejected) and BPCER at APCER = 0.01 (security first: one attack in a hundred accepted).
Worked example: comparing two vendor reports
The figures below are illustrative, not results from any product. They show why the per-species view and the operating point decide the comparison.
| Result in report | Vendor A | Vendor B |
|---|---|---|
| Headline "APCER" (average across species) | 1.0% | 2.0% |
| Worst single species APCER | 12% (silicone mask) | 3% (screen replay) |
| BPCER at the reported threshold | 8% | 2% |
| Non-responses excluded from the rates | Yes, 4% of attacks | No, counted as rejections |
| Injection testing | Not tested | Separate report, virtual camera and emulator scenarios |
Vendor A looks better on the headline but accepts one attack type at a high rate, rejects many more genuine users, and flatters its figures by excluding non-responses. For an onboarding journey, Vendor B is the stronger evidence. Ask Vendor A for results at your intended threshold with non-responses included before going further.
Reading a PAD test report
A conformance letter or "Level 2 compliant" badge is a summary, not evidence. Ask for the full report and check each item below before you rely on it.
| Check | What good looks like | Red flag |
|---|---|---|
| Standard and edition | ISO/IEC 30107-3:2023 named, with the clause used (system-level testing for IAPAR) | "Tested against ISO 30107" with no part, edition or clause |
| Target of evaluation | Exact product version, SDK, capture device or app, and server components | A different version from the one you will deploy |
| PAI species list | Each species described (print, screen replay, paper mask, resin or silicone mask) with numbers of instruments and subjects | Species counted but not described |
| Per-species results | APCER per species and the maximum across species; IAPAR where the full system was tested | Only an average or a pass/fail |
| Bona fide testing | BPCER with the number of genuine subjects, conditions and demographics | No BPCER, or a few dozen staff volunteers |
| Threshold | The operating threshold stated and the same one you will configure | Results at a tuned threshold not available in production |
| Attacker effort | Time, skill and cost budget for creating instruments described | No description of how hard the testers tried |
| Injection | Separate injection attack testing (for example to CEN/TS 18099) or an explicit statement it was out of scope | Silence, implying PAD results cover injection |
| Date and independence | Recent, from an accredited independent laboratory | Self-test by the vendor, or several releases old |
Certification schemes help by fixing some of these choices. The FIDO Alliance biometric component certification sets requirements and test procedures, including presentation attack detection, in its Biometric Certification Requirements (v5.0 active, v4.1 active until 31 March 2027 when checked). Use the scheme's published levels to compare products, but still read the report for the species tested.
Setting thresholds and requirements for your journey
There is no single right threshold. NIST's report notes that the appropriate threshold depends on whether a system prioritises convenience or security. Decide per journey:
| Journey | Priority | Typical approach |
|---|---|---|
| Remote onboarding with document check (account opening, KYC) | Security: an accepted attack creates a synthetic or stolen account | Strict PAD threshold, injection detection, manual review on borderline scores, IAPAR evidence below 0.07, the SP 800-63B-4 deployment target, with the proofing requirements in the linked deepfake guide |
| Account recovery and high-value step-up | Security | Strict threshold plus a second, independent signal (device binding, known phone, in-person fallback) |
| Routine sign-in on a bound device | Convenience, because the device is already a factor | Moderate threshold; local comparison on the device as SP 800-63B-4 prefers; retry limits |
| Physical access or kiosk | Depends on site risk | Controlled capture hardware reduces injection risk; test the species relevant to unattended use (masks) |
SP 800-63B-4 (final, July 2025) adds requirements that shape the design whatever the threshold:
- Biometrics only as part of MFA: biometrics SHALL be used only with a physical authenticator, never as a standalone factor, and a non-biometric alternative SHALL always be offered.
- PAD by modality: PAD SHALL be implemented for facial recognition and SHOULD be for iris and fingerprint; voice-based comparison SHALL NOT be used.
- Retry limits: no more than five consecutive failed attempts, or ten where PAD meets the requirements, then a delay of at least 30 seconds per attempt and an overall limit of 50, or 100 with PAD.
- Accuracy floor: a false match rate of 1 in 10,000 or better across demographic groups at a fixed threshold.
For how these translate into authenticator choices, see the Phishing-Resistant MFA Guide.
Injection attacks: why PAD results do not cover them
A presentation attack happens in front of the camera. An injection attack bypasses the camera: the attacker feeds a pre-recorded or synthesised video into the capture pipeline through a virtual camera driver, an emulator, a hooked or modified app, or by tampering with the upload to the server. A deepfake face that would fail in front of a real sensor can pass PAD perfectly when injected, because PAD never sees an artefact.
The evidence is separate too. NIST's FATE PAD evaluation stated that digital injection attacks were out of scope, while noting they are an important concern on personal devices with virtual cameras. CEN/TS 18099:2024, "Biometric data injection attack detection" (valid from 16 December 2024), covers injection attack instruments, detection, risk mitigation and test plans, and explicitly leaves presentation attack testing to ISO/IEC 30107. MITRE ATLAS lists AML.T0088 Generate Deepfakes and AML.T0015 Evade AI Model as techniques against systems like these.
- Ask for both reports: a PAD report to ISO/IEC 30107-3 and an injection attack detection report, ideally to CEN/TS 18099, for the same product version.
- Control the capture path: app attestation, detection of emulators, rooted or jailbroken devices and virtual camera drivers, signed and timestamped capture sessions, and server-issued challenges that bind the capture to the session.
- Prefer native capture over browser capture for high-assurance journeys, because the browser gives less control over the camera source.
- Test injection yourself as part of authorised testing: include virtual camera and emulator scenarios in your pen test scope, with the vendor's agreement.
Detection and measurement in production
Lab results describe a fixed set of attacks on a fixed day. Production telemetry tells you whether attackers have found something the lab did not test.
- Log: PAD score and decision, injection signals (virtual camera, emulator, integrity failures), device and app attestation result, retries per session and per device, and the final outcome of each journey.
- Review: sample borderline scores and all injection detections; feed confirmed fraud back to the vendor as new attack material.
- Alert: clusters of near-threshold scores from one device family, sudden drops in PAD rejection (attackers succeeding quietly) and spikes in retries from one network.
| Metric | Why it matters |
|---|---|
| Production PAD rejection rate by journey and platform | A proxy for BPCER plus attack volume; a sudden change needs explanation |
| Genuine-user rejection confirmed by support or manual review | Real-world BPCER; high values push users to weaker fallbacks |
| Injection detections per 10,000 sessions | Shows whether attackers are targeting the capture path |
| Confirmed fraud accounts that passed liveness | Real-world IAPAR signal; each one is a test case for the vendor |
| Retry and abandonment rates by demographic segment where lawful | Fairness and accessibility of the threshold |
| Age of the PAD and IAD test reports for the deployed version | Evidence currency; re-test on major releases |
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST SP 800-63B-4 | Section 3.2.3 Use of Biometrics | PAD requirement, IAPAR target, retry limits, biometrics only with a physical authenticator |
| NIST SP 800-53 Rev. 5 | IA-5(12) Biometric Authentication Performance; IA-2; SA-11; CA-8; SI-4 | Set and test biometric performance; developer testing; penetration testing including injection; monitoring |
| NIST CSF 2.0 | PR.AA-02, PR.AA-03 | Identity proofing and authentication |
| ISO/IEC 30107-3:2023 | PAD testing and reporting | APCER, BPCER, IAPAR and report content |
| CEN/TS 18099:2024 | Biometric data injection attack detection | Injection testing and mitigation |
| MITRE ATLAS | AML.T0088, AML.T0015 | Deepfake generation and model evasion against liveness models |
Common failure patterns
- Buying a badge: accepting a conformance letter without the report, then discovering the tested species did not include masks or that the tested version is three releases old.
- Averaged APCER: comparing vendors on a mean error rate that hides one species with a high acceptance rate.
- Treating PAD as injection defence: deploying browser-based capture for onboarding with strong PAD results and no control over virtual cameras, the gap deepfake fraud exploits. See the Deepfake and AI Impersonation Guide.
- Threshold drift: loosening the threshold after customer complaints without re-measuring APCER at the new setting, so the test evidence no longer describes production.
- Weak fallback: users rejected by liveness are routed to knowledge questions or an SMS code, which attackers then target instead. The fallback must meet the same assurance.
- Biometrics as the only factor: using a face match alone for account recovery, against SP 800-63B-4, which allows biometrics only as part of MFA with a physical authenticator.
Practitioner checklist
- Require the full ISO/IEC 30107-3:2023 report for the exact version you will deploy, with per-species APCER, BPCER and IAPAR.
- Require a separate injection attack detection report, or a written statement that injection was not tested.
- Configure the threshold the report used, or re-test at your threshold.
- Set PAD thresholds per journey, strictest for onboarding and recovery.
- Control the capture path with app attestation and virtual camera and emulator detection.
- Apply SP 800-63B-4 retry limits and offer a non-biometric alternative of equal assurance.
- Log PAD scores, injection signals and outcomes, and review borderline and fraud cases monthly.
- Re-test on major releases and at least annually, and include injection scenarios in authorised pen tests.
Standards and references
- NIST: SP 800-63B-4 Digital Identity Guidelines: Authentication and Authenticator Management
- NIST: IR 8491 FATE Part 10, Performance of Passive, Software-Based Presentation Attack Detection Algorithms
- NIST: Face Analysis Technology Evaluation (FATE) PAD
- CEN: CEN/TS 18099:2024 Biometric data injection attack detection
- FIDO Alliance: Biometric Component Certification
- ENISA: Remote Identity Proofing, Attacks and Countermeasures
- MITRE ATLAS: Generate Deepfakes
Related NHI Mgmt Group resources: Biometrics Guide · Deepfake-Resistant Identity Verification Guide · Identity Verification Buyer's Guide · Illinois BIPA and Biometric Privacy Guide