Selfie-based reauthentication fails when teams buy a face matcher and assume it also proves a live person is in front of the camera, or buy liveness and assume it also proves the person is the account holder. Face matching (1:1 face comparison) answers "is this the same person as the enrolled reference?" and is judged on false match and false non-match rates, including how those rates vary across demographic groups. Liveness detection answers "is a real, present human producing this sample?" and splits into presentation attack detection (PAD) against artefacts shown to the camera and injection attack detection against media fed straight into the capture pipeline. This guide is about returning-user journeys, such as step-up, device change and account recovery by selfie, compared against a stored reference; it is for identity, fraud and security architects choosing and tuning these checks. Document onboarding is covered separately in Document Verification vs Liveness Detection.
Key takeaways
- Two errors, two attackers. Face matching fails when it accepts a different person (false match) or rejects the right one (false non-match). Liveness fails when it accepts a photo, mask, replay or injected deepfake. A perfect matcher will happily match a high-quality deepfake of the account holder.
- The threshold is a policy decision. NIST SP 800-63B-4 requires a false match rate of one in 10,000 or better for all demographic groups at a fixed threshold, and recommends a false non-match rate under 5%. Where you set the threshold decides who gets locked out and who gets in.
- Demographic differentials are real and algorithm specific. NIST found false positive rates 10 to 100 times higher for some demographic groups in many 1:1 algorithms, with wide variation between developers. Demand per-group results, not an average.
- PAD testing does not cover injection. ISO/IEC 30107-3:2023 addresses attacks at the capture device only. Injection of synthetic video via virtual cameras or emulators needs separate controls and testing, which CEN/TS 18099:2024 describes.
- A selfie is not an authenticator on its own. SP 800-63B-4 allows biometrics only as part of multi-factor authentication with a physical authenticator. Selfie-only recovery is a re-verification event and should be treated, logged and rate limited as one.
The short answer
You need both checks for any selfie journey, and they are tested, tuned and monitored separately. Face matching is a statistical comparison between a probe image and a reference template; its accuracy is measured under ISO/IEC 19795-1 as false match rate (FMR) and false non-match rate (FNMR) at a chosen threshold. Liveness is a security control against deliberate spoofing; its strength against presentation attacks is measured under ISO/IEC 30107-3 as attack presentation classification error rate (APCER), bona fide presentation classification error rate (BPCER) and, at system level, impostor attack presentation accept rate (IAPAR). The metrics are explained in depth in the Biometric Liveness and Presentation Attack Detection Guide; this page concentrates on how the two halves interact.
| Dimension | Face matching (1:1 comparison) | Liveness detection (PAD and injection detection) |
|---|---|---|
| Question | Same person as the enrolled reference? | Real person present and producing this sample now? |
| Adversary assumed | Zero-effort impostor or lookalike; natural variation of the genuine user | Active attacker with artefacts, replays, masks, deepfakes or injected streams |
| Core metrics | FMR, FNMR (ISO/IEC 19795-1); per-demographic FMR and FNMR (ISO/IEC 19795-10) | APCER, BPCER, IAPAR (ISO/IEC 30107-3); injection attack detection rates (CEN/TS 18099) |
| NIST SP 800-63B-4 bar | FMR one in 10,000 or better for all demographic groups; FNMR under 5% recommended; fixed threshold | PAD required for face; IAPAR under 0.07 recommended; testing per ISO/IEC 30107-3 Clause 13 |
| NIST SP 800-63A-4 bar (proofing) | FMR 1:10,000 or better; FNMR 1:100 or better; demographic results no more than 25% worse than overall | IAPAR under 0.07 for remote collection; Section 3.14 controls for injection and forged media |
| Independent benchmark | NIST FRTE 1:1 Verification (ongoing) | Accredited lab PAD testing; injection testing to CEN/TS 18099 |
| What it cannot see | Whether the sample is live or synthetic | Whether the live person is the account holder |
| Main tuning lever | Match threshold | PAD threshold, capture channel hardening, challenge design |
Face matching: threat model, accuracy and bias
What the matcher is defending against
The matcher's threat model is impersonation without spoofing: someone with physical access to the device, a relative or a lookalike trying their own face, and, at scale, an attacker trying many faces against many accounts. Its other failure is turning away the genuine user because of ageing, lighting, glasses, injury or a poor camera. Both errors move with the threshold: a stricter threshold lowers FMR and raises FNMR. SP 800-63B-4 requires a fixed threshold rather than one tuned per group, and SP 800-63A-4 explains that changing the threshold for each demographic group is not feasible.
Architecture in words
The client captures a probe image; a quality check rejects blurred, occluded or off-angle frames; a template is extracted and compared with the stored reference template; the comparison score is tested against the threshold. The reference is usually the selfie taken at onboarding, sometimes the document portrait. Comparison can run on the device against a locally held template, or on the server against a central template. SP 800-63B-4 prefers local comparison for authentication, and central storage creates a template database that becomes a target in its own right. See Biometric Template Protection.
Accuracy evidence
NIST's FRTE 1:1 Verification evaluation is open indefinitely and ranks algorithms by FNMR at an FMR of 10-6 on its visa-border dataset, with additional results on border, kiosk and twins data and summaries of demographic effects by age, sex and region of birth. Treat these as evidence about an algorithm, not a product: vendors integrate algorithms behind their own capture, quality and threshold choices, so ask which submitted algorithm version is in production and at what threshold.
Demographic differentials
NIST's 2019 demographic effects study (NISTIR 8280) found that, for one-to-one matching, many algorithms produced higher false positive rates for Asian and African American faces than for Caucasian faces, often by factors of 10 to 100 depending on the algorithm, while some algorithms developed in Asian countries showed no dramatic difference between Asian and Caucasian faces. Accuracy varied widely between developers. For a reauthentication journey, a higher false match rate in one group means account takeover is easier for impostors from that group, and a higher false non-match rate means genuine users are pushed into slower, costlier fallback. SP 800-63A-4 sets the disparity limit at no more than 25% worse than overall, across sex, age and skin tone, tested to ISO/IEC 19795-1:2021 and 19795-10:2024.
Liveness detection: presentation versus injection
Presentation attacks
A presentation attack is an artefact shown to a genuine camera: a printed photo, a screen replay, a paper or silicone mask, or a deepfake video played on a second screen. ENISA's remote identity proofing report lists photo, video replay, 3D mask and deepfake attacks against face recognition. PAD analyses texture, depth, reflections, moire and motion. Active liveness asks the user to turn, blink or follow a prompt; passive liveness decides from one or a few frames without instructions. Active challenges add friction and can be answered by real-time face-swap tools; passive checks are smoother but depend entirely on the model's training against current artefacts. ISO/IEC 30107-3:2023 is explicit that it addresses only attacks made at the capture device during presentation.
Injection attacks
An injection attack bypasses the camera: a virtual camera driver, an emulator, a hooked app or an intercepted API call delivers pre-recorded or synthetic frames as if they came from the sensor. Because the PAD model sees a clean, high-quality image of the right face, a lab-certified PAD result says nothing about this route. SP 800-63A-4 Section 3.14 requires technical controls to increase confidence that media comes from a genuine sensor, testing of automated image analysis against available attack artefacts, and notes that a biometric comparison on a captured sample does not prevent these attacks. CEN/TS 18099:2024 defines biometric data injection attacks, the instruments used, detection guidance and test plans. The detection layers are covered in Biometric Injection Attacks.
| Attack | Defeats matcher alone? | Defeats PAD alone? | Primary control |
|---|---|---|---|
| Lookalike or relative presents own face | Sometimes, at a loose threshold | No (they are live) | Matcher threshold and demographic accuracy |
| Printed photo or screen replay of the victim | Yes | Usually caught | PAD |
| 3D mask of the victim | Yes | Depends on PAD level tested | PAD tested at higher attack potential |
| Deepfake of the victim on a second screen | Yes | Often caught as screen replay | PAD plus deepfake artefact analysis |
| Deepfake injected through a virtual camera or emulator | Yes | Not tested by ISO/IEC 30107-3 | Injection attack detection, device and app integrity |
| Real-time face swap answering active challenges | Yes | May pass active challenges | Injection detection, passive artefact analysis, device signals |
Layering with device and session signals
Neither biometric check knows whether the device is new, rooted or an emulator, or whether the session started from a credential-stuffing IP range. For reauthentication, device and session context is often stronger evidence than the face. Useful signals include platform attestation results from the mobile operating system, app integrity checks, detection of virtual camera drivers and emulator artefacts, camera metadata consistency, whether the device is already bound to the account, and the behaviour of the session leading up to the selfie. Bind the capture to the session with a server-issued nonce so a recorded result cannot be replayed into another session. Where a registered device holds a passkey, prefer it: SP 800-63B-4 allows biometrics only as part of multi-factor authentication with a physical authenticator, and requires a non-biometric alternative to be offered. See Biometrics vs Possession Factors.
Decision guidance: which check carries the weight
- Device-bound unlock (platform biometrics): the matcher and PAD run on the device and only release a local key; the server sees a cryptographic assertion, not a face. Put your effort into authenticator assurance rather than face metrics.
- Step-up on a known, bound device: server-side selfie comparison against the enrolled reference with passive PAD is proportionate; injection risk is lower because device integrity is already established.
- Account recovery or new device with selfie only: highest risk. Require injection detection, strict PAD, the strictest matcher threshold your FNMR budget allows, a cooling-off period or notification to existing channels, and a manual review path for failures.
- High-value transactions: combine a passkey or bound device with the selfie, rather than relying on the face alone.
- Small organisations: buy the capability as a service but contract for per-group accuracy figures, PAD and injection test reports, and raw score access. Large organisations: run their own production measurement per demographic segment and per channel.
Trade-off summary: tightening the matcher reduces impostor success but increases lockouts, which are not evenly spread across groups; tightening PAD reduces spoofing but raises BPCER on poor cameras; adding injection detection costs integration effort and occasional false flags on legitimate assistive or virtual camera software. Measure all three together.
Worked scenario: a bank's selfie recovery is abused
A retail bank lets customers recover access on a new phone by taking a selfie that is compared with the onboarding selfie. The supplier reports a strong ISO/IEC 30107-3 result. Fraud losses rise; investigation shows recoveries passing with high match scores from devices that report a virtual camera, at night, shortly after password resets. The matcher worked: the faces were the customers'. PAD worked: there were no artefacts at the camera. The frames were deepfakes built from social media photos and injected. The bank adds injection detection and app attestation to the recovery flow, binds each capture to a session nonce, holds recovered accounts for 24 hours with a notification to the old device and email, and routes virtual camera hits to manual review. It also finds its threshold produced more false non-matches for older customers, and adds an assisted fallback rather than loosening the threshold. See the Deepfake-Resistant Identity Verification Guide.
Vendor evaluation criteria
| Area | Evidence to request | Red flag |
|---|---|---|
| Matcher accuracy | FNMR at stated FMR, at the production threshold, tested to ISO/IEC 19795-1; the algorithm version and any NIST FRTE 1:1 submission | A single "accuracy" percentage with no FMR |
| Demographic performance | Per-group FMR and FNMR by sex, age and skin tone, tested to ISO/IEC 19795-10 | Only population averages |
| PAD | ISO/IEC 30107-3:2023 report from an accredited lab: APCER per species, BPCER, attack potential level, capture devices tested | "Certified" with no report, or a different product version |
| Injection | Test results against virtual cameras, emulators and hooked apps, mapped to CEN/TS 18099 | PAD report offered as the answer to injection |
| Signals and integration | Raw scores, reason codes, device and injection signals returned per transaction; session nonce support | Pass or fail only |
| Data handling | Template storage location, retention, deletion on request, encryption, sub-processors | Images retained for model training by default |
How to implement: phased approach
- Classify journeys: list every selfie use (step-up, device change, recovery) and assign a risk tier.
- Set targets: choose FMR and FNMR targets per tier (use the SP 800-63B-4 bar as a floor for authentication), an IAPAR target, and injection detection requirements.
- Shadow mode: run the new matcher, PAD and injection checks alongside the current flow for several weeks, logging scores without enforcing, and compute per-segment FNMR from genuine users.
- Red team: commission an authorised test with presentation and injection attacks against your own integration, not only the supplier's demo app.
- Enforce by ring: enable for a small share of traffic, then by tier, with a manual review queue and an assisted fallback live before enforcement.
- Roll back: keep the previous thresholds as a configuration you can restore; if FNMR or complaints rise beyond the agreed limit, revert and investigate rather than loosening the threshold silently.
Detection and measurement
- Production FNMR by segment: genuine-user rejections by device class and, where lawful to measure, demographic proxy; compare with the supplier's figures.
- PAD and injection hit rates: rejections by attack type per week; a sudden drop to zero is as suspicious as a spike.
- Post-recovery fraud rate: accounts recovered by selfie that show fraud within 30 days; the most honest measure of combined effectiveness.
- Manual review outcomes: proportion of reviewed failures overturned, which shows whether thresholds are too strict.
- Repeat attempts: consecutive failures per account and device; SP 800-63B-4 caps consecutive failed biometric attempts at five, or ten with PAD.
Log the match score, PAD and injection outcomes, device and attestation signals, reference template version and threshold for every transaction, so investigators can tell which layer failed.
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | PR.AA-02, PR.AA-03, DE.CM-03 | Proofing and binding, authentication, monitoring of activity |
| NIST SP 800-53 Rev 5 | IA-12, IA-5, SI-4, AU-6 | Identity proofing, authenticator management, monitoring, audit review |
| NIST SP 800-63-4 | 800-63B-4 Section 3.2.3; 800-63A-4 Sections 3.11 and 3.14 | Biometric accuracy, PAD, attempt limits; injection and forged media |
| MITRE ATLAS | AML.T0088 Generate Deepfakes; AML.T0015 Evade AI Model | Synthetic face generation; evading the PAD or matching model |
| Testing standards | ISO/IEC 19795-1, 19795-10, 30107-3:2023; CEN/TS 18099:2024 | Matcher accuracy and demographics; PAD; injection detection |
Common failure patterns
- One check bought as both: a matcher with no PAD, or PAD with a loose matcher threshold.
- PAD certificate used to answer injection: ISO/IEC 30107-3 does not test the virtual camera route.
- Threshold loosened to cut complaints: FNMR falls and impostor success rises, often unevenly across groups.
- Average accuracy only: no per-group figures, so demographic differentials surface as complaints or fraud.
- Selfie-only recovery with no cooling-off: the attacker recovers, changes contact details and transacts in one session.
- Scores discarded: pass or fail only, so fraud teams cannot tell which layer failed.
Practitioner checklist
- Map every selfie journey to a risk tier and decide whether a passkey or bound device should carry it instead.
- Set FMR and FNMR targets at a fixed threshold, with per-group FNMR measured in shadow mode.
- Require ISO/IEC 30107-3:2023 PAD reports for the exact product version and capture channels in use.
- Require injection testing mapped to CEN/TS 18099 and enable virtual camera and emulator detection.
- Bind each capture to a session nonce and to device attestation results.
- Add cooling-off and out-of-band notification to selfie-based recovery.
- Provide a non-biometric alternative and an assisted fallback.
- Log scores, signals and thresholds per transaction and review metrics monthly.
Standards and references
- NIST: SP 800-63B-4 Authentication and Authenticator Management
- NIST: SP 800-63A-4 Identity Proofing and Enrollment
- NIST: FRTE 1:1 Verification
- NIST: NISTIR 8280 Face Recognition Vendor Test Part 3: Demographic Effects
- ISO: ISO/IEC 30107-3:2023 Biometric presentation attack detection, Part 3: Testing and reporting
- CEN: CEN/TS 18099:2024 Biometric data injection attack detection
- ENISA: Remote Identity Proofing: Attacks and Countermeasures
- NIST: Study Evaluates Effects of Race, Age, Sex on Face Recognition Software
Related NHI Mgmt Group resources: Biometrics Guide · Biometric Verification vs Identification · Biometric Authentication Platforms Buyer's Guide · Deepfake and AI Impersonation Guide