Glossary
The vocabulary, defined precisely.
Face biometrics runs on terms that vendors use loosely and buyers pay for misunderstanding. Fifteen of them, defined the way the standards — and the error math — actually use them.
Liveness detection
Determining that the face presented to a camera belongs to a live human who is physically present at capture time — not a photo, a screen, a mask or a recording. Liveness comes in three shapes: passive (a single frame, no user action), active (randomized challenges like blinking or head turns) and continuous (scored across a whole video stream). Each stops a different attack tier at a different friction cost; the trade-offs are the subject of our engineer’s guide.
Presentation attack
An attack presented to the biometric sensor itself: a printed photograph, a face displayed on a phone or monitor, a replayed video, a 2D or 3D mask. The defining property is that a real camera captures the artifact. Presentation attacks are the classic spoofing threat that liveness detection was built for, and the ISO/IEC 30107 family of standards defines how defences against them are tested and reported. Contrast with an injection attack, which bypasses the camera entirely.
PAD (presentation attack detection)
The automated mechanism that decides whether a biometric capture is a bona fide presentation or a presentation attack. PAD systems are evaluated on two paired error rates: APCER, the proportion of attacks wrongly accepted, and BPCER, the proportion of genuine users wrongly rejected. Quoting one without the other is meaningless — a PAD system can trivially drive either number to zero by sacrificing the other. Our evaluation methodology for both is documented on the benchmarks page.
Injection attack
An attack that bypasses the camera instead of fooling it: a virtual camera device, a tampered client or an intercepted capture path feeds pre-recorded or synthetic frames directly into the verification flow. Because the injected frames can be genuinely live-looking footage, classical PAD is largely blind to them. Defences are capture-integrity measures — replay hashing, capture-metadata forensics, signed capture payloads, hardware attestation — layered under the liveness check. Injection detection has its own emerging test standard, CEN/TS 18099.
Deepfake
Synthetic media in which a face is generated, swapped or reenacted by a machine-learning model. In identity verification, a deepfake is a payload rather than a delivery mechanism: it reaches the system either as a presentation attack (played on a screen in front of the camera) or as an injection attack (fed in as a virtual camera stream). Defences therefore combine screen-replay and spectral forensics on the image itself with capture-integrity signals on the submission path.
FMR (false-match rate)
The proportion of impostor comparisons — two different people — that a face matcher wrongly scores as the same person. FMR is the security failure rate: every false match is potentially an account opened in someone else’s name. Meaningful FMR claims require enormous impostor sets; observing zero false matches in a few thousand pairs only bounds the true rate, it doesn’t make it zero. Always ask for the dataset size behind an FMR figure.
FNMR (false-non-match rate)
The proportion of genuine comparisons — the same person, twice — that a face matcher wrongly scores as different people. FNMR is the friction failure rate: it is your real customers being rejected, retrying, and abandoning onboarding. FMR and FNMR trade off against each other through the match threshold, which is why an accuracy claim quoting only one of them — or worse, a single “99%” — tells you almost nothing about behaviour in production.
EER (equal-error rate)
The operating point at which FMR and FNMR are equal — the crossover of the two error curves as the match threshold sweeps. EER compresses a system’s accuracy into one comparable number, which makes it useful for benchmarking engines against each other on the same dataset. It is not an operating recommendation: production systems deliberately run away from the crossover, trading a little more friction for much stronger security, or vice versa, depending on the use case.
Face match threshold
The similarity cut-off above which two faces count as the same person. Face embeddings are compared with cosine similarity, and the threshold slides the FMR/FNMR trade-off: raise it and you reject more impostors but also more genuine users. Thresholds should be calibrated on measured data, not intuition — our own first real-data evaluation moved the shipped threshold substantially (v0.2.10 in the changelog) because the old default bought zero security for a large friction cost.
1:N search
Comparing one probe face against a gallery of N enrolled faces, as opposed to 1:1 verification against a single claimed identity. Used for duplicate detection (has this face enrolled before, under another name?) and blocklist screening. The error math changes with scale: each search performs N impostor comparisons, so a false-match rate that is negligible at 1:1 can produce constant false alarms against a large gallery. Per-hit confidence scores and human review matter accordingly.
V-CIP
Video-based Customer Identification Process — the RBI’s framework, added to its Master Direction on KYC in January 2020, allowing regulated entities to complete customer identification over a live, real-time video interaction conducted by a trained official. For the liveness engine, V-CIP implies properties a single-frame check cannot provide: liveness sustained across the whole session, identity continuity, in-stream replay detection and randomized challenges. We wrote up the engineering translation separately.
eKYC
Electronic know-your-customer: completing identity verification digitally instead of with in-person document checks. In India this spans OTP and offline Aadhaar-based verification, PAN validation, DigiLocker documents and V-CIP video identification, under the RBI’s Master Direction on KYC and the DPDP Act’s data obligations. A typical eKYC flow chains several machine checks — document validation, face match against the document portrait, liveness, sanctions screening — which is why verification APIs are usually consumed as an orchestrated workflow rather than single calls.
Biometric template
The stored mathematical representation of a biometric trait — for faces, typically an embedding vector (ours are 512-dimensional) rather than the image itself. Templates are the crown-jewel asset of a biometric system: unlike passwords, a compromised face cannot be changed. Revocable (cancelable) template schemes address this by storing only a non-invertible, keyed transform of the embedding, so a leaked template can be revoked and re-issued under a new key — the approach described by ISO/IEC 24745.
ISO/IEC 30107-3
The international standard governing how presentation attack detection is tested and reported, defining the APCER/BPCER metrics and test levels. Accredited laboratories such as iBeta test products against it. Precision matters in claims: a lab issues a report that a product was tested and found conformant to ISO/IEC 30107-3 at a stated level — ISO itself certifies nothing, so “ISO certified liveness” is a claim that should end the conversation. Our own testing status is stated plainly on the trust page.
C2PA
The Coalition for Content Provenance and Authenticity — an open standard for cryptographically signed provenance metadata (“Content Credentials”) attached to media at capture or edit time. A valid C2PA manifest lets a verifier check where an image came from and what was done to it. In identity verification it is an emerging capture-integrity signal: presence of an intact, camera-signed manifest is evidence for genuine capture, and manifest absence or breakage is one more input to injection-risk scoring — not proof by itself.
Missing a term you had to explain to a vendor? Tell us and we’ll add it.