Atlas — Biometric Authentication, Face Recognition & the AI Layer

A working reference for the bio-authN project. Domain and tech vocabulary first, because you can only be sharp about eval, token economy, infra, and governance once the substrate is precise. Definitions are written for a practitioner, not a newcomer — tight, with the "why it matters" attached where it changes a decision.

Conventions: Term (abbreviation) — definition. Where a term is contested or evolving, the disagreement is named rather than smoothed over. Standards and regulatory facts current as of June 2026.

1. Identity & access fundamentals

The scaffolding biometrics plug into. Biometric auth is never the whole system — it's one factor inside an identity stack.

Identity Provider (IdP) — the system that holds identities and issues assertions about them ("this is user X, authenticated at level Y"). Biometric matching feeds a signal into the IdP; it rarely is the IdP.

Authentication (AuthN) — proving you are who you claim to be. Biometric verification is an AuthN method.

Authorization (AuthZ) — deciding what an authenticated identity is allowed to do. Distinct from AuthN; a face match answers "who," not "what you can touch."

Factor — a category of authentication evidence: something you know (password), something you have (device, passkey), something you are (biometric). Two different categories make a multi-factor.

Multi-Factor Authentication (MFA) — two or more factors from different categories. Two passwords is not MFA; a passkey plus a face is.

Strong Customer Authentication (SCA) — the PSD2 requirement (EU/UK) for two independent factors from different categories on payment and account access, with dynamic linking to the transaction. The open question for any selfie-auth product: does the face plus the device-bound key count as two independent factors, or is the device doing double duty?

Step-up authentication — demanding additional/stronger proof at a high-risk moment (large transfer, credential change, viewing a full account number) rather than on every login. The natural home for selfie-auth: high assurance when it's worth the friction.

Single Sign-On (SSO) — one authentication grants access to many services. Context, not a biometric concept, but it's where the assertion travels.

OAuth 2.0 — a delegation/authorization framework: issues access tokens so an app can act on a resource without holding the user's credentials. It is authorization, frequently misdescribed as login.

OpenID Connect (OIDC) — an authentication layer on top of OAuth 2.0. Adds an ID token (a signed JWT) asserting who the user is. This is the "login with…" pattern.

JSON Web Token (JWT) — a signed, base64 token carrying claims. The container an OIDC assertion rides in.

FIDO2 / WebAuthn — a standard for phishing-resistant, public-key authentication. The device holds a private key; it signs a server challenge; the private key never leaves the device. A biometric (Face ID, fingerprint) unlocks the key locally — the biometric is not transmitted. Proves "you possess and control this device."

Passkey — the consumer-facing name for a FIDO2/WebAuthn credential, usually syncable across a user's devices via a platform keychain. Critical distinction from selfie-auth: a passkey proves device possession; selfie-auth proves the person. They are complementary — passkey for daily login, biometric for step-up and recovery when the device itself is the thing in question.

Device binding — cryptographically tying a credential or session to a specific device (keypair in secure hardware). Underpins the "something you have" factor and most fraud controls.

Level of Assurance (LoA) — graded confidence in an identity claim (e.g. eIDAS Low / Substantial / High; NIST IAL/AAL). Regulators and frameworks speak in LoAs; a biometric method earns or loses LoA based on its PAD strength and binding.

2. Biometric core concepts

Biometric modality — the trait measured: face, fingerprint, iris, voice, palm vein, gait, keystroke dynamics. This project is face.

Behavioural biometric — identity inferred from how you act (typing rhythm, swipe pressure, gait) rather than a static trait. Continuous and passive, lower assurance, used for risk scoring rather than as a hard gate.

Enrollment (enrolment) — the first capture that creates the stored reference. Quality here bounds every later match. In ZKB systems, enrollment must happen through the device SDK or the security property is lost.

Verification (1:1) — "are you who you claim to be?" One probe compared against one stored reference. Authentication is 1:1.

Identification (1:N) — "who are you, among everyone enrolled?" One probe searched against the whole gallery. Watchlists and dedup are 1:N; it is a harder problem and error compounds with N.

Probe — the freshly captured sample presented at auth time.

Gallery / reference — the stored enrolled sample(s) the probe is matched against.

Template — the numerical representation of a biometric extracted by a model. For face, a vector of floats (the embedding). Not the image. See §4.

Genuine vs impostor — a genuine pair is two samples from the same person; an impostor pair is from two different people. All accuracy metrics are built from the score distributions of these two populations.

Matching score / distance — the similarity (or distance) between probe and reference. A threshold turns the score into accept/reject. Everything in §5 is about choosing that threshold well.

3. Face recognition technology

Face detection — locating faces in an image (bounding box) before anything else. Models: MTCNN, RetinaFace, BlazeFace, MediaPipe Face Detection.

Facial landmarks — keypoints on the face (eye corners, nose tip, mouth). 5-point for alignment; 68-point (dlib) or 468-point 3D mesh (MediaPipe Face Mesh) for geometry, pose, and liveness.

Face alignment — warping the detected face to a canonical pose/scale before embedding, so the model sees a consistent input. Skipping it degrades match accuracy.

Convolutional Neural Network (CNN) — the deep network architecture that turns aligned pixels into an embedding. The "proprietary model" every biometric vendor trains is, at core, a CNN (increasingly with transformer components).

Embedding — the output vector (typically 128–2048 dimensions; ArcFace is 512) encoding learned facial features. Two embeddings of the same face sit close together in vector space; different faces sit far apart. Vendor embeddings are not interchangeable — a Mitek vector is meaningless to Onfido.

Embedding space — the high-dimensional space (512 axes for ArcFace) the embeddings live in. The model is trained so that identity becomes geometry: every image is a point, all images of one person form a tight cluster, and different people occupy separate regions. Matching is then just measuring the distance (or angle) between two points, and the threshold is a distance cut-off. Errors concentrate where clusters sit close or overlap — that's where false accepts and the uncertain band live.

Dimensionality reduction (PCA / UMAP / t-SNE) — techniques that project the 512-D embeddings down to 2-D so the cluster structure can be drawn and inspected. PCA is linear and fast (keeps the directions of greatest variance); UMAP and t-SNE are non-linear and better at preserving local clusters. Visualisation only — matching always uses the full-dimensional vectors, never the 2-D projection.

ArcFace — a face-recognition model/loss (Additive Angular Margin) that produces highly discriminative 512-d embeddings. The open, production-grade matcher this project treats as ground truth. Reachable via DeepFace.

FaceNet — earlier influential embedding model (Google, 2015) using triplet loss. Historical reference point; ArcFace generally outperforms it.

InsightFace / Buffalo_L — an open-source face-analysis toolkit; Buffalo_L is a strong packaged model pack. The SOTA-open comparison point against ArcFace.

DeepFace — a Python wrapper exposing several backends (ArcFace, FaceNet, VGG-Face, etc.) behind one API. Convenience layer for the matching engine.

Cosine similarity / cosine distance — the usual way to compare face embeddings: the angle between two vectors. Similarity near 1 = same person; distance = 1 − similarity. Thresholds (e.g. ArcFace ~0.68 cosine distance) are tuned, not universal.

Euclidean / L2 distance — straight-line distance between embeddings; an alternative to cosine. Some models are trained for one metric specifically — using the wrong one quietly degrades accuracy.

Vision-Language Model (VLM) — a model that takes images and text and reasons in natural language (Claude Vision, GPT-4o). In a biometric pipeline it can describe why two faces differ ("lighting and ~30° yaw difference") but is a weak, expensive, non-deterministic decider. The research path exists to map exactly where VLMs help and where they don't.

4. Liveness & presentation attack detection

The hard part. Matching two faces is solved; proving the face in front of the camera is a live human, right now, is where systems are actually attacked.

Presentation Attack (PA) — any attempt to defeat the sensor with a fake: printed photo, screen replay, video, 3D mask, deepfake.

Presentation Attack Detection (PAD) — liveness detection. The countermeasure family. "Liveness" and "PAD" are used interchangeably; PAD is the standards term.

Presentation Attack Instrument (PAI) — the specific artefact used in an attack (the print, the mask, the replay device). PAD is benchmarked per PAI species.

Active liveness — the user performs a randomized challenge (blink, turn head, smile, follow a dot). Defeats static photo and simple replay because a recording can't answer a random prompt. Adds friction and a few seconds of latency. This project's primary defence.

Passive liveness — liveness inferred from a single capture without asking the user to do anything (texture, micro-reflections, depth cues, moiré). Lower friction, harder to build well. Best systems fuse active + passive.

Eye Aspect Ratio (EAR) — ratio computed from eye landmarks; drops sharply (~0.25 → <0.1) on a blink. The cheap, reliable signal behind in-browser blink detection.

Head pose / yaw / pitch / roll — 3D orientation of the head derived from landmark geometry. Yaw crossing ±15° confirms a "turn your head" challenge.

Injection attack — bypassing the camera entirely to feed a synthetic stream straight into the app (virtual camera, compromised SDK). PAD that only inspects pixels won't catch it — this is why capture integrity (signed capture, e.g. Mitek's encryptedPayload) matters separately from liveness.

Replay attack — presenting previously captured legitimate video/audio. Defeated by randomized active challenges, not by static checks.

Deepfake / face swap — AI-generated or face-swapped video of the target. The fastest-moving threat; defeats naive liveness and is the reason "good enough for a prototype" liveness is explicitly not production-grade.

3D mask attack — a sculpted/printed mask with real depth. Defeats 2D and many depth-based checks; the high end of presentation attacks.

ISO/IEC 30107-3 — the international standard for testing PAD. Defines the metrics below and the certification target.

iBeta Level 1 / Level 2 — independent PAD conformance testing against ISO/IEC 30107-3. Level 1 covers low-effort artefacts (print, screen); Level 2 covers high-effort, custom-made PAIs (masks, sophisticated replays). "iBeta Level 2 certified" is the credibility bar for banking liveness.

PAD error metrics (the PAD analogue of FAR/FRR):

Metric	Full name	Meaning
APCER	Attack Presentation Classification Error Rate	Share of attacks wrongly accepted as live. The security number.
BPCER	Bona fide Presentation Classification Error Rate	Share of genuine users wrongly rejected as attacks. The friction number.
ACER	Average Classification Error Rate	Mean of APCER and BPCER. A single headline figure (deprecated in newer revisions but still quoted).

5. Accuracy & evaluation metrics

The spine of this project. If you can't measure a matcher, you can't govern it or price it.

False Accept Rate (FAR) / False Match Rate (FMR) — how often an impostor is accepted. The security failure. FMR is the matcher-only term; FAR includes system effects. Lower is more secure.

False Reject Rate (FRR) / False Non-Match Rate (FNMR) — how often a genuine user is rejected. The usability failure. Every rejected legitimate customer is a support call or an abandonment.

The FAR/FRR trade-off — they move in opposite directions as you slide the threshold. Tighten to cut impostors (lower FAR) and you reject more genuine users (higher FRR). There is no free setting; you pick where on the curve the business sits. Naming this trade-off explicitly is the entire job.

Equal Error Rate (EER) — the threshold where FAR = FRR. A single comparison number between matchers, but rarely the operating point — banks run far tighter than EER on the security side.

Failure to Enrol (FTE) — share of users who can't produce a usable enrollment at all (poor camera, lighting, disability). An inclusion metric; high FTE is an exclusion problem.

Failure to Acquire (FTA) — share of attempts where no usable sample is captured at auth time. Friction and accessibility signal.

Receiver Operating Characteristic (ROC) — curve of true-accept rate against false-accept rate across all thresholds. The standard way to see a matcher's whole behaviour rather than one operating point.

Detection Error Tradeoff (DET) — FRR vs FAR plotted on a normal-deviate scale; the biometrics community's preferred curve because it spreads out the low-error region where decisions actually happen.

Area Under Curve (AUC) — single scalar summarising ROC. Useful for ranking, blind to where on the curve you operate.

TAR @ FAR — True Accept Rate measured at a fixed FAR (e.g. TAR @ FAR=1e-4). How vendors and NIST actually report: "at one-in-ten-thousand false accepts, we let in 98.5% of genuine users." The honest way to state accuracy.

Threshold tuning — choosing the score cutoff for a target FAR on representative data. The hardcoded 0.68 in early notes is a starting guess; sweeping the threshold on a corpus and reading the ROC is the actual method.

Demographic differential / biometric bias — systematic accuracy variation across skin tone, age, sex. A matcher can post a great aggregate FRR and still reject one demographic at several times the rate. This is the metric that gets a bank fined and the reason aggregate numbers lie. The bias panel is a first-class deliverable, not a footnote.

Calibration — whether a model's stated confidence matches its real accuracy (a "0.9 confident" set should be right ~90% of the time). VLMs are often badly calibrated — confidently wrong — which is precisely why their scores can't be trusted as a gate without measuring this.

NIST FRTE / FATE — the US government's ongoing, independent face benchmarks, run by NIST. In August 2023 the long-running FRVT was split and renamed: FRTE (Face Recognition Technology Evaluation) covers verification/identification accuracy; FATE (Face Analysis Technology Evaluation) covers PAD, morph detection, quality, and age estimation. The neutral reference point for "how good is good," and the public source for demographic-differential data.

6. Template protection & privacy architectures

Why biometrics are not passwords: a stolen password is reset; a stolen face is permanent. This section is the security-design core of the project.

Biometric Template Protection (BTP) — techniques that store the template such that a breach doesn't expose the underlying biometric. Standardised in ISO/IEC 24745.

Irreversibility — the property that the stored template can't be turned back into the original biometric. Not absolute for raw embeddings — see model inversion.

Unlinkability — the property that templates of the same person across systems can't be correlated. The defence against cross-bank replay.

Renewability / cancelable biometrics — the ability to revoke a compromised template and issue a fresh, different one from the same face (typically by changing a transform/salt). Restores the "reset" property biometrics otherwise lack.

Salting (per-tenant transform) — applying a non-invertible, tenant-specific transform so the same face yields different templates per bank. Without it, one vendor model shared across banks means a stolen template from Bank A is mathematically comparable at Bank B. The Onfido cross-bank question turns on whether this is applied.

Model inversion attack — reconstructing a usable synthetic face from a stolen embedding by running gradient descent against the known model. The reason "embeddings lose information, so they're safe" is only half true.

GAN reconstruction — training a generative model to map embeddings back to face images. A second, often stronger, route from template to a replayable face.

Cross-bank replay — reusing a stolen template against a different institution that runs the same vendor model. Database isolation doesn't prevent it; only per-tenant template transformation does.

Secure Multi-Party Computation (sMPC) — a cryptographic method letting parties jointly compute a function over inputs none of them fully holds. The basis of split-template biometrics.

Zero-Knowledge Biometrics (ZKB) — a system (e.g. Keyless) where neither device nor server ever holds the complete template. Device holds a mask R; server holds F − R; the match computes distance with R cancelling out, so the server learns the result without seeing the face. Breach of either side alone yields nothing usable, and templates become cancelable by re-masking. Hard dependency: the device SDK must take part in every auth — a backend-API "send the selfie to the server" integration is not ZKB, it's plaintext server-side matching wearing the same brand.

Homomorphic Encryption (HE) — computing directly on encrypted data. An alternative privacy route to sMPC for biometric matching; generally heavier on compute/latency.

Fuzzy vault / fuzzy commitment — classic BTP schemes that bind a secret key to noisy biometric data so a close-enough sample releases the key without storing the biometric in the clear. Older lineage than sMPC/HE; useful conceptual grounding.

Special category data — under UK/EU GDPR, biometric data used for identification is a special category requiring an explicit lawful basis and heightened safeguards. The permanence problem made into a legal obligation.

7. Standards & certifications

Main entries:

ISO/IEC 2382-37 — biometrics vocabulary standard; the authority on precise term definitions when two teams or vendors mean different things by "verification."

ISO/IEC 30107-3 — the PAD testing standard; defines APCER, BPCER, and ACER and sets the methodology any liveness claim must cite.

iBeta Level 1 / Level 2 — NIST/NVLAP-accredited lab for commercial PAD testing against 30107-3; Level 2 (high-effort artefacts, masks, sophisticated replays) is the banking liveness credibility bar.

ISO/IEC 24745 — biometric information protection; sets the three required template properties: irreversibility, unlinkability, and renewability. The gap between "encryption at rest" and actual 24745 compliance is where model-inversion and cross-bank-replay risk lives.

ISO/IEC 19794 / 39794 — biometric data interchange formats; 39794 is the extensible ICAO replacement for ePassports; governs how a face record is structured when data crosses vendor or system boundaries.

NIST FRTE / FATE — US government's independent face benchmarks (renamed from FRVT in August 2023); FRTE covers verification/identification accuracy, FATE covers PAD and bias data — the public yardstick vendors' own marketing won't volunteer.

FIDO2 / WebAuthn — the credential standard behind passkeys; the biometric only unlocks a local key and is never transmitted. Proves device possession, not personhood — complementary to selfie-auth, not a substitute.

8. Regulation (UK/EU focus)

Main entries:

UK GDPR + Data Protection Act 2018 — biometric data used for identification is special-category data; requires explicit lawful basis, a DPIA, and retention/erasure architecture. The sharp edge: what happens to an auth template seeded from an IDV selfie when the source is deleted?

Information Commissioner's Office (ICO) — UK data-protection regulator; has issued specific biometric guidance and enforced against biometric deployments; the body a DPIA is written to satisfy.

PSD2 / Strong Customer Authentication (SCA) — mandates two independent authentication factors from different categories with dynamic transaction linking; the open question is whether selfie-auth plus a device-bound key counts as two independent factors or just one.

Financial Conduct Authority (FCA) — UK financial regulator; any SCA interpretation — what qualifies as a factor, what binding is acceptable — must be validated here before treating selfie-auth as a regulated gate.

UK DIATF / DVS Trust Framework — the UK government scheme certifying digital-identity providers; Gamma (0.4) entered force July 2025, formalised as DVS v1.0 in March 2026; the certification route for reusable digital identity in the UK.

eIDAS 2.0 — EU Regulation 2024/1183, in force May 2024; mandates member-state EUDI Wallets by December 2026, with regulated sectors (banking included) required to accept them by late 2027.

EUDI Wallet — member-state wallet holding verifiable credentials at LoA High; parts of the auth/onboarding flow you build today may need to interoperate with it within a couple of years.

BIPA — Illinois Biometric Information Privacy Act; US-scoped but set the global industry tone on explicit biometric consent and statutory damages; a reference point even for UK/EU product decisions.

9. IDV & onboarding

The front door biometrics often enter through.

Identity Document Verification (IDV) — proving a presented identity document is genuine and belongs to the presenter. The regulatory gate at onboarding; if it fails, onboarding stops.

Know Your Customer / Anti-Money-Laundering (KYC / AML) — the regulatory obligations IDV exists to satisfy.

Document liveness / document PAD — detecting fake or copied documents (screen-of-a-passport, printed copy), analogous to face liveness but for the document.

NFC chip read — reading the cryptographically signed chip in an ePassport/eID for the highest-assurance document check. Defeats most document forgery because the data is signed by the issuing authority.

Face-on-document match — comparing the live selfie to the photo extracted from the ID document. The bridge between IDV and selfie-auth: the onboarding selfie can seed the ongoing auth template.

Optical Character Recognition (OCR) / data extraction* — pulling text fields from the document image. Increasingly where VLMs are tried.

IDV-seeded enrollment — using the onboarding IDV selfie to create the auth reference template, removing a separate enrollment step. Lowest friction, but raises the GDPR question of what happens to the auth template when the source IDV selfie is erased.

10. The AI layer — eval, governance, token economy, infra

The project's reason for being. These are the four dimensions a platform PM owns once the matching substrate is understood.

Eval

LLM-as-judge — using a language/vision model to score or adjudicate. Powerful and cheap to stand up, but needs its own validation (calibration, agreement with ground truth) before its verdicts can be trusted — otherwise you've added a confident, unmeasured oracle to a regulated decision.

Confidence vs calibration — a model's stated confidence is not its accuracy until you've measured calibration (§5). Treat raw VLM confidence as a claim to be tested, not a number to be believed.

Hallucination — a model producing fluent, plausible, wrong output. In a matcher, a hallucinated "these are the same person, high confidence" is a security event.

Uncertain band / second-opinion routing — the region of matcher scores (e.g. ArcFace cosine ~0.55–0.75) where the primary matcher is least reliable. The defensible place to spend a VLM call: route only ambiguous cases to a slower, costlier reasoner, not every transaction. This is where eval and token economy meet.

Drift — degradation of model accuracy as real-world inputs diverge from training data (new phones, camera changes, demographic shifts). The reason eval is continuous, not a one-time gate.

Ground-truth corpus — the labelled set of genuine/impostor pairs every metric is computed against. For this project: synthetic, demographically labelled faces for eval, plus ephemeral live self-capture for the interactive flow.

Governance

Model card — a standardised description of a model: intended use, training data character, measured accuracy by demographic, known limits. The artefact a reviewer or regulator reads before trusting a matcher.

Audit log — an immutable, per-decision record (which matchers fired, scores, threshold applied, outcome, consent state). The difference between "the system decided" and "we can show why the system decided."

Explainability — surfacing, per decision, why an accept/reject happened in terms a human reviewer can act on. For VLMs this is their genuine strength (natural-language reasoning); for embedding matchers it's a distance and a threshold.

Evidence pack — the per-decision bundle a reviewer can open: inputs, every matcher's output, the policy applied, consent and retention state, the model cards in force. Governance rendered as something you can actually click into.

Policy-as-config — thresholds, routing rules, and fallbacks held as version-controlled configuration rather than buried in code. Lets risk/compliance change the operating point without a deploy, and lets the audit log reference exactly which policy version ran.

Consent & retention model — the recorded lawful basis for capturing the biometric and the rules for how long each artefact lives and when it's deleted. A governance object, not a checkbox.

Data lineage — tracking where each piece of biometric data came from, what it seeded, and what must cascade on erasure. The thing that answers "the IDV selfie was deleted — what happens to the auth template?"

Token economy

Token — the unit a VLM/LLM bills and computes in. An image consumes input tokens (scaled by resolution); the model's reasoning consumes output tokens.

Token accounting — measuring input + output tokens per call, per decision. The raw material of unit cost. The project instruments this rather than quoting a flat "~$0.01/call."

Cost-per-verification — the fully loaded cost of one auth decision across whatever matchers fired. The number that decides whether a VLM second opinion is affordable at scale.

Unit economics of assurance — the core economic question the architecture poses: extra assurance costs tokens and latency, so where is it worth paying? The honest answer is selective (uncertain-band routing), not "run everything every time."

Cost/latency/accuracy frontier — the three-way trade. ArcFace is fast/free/deterministic; a VLM is slow/costly/explanatory. There is no matcher that wins all three; the design picks per-decision.

Infrastructure

Provider abstraction — a common interface behind every matcher (ArcFace, VLM-via-API, local VLM, cloud baseline) so they can be swapped, compared, or fallen back to without rewiring. The thing that makes "compare techniques" tractable.

Parallel fan-out — running multiple matchers concurrently on one probe and combining results, rather than serially. Keeps latency near the slowest matcher instead of the sum.

Score fusion — combining multiple matchers' outputs into one decision (weighted, rule-based, or learned). How a multi-matcher or multi-modal system reaches a single accept/reject.

Latency budget — the total time the user will tolerate, allocated across capture, liveness, match, and any second opinion. Active liveness and VLM calls are the expensive line items; the budget forces the routing discipline.

Resilience: active-passive vs intelligent routing — for a regulated gate (IDV) you need a second vendor ready: active-passive keeps a hot standby; intelligent routing sends traffic to whichever vendor is healthy/cheaper/more accurate per request. Selfie-auth tolerates a softer fallback (OTP/passkey); IDV does not.

Fallback — the defined path when the primary biometric fails or is unavailable (OTP, passkey, manual review). Its strength is set by how regulated the use case is.

11. Market & vendors (reference)

Named for orientation; not endorsements. Architecture matters more than brand — the same brand can ship a ZKB integration or a plaintext one (see §6).

Vendor / product	What it is	Notable architectural point
Mitek — MiPass	Selfie-to-auth, bank-focused	Template-based, server-side matching
Mitek — MiSnap	Capture SDK (distinct from MiPass)	`encryptedPayload` proves capture integrity, not template privacy
Onfido (Entrust)	IDV + biometric auth, strong UK banking presence	Single model across clients → the cross-bank replay question turns on per-tenant salting
Keyless (Ping Identity)	Zero-Knowledge Biometrics via sMPC	ZKB property holds only with device-SDK enrollment; backend-API mode loses it
iProov	Specialist liveness/PAD	"Flashmark" challenge; the dedicated-liveness tier prototypes are measured against
FaceTec	Specialist 3D liveness	3D face mapping; common high-assurance liveness benchmark
AWS Rekognition	Cloud face compare/detect	Cloud production baseline; cheap per call, no on-prem control
Jumio / Persona / Idemia / NEC	IDV / biometric suites	Reference points for breadth of the IDV+auth market

12. Quick disambiguations (the ones people get wrong)

AuthN ≠ AuthZ. Who you are vs what you may do.
FAR/FMR ≠ APCER. FAR is impostor-accepted-as-genuine (matching); APCER is attack-accepted-as-live (liveness). Different failures.
Passkey ≠ selfie-auth. Device possession vs personhood. Complementary, not substitutes.
Capture integrity ≠ template privacy. Signed capture (anti-injection) says nothing about how the template is stored.
"Keyless/ZKB integration" ≠ ZKB. Without device-SDK enrollment there is no mask, so it's plaintext server-side matching.
Aggregate accuracy ≠ fair accuracy. A strong overall FRR can hide a demographic rejected several times more often.
VLM confidence ≠ VLM accuracy. Not until calibration is measured.
EER ≠ operating point. Banks run far tighter than equal-error on the security side.