H33›Blog›Biometric Enrollment Best Practices: Setting Up for Success
Biometrics February 24, 2026 · 40 min read

Biometric Enrollment Best Practices:
Setting Up for Success

Enrollment is the most consequential step in any biometric system. A poor enrollment produces a poor template, which guarantees poor matching—forever. This guide covers every modality, every failure mode, and the architectural decision that eliminates the plaintext template problem entirely.

3–5x
Captures/Enroll
NFIQ 1–2
Quality Target
0
Plaintext Templates
99.97%
Match Accuracy

Every biometric authentication system follows the same fundamental pipeline: enroll, store, match. Of these three stages, enrollment receives the least attention and causes the most problems. Engineering teams spend months optimizing matching algorithms and template storage security, then lose all of that effort to a sloppy enrollment process that captures low-quality samples, generates weak templates, and ships them into production where they silently degrade every authentication attempt for the lifetime of the deployment.

This is not hypothetical. NIST's Fingerprint Vendor Technology Evaluation (FpVTE) has consistently shown that enrollment quality is the single largest predictor of operational accuracy. A system achieving 0.001% FAR and 0.1% FRR in laboratory testing can degrade to 0.01% FAR and 5% FRR in field deployment—a 50x increase in false rejections—when enrollment conditions are uncontrolled. The matching algorithm did not change. The templates did.

This guide covers the complete enrollment lifecycle: what enrollment actually is at a technical level, how quality impacts matching metrics, modality-specific best practices for fingerprint, face, iris, and voice, how to encrypt templates at enrollment time using Fully Homomorphic Encryption so plaintext biometrics never persist, multi-modal enrollment strategies, UX design that reduces friction without sacrificing quality, re-enrollment policies, regulatory compliance, common failure modes, enterprise deployment considerations, and the H33 API enrollment flow.

Table of Contents

  1. What Is Biometric Enrollment?
  2. Enrollment Quality and Matching Accuracy
  3. Fingerprint Enrollment Best Practices
  4. Face Enrollment Best Practices
  5. Iris Enrollment Best Practices
  6. Voice Enrollment Best Practices
  7. Template Generation and Storage Security
  8. FHE-Based Enrollment
  9. Multi-Modal Enrollment Strategies
  10. Enrollment UX: Friction vs. Quality
  11. Re-Enrollment and Template Update Policies
  12. Compliance Considerations
  13. Common Enrollment Failures
  14. Enterprise Deployment Considerations
  15. H33 API Enrollment Flow

1. What Is Biometric Enrollment?

Biometric enrollment is the process of capturing a person's biometric sample, extracting distinguishing features from that sample, generating a mathematical template, and storing that template for future authentication comparisons. It is the foundational step that every subsequent authentication depends on. If enrollment fails—even subtly—every future match against that template inherits the failure.

The enrollment pipeline has four distinct stages:

1

Identity Verification

Before any biometric capture, the system must verify that the person enrolling is who they claim to be. This typically involves government-issued ID verification, in-person proofing by an authorized operator, or multi-factor authentication against an existing identity system. Without this step, an attacker can enroll their own biometrics under someone else's identity—the most devastating attack on any biometric system.

2

Sample Capture

The biometric sensor captures one or more raw samples: fingerprint images, face photographs, iris scans, or voice recordings. The capture environment must be controlled to produce samples of sufficient quality. Multiple captures are standard practice across all modalities—typically 3 to 5 per enrollment session—to enable quality selection and template fusion.

3

Feature Extraction and Template Generation

The raw sample is processed to extract distinctive features: minutiae points from fingerprints, embedding vectors from faces, IrisCodes from iris scans, or spectral coefficients from voice. These features are encoded into a compact numerical template—typically a fixed-length vector—that serves as the mathematical representation of the person's biometric. For a deeper explanation of how templates work, see our complete biometric authentication guide.

4

Template Storage

The generated template must be stored securely for use in future authentication. This is where the most consequential architectural decision occurs: whether to store templates in plaintext (encrypted at rest but decrypted for matching), in a transformed form (cancelable biometrics), or in a fully encrypted form that never requires decryption (FHE-based template protection). We will examine each approach in detail later in this guide.

The critical insight is that enrollment is not just data collection—it is the process that determines the ceiling on your system's operational accuracy. A perfect matching algorithm operating on a poor template will always produce poor results. The best enrollment systems treat sample quality as a hard gate: if the captured sample does not meet minimum quality thresholds, enrollment fails and the user is asked to try again.

Key Principle

Enrollment quality establishes the upper bound on authentication accuracy. No amount of matching algorithm optimization can compensate for a poor enrollment template. Invest disproportionate engineering effort in the enrollment pipeline.

2. Enrollment Quality and Its Impact on Matching Accuracy

The relationship between enrollment quality and matching accuracy is not linear—it is exponential at the extremes. Small quality improvements in already-good templates produce marginal accuracy gains. But small quality degradations in borderline templates produce catastrophic accuracy losses. This asymmetry means that quality control at enrollment time has an outsized impact on operational performance.

Understanding FAR, FRR, and EER

Two metrics define biometric system accuracy:

These two metrics are inversely related through the matching threshold. Lowering the threshold (accepting more matches) reduces FRR but increases FAR. Raising the threshold reduces FAR but increases FRR. The Equal Error Rate (EER)—the point where FAR equals FRR—provides a single number for comparing systems.

How Quality Degrades These Metrics

When enrollment quality drops, both metrics degrade simultaneously—a phenomenon that seems counterintuitive given their inverse relationship. The explanation is that a low-quality template shifts the entire score distribution for genuine attempts leftward (toward lower similarity scores) while leaving the impostor distribution unchanged. This means:

Enrollment QualityTypical FARTypical FRREERUser Experience
Excellent (NFIQ 1) 0.001% 0.1% 0.05% Seamless
Good (NFIQ 2) 0.005% 0.5% 0.2% Acceptable
Fair (NFIQ 3) 0.02% 2.0% 0.8% Noticeable friction
Poor (NFIQ 4) 0.1% 8.0% 3.0% Frequent failures
Very Poor (NFIQ 5) 0.5% 20%+ 8.0%+ Unusable

The NIST Fingerprint Quality (NFIQ) scale is shown here for illustration, but the principle applies universally across modalities. A face enrollment captured in harsh backlighting, an iris scan with excessive pupil dilation, or a voice sample recorded in a noisy environment all produce the same kind of quality degradation—and the same kind of operational accuracy loss.

Warning: Quality Degradation Compounds

If both the enrollment template and the authentication sample are low quality, the accuracy degradation compounds. An NFIQ 4 enrollment matched against an NFIQ 4 probe can produce FRR above 30%—meaning nearly one in three legitimate authentication attempts will fail. This is why enrollment quality control is more important than authentication-time quality control: you only enroll once, but that template affects every future authentication.

Quality Scoring Standards

Every production enrollment system must implement real-time quality scoring. The major standards are:

The key principle is reject early, reject clearly. If a sample does not meet quality thresholds, reject it immediately and provide specific, actionable feedback to the user: "Move your finger to the center of the sensor," "Turn to face the light source," "Please speak louder." Generic "try again" messages cause frustration and rarely produce better samples.

3. Fingerprint Enrollment Best Practices

Fingerprint remains the most widely deployed biometric modality. Billions of fingerprint sensors are embedded in smartphones, access control terminals, border kiosks, and time-and-attendance systems. The technology is mature, well-understood, and supported by decades of standardization. But fingerprint enrollment is also the most sensitive to environmental conditions, user behavior, and sensor quality.

Capture Requirements

Environmental Controls

Fingerprint quality is acutely sensitive to moisture, temperature, and cleanliness:

Placement Guidance

Incorrect finger placement is the second most common cause of enrollment failure after dry skin. Effective enrollment systems provide real-time visual guidance:

Fingerprint Enrollment Checklist

Template Fusion vs. Best-Sample Selection

When multiple samples are captured, two strategies exist for template generation:

Best-sample selection picks the single highest-quality sample and generates the template from it alone. This is simpler to implement and avoids the risk of fusing a good sample with a bad one. It is the recommended approach when quality scoring is reliable.

Multi-sample fusion combines information from multiple samples into a single super-template. This can produce a template with more minutiae (combining partial-overlap captures) and higher discriminative power. However, fusion requires careful alignment of the component samples, and fusing a high-quality sample with a low-quality one can actually degrade the result. Use fusion only when all component samples meet quality thresholds.

4. Face Enrollment Best Practices

Face recognition has overtaken fingerprint as the fastest-growing biometric modality, driven by smartphone adoption (Apple Face ID, Android face unlock) and contactless post-pandemic authentication requirements. Face enrollment is more accessible to users than fingerprint (no special sensor required beyond a camera) but more sensitive to environmental conditions.

Lighting Requirements

Lighting is the single most important factor in face enrollment quality. Poor lighting produces shadows that obscure facial features, harsh highlights that wash out texture detail, and color casts that confuse skin detection algorithms.

Pose and Expression

Face recognition systems trained on frontal poses degrade rapidly as the face rotates away from the camera. Enrollment should capture the face under controlled pose conditions:

Liveness Detection During Enrollment

Enrollment is the highest-value target for presentation attacks. If an attacker can enroll a synthetic face (a printed photo, a 3D mask, or a deepfake video), they gain permanent access to the identity. Liveness detection during enrollment is not optional—it is a hard security requirement.

ISO 30107-3 Compliance

Presentation Attack Detection (PAD) should comply with ISO/IEC 30107-3, which defines testing methodology for evaluating PAD mechanisms. The standard defines Attack Presentation Classification Error Rate (APCER) and Bona Fide Presentation Classification Error Rate (BPCER) as the key metrics. For enrollment, target APCER < 1% and BPCER < 5%.

Face Enrollment Checklist

5. Iris Enrollment Best Practices

Iris recognition offers the highest uniqueness of any biometric modality. The probability of two irises producing the same IrisCode is approximately 1 in 1078, making it effectively impossible for false acceptance to occur from random chance. The iris pattern is stable from approximately 18 months of age through the entire lifetime, making it the most persistent biometric as well.

However, iris enrollment is the most technically demanding modality. The iris is a small target (approximately 11mm in diameter), located behind a curved, reflective cornea, often partially occluded by the eyelid, and subject to dilation changes from ambient light.

Capture Hardware Requirements

User Positioning and Gaze

Environmental Controls for Iris

Iris Enrollment Checklist

6. Voice Enrollment Best Practices

Voice recognition is the most accessible biometric modality—it requires no specialized hardware beyond a microphone, which is present in every smartphone, laptop, and smart speaker. This accessibility makes voice enrollment attractive for remote onboarding, call center authentication, and IoT device integration. However, voice is also the least stable biometric: it changes with health (colds, allergies), emotion, fatigue, aging, and environmental noise.

Recording Requirements

Environmental Controls for Voice

Voice Anti-Spoofing at Enrollment

Voice spoofing has become significantly easier with modern AI speech synthesis. Deepfake voice generation can produce convincing replicas from as little as 3 seconds of reference audio. Enrollment must include anti-spoofing measures:

Voice Enrollment Limitation

Voice is the weakest biometric modality for standalone authentication due to its high FAR (0.1–1.0%), susceptibility to spoofing, and variability with health and environment. It is best used as a supplementary factor in multi-modal systems or for continuous authentication during phone calls, not as a primary authentication mechanism for high-security applications.

7. Template Generation and Storage Security

Once a quality biometric sample has been captured and features extracted, the resulting template must be stored securely. This is where most biometric systems fail catastrophically. The fundamental problem is simple: biometric templates are permanent, irrevocable credentials. Unlike passwords, they cannot be changed if compromised. A leaked fingerprint template is compromised forever.

For a comprehensive treatment of template protection mechanisms, see our dedicated guide on biometric template protection. Here, we focus on the storage decisions made at enrollment time.

Storage Architecture Options

ApproachSecurityRevocabilityPerformanceLimitation
Plaintext (encrypted at rest) Low None Highest Templates exposed during matching; single breach compromises all users
Salted hash Medium Partial High Biometrics are fuzzy—hashing destroys the tolerance needed for matching
Cancelable biometrics Medium Yes High Transformed templates still vulnerable to inversion attacks if transform leaks
Biometric cryptosystem Medium-High Yes Medium Helper data can leak partial template information; limited error tolerance
FHE-encrypted templates Highest Yes High (with optimization) Requires FHE infrastructure; historically slow, now solved at scale

The critical decision point at enrollment is: at what stage does encryption occur? In traditional architectures, the raw sample is sent to the server, features are extracted server-side, and the resulting template is encrypted for storage. This means the plaintext biometric exists on the server, even if only transiently. In an FHE architecture, encryption occurs on the client before the template ever leaves the device.

Template Isolation Requirements

Regardless of encryption approach, enrollment systems must implement template isolation:

8. FHE-Based Enrollment: Eliminating Plaintext Templates

Fully Homomorphic Encryption fundamentally changes the enrollment architecture. Instead of encrypting the template after it reaches the server, FHE encrypts the template on the client device, before it ever leaves the user's possession. The server receives, stores, and performs matching on ciphertexts. The plaintext template never exists outside the client device.

For a deep dive into how FHE works mathematically, see our guide on Fully Homomorphic Encryption. Here, we focus on how FHE changes the enrollment workflow.

FHE Enrollment Architecture

1

Client-Side Capture and Feature Extraction

The biometric sample is captured on the user's device (phone, laptop, enrollment kiosk). Feature extraction runs locally, producing a numerical template vector. This is identical to traditional enrollment so far.

2

Client-Side Encryption

The template vector is encrypted using the system's FHE public key before leaving the device. The resulting ciphertext is computationally indistinguishable from random data. No entity—including the server, the network, and any intermediary—can recover the plaintext template from this ciphertext without the private key.

3

Encrypted Template Transmission and Storage

The encrypted template is transmitted to the server over a standard TLS channel (defense in depth) and stored in the template vault. The server has no mechanism to decrypt this template. The stored ciphertext serves as the reference template for all future authentication.

4

Encrypted Matching (Authentication Time)

When the user authenticates, a fresh biometric sample is captured, a template is extracted, and the template is encrypted client-side using the same FHE public key. This encrypted probe is sent to the server, which performs the matching computation entirely on ciphertexts—the inner product of two encrypted vectors—and returns an encrypted match score. The score is decrypted only by an authorized entity to make the accept/reject decision.

Why FHE Changes Everything

In a traditional system, a server breach exposes every template in the database. In an FHE system, a server breach exposes encrypted ciphertexts that are computationally infeasible to decrypt—even with quantum computers, since the underlying lattice problems are post-quantum secure. The entire "store securely, decrypt to match, re-encrypt" cycle is eliminated. Templates are born encrypted and die encrypted.

Enrollment-Time Quality Gating with FHE

One concern with client-side encryption is: how does the server verify template quality if it cannot see the template? The answer is that quality assessment must happen client-side, before encryption. The client runs quality scoring on the raw sample, and only proceeds to encryption and transmission if the sample meets quality thresholds. A quality attestation (a signed quality score or a zero-knowledge proof of quality) can accompany the encrypted template, allowing the server to verify that quality gating was performed without seeing the template itself.

This is architecturally superior to server-side quality gating because it keeps the plaintext entirely on the client. The trade-off is that you must trust the client-side quality assessment—or verify it cryptographically. In the H33 architecture, quality attestation is included as part of the enrollment payload.

9. Multi-Modal Enrollment Strategies

Multi-modal biometric systems combine two or more biometric modalities (fingerprint + face, face + voice, iris + fingerprint, etc.) to achieve higher accuracy, broader population coverage, and stronger spoofing resistance than any single modality alone.

Why Multi-Modal?

Enrollment Strategies for Multi-Modal Systems

Sequential enrollment captures each modality one at a time. The user first enrolls their fingerprint, then their face, then their voice (as applicable). This is simpler to implement and allows modality-specific quality gating at each step. It is the recommended approach for most deployments.

Simultaneous enrollment captures multiple modalities in a single interaction. For example, a kiosk with a fingerprint sensor and a camera captures the fingerprint while simultaneously taking a face photo. This reduces enrollment time but requires more complex hardware and makes it harder to provide modality-specific quality feedback.

Progressive enrollment starts with a single modality and adds additional modalities over time. The user might enroll face on day one, then be prompted to add a fingerprint enrollment the next time they authenticate. This minimizes initial enrollment friction while building toward a more secure multi-modal profile over time. It is particularly effective for consumer applications where a lengthy initial enrollment would cause drop-off.

Face + Fingerprint

The most common combination. Face provides contactless convenience; fingerprint provides high-precision fallback. Widely supported by smartphone hardware. Excellent for mobile banking and device unlock.

Face + Voice

Fully contactless and requires no specialized hardware beyond camera and microphone. Ideal for remote onboarding and call center authentication. Voice compensates for face failures in poor lighting.

Iris + Fingerprint

The highest-security combination for physical access control and border security. Both modalities have extremely low FAR. Requires specialized iris capture hardware.

Face + Behavioral

Face for explicit authentication; behavioral biometrics (keystroke dynamics, mouse patterns) for continuous session verification. Provides both point-in-time and ongoing assurance.

Fusion Strategy Selection

Multi-modal systems must decide how to combine the signals from different modalities. The three primary fusion levels are:

For most deployments, score-level fusion with learned weights provides the best trade-off between accuracy and implementation complexity.

10. Enrollment UX: Reducing Friction While Maintaining Quality

Enrollment user experience is a critical success factor, particularly in consumer-facing deployments where participation is voluntary. A frustrated user who abandons enrollment is a user who will never authenticate. The challenge is that enrollment inherently requires user cooperation—holding still, presenting biometrics to sensors, following instructions—and quality requirements impose constraints that conflict with speed and convenience.

Principles of High-Quality Enrollment UX

Enrollment Time Benchmarks

These are target enrollment times for a well-designed system. Exceeding these times significantly increases abandonment rates:

ModalityTarget TimeMax AcceptableCaptures Required
Fingerprint (single finger) 15–20 sec 45 sec 3–5 captures
Face 5–10 sec 30 sec 1–3 captures + liveness
Iris (both eyes) 10–15 sec 45 sec 2–3 captures per eye
Voice 30–45 sec 90 sec 3 utterances of 10–15 sec
Multi-modal (face + finger) 25–35 sec 75 sec Combined captures

Accessibility Considerations

Enrollment systems must accommodate users with disabilities:

11. Re-Enrollment and Template Update Policies

Biometric templates are not permanent. Physical changes (aging, injury, surgery, weight changes), environmental changes (new capture device with different characteristics), and quality improvements (new feature extraction algorithms) all create scenarios where re-enrollment is necessary or beneficial.

When to Trigger Re-Enrollment

Re-Enrollment Security Requirements

Re-enrollment is a high-risk operation because it replaces the reference template. An attacker who can trigger and control a re-enrollment effectively takes over the identity. Re-enrollment must be protected with strict identity verification:

Template Update vs. Full Re-Enrollment

Template update (adaptive enrollment) silently improves the template using high-quality authentication samples. When a user authenticates with a sample that scores well above the match threshold, the system blends the new sample's features into the stored template, gradually improving it over time. This avoids the friction of explicit re-enrollment but requires careful implementation to prevent template drift from adversarial inputs.

Full re-enrollment replaces the template entirely from a new capture session. This is necessary when the template has degraded significantly or when the feature extraction algorithm has changed. It requires explicit user participation and identity verification.

12. Compliance Considerations

Biometric data is classified as sensitive personal data under virtually every major privacy regulation. Enrollment is the point at which biometric data is first collected, making it the primary compliance touchpoint. Getting enrollment consent and data handling wrong exposes the organization to significant legal liability.

For deep dives into specific regulations, see our dedicated guides on BIPA compliance and CCPA compliance for biometrics.

GDPR (European Union)

The General Data Protection Regulation classifies biometric data as a "special category" of personal data under Article 9, which means it is prohibited by default unless one of the limited exceptions applies:

BIPA (Illinois, United States)

The Illinois Biometric Information Privacy Act is the most aggressive biometric privacy law in the United States and the source of the largest biometric privacy settlements (Facebook: $650M, Google: $100M, TikTok: $92M):

CCPA/CPRA (California, United States)

Additional Regulations

Consent Must Be Pre-Enrollment

Under every regulation listed above, consent must be obtained before the first biometric sample is captured—not during or after. Displaying a consent screen while the camera is already capturing, or requesting consent after the fingerprint has been scanned, violates the consent requirements of GDPR, BIPA, CCPA, and every other biometric privacy law. The enrollment system must present the consent request, receive affirmative consent, log the consent, and only then activate the biometric sensor.

Consent Implementation at Enrollment

A compliant enrollment consent flow includes these elements:

  1. Clear disclosure: Explain in plain language what biometric data will be collected, how it will be used, how long it will be retained, and who will have access to it.
  2. Specific consent: The consent must be specific to biometric data processing, not bundled with general terms of service.
  3. Affirmative action: Require an affirmative action (checking an unchecked box, clicking "I consent," signing) that cannot be confused with passive agreement.
  4. Record the consent: Store the consent record (timestamp, user identity, consent text version, IP address) in an immutable audit log.
  5. Provide a copy: Offer the user a copy of their consent (email, downloadable PDF).
  6. Easy withdrawal: Provide a clearly accessible mechanism for withdrawing consent, which triggers template deletion.

13. Common Enrollment Failures and How to Handle Them

Enrollment failures fall into two categories: hard failures (the system cannot capture a usable sample at all) and soft failures (the system captures a sample and generates a template, but the template quality is too low to support reliable authentication). Soft failures are more dangerous because they often go undetected until authentication-time problems emerge.

Hard Enrollment Failures

FailureModalityCauseResolution
Failure to capture Fingerprint Worn fingerprints (manual labor, elderly), extremely dry skin, skin conditions (eczema) Try alternate fingers; apply moisturizer; switch to face modality
Failure to detect Face Extreme lighting (backlit, very dark), face fully covered, camera obstruction Reposition lighting; ensure face is visible; clean camera lens
Failure to segment Iris Excessive eyelid occlusion, reflections on glasses, cosmetic contact lenses Open eyes wider; remove glasses; remove cosmetic lenses
Failure to extract Voice Extreme background noise, user whispering, microphone failure Move to quiet environment; speak at normal volume; check hardware
Liveness failure Face/Iris Presentation attack detected, or false positive from poor sensor data Investigate for spoofing; if legitimate, retry with better lighting/positioning

Soft Enrollment Failures

Failure Handling Workflow

A robust enrollment system implements a three-tier failure handling strategy:

  1. Tier 1 — Automatic retry: For quality failures, automatically prompt the user to retry with specific corrective guidance. Allow up to 3 automatic retries per capture.
  2. Tier 2 — Alternative modality: If the user cannot produce an acceptable sample after retries, offer an alternative biometric modality. If fingerprint fails, try face. This requires multi-modal enrollment capability.
  3. Tier 3 — Operator escalation: If all automated paths fail, escalate to a human operator who can assess the situation, provide hands-on assistance, adjust environmental conditions, or authorize an exception (such as accepting a lower-quality template with a note in the user's record).
Never Silently Degrade

The worst possible failure handling is to silently accept low-quality templates to avoid user friction. This creates a population of users who experience chronic authentication failures, generate support tickets, lose trust in the system, and ultimately abandon biometric authentication. It is always better to fail enrollment explicitly and try again than to succeed with a bad template.

14. Enterprise Deployment Considerations

Enterprise biometric enrollment introduces scale, logistics, and organizational challenges that do not exist in consumer or small-deployment contexts.

Enrollment Station Design

For organizations enrolling hundreds or thousands of employees, dedicated enrollment stations provide controlled conditions and high throughput:

Remote Enrollment

Post-pandemic, many organizations need to support remote enrollment for distributed workforces. Remote enrollment is inherently less controlled than in-person enrollment, which means quality will be more variable. Mitigations include:

Rollout Strategy

15. H33 API Enrollment Flow

H33's enrollment API is designed to make FHE-based biometric enrollment as straightforward as a traditional enrollment while providing the cryptographic guarantees described throughout this guide. The key architectural principle is that the plaintext biometric template never reaches H33's servers.

High-Level Enrollment Flow

1

Initialize Enrollment Session

Your application calls the H33 API to create an enrollment session. The API returns a session token and the FHE public key parameters needed for client-side encryption.

2

Client-Side Capture and Encryption

Your application captures the biometric sample on the user's device, extracts the feature template using your feature extraction pipeline (or H33's client SDK), encrypts the template using the provided FHE public key, and packages the encrypted template with a quality attestation.

3

Submit Encrypted Template

Your application submits the encrypted template to the H33 enrollment endpoint. The server stores the ciphertext without ever decrypting it.

4

Enrollment Confirmation

The API returns a confirmation including the enrollment ID and a cryptographic receipt. Optionally, your application can immediately perform a test verification to confirm the enrollment is working.

API Example

# Step 1: Initialize enrollment session
curl -X POST https://api.h33.ai/v1/enroll/init \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "usr_a1b2c3d4",
    "modality": "face",
    "quality_threshold": 0.85
  }'

# Response includes session_id and FHE public key parameters

# Step 2: Client-side encryption (using H33 client SDK)
# Template extraction and FHE encryption happen on-device

# Step 3: Submit encrypted template
curl -X POST https://api.h33.ai/v1/enroll/submit \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "ses_x9y8z7",
    "encrypted_template": "<base64-encoded FHE ciphertext>",
    "quality_attestation": "<signed quality score>",
    "modality": "face"
  }'

# Step 4: Enrollment confirmed
# Response:
# {
#   "enrollment_id": "enr_m4n5o6",
#   "status": "active",
#   "modality": "face",
#   "receipt": "<cryptographic enrollment receipt>",
#   "created_at": "2026-02-24T12:00:00Z"
# }

Batch Enrollment

For enterprise deployments enrolling large numbers of users, the H33 API supports batch enrollment. Multiple encrypted templates can be submitted in a single API call, and H33's SIMD batching processes up to 32 users per ciphertext operation. This means enrolling 32 users is approximately the same computational cost as enrolling 1 user.

# Batch enrollment for enterprise deployments
curl -X POST https://api.h33.ai/v1/enroll/batch \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "enrollments": [
      {
        "user_id": "usr_001",
        "encrypted_template": "<ciphertext>",
        "quality_attestation": "<attestation>",
        "modality": "face"
      },
      {
        "user_id": "usr_002",
        "encrypted_template": "<ciphertext>",
        "quality_attestation": "<attestation>",
        "modality": "face"
      }
    ]
  }'

Enrollment Security Guarantee

At no point in the H33 enrollment flow does a plaintext biometric template exist on H33 infrastructure. The template is encrypted on the client device using lattice-based FHE that is secure against both classical and quantum attacks. The server stores, indexes, and matches on ciphertexts exclusively. Even a complete compromise of H33's servers would not expose any user's biometric data. For a detailed explanation of the underlying cryptography, see our guide on Fully Homomorphic Encryption.

Conclusion: Enrollment Is the Foundation

Biometric enrollment is not a one-time setup step that you build once and forget. It is the foundation that determines the ceiling on your system's accuracy, security, and user experience for the entire lifetime of every enrolled user. The key principles are:

The enrollment process you build today will determine the authentication experience your users have for years to come. Build it right.

Ready to Build Quantum-Secure Biometric Enrollment?

Start with H33's FHE-encrypted enrollment API. Zero plaintext templates. Sub-millisecond matching. Post-quantum secure. Free tier available.

Get Free API Key →

Build With Post-Quantum Security

Enterprise-grade FHE, ZKP, and post-quantum cryptography. One API call. Sub-millisecond latency.

Get Free API Key → Read the Docs
Free tier · 10,000 API calls/month · No credit card required

Related Articles

Biometrics
Biometric Authentication: Complete Guide
Security
Biometric Template Protection
Biometrics
Liveness Detection & Anti-Spoofing
Verify It Yourself