Liveness detection APIs: 3 providers compared

TL;DR

Choosing a facial liveness API requires product-specific evidence: capture labels alone do not establish security, and attack coverage, user performance, data handling, workflow fit, and cost should drive the decision.

• Facial liveness detection determines whether a face presented to a camera comes from a live, physically present person rather than a photo, mask, replay, or deepfake. Standards call this function presentation attack detection.
• Liveness detection confirms live presence. Identity verification separately matches the face against a trusted credential or reference.
• Passive checks require no prompted action. Active checks request a response, while hybrid checks combine prompted responses with passive analysis.
• Product design and configuration affect performance more than the capture-mode label alone.
• Unlike systems that maintain persistent identity profiles, Stile uses a privacy-first processing model.

What facial liveness detection is

Facial liveness detection evaluates whether a biometric presentation to a camera is bona fide or is an attack presentation, such as a print, replay, or mask. Standards call this function presentation attack detection, or PAD, and ISO/IEC 30107-3 defines how PAD performance is tested and reported.

Liveness detection does not establish who the person is. Identity verification must also match the live face against a trusted reference, such as an identity document portrait or enrolled biometric.

A deepfake is manipulated or synthetic content, while presentation and injection describe how an attacker delivers that content. A presentation attack places an artifact, screen, or mask before the camera. An injection attack feeds synthetic or prerecorded media directly into the capture pipeline, bypassing the physical camera.

Active and passive modes describe the interaction model rather than the full detection stack. A liveness detection API can apply several underlying defenses within any of these modes.

DHS testing found wide performance variation within both active and passive categories. Some tested configurations in each category stopped every attack in the relevant test set. A well-implemented passive product can therefore outperform a weak active product, so evaluators should choose the capture mode for interaction fit and assess security through product-specific evidence.

Challenge-response can use a visible head-turn prompt or a controlled illumination sequence that requires no conscious action. The technique therefore does not determine whether the overall capture mode is active or passive.

Why performance varies by product and configuration

Products within the same capture category can produce sharply different security and genuine-user results. In the DHS RIVTD 2023-24 evaluation, worst-case bona fide acceptance across six active presentation attack detection subsystems ranged from approximately 41.4% to 93.9%. DHS calculated this range as one minus each subsystem's maximum bona fide presentation classification error rate, or BPCER. It should not be read as a generic success rate.

For screen-and-printout attacks, the median attack error rate was 2%, while one outlier reached 88%. The 2% figure represents the median rather than the strongest result. DHS's successor RIVR evaluation, conducted in 2025 with results published in 2026, continued to find substantial product-level variation. Buyers should therefore request independent, dated evidence that identifies the tested configuration, devices, attack types, and genuine-user performance.

Stile's privacy-first processing model

Stile uses a transient-processing architecture designed to minimize persistent identity records rather than claiming that it never handles personal data. Stile processes identity and biometric data during each check. By default, it deletes raw ID images and biometric captures when the check finishes.

Stile may retain signed verification outcomes, audit metadata, and one-way hashed anchors. The anchors support deduplication and returning-user checks. They reduce the need to retain raw captures, but they should not be treated as anonymous or immune to misuse. Organizations can therefore support deduplication and returning-user checks without retaining reusable face or ID images by default.

Stile holds iBeta PAD Level 1 and Level 2 certification under ISO/IEC 30107-3. iBeta is a NIST NVLAP-accredited independent testing lab, and Stile's certification covers resistance to tested presentation attacks, such as photos, replays, and masks presented to a camera. Its scope does not cover injection attack detection, face-matching performance or bias, general system security, or SOC 2 controls.

Choosing a liveness detection API: five filters

Evaluate a liveness detection API against evidence, attack coverage, user experience, data handling, and cost.

• Check evidence specificity by looking for dated third-party testing that names the tested attack types, devices, configurations, and error measures.
• Check threat coverage by confirming whether the API detects presentation attacks, injection attacks, or both.
• Measure genuine-user performance by reviewing rejection and completion rates under your expected operating conditions and across demographic groups.
• Review the data lifecycle by asking what the vendor processes, deletes, retains, and reuses after each check.
• Assess commercial fit by comparing integration effort, pricing, expected volume, support, and the other verification capabilities you need.

Products with different scopes also differ in cost, integration effort, and suitable use cases. Stile's certification scope shows why buyers should separate verified PAD coverage from untested capabilities such as injection detection.

Conclusion

Choose a liveness API by verified attack coverage and measured workflow performance, not by its passive, active, or hybrid label. Shortlist products with dated third-party evidence and clear retention terms, then pilot the exact configuration on expected devices, users, and attack paths before deployment.