Liveness detection is one of the fastest-moving components of the identity verification stack. The gap between what current production systems can do and what a well-resourced adversary can attempt with synthetic media has narrowed considerably over the past two years. But the narrative that "deepfakes have broken biometric liveness" is as inaccurate as the opposite claim that "modern liveness is unbeatable." The actual situation is more nuanced and operationally relevant than either framing suggests.
This piece is about what the current generation of liveness technology does well, where it still has meaningful gaps, and how to think about liveness detection as one component in a layered identity assurance stack, not as a standalone verification solution.
Active vs. Passive Liveness: The Core Architecture Distinction
Liveness detection approaches divide broadly into active and passive categories. Understanding the distinction matters because the tradeoffs are fundamentally different.
Active liveness checks require the user to perform a specific action: turn their head left and right, blink twice, follow a moving target, speak a phrase. The action requirement creates a challenge-response dynamic that makes it harder for an adversary to pre-generate a synthetic media file that passes. The attack surface for active liveness is primarily real-time video injection: feeding a synthetic video stream at the moment of capture, rather than pre-recorded manipulation. Active liveness is more resistant to photo attacks and video replay attacks. It is more vulnerable to real-time injection attacks using face-swap pipelines.
Passive liveness checks ask the user only to hold still for a selfie capture. There is no challenge response. The model infers liveness from texture analysis, depth cues, micro-expression artifacts, and other signals from a single or multi-frame capture without requiring user interaction. Passive liveness is inherently more user-friendly and has lower abandonment rates. It is more vulnerable to high-quality still-photo attacks and to early-generation synthetic media, though modern passive models have improved substantially on detecting generated images.
Most production deployments in regulated financial services use active liveness for high-assurance steps. Passive liveness is more common in lower-risk verification contexts and in situations where user friction is a primary concern. Some systems use a passive check first and escalate to active liveness only for sessions that the passive check does not clearly resolve, which is a reasonable risk-tiered approach.
What Current Models Handle Well
Print attack detection (holding up a printed photo to the camera) has been solved to a high degree of reliability for several years. The texture and depth artifacts of a printed image are detectable by modern classifiers with very high accuracy even under variable lighting conditions.
Screen replay attacks (playing a video of the subject on a phone or tablet held in front of the camera) are similarly well handled by current systems. Screen moire patterns, color profiles, and motion artifacts at frame boundaries are consistent signals that production models detect reliably.
3D mask attacks (using a physical mask) are handled by active liveness more effectively than passive, because the motion dynamics of a rigid mask are difficult to match with the kind of natural head movement variation that active challenges require. Passive liveness is more vulnerable to high-quality 3D masks, though the barrier to constructing a convincing 3D mask remains high for most adversaries.
Generic deepfake faces generated by publicly available models (face swap tools, open source GAN-based synthesis) are detectable by current production systems with high reliability. This category of attack uses compression artifacts, blending edge anomalies, and temporal inconsistency patterns that well-trained classifiers catch.
Where the Gaps Remain
The harder attack category for current liveness models is high-quality real-time face injection targeting active liveness flows. A technically capable adversary with access to a high-quality face-swap pipeline running on adequate hardware can, in principle, pass some active liveness checks by feeding synthetic video that responds to the challenge in real time. This is not a commodity attack: it requires specific technical capability and a real-time inference pipeline fast enough to respond to challenge prompts within normal interaction timing. But the capability exists, and the barrier to entry has been declining.
The countermeasures deployed by liveness vendors against injection attacks include injection detection (detecting that the camera feed itself is synthetic rather than hardware-captured), environment consistency analysis (checking that lighting, background, and metadata are consistent with a real capture), and device attestation (validating that the capture came from a known hardware camera, not a virtual camera driver). These countermeasures work with varying effectiveness depending on the deployment environment.
The coverage gap that gets less attention is demographic variation in passive liveness accuracy. Passive liveness models trained predominantly on certain demographic populations have historically shown higher error rates on skin tones underrepresented in training data. This is a documented problem in biometric systems generally. The degree to which it affects current production systems depends on how rigorously specific vendors have addressed it in their training data and evaluation methodology. Vendors who cannot show you disaggregated accuracy metrics by skin tone are not vendors you should deploy for a user population with significant demographic diversity without further scrutiny.
ISO 30107 and What It Does and Doesn't Tell You
ISO/IEC 30107 is the international standard for biometric presentation attack detection. It defines attack types, evaluation methodologies, and metrics for comparing liveness detection performance. Part 3 (ISO 30107-3) specifies the testing methodology for presentation attack detection systems.
iBeta Quality Assurance runs conformance testing programs against ISO 30107-3 and certifies whether a given liveness system passes at Level 1 or Level 2. Level 1 tests address artifact attacks (printed photos, screen replay). Level 2 tests address more sophisticated attack types including 3D mask attacks.
ISO 30107 certification is a meaningful baseline signal. But it has important limitations. The standard was developed before high-quality face-swap injection attacks became a practical threat. A system that passes ISO 30107-3 Level 2 is not necessarily resistant to real-time injection attacks, because that attack category was not in scope when the testing methodology was designed. ISO 30107 compliance is a necessary condition for a production liveness system, not a sufficient one for high-assurance deployments.
We're not saying ISO 30107 certification is worthless. It is a real quality signal and a reasonable first filter when evaluating liveness vendors. The point is that it should be understood as addressing one set of attack types, not all of them.
Liveness in Context: What It Is and Isn't Doing
A liveness check confirms that the session involves a live person in front of the camera at the time of capture. That is a specific and limited assertion. It does not confirm that the person is who they claim to be. That confirmation requires comparing the captured face against a document-extracted face, which is a separate step (face match) that occurs after liveness is established.
The face match step has its own accuracy characteristics and demographic distribution considerations. The overall identity assurance provided by the biometric component of a verification session is a product of liveness detection accuracy and face match accuracy, not just liveness detection accuracy alone. Marketing materials that describe liveness performance without discussing face match performance are presenting only half of the relevant accuracy picture.
Beyond the biometric component, liveness and face match are one layer of a broader identity assurance stack that includes document verification, data matching (name, date of birth, address), database checks against fraud signals, device intelligence, and behavioral analytics. Liveness is an important layer for its specific attack resistance properties, but it is not the whole answer. Designing a KYC flow around liveness as the primary assurance mechanism, with weak data matching or no fraud signal database checks, will stop certain attack types while leaving others completely unaddressed.
Practical Guidance for 2026 Deployments
For most regulated onboarding use cases, a combination of passive liveness with active escalation for uncertain sessions remains the right architecture. Full active liveness for every session adds friction that suppresses legitimate user conversion without proportionally improving fraud outcomes for the typical fraud profile of consumer financial services.
Device attestation and injection detection should be evaluated as first-class requirements for any vendor you deploy in higher-assurance segments. The hardware camera validation question is increasingly central as injection attack tooling becomes more accessible.
Require disaggregated accuracy data from vendors before deployment decisions. A vendor who cannot show you accuracy metrics broken down by demographic group has either not measured it or does not want to share the results. Both possibilities are relevant to your deployment decision and potentially to your model risk management documentation obligations.
Finally, think about liveness decay: a liveness check performed at account opening does not provide ongoing assurance about who is operating the account. Periodic step-up verification at high-value transaction moments or unusual access patterns is the operational mechanism for extending identity assurance beyond the point of onboarding. How your orchestration layer handles those step-up triggers is as important as the quality of the initial liveness check at registration.