Liveness Detection in 2026: Four Approaches For Stopping Deepfake and Injection Attacks

Quick answer: If deepfakes and synthetic identities are getting past your onboarding, you have four realistic ways to respond: active liveness, passive liveness, certified PAD-grade liveness on its own, or a layered stack that also inspects the capture path. The first three all analyze the face. The attacks growing fastest in 2026 fake the camera, not the face. Only the fourth approach checks for that.

Between January and August 2025, investigators at Group-IB documented 8,065 attempts to bypass the liveness checks of a single financial institution's digital loan onboarding. Not one attempt used a printed photo held up to a webcam. Every one used a virtual camera to inject an AI-generated face directly into the verification stream.

The liveness software did its job. It found blinking, natural micro-movement, plausible skin texture, and depth. The faces were live in every sense the software could measure. The camera was not real.

If you are reading this, you probably already know you have some version of this problem. Fraud that should not have passed is passing. The question is no longer whether to deploy liveness detection. It is which of the available approaches actually closes the gap you have. This guide compares the four, with the numbers each one is tested against.

What liveness detection actually verifies in identity checks

Liveness detection confirms the face captured during onboarding comes from a physically present human. It runs alongside face matching, which confirms the live face and the ID document photo belong to the same person. Liveness answers "is someone there." Face match answers "is it the right someone."

Every approach below answers those two questions differently. The differences that matter are three: what friction it adds, what it catches, and what it structurally cannot see.

Active vs passive liveness detection for KYC

These are the two baseline methods, and the trade-off between them is operational, not academic.

For most institutions the choice between these two has already been made by conversion data: passive by default, active as step-up. But read the second-to-last row twice. Neither one catches injection, because both analyze whatever frames they are handed. If your fraud pattern looks like the Group-IB case, choosing between these two is rearranging the same blind spot.

What certified liveness detection does and does not stop

The natural upgrade path is to buy better liveness: a product with an iBeta conformance letter under ISO/IEC 30107-3. That testing is real and worth having. It is also narrower than most buyers assume.

The last row is the reason certified liveness alone is still a single-layer approach. ISO/IEC 30107-3, including the 2023 revision, defines presentation attacks as attacks presented to the biometric sensor. A virtual camera never touches the sensor. A vendor can hold a flawless Level 2 letter and remain fully blind to a stream of injected deepfakes.

This is not speculation. The World Economic Forum's Cybercrime Atlas work published in January 2026 tested 17 face-swap tools against 8 camera-injection tools and found that even moderate-quality face swaps, delivered through injection, deceived biometric systems holding formal PAD certification.

Why layered liveness detection performs better against injection attacks

An injection attack removes the camera from the equation. The fraudster feeds synthetic video into the pipeline through a virtual camera driver, an emulator, a tampered SDK, or direct API manipulation. The verification system receives a technically perfect stream of a face that does not exist. We wrote a separate deep dive on what injection attack detection involves and what to ask vendors.

The pair that should drive your decision is 58% versus 0.3%. The same attacks that walk through a lone liveness check almost never survive a stack that also inspects the capture path. The defense exists. The gap is in which approach institutions bought.

Regulators have caught up faster than procurement checklists. FinCEN issued an advisory in late 2025 directing institutions to evaluate identity controls against AI-generated media. RBI's V-CIP regime for Video KYC in India already demands in-house capture applications, IP controls, and trained agents, and industry guidance now reads deepfake resistance as the baseline. The EBA's remote onboarding guidelines in the EU point the same direction.

How layered liveness detection stops deepfake injection attacks

At Signzy we process 10 million-plus onboardings every month across 1,800+ financial institutions, and that scale forced a conclusion on us early: we stopped treating "is the face live" as the final question and started asking "is the camera real." A liveness verdict is only as trustworthy as the capture path that produced the frames.

The layered approach adds three checks around liveness. Each inspects something different, and each catches what the previous one structurally cannot.

This layered design is what Signzy ships as FaceTrust: passive liveness, face match, deepfake detection, and capture-path integrity checks delivered as one verdict in one call, rather than liveness sold alone and the other three layers left as your problem. Because every layer runs on the same session, a face that passes liveness but arrives through a virtual camera driver fails the check anyway. That is the difference between certifying the face and trusting the capture.

FaceTrust sits inside Signzy's Biometric Verification, powers the face checks in One Touch KYC, and is available through our KYC API suite for the US market. For a deeper look at how deepfakes specifically target KYC flows, see our guide on preventing deepfake fraud in KYC, and for the selfie-based flow specifics, how selfie ID verification stays compliant.

The common objection to layering is drop-off. In practice the opposite holds: passive checks carry the honest majority through untouched, and step-up friction lands only on sessions that earned it. Signzy customers see a 33% reduction in onboarding drop-offs and a 30% reduction in manual operations, because the layers filter rather than accumulate.

How to choose the right liveness detection approach

There is no universal answer, but there is a floor for each fraud reality.

One diagnostic before you decide anything. Take your fraud team's most recent confirmed synthetic identity case and trace how it entered. If the session passed liveness cleanly and nobody can tell you whether the capture path was inspected, you already know which row of that table you are in.

And when you evaluate any vendor, ask two questions in order. First: "Show me your PAD conformance letter." They will have one. Second: "Now show me what happens when the attack never touches the camera." The pause after the second question tells you which year their product is defending. If you want to see all four layers run against a virtual camera in a sandbox, talk to us.