Deepfakes vs liveness detection: how presence checks hold the line

Not long ago, opening a bank account meant standing at a counter. A staff member compared your face to your photo ID in real time. Physical presence was the silent control underneath every identity check.

Remote onboarding removed that control. The applicant is now a video stream, and the document is an image on a screen. Fraud teams must prove something new: that a live human is genuinely on the other end.

Two families of spoof dominate this new battlefield. They fail in different ways, and they demand different defences.

Presentation attacks: fakes shown to a real camera

The older family is the presentation attack. Something fake is physically presented to a genuine camera. The camera works exactly as intended; the object in front of it is the lie.

• Printed photos, held flat or curved to fake depth.
• Screen replays: a video of the victim played back on another phone or tablet.
• Paper cut-outs with eye holes, worn over a real face.
• Silicone or resin masks built to mimic skin texture and shape.

These attacks are cheap, scalable and still everywhere. Many need nothing more than a photo scraped from social media. That is why testing standards focused on them first.

Injection attacks: bypassing the lens entirely

The second family never touches the lens. In an injection attack, footage is pushed straight into the verification session as data. Virtual cameras, tampered apps or intercepted traffic present the fake feed as if real hardware captured it.

Deepfakes turned this from a niche trick into the sharper threat. Generative tools can animate a stolen portrait into a blinking, talking face. That synthetic video can respond to prompts in real time.

Injection also undermines naive liveness prompts. If the check asks for a blink, a live deepfake can blink on cue. Defences must look below the content, at how the video arrives and behaves.

Regulators are watching. In November 2024, the US financial crimes agency FinCEN issued an alert on deepfake fraud schemes. It reported a rise since 2023 in suspicious activity reports describing suspected deepfake media. It also flagged fraudsters using third-party webcam plugins to sidestep live verification checks.

For fraud and KYC leads, the practical message is blunt. The camera feed can no longer be trusted by default. Presence must be proven, not assumed.

How modern liveness detection works

Liveness detection is designed to answer one narrow question. Is a live human present in front of this device, right now? It does not decide who the person is; that is the job of face matching.

Most modern checks run as a short, guided selfie. The user holds the phone up, follows a prompt, and the system analyses the video live. Several kinds of signal feed that analysis.

• Active illumination: the screen flashes changing colours, and the system checks how a real face reflects them.
• Randomised prompts: turn, blink or move closer, in an order an attacker cannot pre-record.
• Texture and depth cues: skin, paper and screens reflect light in measurably different ways.
• Session integrity signals designed to spot virtual cameras, emulators and replayed streams.

Passive analysis runs underneath the whole session. The system inspects frames for tell-tale screen patterns, unnatural edges and rendering artefacts. No single check is decisive; the strength is in combining them.

The output is a confidence score, not a verdict. Your business sets the threshold that turns that score into a decision. A platform onboarding thousands daily may tune it differently to a bank issuing credit.

For genuine customers, the whole flow should take seconds. Clear prompts, quick capture and instant feedback keep drop-off low. Security that frustrates real applicants quietly costs you growth.

Testing that counts: ISO/IEC 30107-3 and independent labs

Vendor claims are easy to make and hard to compare. ISO/IEC 30107-3 exists to fix that. It is the international standard for testing and reporting presentation attack detection, with defined attack types and error metrics.

Its core metrics are worth knowing. One measures how often attacks get through; another measures how often genuine users are wrongly rejected. Both matter, because a system that blocks everyone is trivially "secure".

Independent testing puts those numbers under pressure. The best-known lab is iBeta, accredited through NIST's NVLAP scheme. Its Level 1 testing uses cheap artefacts such as printed photos and screen replays. Level 2 escalates to costlier attacks, including three-dimensional masks.

Read lab results carefully, though. A confirmation letter shows conformance under the tested configuration and conditions. It is not a blanket certification of everything a product does.

Regulator expectations point the same way. NIST's latest Digital Identity Guidelines require presentation attack detection for remote biometric capture at higher assurance levels. They also call for injection defences: detecting virtual cameras, emulators and signs of manipulated media.

An honest word about limits

No liveness system ends the contest. This is an arms race, and generative tools improve every few months. A control that holds today needs retesting as attacks evolve.

Scores are probabilities, not certainties. Any system will pass some attacks and flag some genuine customers. Treat every threshold as a risk decision you own, review and adjust.

Design for the grey zone. Low-confidence sessions should step up to manual review or another channel, not silently fail people. Liveness works best as one layer among several: document checks, device signals and behavioural monitoring.

Where Smart Solutions fits

Smart Solutions builds presence checks into remote identity flows end to end. Our liveness detection is designed to catch presentation and injection attacks during a short guided selfie. It returns a confidence score your team can threshold against your own risk appetite.

It pairs naturally with our face recognition, which matches the live face to the document portrait. Together they are built to answer both onboarding questions: is someone really there, and are they who they claim? If you are weighing up presence checks, we would be glad to talk it through.