Blink
A deliberate eye closure and reopen, measured from eye shape over the sequence of frames.
Liveness Detection
A live person. Not a photo of one.
Before Face Match ever runs, ShieldID asks the applicant to complete a short, guided challenge on their own phone: it proves a real, present, three-dimensional person is holding the camera, tracked on 478 facial landmarks over at least ten frames. A still photo, a video replay or a mask does not pass.
The active challenge
A challenge only a live face can complete.
The applicant is asked to do one or more of a small set of natural movements, in a random order, while the front camera tracks 478 points across the face in 3D.
Turn left / turn right
The head yaws away from centre, then — this part matters — comes back.
Nod
The head pitches down and up, a movement a flat image can't produce.
Smile
A genuine expression change, scored from the face's own movement, not a static grin.
At least ten frames are required before a challenge can be scored. A single photo, however it's presented, never has enough frames to pass.
Why the return matters
The return to centre is the whole point.
A turn challenge doesn't stop at "did the head move sideways." It requires the head to come back past the midpoint of that turn, on camera, before the challenge counts.
A printed photo or a phone held up sideways can fake the first half of a turn — tilt it, and the angle changes. It cannot then smoothly turn itself back toward the camera, because there's no face behind it to move. Requiring the return is what separates a live head turning on a neck from a flat image being tilted in front of the lens.
Every selfie, three separate checks
Three independent checks, run on every selfie, every time.
Alongside the active challenge, three more checks look at the selfie itself for signs it isn't a live person at all. Each is reported on its own — never bundled into a single opaque score.
Anti-spoof
Presentation-attack detection aimed at the most common attack: holding up a printed photo or a phone screen showing someone else's face.
Deepfake detection
A two-model ensemble looks for the artifacts AI-generated or face-swapped video leaves behind — signals a single classifier is more likely to miss.
3D-mask / print / replay detection
A separate check purpose-built for the harder cases: a printed mask, a moulded 3D mask, or a screen replaying a recorded video of the real applicant.
Tolerant of noise, strict on spoofs
A grainy phone camera isn't a security threat. We treat it that way.
Phone cameras compress, mislight and blur real faces every day. Rejecting every borderline reading would fail honest applicants far more often than it stops attacks.
- Scores in the middle of the range — the kind ordinary compression and lighting produce — are tolerated rather than treated as a failure.
- Only a high-confidence spoof or fake call fails the check. A borderline reading doesn't get the benefit of the doubt taken away from a real applicant.
- The result: fewer people bounced back for a retake because of bad lighting, and a harder bar for anyone actually trying to spoof the camera.
Every check runs, every time
Not a paid tier. Not a toggle. Every verification, every time.
Active liveness, anti-spoof, deepfake detection and mask detection run on every single verification, on every plan. There is no cheaper tier that skips one of them, and no setting that turns one off to save money.
4checks run on every verificationActive liveness, anti-spoof, deepfake and mask detection. Every plan.
Threat coverage
What each attack looks like, and what stops it.
Plain language, not a scorecard. Where a threat isn't a liveness problem at all, we say so.
| Threat | What it is | What stops it |
|---|---|---|
| Photo | A printed or on-screen still photo held up to the camera. | Stopped by the active challenge — a flat image can't blink, turn and return, nod or smile on request — and backed up by anti-spoof detection. |
| Screen replay | A recorded video of the real applicant, played back on a second screen. | Targeted directly by the 3D-mask/print/replay check, alongside anti-spoof and deepfake detection. |
| Printed mask | A flat, printed face mask worn or held over the attacker's own face. | Targeted directly by the 3D-mask/print/replay check. |
| 3D mask | A moulded, three-dimensional mask made to match someone else's face. | Targeted directly by the 3D-mask/print/replay check — the hardest case it's built for. |
| Deepfake video | Real-time, AI-generated or face-swapped video fed into the camera. | Targeted directly by the two-model deepfake ensemble, alongside anti-spoof detection. |
| Borrowed document | A live, real applicant completes the challenge honestly — holding someone else's genuine ID. | Not a liveness problem. Liveness proves a real person is present, not which person. That's what Face Match checks, on the same verification. |
Certification
We're not going to print a badge we haven't earned.
Liveness testing standards like ISO/IEC 30107-3 exist, and we haven't completed that testing yet. Rather than imply a certification with a vague "industry-leading" line, here's where we actually stand.
- No liveness or anti-spoofing certification is held today.
- We'll publish our testing status, and the date, once it exists — not a marketing claim in the meantime.
- Until then, this page describes exactly what runs and what each check targets, so you can judge the coverage yourself.
Questions
Straight answers about the challenge.
See the challenge for yourself
Run a sandbox verification and try to beat it.
Blink, turn, nod, smile — on your own phone, in your own sandbox session. Sandbox sessions are never billed.