Cognau - Deepfake-proof security

Safety for what arrives from outside

Proof a real person is there.

A live camera check that a recording or a generated face cannot answer.

Fresh random prompts every session, so there is nothing to rehearse.

Start free
AnythingAnythingaacameracameracancanshowshowcancannownowbebegenerated.generated.SoSoaaCognauCognausessionsessiondoesdoesnotnotwatchwatchandandguess.guess.ItItasks,asks,live,live,ininaasequencesequencenobodynobodycouldcouldhavehaverehearsed,rehearsed,andandaarecordingrecordinghashasnonoanswer.answer.

Passive checks ask a model whether a video looks generated. Every generator is then trained until it does not.

97.6%

Of the generated faces we tested satisfied a simple instruction to turn the head. The most common liveness challenge in the industry is close to free for an attacker.

Cognau internal evaluation.

So stop asking whether the video is fake. Ask it something only a live person can answer.

A short camera check, drawn at random, that a recording cannot satisfy.

Commit

Seal the prompts

Before the camera opens, the session draws its prompts at random and seals them. Your integration receives the sealed commitment when the session starts.

Reveal

One step at a time

Prompts are released live, each one moments before it appears on screen. Nothing downstream can be prepared in advance, and no recording made earlier can contain an answer to it.

Measure

Decide out of reach

Each response is judged against the prompt that was actually issued, away from the device. Your backend receives a pass or a fail and a confidence score, signed.

When a session finishes, you can open its commitments through the API and recompute them yourself: proof that the prompts were fixed before your user saw anything. Verifiable by you, not a claim you have to take on trust.

Under 1 min
for a typical check
3 ways in
a one-line embed, a server API, or a shareable link
$1.25
per liveness check, billed on results
A human eye, photographed close up

The verdict has value. The footage is a liability.

Identity documents and face images are exactly what attackers went looking for when they targeted a verification provider rather than its customers. We would rather not be holding them, so by default we are not.

Biometric data is special-category personal data under the GDPR, and in some US states individuals can sue over it directly. Keeping less is a product decision before it is a legal one.

Resting state

Deletion is the default

Nothing is kept because it might be useful later. Session material is written only where a step in the check has to read it back, and everything written carries a deletion deadline from the moment it exists.

Within 24 hours

An unfinished check leaves nothing

If somebody abandons a check, or it times out, or it is never scored, everything captured is deleted within a day. It backs no decision, so every hour it survived would be liability bought for nothing.

Not the recording

A finished check keeps a sample

A completed check keeps a few still images per prompt rather than the hundreds captured, because those few are the evidence behind a result somebody acted on. They expire on a published schedule.

Off by default

No face database

Proving that a person is present is a different question from recording which person. Recognising a returning face is an option a customer switches on for a stated reason, and it is off until they do.

Ready to secure your platform?

Confirm a real person is on the other end, before the decision that depends on it. Create an account and run your first check in the sandbox, free.