Prove a
live human
was there
Axiom asks the screen a question only a real face can answer. It emits a keyed sequence of light and measures how a three-dimensional human reflects it — then binds the result to a hardware key on an enrolled device.
A recording cannot answer a challenge that did not exist when it was made.
The Challenge
Seeing a face used to be enough.
It is not enough any more.
Detectors try to spot the fake after the fact, and they degrade as generators improve. On contemporary in-the-wild video the best published open-source detector reaches 0.63 AUC, and at a usable false-positive rate it catches under a quarter of what it is shown.
So Axiom does not guess whether an image is fake. It asks for something only physics and a private key can supply.
Step 1
Issue a challenge
nobody can predict
A server nonce derives a schedule of illumination patterns. Every pattern lights exactly half the screen, so the total light never changes and the camera's auto-exposure has nothing to fight — while the direction the light arrives from shifts continuously.
Total emitted luminance is constant across every pattern, so auto-exposure stays put while the light's direction changes.
Step 2
Measure what
only a face can do
A flat screen returns light uniformly; a real face returns it differently across the forehead, cheeks and nose, because those surfaces point in different directions. We score the response against the real key and against two hundred keys we never displayed. The real one has to win.
A recording responds to the key that was on screen when it was made. It cannot respond to this one.
Step 3
Bind it to
a device and a decision
Presence and identity are separate claims, so we report them separately. The optical challenge shows a live human was present; a hardware-backed key shows which enrolled device authorised the action. Both are bound to one nonce in one signed receipt.
presence: "live human"
What this does and does not prove
We would rather tell you the limits.
Proven
That a live three-dimensional face responded to this session's key, and that a specific enrolled device signed the result. A previously recorded or injected stream fails both.
Not proven
That the face belongs to a particular person. A real human standing in for someone else is a live human. Identity is the hardware key's job, which is exactly why we do not ask the camera to do it.
Try to fool it
The demo runs in your browser and reports every channel it measured. Point a recording at it and watch the freshness check fail.
Open the demo →