Next-generation mobile biometric defenseVerified by physiology,
not by appearance.
The biological identity layer for mobile — passive authentication from involuntary eye movement, on the phone already in the user's hand.
The threatGenerative AI has been weaponised.
Facial recognition proves likeness. It can no longer prove life — and deepfake fraud attempts rose roughly 3,000% in 2023 alone as the tools became free and trivial to use.
The visual trust gapGenerative AI can replicate any face and any voice in real time. Standard biometrics prove likeness — they cannot prove life.
Facial recognition did not fail because it was badly implemented. It failed because it answers the wrong question. And liveness detection fails at the moment it matters most: in a real-time faceswap, the attacker is live.
So we went looking for a signal that generative models cannot reach — one the human nervous system produces continuously and cannot be told to stop producing.
The signalYour eyes never hold still.
Even locked on a single point, the oculomotor system produces continuous involuntary motion. These are reflexes, not behaviours. They cannot be coached, rehearsed or performed on command — and current generative models cannot synthesise them in real time at the correct kinematic profile.
Involuntary corrective jumps you cannot feel, suppress, or produce on cue.
Slow, smooth wander between corrections. It never stops while you are alive.
Low-amplitude oscillation riding on top of drift — the noise floor of a living eye.
Response timing to a light stimulus, governed by the autonomic nervous system.
Try itYou cannot perform a reflex.
Move as fast as you physically can between the two targets. Your movement gets measured and placed on the saccadic main sequence — the amplitude-to-peak-velocity relationship every real eye movement obeys.
Your fastest voluntary movement
Press the magenta target, then move to the cyan one as fast as you can.
The main sequence
Real eye movements land inside the band. Microsaccades live in the cyan strip.
This measures a hand, not an eye — it is an illustration of the principle, not a measurement of our technology. That is the point: a deliberate movement cannot reach the kinematics of an involuntary one, and the gap is large enough that you can feel it.
Three questionsStandard biometrics answer one of them.
Our two layers together answer all three — from a single front-facing camera, with no hardware addition.
Face / iris matching
- ✕Spoofed by photographs
- ✕Beaten by masks and faceswaps
- ✕Static checkpoint only
Reflex detection
- ✓Defeats synthetic deepfakes
- ✓Defeats replay and static images
- !A live attacker behind a faceswap can pass liveness
Behavioural signature
- ✓Defeats impostors and account takeover
- ✓Continuous, session-wide
- ✓Catches the faceswap liveness lets through
A real-time faceswap preserves the attacker's own genuine microsaccadic activity — liveness detects a living brain, but not the right one. This is why the two layers run as parallel classifiers on the same stream, not as stages of a pipeline. Liveness is not a prerequisite for identity.
Two productised layersLiveness is the shield. Identity is the lock.
Trajence Liveness
Independent verification of any video stream — a live session, a captured call, or recorded footage under forensic review. Frame-by-frame microsaccade detection produces a per-session liveness verdict with a confidence score.
A layered active mode raises a precise visual stimulus and measures the involuntary sub-30 ms pupillary response — defeating any model that learns to synthesise microsaccades alone.
Trajence Identity
Authentication and identification for high-value transactions and remote video. A per-person gaze-print built from saccadic kinematics, drift structure and fixation sequencing — biological, not behavioural, so it cannot be coached, copied or rehearsed.
After the initial check, a continuous-identity loop re-verifies the session through periodic high-frame-rate bursts on the same camera. Assurance for the whole session, not just the front door.
Why frame rate is the productA microsaccade lasts about 20 milliseconds.
How much of that event you capture decides what you can measure — and what you can measure decides whether you get liveness only, or liveness and identity.
| Capture rate | Liveness | Identity parameters | What becomes measurable |
|---|---|---|---|
| 30 fps | Drift only | 0 of 13 | The ~20 ms microsaccade window is unresolvable at 33 ms per frame. |
| 60 fps | Binary detection | 0 of 13 | Position jump visible — enough to say a reflex occurred, nothing about its shape. |
| 120 fps | Strong | 7 of 13 | Amplitude, angle, duration, average velocity and average acceleration / deceleration. |
| 200–240 fps | Full | 13 of 13 | Adds peak velocity, peak acceleration / deceleration and all four derived ratios. |
The two layers have two different frame-rate requirements, which is why they are two tracks. Liveness needs only binary microsaccadic detection. Identity needs the full kinematic fingerprint: two people can produce the same amplitude saccade and reach peak velocity in measurably different ways — the shape of the velocity curve is individual.
Core technologyFrom camera frame to signed assertion.
A deterministic real-time pipeline running on the smartphone's existing front camera — no new sensor, no hardware addition. Microsaccades are 20–150 ms events, so latency is not a performance detail: it decides whether the signal exists at all.
Five fused signal layers
If one signal weakens, the others carry the weight. No single failure breaks authentication — and every layer is evaluated on-device.
Involuntary eye micro-movements. The hard liveness signal.
Grip, device tilt and hand motion during the authentication event.
Hardware fingerprint of the user's own trusted phone.
The network and location environment expected for that user at that time.
Habit signature matching the user's normal daily and weekly rhythm.
Privacy by architectureNo biometric template ever leaves the device.
Enrolled on-device
The personal model is built and matched locally. Privacy is a property of the architecture, not a policy promise.
Short-lived assertion
The output is a signed, expiring assertion — not a template, not raw video, not a face embedding. Template storage is HSM-compatible and never held in plaintext.
Drops into your stack
Built for existing OAuth 2.0 / OIDC, FIDO2 / WebAuthn and FAPI infrastructure, with mTLS and certificate pinning. No new stack.
Where it mattersBuilt for the moments that carry liability.
Banks carry the loss on a fraudulent wire transfer — not the customer, not the AI vendor. That is where passive biometrics land first.
Step-up authentication
Transaction over a threshold → a short passive check → proceed or block. No prompt, no extra tap, no SMS that a SIM swap can intercept.
Remote video and enterprise calls
A live signal that the participants are present humans, not synthesised ones — verified continuously, not once at join time.
Continuous session assurance
Hands-free re-verification that the person who authenticated is still the person at the device.
KYC and remote identity proofing
Onboarding checks that survive a screen, a printout or a real-time generative model held up to the camera.
ResearchGrounded in peer-reviewed oculomotor science.
Trajence is developed in collaboration with academic researchers working on peer-reviewed microsaccadic methodology. Detection follows established velocity-threshold analysis, extended into a 13-parameter kinematic profile — geometric, velocity, acceleration and derived ratios — that separates the person from the reflex.
Every commercial eye tracker filters involuntary motion out as noise. For us, it is the entire signal.
ContactSeeing is no longer believing.
Something else has to be.
We are opening partnerships and pilot programmes with financial institutions, identity providers and device manufacturers.
gabriel@trajence.com