Technology
AI can fake data. It cannot fake physics.
During manufacture, each photodetector ends up microscopically different. Each pixel converts photons into electrons at a slightly different rate — a fixed, random pattern called PRNU. It is baked into the silicon and cannot be manufactured to order.
A live authentication demo — open access, no sign-up.
Inside the camera
How light becomes data — and where identity is born.
Everything to the right of the sensor is data — and data can be copied. The sensor itself is matter. PRNU is created at the sensor, and nowhere else.
- Subjectreflected light
- Lensfocuses the beam
- Colour filterBayer mosaic
- Image sensorphotons → electrons
- A/D convertersignal → numbers
- Processingdemosaic, encode
- Digital filecopyable forever
PRNU is created here — and nowhere else.
Everything to the right of the sensor is data — and data can be copied. The sensor itself is matter.
Why every sensor is unique
Same model. Same factory. Same line. Different pattern — always.
Not assigned by anyone. Not chosen by anyone. An accident of physics, unique to one device.
The pattern is read from the camera on the device, converted into key material, and never transmitted or stored.
Read
Factory-random silicon noise is read via the standard camera permission — no photo is captured or kept.
Rebuild
A standard elliptic-curve private key is regenerated in ~30 ms to sign the challenge.
Vanish
The private key is never exposed in clear text — not even in RAM. It exists only inside the protected execution context for the instant it signs, and is then destroyed.
Cryptographic sequence
Challenge–response, standard by design.
Nothing is stored, so nothing can be stolen — and a failed read degrades, it does not break.
Standard elliptic curve, FIDO2-compliant signatures.
Never exposed in clear text, not even in RAM — protected for the instant it signs, then destroyed.
No keys, seeds or templates on handset or server.
Camera silicon → SDK → App → Server / IdP
- 1. Authentication request
- 2. Challenge (nonce)
- 3. Sign (keyID, challenge)
- 4. Acquire sensor RAW frames
- 5. Extract PRNU sensor noise
- 6. Reconstruct EC key · sign · zeroize
- 7. Signature returned to the app
- 8. Signature sent to server / IdP
- 9. Result — verified against the stored public key
Threat model
Multi-layer anti-tampering.
Every layer above the silicon can be attacked — the silicon layer cannot be simulated.
| Threat vector | Attack mechanism | ToothPic mitigation |
|---|---|---|
| Rooted / jailbroken devices | Admin access to OS memory and key stores. | Ephemeral keys never touch the OS keystore, disk or persistent memory. |
| Emulator and hooking attacks | Reverse engineering via Frida, Xposed or virtual environments. | Anti-debugging and binary obfuscation. Virtual devices carry no PRNU, so key reconstruction fails deterministically. |
| Synthetic frame injection | Synthetic video injected straight into vendor APIs or SDKs. | Operates below software liveness. Injected frames carry no photodetector noise, so signing fails at the physics layer. |
| Physical camera spoofing | Deepfake replayed on a high-resolution second screen. | Screen detection analyses moiré patterns, reflection glare and spatial noise profiles to reject re-photographed presentations. |
Integration
Beside your fraud stack — not instead of it.
A signal into the rules, scores and case queues you already run. Standard FIDO2 and elliptic-curve under the hood — no rip-and-replace, no migration.
- No new platform — a signal into the rules, scores and case queues you already run. No rip-and-replace, no migration.
- No database changes — one public key per device in a directory you control. No new identity store, no user secrets.
- 100% on-device — no biometric template or private key ever leaves the handset. GDPR-clean.
- Weeks, not quarters — a few days of mobile work, feature-flagged, with a kill switch from day one.
Where the signal goes is your choice — a score input, a hard rule, or a flag on a case queue. It also plugs into the identity provider you already run (Okta, Keycloak, Duende, Microsoft, F5, Shibboleth or any standards-based IdP; the list is illustrative). ECDSA · NIST P-256 signatures, FIDO2-compliant.
Comparison
The only signal that is deterministic.
Every other row produces a score you must tune. This one produces an answer.
| Signal | DeterministicYes / no — not a score | Survives resetSame ID after a wipe | Defeats injectionBlocks faked camera feeds | Account linkingMany accounts, one phone |
|---|---|---|---|---|
| Device intelligenceProbabilistic telemetry | ✕ | ✕ | ✕ | ✕ |
| Behavioural biometricsStatistical model | ✕ | ✕ | ✕ | ✕ |
| Document & livenessPresentation attacks only | ✕ | ✓ | ✕ | ✕ |
| SMS OTP / SIMTransferable by design | ✕ | ✕ | ✕ | ✕ |
| Cloud passkeysSynced by design | ✕ | ✓ | ✕ | ✕ |
| ToothPicBound to camera silicon | ✓ | ✓ | ✓ | ✓ |
Edge cases & FAQ
Physics over optics.
Optics can be dark, dirty or damaged — the silicon underneath does not change.
Does it work in low light?
Yes. PRNU measures photodetector-level silicon noise — photon-to-electron conversion variance — not scene brightness or image detail. Normalisation across gain levels keeps key reconstruction accurate in the dark.
What about dirty, scratched or cracked lenses?
Scratches, dust and damaged glass exist purely on the optical plane. Signal processing strips optical artefacts and visual content, isolating noise on the sensor die itself.
Do OS updates or hardware ageing change the key?
No. PRNU is a physical property baked into the silicon at manufacture. It stays invariant across the multi-year device lifecycle — unaffected by OS updates, thermal wear or physical ageing.
What if camera permission is denied?
ToothPic runs on the standard OS camera permission and never stores or captures a photo. If permission is denied or revoked, the SDK raises an error so the host app can fall back gracefully.
Is any personal or biometric data processed?
No biometric template and no personal data leave the device — only a standard public key is registered with your IdP.
Is ToothPic quantum-safe?
ToothPic combines a physical component with a standard digital-signature algorithm. Quantum-safe signature algorithms exist and can be integrated with ToothPic. Read the full answer.
Want the full technical deep-dive?
Our engineers will walk your security architects through the threat model, integration guide and certification documents.