Physical-layer authentication vs watermarking vs C2PA: what each can prove
Lightmark · · 3 min read

In short: most authentication methods work on the file. Physical-layer authentication works on the scene. That difference decides what can be proven, and when.
Four families of approaches
1. Deepfake detection
Classifiers look for artefacts that suggest content was generated or edited. Detection needs no cooperation from the creator, but it produces a probability, not proof, and its accuracy drops as generators improve and as content is compressed and re-shared.
2. Watermarking
A pattern is embedded into the pixels or audio of a file, usually by the tool that created it. Watermarks are useful for labelling AI output, but they are added after the content exists. They can be weakened or removed, and they say nothing about authentic footage that was never watermarked.
3. C2PA content credentials
C2PA attaches signed metadata that records where a file came from and how it was edited. It creates a useful audit trail when every tool in the chain supports it. But metadata is frequently stripped by platforms, screenshots and re-encoding, and security researchers have cautioned against relying on credentials alone in high-stakes settings.
4. Camera-embedded signing
Some cameras sign images at capture with a hardware key. This is strong evidence for that device, but it requires new hardware for every camera, and it proves which device took a picture, not where the scene was.
Physical-layer optical authentication
Lightmark takes a different starting point: the space where footage is captured. Beacons broadcast a dynamic signature, a pattern of light that changes continuously and is imperceptible to people. Any camera in the space captures it as part of the scene. Verification reads the signature from the footage, per pixel.
Because the signature lives in the light rather than in the file:
- It works with existing cameras, including third-party and bystander devices.
- It does not depend on metadata surviving distribution.
- Content that was never physically captured in the signed space, including AI-generated footage, does not carry it.
- Per-pixel reading shows where content has been altered: a region without the signature reads as an absence of signed light.
Side by side
| What it proves | Needs new cameras? | Survives re-sharing? | Proves authentic footage? | |
|---|---|---|---|---|
| Detection | Probability of manipulation | No | Degrades | No |
| Watermarking | That a tool marked this file | No | Variable | No |
| C2PA | A signed history of the file | No (needs tool support) | Often stripped | Partially |
| Camera signing | Which device captured it | Yes | Often stripped | For that device |
| Physical-layer (Lightmark) | Where and when the scene was captured | No | Designed to (validated in pilots) | Yes |
They work best together
None of these replaces the others. A robust stack labels AI output with watermarks, documents edits with C2PA, uses detection for triage, and anchors authentic footage in the physical world with a dynamic signature.
Want to see where physical-layer authentication fits in your stack? Book a briefing or read how the technology works.
See it on your footage.
Book a briefing
