E-Paper
Deepfakes are really dangerous but this small process could make them easier to catch
Deepfakes are getting harder to spot. A simple lighting step during recording hides a code the camera picks up. Later, edits can show up as breaks in that code.
Worried about deepfakes and how they can impact our societies, systems, and political processes? This simple fix could be the key to address this issue. A team at Cornell University has shown that you can watermark reality itself by using light, not software. Instead of embedding a signature in a file that a bad actor can strip or ignore, they embed a quiet code in the scene while you record. Lamps or panels fitted with a tiny controller nudge brightness in patterns that people do not notice, yet cameras do. The camera captures those fluctuations as part of the image. Later, anyone with the matching key can recover a low fidelity code stream from the footage and check whether it lines up with the scene. If a face was swapped, an object pasted in, or a section reshot, the code in that region will not match. What you get is a built in authenticity check that travels with the frames and does not rely on downstream cooperation from platforms or models.
How the light code works and why it helps
At capture, the system gently modulates one or more lights with a pseudorandom sequence. The variations sit below human perception, so the scene looks normal to viewers in the room and on camera. Because the camera sensor integrates that light, the code becomes part of every frame. During verification, software extracts a reference signal from the footage and compares it with the expected pattern. A clean match says the scene was recorded under the coded lights. A mismatch highlights regions that do not belong. The clever twist is that you can run different codes on different fixtures in the same scene. That makes life difficult for forgers because any edit has to respect multiple overlapping light signatures, frame by frame, across moving subjects and changing shadows, according to Interesting Engineering. File based watermarks and metadata have never solved this. They depend on compliant software and can be stripped, reencoded, or never added. A light borne signature raises the bar in settings where truth matters most, such as interviews, debates, press briefings, and courtroom recordings. It does not stop every attack, and it will not fix content that was never lit with the system, but it shifts trust earlier in the chain and makes convincing edits costlier and slower to produce.
Where this could land next
The practical upside is that you do not need to replace cameras. You retrofit lights. A postage stamp sized controller can live inside a studio panel, a conference room downlight, or a stage fixture. Newsrooms can light sets with coded patterns that look normal on air. Event organisers can enable coded lighting for high stakes appearances without changing run sheets. Fact checkers can ask sources to supply a short verification clip alongside raw footage, which speeds reviews and reduces guesswork. Standards bodies can define open keys and audit trails so that verification scales beyond a single lab and works across vendors. None of this is a silver bullet. Lighting can drift. Keys can leak. Outdoor scenes are harder to control, and the method needs care around skin tone rendering and flicker. The Cornell team frames it as a layer, not a lock. Pair it with provenance logs, capture time attestations, and robust forensic models, and you get a defence in depth that prioritises trust at the moment of recording rather than a late scramble after a video goes viral. A watermark carried by photons is a refreshingly simple way to make fakes work harder and truth easier to prove.
