Hiding in light
The video was taken in a conference hall with two cryptocurrencies.
Peter Michael and others, 2025
Preparation for outdoor capture
Peter Michael and others, 2025
Preparation for outdoor capture
Peter Michael and others, 2025
The video was taken in a conference hall with two cryptocurrencies.
Peter Michael and others, 2025
Preparation for outdoor capture
Peter Michael and others, 2025
Previously, Cornell team discovered how small changes are made on the selected pixel units to see if a video has been processed or created by artificial intelligence. But its success depends on the video creator using a specific camera or AI. Their new method, “Clear Lighting” (NCI), deals with these extreme aspects and other shortcomings by hiding watermarks in the noise shown of light sources. A small program of software can do this for computer screens and certain types of room lighting, while ready -made lamps can be coded through a small connected computer slice.
“Each watermark carries a low version of the non -reliable video time under a slightly different lighting. We call these code videos,” Davis said. “When someone manipulates a video clip, the parts that are dealt with in contradiction begin with what we see in these symbol videos, which allow us to know where the changes were made. If someone tries to create a fake video with artificial intelligence, the resulting videos seem to be random differences.” Since the watermark is designed to look like a noise, it is difficult to discover it without knowing the secret symbol.
Cornell team tested its way with a wide range of manipulations: changing the cuts in warp, speed and acceleration, for example, and deep items. Their technology has proven strong to things like signal levels without human perception; The subject and movement of the camera; Flash Camera Human topics with different skin tones; Various levels of video pressure; And internal and external settings.
“Even if the opponent knows that the technique is used and spent in one way or another, their job is still much more difficult,” Davis said. “Instead of falsifying the light for only one video, they must falsify each video symbol separately, and all of that fake must agree with each other.” However, Davis added, “This is an important continuous problem. It will not disappear, and in fact it will be more difficult.”
DOI: ACM transactions on graphics, 2025. 10.1145/3742892 (About Dois).