Cornell Researchers Develop Light-Based Watermark to Combat AI Deepfakes
A team of computer scientists from Cornell University has introduced a novel method to detect manipulated videos, including AI-generated deepfakes and altered clips. The technique, known as noise-coded illumination, embeds verification data into light sources, making it possible to identify tampered footage.
The system works by introducing a subtle flicker in lighting that cameras capture but human eyes cannot perceive. Each light source emits a unique code, creating a hidden watermark within the video. If the footage is later altered, investigators can decode these light patterns to determine whether the content has been manipulated.
For instance, in a scenario like a press conference under controlled lighting, the studio lights would flicker with distinct codes. If a clip from that event is later shared with suspicious edits, forensic analysis could reveal inconsistencies in the light patterns, exposing the tampering.
Abe Davis, assistant professor of computer science at Cornell, explained that the watermark contains time-stamped, low-quality versions of the original video under different lighting conditions. When a video is altered, these embedded code videos show discrepancies, highlighting where changes were made. AI-generated fake videos, meanwhile, produce random, nonsensical code videos, making them easier to detect.
While the method is not foolproof—rapid motion and bright sunlight can interfere with its effectiveness—it shows promise for controlled environments like interviews, speeches, or conference recordings. The research was published in ACM Transactions on Graphics and will be presented at SIGGRAPH in August.
This innovation could become a valuable tool in the ongoing battle against digital misinformation, offering a way to verify the authenticity of video content in an era of increasingly sophisticated AI manipulation.
