Flickering Lights Could Help Fight Misinformation

A group of Cornell machine scientists has unveiled what they judge could beryllium a caller instrumentality successful nan conflict against AI‑generated video, deepfakes and doctored clips.

The watermarking technique, called “noise‑coded illumination,” hides verification information successful ray itself to thief investigators spot doctored videos. The approach, devised by Peter Michael, Zekun Hao, Serge Belongie and adjunct professor Abe Davis, was published successful nan June 27 rumor of ACM Transactions connected Graphics and will beryllium presented by Michael astatine SIGGRAPH connected August 10.

The strategy adds a hardly perceptible flicker to ray sources successful a scene. Cameras grounds this pseudo-random shape moreover though viewers cannot observe it, and each lamp aliases surface that flickers carries its ain unsocial code.

As an example, ideate a property convention filmed successful nan White House briefing room. The workplace lights would beryllium programmed to flicker pinch unsocial codes. If a viral clip from that property convention later circulates pinch what appears to beryllium an inflammatory statement, investigators tin tally it done a decoder, and by checking whether nan recorded ray codes statement up, could find whether nan footage was doctored.

“Each watermark carries a low‑fidelity, time‑stamped type of nan unmanipulated video nether somewhat different lighting. We telephone these codification videos,” said Abe Davis, adjunct professor of machine subject astatine Cornell. “When personification manipulates a video, nan manipulated parts commencement to contradict what we spot successful these codification videos, which lets america spot wherever changes were made. And if personification tries to make clone video pinch AI, nan resulting codification videos conscionable look for illustration random variations."

While nan scientists admit that accelerated mobility and beardown sunlight tin inhibit nan technique’s efficacy, they are bullish connected its inferior successful settings for illustration conference‑room presentations, tv interviews aliases lecture‑hall speeches.