Meta Open-Sources Video Watermarking Tool to Combat Rising Deepfake Fraud
Meta introduces Video Seal, an open-source tool that applies imperceptible watermarks to AI-generated videos to help trace their origins. The technology aims to withstand common compression and editing techniques better than current competing solutions.
Meta releases Video Seal, a new open-source tool designed to apply imperceptible watermarks to AI-generated video clips. This launch comes as deepfake fraud skyrockets, with fake content accounting for 7% of all fraud cases in 2024. By making Video Seal open source, Meta hopes to provide developers with an accessible way to tag and identify AI videos before they spread across social platforms.
The new tool sets itself apart from existing video watermarking methods like DeepMind's SynthID by offering superior resilience against common video compression algorithms used by social media sites. In addition to a standard watermark, Video Seal embeds a hidden message within the video file that researchers can later uncover to determine the clip's exact origins. Meta researcher Pierre Fernandez notes that deriving video watermarks from image watermarking techniques often falls short, making a video-native approach necessary.
Despite its promising features, Video Seal faces limitations regarding the balance between watermark visibility and robustness against heavy manipulation. Extreme compression or significant editing potentially renders the hidden markers unrecoverable. Furthermore, Meta faces the challenge of convincing the industry to adopt this tool over existing proprietary solutions, prompting the company to launch a public leaderboard to benchmark watermarking performance.