Meta has unveiled a new tool designed to combat the growing problem of deepfakes, which are generated by artificial intelligence (AI), by applying imperceptible watermarks to AI-created videos.
The tool, called Video Seal, was announced and is available in open-source format.
It is intended to be integrated into existing software systems and aims to provide a robust solution for identifying AI-generated content.
Meta's other watermarking tools, including Watermark Anything and Audio Seal, have also been re-released under a permissive license.
“We developed Video Seal to provide a more effective video watermarking solution, particularly for detecting AI-generated videos and protecting originality,” Pierre Fernandez, an AI research scientist at Meta, said in an interview with TechCrunch.
Video Seal is not the first of its kind; DeepMind’s SynthID and Microsoft also offer video watermarking solutions. However, Fernandez argues that many existing tools fall short in key areas.
“While other watermarking tools exist, they don’t offer sufficient robustness to video compression, which is very prevalent when sharing content through social platforms; weren’t efficient enough to run at scale; weren’t open or reproducible; or were derived from image watermarking, which is suboptimal for videos,” Fernandez explained.
In addition to embedding a watermark, Video Seal can also insert a hidden message into videos.
This message can later be uncovered to determine the video’s origin. Meta claims that the watermarking method is resilient to common edits such as blurring, cropping, and popular compression algorithms.
Despite its promise, Fernandez acknowledged that Video Seal does have limitations, especially in cases of heavy compression or significant edits. In these instances, the watermark may be altered or lost altogether.
The wider adoption of Video Seal could be challenging, as developers and industry professionals may already be using proprietary solutions.
To encourage its uptake, Meta is launching the Meta Omni Seal Bench, a public leaderboard that compares the performance of various watermarking methods.
It will also organise a workshop on watermarking at ICLR, a leading AI conference, later this year.
“We hope that more and more AI researchers and developers will integrate some form of watermarking into their work,” Fernandez added. “We want to collaborate with the industry and the academic community to progress faster in the field.”
