If you throw a stone, you'll probably come across a deepfake. The commercialization of generative AI has led to a veritable explosion of pretend online content: According to ID verification platform Sumsub, there was a fourfold increase in deepfakes worldwide from 2023 to 2024. In 2024, deepfakes accounted for 7% of all scams, starting from identity theft and account takeovers to stylish social engineering campaigns, in response to Sumsub.
Hoping to make a meaningful contribution to the fight against deepfakes, Meta is releasing a tool to use imperceptible watermarks to AI-generated video clips. The tool, called Meta Video Seal, announced Thursday, is offered as open source and designed to be integrated into existing software. The tool complements Meta's other watermarking tools, Watermark Anything (re-released today under a permissive license) and Audio Seal.
“We developed Video Seal to offer a more practical video watermarking solution, especially for recognizing AI-generated videos and protecting originality,” said Pierre Fernandez, AI researcher at Meta, in an interview with TechCrunch.
Video Seal isn’t the primary technology of its kind. DeepMind's SynthID can watermark videos, and Microsoft has its own methods for video watermarking.
However, Fernandez contends that many existing approaches are inadequate.
“Although other watermarking tools exist, they don’t provide sufficient robustness to video compression, which is widely used when sharing content across social platforms. weren’t efficient enough to run at scale; weren’t open or reproducible; or were derived from image watermarks, which isn’t optimal for videos,” Fernandez said.
In addition to a watermark, Video Seal can tag videos with a hidden message that may later be revealed to find out their origin. Meta claims that Video Seal is proof against common edits similar to blurring and cropping, in addition to common compression algorithms.
Fernandez acknowledges that Video Seal has certain limitations, most notably the trade-off between the detectability of the tool's watermarks and their general resistance to tampering. Heavy compression and significant editing could alter the watermarks or render them irretrievable, he added.
Of course, the larger problem for Video Seal is that developers and industry have little reason to adopt it, especially those already using proprietary solutions. To address this, Meta is launching a public leaderboard, Meta Omni Seal Bench, dedicated to comparing the performance of various watermarking methods, and is organizing a workshop on watermarking at ICLR, a significant AI conference, this yr.
“We hope to see increasingly AI researchers and developers incorporate some type of watermarking into their work,” Fernandez said. “We need to work with industry and the tutorial community to make faster progress on this area.”
