If you throw a rock, there's a good chance you'll hit a deepfake. The commoditization of generative AI has led to an explosion of fake content online. According to identity verification platform Sumsub, deepfakes have quadrupled worldwide from 2023 to 2024. According to Sumsub, deepfakes accounted for 7% of all fraud in 2024. , ranging from identity theft and account takeover to sophisticated social engineering campaigns.
Meta is releasing a tool that applies imperceptible watermarks to AI-generated video clips in hopes of making a meaningful contribution to the fight against deepfakes. The tool, announced Thursday, is called Meta Video Seal, is available as open source, and is designed to be integrated into existing software. This tool joins Meta's other watermarking tools, Watermark Anything (re-released today under a permissive license) and Audio Seal.
“We developed Video Seal to provide a more effective video watermarking solution, especially for the detection and originality protection of AI-generated videos,” said Meta's AI Research Scientist. said Pierre Fernandez in an interview with TechCrunch.
Video Seal is not the first technology of its kind. DeepMind's SynthID can watermark videos, and Microsoft has its own video watermarking method.
But Fernandes argues that many existing approaches are insufficient.
“While other watermarking tools exist, they do not provide sufficient robustness against video compression. Video compression is very common when sharing content through social platforms. It wasn't efficient enough to do. It wasn't open or reproducible or derived from image watermarks, which isn't great for video,” Fernandez said.
In addition to watermarks, Video Seal can add hidden messages to videos that can be revealed later to identify their origin. Meta claims that Video Seal is resistant to common compression algorithms as well as common edits such as blurring and cropping.
Fernandez acknowledged that Video Seal has certain limitations. It's primarily a trade-off between the tool's watermark recognizability and its overall resistance to manipulation. He added that advanced compression or heavy editing can alter or irrecoverably alter watermarks.
Of course, the bigger problem facing Video Seal is that developers and industries, especially those already using proprietary solutions, don't have much reason to adopt it. To address this, Meta launched Meta Omni Seal Bench, a public leaderboard dedicated to comparing the performance of different watermarking methods, and hosted a watermarking workshop this year at ICLR, a leading AI conference. Masu.
“We hope that more AI researchers and developers will incorporate some form of watermarking into their work,” Fernandez said. “We look forward to collaborating with industry and the academic community to accelerate progress in this area.”