Nowadays, Deepfake is everywhere. With the rise of generative AI, the amount of false content online has exploded. According to the verification platform Sumsub, the number of Deepfakes globally increased fourfold from 2023 to 2024. In 2024, Deepfakes accounted for 7% of all fraudulent activities, involving everything from identity theft and account hijacking to complex social engineering attacks.
To effectively combat Deepfake, Meta recently released a tool that adds nearly undetectable watermarks to AI-generated video clips. This tool, named Meta Video Seal, was announced as open-source on Thursday and is intended to be integrated into existing software. The tool works alongside Meta's other watermarking tools, Watermark Anything (which has been re-released today under a permissive license) and Audio Seal, forming a comprehensive watermarking solution.
Meta AI research scientist Pierre Fernandez stated in an interview with TechCrunch, "We developed Video Seal to provide a more effective video watermarking solution, especially for detecting AI-generated videos and protecting originality."
Video Seal is not the first technology of its kind. DeepMind's SynthID can add watermarks to videos, and Microsoft has its own video watermarking methods.
However, Fernandez believes that many existing methods have shortcomings.
"While there are other watermarking tools, they lack robustness in video compression (which is very common when sharing content through social platforms); their operational efficiency is insufficient for large-scale applications; they are not open or reproducible enough; or they originate from image watermarking, which is not the best choice for video," Fernandez said.
In addition to watermarks, Video Seal can also embed hidden messages in videos to reveal the source later. Meta claims that Video Seal can withstand common editing operations such as blurring and cropping, as well as common compression algorithms.
Fernandez acknowledged that Video Seal has certain limitations, mainly the trade-off between the perceptibility of the watermark and its overall resistance to manipulation. He added that high-intensity compression and significant editing may alter the watermark or make it irretrievable.
Of course, a bigger challenge for Video Seal is that developers and the industry have little incentive to adopt it, especially those companies already using proprietary solutions. To address this issue, Meta is launching a public leaderboard called Meta Omni Seal Bench, specifically for comparing the performance of various watermarking methods. Additionally, Meta will organize a workshop on watermarking at this year's ICLR (International Conference on Learning Representations), which is a major AI conference.
"We hope that more and more AI researchers and developers will integrate some form of watermarking into their work," Fernandez said. "We want to collaborate with industry and academia to accelerate progress in this field."
Meta's initiative undoubtedly provides new ideas for combating the proliferation of Deepfakes. Open-source and collaborative efforts may be effective ways to tackle this tricky problem. However, the future of Video Seal still depends on its widespread acceptance and application in the industry, as well as its ability to counter the ever-evolving threats posed by Deepfakes.
