An invisible watermark can be added to videos produced by artificial intelligence (AI) with a new tool that Meta is releasing. The new product, called Video Seal, complements the company’s already-available watermarking tools, Watermark Anything and Audio Seal. Although it has not yet released the code, the business indicated that the tool would be open-sourced. Fascinatingly, the company asserts that the watermarking approach will be resistant to popular techniques for removing them from recordings while having no effect on the quality of the video.
Since the advent of generative AI, deepfakes have proliferated online. Deepfakes are artificial intelligence (AI)-generated synthetic content that depicts fictitious and deceptive things, people, or situations. Such content is frequently used to perpetrate fraud and frauds, produce fictitious sexual content, or disseminate false information about public figures.
Furthermore, deepfake content will become increasingly difficult to identify as AI systems advance, making it even more challenging to distinguish it from authentic content. Seventy percent of respondents to a McAfee survey already express uncertainty about their ability to distinguish between an AI-generated speech and a real voice.
The proliferation of fraudulent content on the internet is a direct result of the commercialization of generative AI. According to Sumsub, the number of deepfakes worldwide increased fourfold from 2023 to 2024. According to Sumsub, the internal data shows deepfakes will account for 7% of all fraud in 2024, with attacks ranging from impersonations and account takeovers to sophisticated social engineering efforts.
As concerns about deepfakes grow, many AI companies have begun to release watermarking technologies that can distinguish between synthetic and real content. Google introduced SynthID earlier this year, which can be used to watermark AI-generated text and films. Microsoft has also offered similar tools. In addition, the Coalition for Content Provenance and Authenticity (C2PA) is developing new standards to identify AI-generated content.
Meta has released a tool for applying invisible watermarks to AI-generated video clips, which it expects will make a significant contribution to the fight against deepfakes. The program, known as Meta Video Seal, was announced on Thursday and is free source. The researchers emphasize that the program can watermark each frame of a film with an imperceptible tag that cannot be altered. It is claimed to be resistant or can withstand common alterations to tactics like blurring, cropping, and compression software. It is also designed to be integrated into existing software. The program joins Meta’s two watermarking utilities, Watermark Anything (which was re-released today with a permissive license) and Audio Seal. Despite applying the watermark, the researchers say that the video’s quality will not be damaged.
Meta disclosed that Video Seal will be open-sourced under a permissive license, however the tool and its coding have yet to be released to the public.
Pierre Fernandez, AI research scientist at Meta, told TechCrunch, “They developed Video Seal to provide a more effective video watermarking solution, particularly for detecting AI-generated videos and protecting originality,”.
The Video Seal technology is not the first of its sort. DeepMind’s SynthID is capable of watermarking videos, and Microsoft has its own video watermarking methods.
However, Fernandez claims that many present approaches fall short.
“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,” says Fernandez.
In addition to a watermark, Video Seal can add a hidden message to videos that can subsequently be deciphered to establish their origin.
Fernandez acknowledges that Video Seal has several limits, most notably the trade-off between how visible the tool’s watermarks are and their general resistance to manipulation. He added that heavy compression and major changes could alter or leave the watermarks unrecoverable.
Of course, the main issue with Video Seal is that developers and industry will have little motive to use it, especially those who currently use proprietary solutions. To address this, Meta is releasing a public leaderboard, Meta Omni Seal Bench, to compare the performance of various watermarking approaches, as well as organizing a watermarking workshop at this year’s ICLR, a prominent AI conference.
Finally, Fernandez said, “they hope that more and more AI researchers and developers will integrate some form of watermarking into their work, and they want to collaborate with the industry and the academic community to progress faster in the field.”