Meta Launches AI-Generated Video Watermarking Tool

The Proliferation of Deepfakes and Meta's Response

The prevalence of deepfakes is rapidly increasing. The increased accessibility of generative AI technologies has resulted in a significant surge of fabricated content online. Data from Sumsub, an ID verification platform, indicates a fourfold rise in deepfakes globally between 2023 and 2024.

In 2024, deepfakes were implicated in 7% of all fraudulent activities, according to Sumsub. These instances encompassed identity theft, unauthorized account access, and complex social engineering schemes.

Introducing Meta Video Seal

Meta is introducing a new tool designed to combat the spread of deepfakes. Meta Video Seal is intended to apply subtle, nearly invisible watermarks to videos created by artificial intelligence. This tool, released as open-source software, is meant for integration into existing video editing and creation platforms.

It complements Meta’s existing watermarking technologies, including Watermark Anything (re-released with a permissive license) and Audio Seal.

“Video Seal was developed to offer a more robust solution for video watermarking, specifically for identifying AI-generated videos and safeguarding content authenticity,” explained Pierre Fernandez, an AI research scientist at Meta, in a recent interview with TechCrunch.

Comparison to Existing Technologies

Video Seal is not the first attempt at video watermarking. DeepMind’s SynthID and Microsoft also have their own methodologies for embedding watermarks in videos.

However, Fernandez argues that current solutions often have shortcomings. He notes that many existing tools lack resilience against video compression – a common process when sharing content online.

Furthermore, some are not efficient enough for large-scale implementation, are not openly available for scrutiny, or are adapted from image watermarking techniques, which are less effective for video.

Key Features of Video Seal

Beyond the watermark itself, Video Seal can embed a concealed message within videos. This hidden message can be extracted to trace the video’s origin.

Meta asserts that Video Seal is resistant to typical video manipulations, such as blurring and cropping, and remains effective even after compression using popular algorithms.

Fernandez acknowledges that Video Seal isn’t without limitations. A trade-off exists between the watermark’s visibility and its ability to withstand alterations. Extensive compression or significant editing could potentially compromise or erase the watermark.

Encouraging Adoption and Collaboration

A significant challenge for Video Seal is encouraging widespread adoption, particularly among those who already utilize proprietary watermarking solutions.

To address this, Meta is launching Meta Omni Seal Bench, a public leaderboard for comparing the performance of different watermarking methods.

Additionally, Meta will host a workshop on watermarking at ICLR, a leading AI conference, later this year.

“Our goal is to encourage the integration of watermarking into the work of more AI researchers and developers,” Fernandez stated. “We are committed to collaborating with both industry and academia to accelerate progress in this critical field.”