TLDR: Meta has launched ‘Video Seal’, an open-source tool to embed invisible watermarks in AI-generated videos to combat deepfakes and enable verification of content origin. This technology aims to build trust, protect intellectual property, and is freely available, encouraging wide adoption and collaborative improvement.
The internet faces a growing challenge: the rapid spread of AI-generated deepfakes. These sophisticated fake videos are becoming harder to spot and pose a real threat to the reliability of online information. To tackle this, Meta has launched ‘Video Seal’, an open-source tool designed to add invisible watermarks to videos created by artificial intelligence. The aim is to make it easier to verify the origin and trustworthiness of video content.
How Meta Video Seal Works:
Meta’s Video Seal embeds a hidden watermark into each and every frame of a video. This process is designed to be completely imperceptible to the human eye and, according to researchers, does not negatively affect the video’s quality. Furthermore, Video Seal allows for the embedding of a secret message within the video, which can be later revealed to confirm where the content originated. The technology is built to be strong and resilient against common video editing techniques, including:
- Blurring
- Cropping
- Video compression
The system behind Video Seal has two main parts: an ’embedder’, which inserts the watermark and optional hidden message, and an ‘extractor’, which retrieves this information for verification. Meta has made the Video Seal model, along with a research paper, training code, inference code, pre-trained models, and a public web demonstration, openly available. The system uses a technique called “temporal watermark propagation,” which efficiently watermarks videos by processing only certain frames.
Meta has released a public demo of Video Seal so that you can try embedding watermarks into videos and testing its features.
The increasing ease of access to AI tools has led to a surge in deepfakes, which can be used to spread misinformation, commit fraud, and impersonate people. As AI gets even better, it will become even more difficult to distinguish real videos from fakes.
Meta’s creation and open-sourcing of Video Seal is a proactive step to fight this trend by providing a way to trace and verify video content. This effort is part of a wider push in the industry to address the issues caused by AI-generated media, with similar tools being developed by companies like Google (SynthID) and Microsoft, and collaborative initiatives such as the Coalition for Content Provenance and Authenticity (C2PA).
Meta’s decision to release Video Seal as open source under a permissive MIT license is intended to:
- Encourage widespread use across the tech industry and by content creators.
- Foster collaboration and innovation within the research community through open access to the code and models.
- Promote a more transparent and collaborative approach to tackling the challenges of AI-generated content.
To further this aim, Meta has also launched the ‘Meta Omni Seal Bench’, a public leaderboard for comparing the effectiveness of different neural watermarking methods. Additionally, Meta plans to host a workshop on watermarking at ICLR in 2025, with the goal of encouraging collaboration between researchers and developers in this vital area.
While Video Seal is a significant advancement in video watermarking, Meta recognises that it has limitations. A key challenge is the natural trade-off between how invisible the watermark is and how well it resists tampering. Strong video compression and significant editing might alter or make the watermarks unrecoverable. Meta has identified areas for future improvement, including making the watermarks more visually consistent across video frames and increasing the amount of information that can be embedded. Future work may also focus on embedding watermarks in formats more suited to video compression and further improving the robustness and security of the system.
It’s important to remember that Video Seal is part of Meta’s larger set of ‘Seal’ tools aimed at boosting content authenticity. This includes Audio Seal, which serves a similar purpose for watermarking audio files, and Watermark Anything.
Benefits for Small and Medium-Sized Businesses (SMBs):
For SMBs, especially those involved in creating content, media, and marketing, Video Seal offers several potential advantages:
- Protecting Intellectual Property: SMBs creating original videos can use Video Seal to embed watermarks, helping to protect their copyright and prevent unauthorised use of their visual content.
- Combating Misinformation: Businesses can use Video Seal to authenticate their official video communications, assuring customers and stakeholders of the content’s legitimacy and helping to reduce the spread of misinformation associated with their brand.
- Building Digital Trust: By using watermarking technology, SMBs can demonstrate a commitment to transparency and authenticity, potentially building greater trust with their audience in an age of increasing digital manipulation.
- Cost-Effective Solution: As an open-source tool, Video Seal is available without licensing fees, making advanced video watermarking technology more accessible to SMBs compared to proprietary options.
- Flexible Integration: The open-source nature of Video Seal allows developers within SMBs or those they hire to integrate the technology into their existing content workflows and software, potentially simplifying the process of watermarking their video content.
By making Video Seal an open-source project, Meta aims to contribute to a more trustworthy digital environment, offering benefits not only to large platforms but also to individual creators and businesses of all sizes.
Sources: Meta’s ‘Video Seal: Open and Efficient Video Watermarking’ Paper, Meta Blog Article.
