Throw a stone and also you’ll doubtless hit a deepfake. The commoditization of generative AI has led to an absolute explosion of pretend content material on-line: In keeping with ID verification platform Sumsub, there’s been a 4x enhance in deepfakes worldwide from 2023 to 2024. In 2024, deepfakes accounted for 7% of all fraud, per Sumsub, starting from impersonations and account takeovers to stylish social engineering campaigns.
In what it hopes shall be a significant contribution to the combat in opposition to deepfakes, Meta is releasing a software to use imperceptible watermarks to AI-generated video clips. Introduced on Thursday, the software, referred to as Meta Video Seal, is on the market in open supply and designed to be built-in into current software program. The software joins Meta’s different watermarking instruments, Watermark Something (re-released at the moment below a permissive license) and Audio Seal.
“We developed Video Seal to offer a simpler video watermarking resolution, significantly for detecting AI-generated movies and defending originality,” Pierre Fernandez, AI analysis scientist at Meta, advised TechCrunch in an interview.
Video Seal isn’t the primary know-how of its sort. DeepMind’s SynthID can watermark movies, and Microsoft has its personal video watermarking methodologies.
However Fernandez asserts that many current approaches fall quick.
“Whereas different watermarking instruments exist, they don’t supply ample robustness to video compression, which may be very prevalent when sharing content material by means of social platforms; weren’t environment friendly sufficient to run at scale; weren’t open or reproducible; or have been derived from picture watermarking, which is suboptimal for movies,” Fernandez mentioned.
Along with a watermark, Video Seal can add a hidden message to movies that may later be uncovered to find out their origins. Meta claims that Video Seal is resilient in opposition to frequent edits like blurring and cropping, in addition to common compression algorithms.
Fernandez admits that Video Seal has sure limitations, primarily the trade-off between how perceptible the software’s watermarks are and their total resilience to manipulation. Heavy compression and vital edits could alter the watermarks or render them unrecoverable, he added.
In fact, the larger downside dealing with Video Seal is that devs and trade received’t have a lot cause to undertake it, significantly these already utilizing proprietary options. In a bid to deal with that, Meta is launching a public leaderboard, Meta Omni Seal Bench, devoted to evaluating the efficiency of assorted watermarking strategies, and organizing a workshop on watermarking this yr at ICLR, a serious AI convention.
“We hope that an increasing number of AI researchers and builders will combine some type of watermarking into their work,” Fernandez mentioned. “We need to collaborate with the trade and the educational group to progress quicker within the discipline.”