Digital watermarking for generative AI
summary
Meta Seal, an open-source framework, implements invisible, robust digital watermarks for images, video, audio and text using components named PixelSeal (adversarially trained image/video watermarking), AudioSeal (real-time streaming-compatible audio watermarking) and TextSeal (text watermarking for research). The talk explains how these systems were moved from research into large-scale production for provenance and transparency on Instagram and Facebook, discusses current limitations, and outlines future research directions like dataset watermarking to trace training-data use.
AI-written summary · about
chapters (8)
- 0:00Watermarking challenges
- 5:59Watermarking principles
- 8:46Post-generation methods
- 18:44Text watermarking
- 29:13Image model watermarking
- 33:06Advanced applications
- 39:32Watermark robustness
- 43:21Industry adoption
related talks
08mSynthID – Watermarking and identifying AI-generated text
27mTurning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdooring
13mlo-fi spy: video, steganography, and you
46mTackling Deepfakes, Wars, and Other Security Threats in the GenAI World
45mGenerative AI for Image and Video Production
50mFace the Music: Cybersecurity and the Music Industry
27mTruth or Fiction The Art of Unveiling DeepFakes in Text, Audio, Image
13mNavigating Generative AI and Its Risks