Watermarking across Modalities for Content Tracing and Generative AI

List of contributions

Images

• Watermarking Images in Self-Supervised Latent Spaces
  An approach to embed invisible watermarks in images using self-supervised learning features.
  • 📄 arxiv.org/abs/2112.09581
  • 💻 github.com/facebookresearch/ssl_watermarking
• Active Image Indexing
  A method for efficient better image fingerprinting and retrieval by modifying images before their release.
  • 📄 arxiv.org/abs/2210.10620
  • 💻 github.com/facebookresearch/active_indexing
• The Stable Signature
  A framework for watermarking diffusion models by embedding binary messages in their latent decoder.
  • 📄 arxiv.org/abs/2303.15435
  • 💻 github.com/facebookresearch/stable_signature
• Watermark Anything With Localized Messages
  Models to embed localized watermarks in images, enabling the extraction of one or multiple messages from different regions of an image, and the segmentation of the watermark’s presence.
  • 📄 arxiv.org/abs/2411.07231
  • 💻 github.com/facebookresearch/watermark-anything

Audio

• Proactive Detection of Voice Cloning with Localized Watermarking
  Model for localized watermarking to detect AI-generated speech with precision. The detector directly predicts watermark presence for each time step, making it fast and suitable for voice cloning applications.
  • 📄 arxiv.org/abs/2401.17264
  • 💻 github.com/facebookresearch/audioseal
• Latent Watermarking of Audio Generative Models
  A method to watermark audio generative models by watermarking their training data in a way that is robust to the audio tokenizer.
  • 📄 arxiv.org/abs/2409.02915

Text

• Three Bricks to Consolidate Watermarks for Large Language Models
  Three key improvements to watermarking methods for LLMs: theoretically grounded statistical tests that guarantee false positive rates, evaluation on classical NLP benchmarks, and extension to multi-bit watermarking.
  • 📄 arxiv.org/abs/2308.00113
  • 💻 github.com/facebookresearch/three_bricks
• Watermarking Makes Language Models Radioactive
  Demonstrates that training on watermarked text can be easily detected, linking contamination level to watermark robustness and fine-tuning process. Shows that training on watermarked synthetic instructions can be detected with high confidence.
  • 📄 arxiv.org/abs/2402.14904
  • 💻 github.com/facebookresearch/radioactive-watermark

Model

• Functional Invariants to Watermark Large Transformers
  A method to watermark transformer models using functional invariants while preserving model utility.
  • 📄 arxiv.org/abs/2310.11446

Miscellaneous

• DINOv2
  Self-supervised vision model that learns robust visual features without any supervision.
  • 📄 arxiv.org/abs/2304.07193
  • 💻 github.com/facebookresearch/dinov2
• A Cookbook of Self-Supervised Learning
  A comprehensive guide to self-supervised learning methods, best practices, and implementation details.
  • 📄 arxiv.org/abs/2304.12210
• Seamless: Multilingual Expressive and Streaming Speech Translation
  A unified framework for multilingual speech translation supporting streaming and expressive speech.
  • 📄 arxiv.org/abs/2312.05187
  • 💻 github.com/facebookresearch/seamless_communication

Links

• arXiv

Slides