THU-BPM/MarkDiffusion

What it solves

MarkDiffusion addresses the challenge of verifying the authenticity and origin of media generated by latent diffusion models (LDMs). As AI-generated images and videos become more common, it provides a standardized way to embed and detect invisible watermarks to track provenance and prevent misuse.

How it works

The toolkit provides a unified framework to integrate various watermarking algorithms into the generation process of diffusion models. It supports two primary categories of methods:

• Pattern-based methods: These embed specific patterns into the generated media (e.g., Tree-Ring, SFW).
• Key-based methods: These use cryptographic keys to embed watermarks (e.g.,GaussMarker, VideoMark).

Beyond embedding, it includes a visualization suite to show how watermarks are added and extracted, and an evaluation module that tests the watermarks against various "attacks" (like cropping, compression, or noise) to measure their robustness and detectability.

Who it’s for

• AI Researchers: Those developing new watermarking algorithms for generative models.
• Developers: Those needing to implement provenance tracking for AI-generated media.
• Security Auditors: People testing the robustness of generative watermarks against removal or modification.

Highlights

• Broad Algorithm Support: Implements 11 state-of-the-art watermarking algorithms for both images and videos.
• Robust Evaluation Suite: Includes 33 tools to measure detectability, output quality (using metrics like FID and CLIP score), and robustness against 16 different image and video attacks.
• Automated Pipelines: Features 8 specialized pipelines for streamlined quality and detection analysis.
• Visualization Tools: Custom tools to help users intuitively understand the internal mechanisms of different watermarking algorithms.