ML Security PapersMarkDiffusion: An Open-Source Toolkit for Generative Watermarking of Latent Diffusion Models
Leyi Pan 1, Sheng Guan 1,2, Zheyu Fu 1, Luyang Si 1, Huan Wang 1, Zian Wang 1, Hanqian Li 3, Xuming Hu 3, Irwin King 4, Philip S. Yu 5, Aiwei Liu 1, Lijie Wen 1
2 Beijing University of Posts and Telecommunications
3 The Hong Kong University of Science and Technology
Published on arXiv
2509.10569
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Provides 24 evaluation tools and 8 automated pipelines covering detectability, robustness, and output quality for generative watermarking of latent diffusion models
MarkDiffusion
Novel technique introduced
We introduce MarkDiffusion, an open-source Python toolkit for generative watermarking of latent diffusion models. It comprises three key components: a unified implementation framework for streamlined watermarking algorithm integrations and user-friendly interfaces; a mechanism visualization suite that intuitively showcases added and extracted watermark patterns to aid public understanding; and a comprehensive evaluation module offering standard implementations of 24 tools across three essential aspects - detectability, robustness, and output quality - plus 8 automated evaluation pipelines. Through MarkDiffusion, we seek to assist researchers, enhance public awareness and engagement in generative watermarking, and promote consensus while advancing research and applications.
Key Contributions
- Unified implementation framework for streamlined integration of generative watermarking algorithms with user-friendly interfaces
- Mechanism visualization suite that displays watermark embedding and extraction patterns for transparency and public understanding
- Comprehensive evaluation module with 24 standard tools across detectability, robustness, and output quality, plus 8 automated evaluation pipelines
🛡️ Threat Analysis
Focuses on generative watermarking of latent diffusion model outputs — embedding and detecting marks in generated images for content provenance and authenticity. Evaluation covers detectability, robustness (to attacks on the watermark), and output quality, which are all output integrity concerns.