Welcome to the Watermarking LLMs 101 repository! This repository serves as a central hub for educational resources, research papers, and community contributions focused on the study and application of watermarking techniques in Large Language Models (LLMs). Whether you're a student, researcher, or enthusiast, this repository is designed to provide you with comprehensive materials to deepen your understanding of LLM watermarking.
Explore our interactive Watermarking LLMs 101 Course to engage with the material in a dynamic learning environment.
This section lists the essential research papers on LLM watermarking. For a more detailed summary and discussion, refer to the respective links.
| Date | Paper Title | Authors | Summary | Link |
|---|---|---|---|---|
| April 16, 2024 | Topic-based Watermarks for LLM-Generated Text | Alexander Nemecek et al. | Introduces topic-based watermarking for differentiating between LLM- and human-generated text. | Read more |
| April 5, 2024 | Have You Merged My Model? | Tianshuo Cong et al. | Discusses the robustness of LLM IP protection against model merging. | Read more |
| April 2, 2024 | An Entropy-based Text Watermarking Detection Method | Yijian Lu et al. | Proposes an Entropy-based Watermark Detection (EWD) method that adjusts the detection weights based on token entropy, improving performance in low-entropy scenarios. | Read more |
| April 2, 2024 | A Statistical Framework of Watermarks for LLMs | Xiang Li et al. | Proposes a statistical framework for watermarking LLMs, focusing on detection efficiency and optimal rules. | Read more |
| March 19, 2024 | Bypassing LLM Watermarks with Color-Aware Substitutions | Qilong Wu, Varun Chandrasekaran | Introduces SCTS, a color-aware attack that effectively bypasses state-of-the-art watermarking by substituting tokens based on their "color" information. | Read more |
| March 18, 2024 | Towards Better Statistical Understanding of Watermarking LLMs | Zhongze Cai et al. | Discusses an optimization-based approach for watermarking LLMs, balancing model distortion and detection ability with new insights into the green-red algorithm. | Read more |
| March 15, 2024 | WaterJudge: Quality-Detection Trade-off | Piotr Molenda et al. | Analyzes the trade-offs between watermark detectability and the quality of generated texts. | Read more |
| March 12, 2024 | Duwak: Dual Watermarks in Large Language Models | Chaoyi Zhu et al. | Introduces Duwak, a method that embeds dual watermarks enhancing detection efficiency and text quality, significantly reducing the tokens needed for reliable detection. | Read more |
| March 10, 2024 | PiGW: A Plug-in Generative Watermarking Framework | Rui Ma et al. | Proposes PiGW, a framework that integrates watermarks into generative images with minimal quality loss and high security against noise attacks. | Read more |
We encourage contributions from the community! If you have suggestions, research papers, or educational materials related to LLM watermarking, please feel free to contribute.