Awesome LLM Watermark：大语言模型水印技术全景资料库

This post introduces the Awesome-LLM-Watermark project, a GitHub repository that systematically collects and organizes research papers, open-source projects, and technical resources related to large language model (LLM) watermarking. It covers various types of watermarking (Token-level, Sentence-level, Model-level, etc.) as well as attack and defense strategies. The repo aims to help address issues like academic integrity, fake news identification, copyright归属, and content溯源 in the age of AI-generated content (AIGC).

With the普及 of LLMs like ChatGPT and Claude, AIGC has渗透到 many aspects of life (student assignments, news, code, academic papers). This brings several problems:

1. Academic integrity: Detecting AI-written papers.
2. Fake information: Identifying sources of AI-generated fake news.
3. Copyright归属: Who owns AI-generated content?
4. Content溯源: Tracking which model generated a text. LLM watermarking solves these by embedding invisible "fingerprints" during generation, enabling source identification without affecting readability.

The repo categorizes LLM watermarking into 7 main types:

1. Token-level: Modify token sampling (e.g., green/red lists in ICML2023 paper, publicly detectable schemes, lossless via lexical redundancy).
2. Sentence-level: Use sentence embeddings (e.g., SemStamp with paraphrastic robustness).
3. Model-level: Embed in model parameters (e.g., weight quantization for IP protection).
4. Multi-modal: For multi-modal models (image+text).
5. Attack & Defense: Types like stealing/removal/spoofing attacks, and robust/anti-spoofing/multi-bit defenses.
6. CoT Watermark: For models with Chain-of-Thought reasoning.
7. Low Entropy: For low-entropy scenarios like code generation.

The repo shows the evolution path:

• 1st Gen (2023 early): Basic statistical (e.g., Kirchenbauer's work, simple but sensitive to rewriting).
• 2nd Gen (2023-2024): Semantic robust (e.g., SemStamp, resistant to paraphrasing).
• 3rd Gen (2024): Adaptive/lossless (e.g., WatME, minimal quality loss).
• 4th Gen (2024-2025): Model-level & multi-modal (focus on model IP and multi-modal content).

Key application scenarios:

1. Academic integrity: Detect AI-generated student assignments (higher accuracy than traditional detectors).
2. Content platform溯源: Embed watermarks in user content to track sources and fight fake news.
3. Model copyright: Protect model IP by embedding watermarks in parameters.
4. Compliance audit: Record content sources for enterprise AI use to meet audit requirements.

Current challenges and future focus:

1. Robustness vs Quality: Balancing resistance to attacks and text quality.
2. Multilingual Support: Improving support for non-English languages (e.g., Chinese, Arabic).
3. Long Text: Ensuring consistent detectability in long documents.
4. Adversarial Attacks: Updating schemes to counter new attack methods.
5. Standardization: Establishing industry standards for interoperability between different watermarking schemes.

Recommended reading paths for different users:

• Beginners: Start with "Survey" sections → read Kirchenbauer's paper → try open-source implementations.
• Researchers: Choose relevant categories → follow latest SOTA papers → understand attack/defense methods.
• Developers: Check open-source projects → select algorithms based on needs (quality vs robustness) → focus on performance optimization.

Awesome-LLM-Watermark is one of the most comprehensive resources in the LLM watermarking field, offering a systematic classification framework. As AIGC becomes more prevalent, watermarking will play a crucial role in content溯源, copyright protection, and compliance. It's an ideal time for researchers and developers to enter this field (mature tech, clear applications, not overly competitive). The repo is a valuable resource to bookmark and revisit regularly.