This repository serves as a curated, up-to-date collection of research papers on watermarking techniques for Large Language Models (LLMs) and image generation models. It aims to provide researchers and practitioners with a comprehensive overview of the latest advancements, challenges, and methodologies in ensuring the provenance and intellectual property of AI-generated content.

How It Works

The repository organizes papers by domain (text and image watermarking), featuring a wide array of approaches. These include statistical methods, embedding watermarks directly into model weights or generation processes, and exploring robustness against adversarial attacks and paraphrasing. The collection highlights research into the trade-offs between watermark detectability, robustness, and output quality, as well as novel evaluation metrics and frameworks.

Quick Start & Requirements

This repository is a collection of research papers and does not involve direct installation or execution. Links to preprints (arXiv) and conference publications are provided for each paper.

Highlighted Details

• Comprehensive coverage of text watermarking techniques, including statistical frameworks, semantic invariants, and robustness against various attacks.
• Extensive collection of image watermarking methods, focusing on diffusion models, latent spaces, and concept protection.
• Inclusion of papers addressing theoretical limitations, evaluation benchmarks (e.g., WaterBench), and practical applications like copyright protection and API security.
• Features recent work on multi-bit watermarking, distortion-free methods, and cross-lingual watermark consistency.

Maintenance & Community

The repository is maintained by hzy312. It is a static collection of links to external research papers, with contributions encouraged to add new relevant works.

Licensing & Compatibility

The repository itself does not have a specified license. The licensing of the individual papers referenced within the repository varies by their respective publication venues (e.g., arXiv preprints, conference proceedings).

Limitations & Caveats

This repository is a literature aggregator and does not provide implementations or code for the watermarking techniques discussed. Users must refer to the individual papers for implementation details and potential code availability. The field is rapidly evolving, and the "up-to-date" status is dependent on ongoing curation.