[Paper Review] A Survey of Text Watermarking in the Era of Large Language Models

Aiwei Liu, Leyi Pan+8 more|arXiv (Cornell University)|Dec 13, 2023

Advanced Steganography and Watermarking Techniques4 citations

TL;DR

This survey provides a comprehensive analysis of text watermarking in the era of large language models (LLMs), examining how LLMs enhance watermarking techniques and are themselves protected via watermarking. It proposes a unified benchmark for evaluation, identifies key challenges in adoption, and outlines future research directions for robust, scalable, and trustworthy watermarking in AI-generated text.

ABSTRACT

Text watermarking algorithms are crucial for protecting the copyright of textual content. Historically, their capabilities and application scenarios were limited. However, recent advancements in large language models (LLMs) have revolutionized these techniques. LLMs not only enhance text watermarking algorithms with their advanced abilities but also create a need for employing these algorithms to protect their own copyrights or prevent potential misuse. This paper conducts a comprehensive survey of the current state of text watermarking technology, covering four main aspects: (1) an overview and comparison of different text watermarking techniques; (2) evaluation methods for text watermarking algorithms, including their detectability, impact on text or LLM quality, robustness under target or untargeted attacks; (3) potential application scenarios for text watermarking technology; (4) current challenges and future directions for text watermarking. This survey aims to provide researchers with a thorough understanding of text watermarking technology in the era of LLM, thereby promoting its further advancement.

Motivation & Objective

• To analyze the synergistic relationship between large language models (LLMs) and text watermarking technologies.
• To identify key challenges in deploying text watermarking for LLM-generated content, including quality degradation and lack of provider engagement.
• To address the lack of public trust and transparency in watermarking detection mechanisms.
• To propose a standardized benchmark for evaluating watermarking algorithms under unified metrics.
• To guide future research toward robust, scalable, and ethically sound watermarking solutions for AI-generated text.

Proposed method

• Systematically categorizes text watermarking techniques into three paradigms: watermarking LLM-generated text, using LLMs to generate watermarks, and integrating watermarking directly into LLM architectures.
• Proposes a benchmarking framework with standardized evaluation metrics, including robustness, payload capacity, and semantic quality preservation.
• Analyzes existing watermarking algorithms through the lens of semantic integrity, imperceptibility, and resistance to common attacks.
• Reviews the integration of watermarking in real-world applications such as copyright protection, plagiarism detection, and fake news mitigation.
• Evaluates the role of LLMs in enabling more semantically aware watermarking by leveraging contextual understanding and generation control.
• Highlights the need for third-party verification and regulatory frameworks to enhance transparency and public trust in detection systems.

(a) A description of how LLMs promote the development of text watermarking techniques and broaden their application scenarios.

Experimental results

Research questions

• RQ1 How can text watermarking be effectively applied to content generated by large language models to ensure traceability and ownership?
• RQ2 In what ways do large language models enhance the robustness and semantic fidelity of text watermarking techniques?
• RQ3 What are the key technical and non-technical barriers limiting the adoption of text watermarking by LLM providers?
• RQ4 How can public trust in watermarking detection be improved through transparency and independent verification?
• RQ5 What future research directions are needed to ensure watermarking remains effective against evolving LLM-based attacks?

Key findings

• LLMs significantly improve watermarking by enabling semantic-aware embedding, minimizing distortion while preserving meaning.
• Current watermarking algorithms often reduce text quality, creating a trade-off between robustness and fluency that remains unresolved.
• Large language model providers show limited engagement in adopting watermarking, primarily due to concerns over service quality and unclear ROI.
• Public trust in watermarking is hindered by lack of transparency in detection algorithms and potential conflicts of interest in proprietary systems.
• A standardized benchmark for evaluating watermarking algorithms is essential to enable fair comparison and accelerate research progress.
• Future watermarking systems must be resilient to sophisticated attacks and adaptable to evolving LLM capabilities, especially in high-stakes domains like journalism and education.

(b) Number of publications in the field of text watermarking and LLMs (the data for "Number of Publications in the field of LLMs" is sourced from Zhao et al . ( 2023c ) )