大语言模型水印技术研究进展

doi:10.3969/j.issn.1671-1122.2025.02.001

摘要/Abstract

摘要：

目前大语言模型LLM在文本生成、机器翻译和情感分析等领域取得了显著的成果。为了保护模型数据集与参数版权，防止未经授权的复制和使用，并验证消息的真实性，需要通过水印技术确保LLM的安全性和可信度。根据LLM运行的不同时间点，文章将当前水印技术分为嵌入模型训练的水印、推理阶段插入的水印和文本生成后的追加水印3类。针对水印的鲁棒性、保密性和有效性需求，文章对水印技术的评价指标进行了整理，并对现存的抗水印攻击进行综述，旨在进一步推动大语言模型水印技术的发展和应用。

关键词: 大语言模型, 文本水印, 权重保护, AI鉴别

Abstract:

Currently Large Language Model (LLM) has achieved remarkable results in the fields of text generation, machine translation and sentiment analysis. In order to protect the model dataset and parameter copyrights, prevent unauthorized copying and use, and verify the authenticity of messages, watermarking techniques are needed to ensure the security and trustworthiness of LLM. According to the different points in time when LLM operates, this paper categorized the current watermarking techniques into three types, watermarks embedded in model training, watermarks inserted in the inference phase and additional watermarks after text generation. For the robustness, confidentiality and effectiveness needs of watermarking, this paper also organized the evaluation metrics of watermarking techniques and reviewed the existing anti-watermarking attacks. This paper provides a comprehensive overview of LLM watermarking techniques with the aim of further promoting their development and application.

Key words: large language model, text watermarking, weight protection, AI forensics

中图分类号:

TP309

秦中元, 王田田, 刘伟强, 张群芳. 大语言模型水印技术研究进展[J]. 信息网络安全, 2025, 25(2): 177-193.

QIN Zhongyuan, WANG Tiantian, LIU Weiqiang, ZHANG Qunfang. Advances in Watermarking Techniques for Large Language Models[J]. Netinfo Security, 2025, 25(2): 177-193.

图/表 8

图1

图2

图3

图4

图5

表1

表2

表3

参考文献 63

[1]	RADFORD A, NARASIMHAN K, SALIMANS T, et al. Improving Language Understanding by Generative Pre-Training[EB/OL]. (2018-06-11)[2024-03-21]. https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
[2]	RADFORD A, WU J, CHILD R, et al. Language Models are Unsupervised Multitask Learners[J]. OpenAI Blog, 2019, 1(8): 9-32.
[3]	BROWN T B, MANN B, RYDER N, et al. Language Models are Few-Shot Learners[J]. Neural Information Processing Systems, 2020, 33: 1877-1901.
[4]	OPENAI, ACHIAM J, ADLER S, et al. GPT-4 Technical Report[EB/OL]. (2024-03-04)[2024-03-21]. https://arxiv.org/pdf/2303.08774.
[5]	TOUVRON H, MARTIN L, STONE K, et al. Llama 2: Open Foundation and Fine-Tuned Chat Models[EB/OL]. (2023-07-19)[2024-03-21]. https://arxiv.org/pdf/2307.09288.
[6]	DEVLIN J, CHANG M W, LEE K, et al. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding[C]//ACL. 17th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2019). Minneapolis: ACL, 2019: 4171-4186.
[7]	ZHANG S, ROLLER S, GOYAL N, et al. OPT: Open Pre-Trained Transformer Language Models[EB/OL]. (2022-06-21)[2024-03-21]. https://arxiv.org/abs/2205.01068.
[8]	YANG Zhilin, DAI Zihang, YANG Yiming, et al. XLNet: Generalized Autoregressive Pretraining for Language Understanding[J]. Neural Information Processing Systems, 2019, 32: 1-11.
[9]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is All You Need[J]. Advances in Neural Information Processing Systems, 2017, 30: 1-11.
[10]	LEVIATHAN Y, KALMAN M, MATIAS Y. Fast Inference from Transformers via Speculative Decoding[C]// JMLR. Proceedings of the 40th International Conference on Machine Learning. Hawaii: JMLR, 2023: 19274-19286.
[11]	KAlYAN K S, RAJASEKHARAN A, SANGEETHA S. Ammus: A Survey of Transformer-Based Pretrained Models in Natural Language Processing[EB/OL]. (2021-08-28)[2024-03-21]. https://arxiv.org/pdf/2108.05542.
[12]	OUYANG Long, WU J, JIANG Xu, et al. Training Language Models to Follow Instructions with Human Feedback[J]. Neural Information Processing Systems, 2022, 35: 27730-27744.
[13]	RAFAILOV R, SHARMA A, MITCHELL E, et al. Direct Preference Optimization: Your Language Model is Secretly a Reward Model[J]. Neural Information Processing Systems, 2024, 36: 1-14.
[14]	ETHAYARAJH K, XU W, MUENNIGHOFF N, et al. KTO: Model Alignment as Prospect Theoretic Optimization[EB/OL]. (2023-09-03)[2024-03-21]. https://arxiv.org/pdf/2402.01306v3.
[15]	CHANG Yupeng, WANG Xu, WANG Jindong, et al. A Survey on Evaluation of Large Language Models[J]. ACM Transactions on Intelligent Systems and Technology, 2024, 15(3): 1-45.
[16]	LI Linyang, JIANG Botian, WANG Pengyu, et al. Watermarking LLMs with Weight Quantization[EB/OL]. (2023-10-17)[2024-03-21]. https://arxiv.org/pdf/2310.11237.
[17]	WANG L, YANG Wenkai, CHEN Deli, et al. Towards Codable Watermarking for Injecting Multi-Bit Information to LLM[C]// ICLR. The Twelfth International Conference on Learning Representations. Vienna: OpenReview, 2024: 1-27.
[18]	YOO K, AHN W, JANG J, et al. Robust Multi-Bit Natural Language Watermarking through Invariant Features[C]// ACL. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Toronto: ACL, 2023: 2092-2115.
[19]	LI Shuai, CHEN Kejiang, TANG Kunsheng, et al. Function Marker: Watermarking Language Datasets via Knowledge Injection[EB/OL]. (2023-11-17)[2024-03-21]. https://arxiv.org/pdf/2311.09535v2.
[20]	YAO Hongwei, LOU Jian, QIN Zhan, et al. PromptCARE: Prompt Copyright Protection by Watermark Injection and Verification[EB/OL]. (2023-11-17)[2024-03-21]. https://arxiv.org/pdf/2311.09535v2.
[21]	ZHANG Ruisi, HUSSAIN S S, NEEKHARA P, et al. REMARK-LLM: A Robust and Efficient Watermarking Framework for Generative Large Language Models[C]// IEEE. 2024 IEEE Symposium on Security and Privacy (SP). New York: IEEE, 2024: 845-861.
[22]	LI Zongjie, WANG Chaozheng, WANG Shuai, et al. Protecting Intellectual Property of Large Language Model-Based Code Generation APIs via Watermarks[C]// ACM. Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2023: 2336-2350.
[23]	SATO R, TAKEZAWA Y, BAO H, et al. Embarrassingly Simple Text Watermarks[EB/OL]. (2023-10-13)[2024-03-21]. https://arxiv.org/pdf/2310.08920.
[24]	LEE T, HONG S, AHN J, et al. Who Wrote this Code? Watermarking for Code Generation[EB/OL]. (2023-07-03)[2024-03-21]. https://arxiv.org/pdf/2305.15060.
[25]	YANG Borui, LI Wei, XIANG Liyao, et al. Towards Code Watermarking with Dual-Channel Transformations[EB/OL]. (2024-01-02)[2024-03-21]. https://arxiv.org/pdf/2309.00860.
[26]	TAKEZAWA Y, SATO R, BAO H, et al. Necessary and Sufficient Watermark for Large Language Models[EB/OL]. (2023-10-02)[2024-03-21]. https://arxiv.org/pdf/2310.00833.
[27]	LIU Aiwei, PAN Leyi, HU Xuming, et al. An Unforgeable Publicly Verifiable Watermark for Large Language Models[C]// ICLR. The Twelfth International Conference on Learning Representations. Vienna: OpenReview, 2024: 1-17.
[28]	KIRCHENBAUER J, GEIPING J, WEN Y, et al. On the Reliability of Watermarks for Large Language Models[C]// ICLR. The Twelfth International Conference on Learning Representations. Vienna: OpenReview, 2024: 1-45.
[29]	TANG Ruixiang, CHUANG Yuneng, HU Xia. The Science of Detecting LLM-Generated Texts[J]. Communications of the ACM, 2024, 67(4): 50-59.
[30]	LIU Aiwei, PAN Leyi, LU Yijian, et al. A Survey of Text Watermarking in the Era of Large Language Models[J]. ACM Computing Surveys, 2024, 57(2): 1-36.
[31]	FERNANDEZ P, CHAFFIN A, TIT K, et al. Three Bricks to Consolidate Watermarks for Large Language Models[C]// IEEE. 2023 IEEE International Workshop on Information Forensics and Security (WIFS). New York: IEEE, 2023: 1-6.
[32]	BALDASSINI F B, NGUYEN H H, CHANG C C, et al. Cross-Attention Watermarking of Large Language Models[C]// ICASSP. 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE, 2024: 4625-4629.
[33]	LI Shen, YAO Liuyi, GAO Jinyang, et al. Double-I Watermark: Protecting Model Copyright for LLM Fine-Tuning[EB/OL]. (2023-06-05)[2024-03-21]. https://arxiv.org/pdf/2402.14883.
[34]	XU Xiaojun, YAO Yuanshun, LIU Yang. Learning to Watermark LLM-Generated Text via Reinforcement Learning[EB/OL]. (2024-03-13)[2024-03-21]. https://arxiv.org/pdf/2403.10553.
[35]	LIU Yepeng, BU Yuheng. Adaptive Text Watermark for Large Language Models[EB/OL]. (2024-01-25)[2024-03-21]. https://arxiv.org/pdf/2401.13927v1.
[36]	LIU Aiwei, PAN Leyi, HU Xuming, et al. A Semantic Invariant Robust Watermark for Large Language Models[EB/OL]. (2023-10-10)[2024-03-21]. https://arxiv.org/pdf/2310.06356.
[37]	AARONSON S. My AI Safety Lecture for UT Effective Altruism[EB/OL]. (2022-11-28)[2024-03-21] https://scottaaronson.blog/?p=6823.
[38]	KIRCHENBAUER J, GEIPING J, WEN Y, et al. A Watermark for Large Language Models[C]// ICML. Proceedings of the 40th International Conference on Machine Learning. Hawaii: JMLR, 2023: 17061-17084.
[39]	WU Yihan, HU Zhengmian, ZHANG Hongyang, et al. DiPmark: A Stealthy, Efficient and Resilient Watermark for Large Language Models[EB/OL]. (2023-10-11)[2024-03-21]. https://arxiv.org/pdf/2310.07710v1.
[40]	REN Jie, XU Han, LIU Yiding, et al. A Robust Semantics-Based Watermark for Large Language Model against Paraphrasing[EB/OL]. (2023-11-15)[2024-03-21]. https://arxiv.org/pdf/2311.08721v1.
[41]	LI Yuhang, WANG Yihan, SHI Zhouxing, et al. Improving the Generation Quality of Watermarked Large Language Models via Word Importance Scoring[EB/OL]. (2023-11-16)[2024-03-21]. https://arxiv.org/pdf/2311.09668.
[42]	CHEN Liang, BIAN Yatao, DENG Yang, et al. X-Mark: Towards Lossless Watermarking through Lexical Redundancy[EB/OL]. (2023-11-16)[2024-03-21]. https://arxiv.org/pdf/2311.09832v1.
[43]	WOUTERS B. Optimizing Watermarks for Large Language Models[EB/OL]. (2023-12-28)[2024-03-21]. https://arxiv.org/pdf/2312.17295.
[44]	HUO Mingjia, SOMAYAJULA S A, LIANG Youwei, et al. Token-Specific Watermarking with Enhanced Detectability and Semantic Coherence for Large Language Models[EB/OL]. (2024-02-28)[2024-03-21]. https://arxiv.org/pdf/2402.18059v1.
[45]	QU Wenjie, YIN Dong, HE Zixin, et al. Provably Robust Multi-Bit Watermarking for AI-Generated Text via Error Correction Code[EB/OL]. (2024-01-30)[2024-03-21]. https://arxiv.org/pdf/2401.16820v1.
[46]	BOROUJENY M K, JIANG Ya, ZENG Kai, et al. Multi-Bit Distortion-Free Watermarking for Large Language Models[EB/OL]. (2024-02-26)[2024-03-21]. https://arxiv.org/pdf/2402.16578.
[47]	LI Boquan, ZHANG Mengdi, ZHANG Peixin, et al. Resilient Watermarking for LLM-Generated Codes[EB/OL]. (2024-02-12)[2024-03-21]. https://arxiv.org/pdf/2402.07518v1.
[48]	MUNYER T, ZHONG Xin. DeepTextMark: Deep Learning Based Text Watermarking for Detection of Large Language Model Generated Text[EB/OL]. (2023-05-09)[2024-03-21]. https://arxiv.org/pdf/2305.05773v1.
[49]	WEI J T Z, WANG R Y, JIA R. Proving Membership in LLM Pretraining Data via Data Watermarks[EB/OL]. (2024-02-16)[2024-03-21]. https://arxiv.org/pdf/2402.10892v1.
[50]	PANG Kaiyi, QI Tao, WU Chuhan, et al. Adaptive and Robust Watermark against Model Extraction Attack[EB/OL]. (2023-05-03)[2024-03-21]. https://arxiv.org/pdf/2405.02365v1.
[51]	SANDER T, FERNANDEZ P, DURMUS A, et al. Watermarking Makes Language Models Radioactive[EB/OL]. (2024-02-22)[2024-03-21]. https://arxiv.org/pdf/2402.14904v1.
[52]	GIBOULOT E, TEDDY F. WaterMax: Breaking the LLM Watermark Detectability-Robustness-Quality Trade-Off[EB/OL]. (2024-03-06)[2024-03-21]. https://arxiv.org/pdf/2403.04808v1.
[53]	PANG Qi, HU Shengyuan, ZHENG Wenting, et al. Attacking LLM Watermarks by Exploiting Their Strengths[EB/OL]. (2024-02-15)[2024-03-21]. https://openreview.net/forum?id=P2FFPRxr3Q.
[54]	HOU A B, ZHANG Jingyu, HE Tianxing, et al. SemStamp: A Semantic Watermark with Paraphrastic Robustness for Text Generation[EB/OL]. (2023-10-06)[2024-03-21]. https://arxiv.org/pdf/2310.03991v1.
[55]	ZHANG Hanlin, EDELMAN B L, FRANCATI D, et al. Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models[EB/OL]. (2023-11-07)[2024-03-21]. https://arxiv.org/pdf/2311.04378v1.
[56]	KRISHNA K, SONG Yixiao, KARPINSKA M, et al. Paraphrasing Evades Detectors of AI-Generated Text, but Retrieval is an Effective Defense[J]. Advances in Neural Information Processing Systems, 2024, 36: 1-32.
[57]	SADASIVAN V S, KUMAR A, BALASUBRAMANIAN S, et al. Can AI-Generated Text be Reliably Detected?[EB/OL]. (2023-03-17)[2024-03-21]. https://arxiv.org/pdf/2303.11156v1.
[58]	JOVANOVIC N, STAAB R, VECHEV M. Watermark Stealing in Large Language Models[EB/OL]. (2024-02-29)[2024-03-21]. https://arxiv.org/pdf/2402.19361v1.
[59]	ZHAO Xuandong, ANANTH P, LI Lei, et al. Provable Robust Watermarking for AI-Generated Text[EB/OL]. (2023-06-30)[2024-03-21]. https://arxiv.org/pdf/2306.17439v1.
[60]	TANG L, UBERTI G, SHLOMI T. Baselines for Identifying Watermarked Large Language Models[EB/OL]. (2023-05-29)[2024-03-21]. https://arxiv.org/pdf/2305.18456.
[61]	SINGH K, ZOU J. New Evaluation Metrics Capture Quality Degradation due to LLM Watermarking[EB/OL]. (2024-03-18)[2024-03-21]. https://openreview.net/forum?id=PuhF0hyDq1.
[62]	AJITH A, SINGH S, PRUTHI D. Performance Trade-Offs of Watermarking Large Language Models[EB/OL]. (2024-01-16)[2024-03-21]. https://arxiv.org/pdf/2311.09816v1.
[63]	LUO Yiyang, LIN Ke, GU Chao. Lost in Overlap: Exploring Watermark Collision in LLMs[EB/OL]. (2024-03-15)[2024-03-21]. https://arxiv.org/pdf/2403.10020v1.

编辑推荐 0

Metrics

阅读次数

全文

152

HTML			PDF

最新录用	在线预览	正式出版	最新录用	在线预览	正式出版
0	0	22	0	0	130

来源	本网站	其他网站

次数	152	0
比例	100%	0%

摘要

166

最新录用	在线预览	正式出版

0	0	166

	来源	本网站

	次数	167
	比例	100%

推理时间点	算法类别	算法名称	算法概述	水印信息
Logits 生成	Logits 缩放	Adaptive Text Watermark^[35]	生成logits缩放向量和自适应温度系数	zero-bit
	添加message logits	CTWL^[17]	概率公式推导的可编码水印	multi-bits
	添加message logits	Robust Watermark^[36]	额外LLM生成logits偏置项	zero-bit
Token 采样	调整同义词分布	LLCG API WM^[22]	编程语言token-level同义词替换	zero-bit
	含散列密钥的伪随机函数	OpenAI WM^[37]	OpenAI水印方案(未详细说明)	未说明
		KGW Watermark^[38]	伪随机函数划分红绿列表	zero-bit
		NS-Watermark^[26]	优化绿色单词插入位置
		Unforgeable Publicly Verifiable Watermark^[27]	生成器和检测器共享嵌入网络
		SWEET^[24]	仅在高熵位置促进绿色token
		DiPmark^[39]	重新加权策略促进采样
		SemaMark^[40]	用token语义替代哈希
		WIS^[41]	token重要性估算
		X-Mark^[42]	添加单词间互斥规则
		OPT Watermark^[43]	针对困惑度变化的多目标优化
		MOO^[44]	感知器网络生成分割比率
		Advanced Detection Schemes^[31]	使用不同哈希密钥	multi-bits
		Error-Correction Code Based Watermark^[45]	通过ECC编码选定红或绿列表
	制定水印映射法则	Multi-Bit Distortion-Free Watermark^[46]	分布间隔伪随机编码

方案名称	水印嵌入方案	水印信息	方案类别
ACW^[47]	有选择地应用一组精心设计的语义保留、幂等代码转换	multi-bits	信息代换
EASYMARK^[23]	利用Unicode 具有相同或相似外观的代码点进行字符替换	multi-bits	信息代换
Random Sequence/Unicode Watermark^[49]	研究了两种水印：在文档末尾插入随机序列，用Unicode相似字符随机替换字符	zero-bit	信息增添/代换
DeepTextMark^[48]	基于深度学习，应用 Word2Vec 和句子编码进行水印插入，并使用基于转换器的分类器进行水印检测	zero-bit	神经网络
Robust Multi-bit Natural Language Watermark^[18]	通过语义或句法上的特征精确定位水印的位置，并通过infill模型嵌入水印	multi-bits	神经网络
REMARK-LLM^[21]	额外的LLM将生成的文本及其相应的签名编码到潜在特征空间中，添加它们的特征表示并产生词汇表上的带水印的分布	multi-bits	神经网络

攻击类型	典型攻击模型	攻击方法概述	攻击算法是否通用	作者提出的对抗手段
擦除攻击	Generic Attack Methodology^[55]	给定带水印的文本，在每次迭代时，使用小型屏蔽 LLM 生成片段替换，同时确保根据奖励模型判断的响应质量下降幅度轻微	通用	强水印方案中，计算有限的攻击者擦除水印必然导致文本质量显著下降
释义攻击	DIPPER^[56]	训练一个11B 参数释义生成模型：（1）在上下文中释义长文本（2）控制输出多样性		给定候选文本，在序列数据库检索语义相似的生成
释义攻击	Bigram Paraphrase Attack^[54]	解码大量来自束搜索的top-raking序列，获得 k 个释义候选，然后选择与原始句子具有最小二元组重叠的候选者		基于LSH的鲁棒句子级语义水印算法：对候选句子进行编码和LSH哈希，进行句子级拒绝采样，直到采样的句子落入语义嵌入空间中的水印分区中
欺骗攻击	Spoofing Attacks on Watermarked Models^[57]	多次查询带水印的 LLM，以观察其输出中成对token的出现情况，从而估计 N 个token的绿名单分数	非通用	未提及
欺骗攻击	Automated Watermark Stealing^[58]	建立水印规则模型，通过频繁出现的token序列来推断基于红绿列表的水印规则	非通用	设计鲁棒性更高的水印方案，如使用多个密钥、采样修改等