DMark: Order-Agnostic Watermarking for Diffusion Large Language Models

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Wu, Linyu, Zhong, Linhao, Qu, Wenjie, Li, Yuexin, Liu, Yue, Zhai, Shengfang, Shen, Chunhua, Zhang, Jiaheng
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866918153740615680
author Wu, Linyu
Zhong, Linhao
Qu, Wenjie
Li, Yuexin
Liu, Yue
Zhai, Shengfang
Shen, Chunhua
Zhang, Jiaheng
author_facet Wu, Linyu
Zhong, Linhao
Qu, Wenjie
Li, Yuexin
Liu, Yue
Zhai, Shengfang
Shen, Chunhua
Zhang, Jiaheng
contents Diffusion large language models (dLLMs) offer faster generation than autoregressive models while maintaining comparable quality, but existing watermarking methods fail on them due to their non-sequential decoding. Unlike autoregressive models that generate tokens left-to-right, dLLMs can finalize tokens in arbitrary order, breaking the causal design underlying traditional watermarks. We present DMark, the first watermarking framework designed specifically for dLLMs. DMark introduces three complementary strategies to restore watermark detectability: predictive watermarking uses model-predicted tokens when actual context is unavailable; bidirectional watermarking exploits both forward and backward dependencies unique to diffusion decoding; and predictive-bidirectional watermarking combines both approaches to maximize detection strength. Experiments across multiple dLLMs show that DMark achieves 92.0-99.5% detection rates at 1% false positive rate while maintaining text quality, compared to only 49.6-71.2% for naive adaptations of existing methods. DMark also demonstrates robustness against text manipulations, establishing that effective watermarking is feasible for non-autoregressive language models.
format Preprint
id arxiv_https___arxiv_org_abs_2510_02902
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DMark: Order-Agnostic Watermarking for Diffusion Large Language Models
Wu, Linyu
Zhong, Linhao
Qu, Wenjie
Li, Yuexin
Liu, Yue
Zhai, Shengfang
Shen, Chunhua
Zhang, Jiaheng
Machine Learning
Artificial Intelligence
Cryptography and Security
Diffusion large language models (dLLMs) offer faster generation than autoregressive models while maintaining comparable quality, but existing watermarking methods fail on them due to their non-sequential decoding. Unlike autoregressive models that generate tokens left-to-right, dLLMs can finalize tokens in arbitrary order, breaking the causal design underlying traditional watermarks. We present DMark, the first watermarking framework designed specifically for dLLMs. DMark introduces three complementary strategies to restore watermark detectability: predictive watermarking uses model-predicted tokens when actual context is unavailable; bidirectional watermarking exploits both forward and backward dependencies unique to diffusion decoding; and predictive-bidirectional watermarking combines both approaches to maximize detection strength. Experiments across multiple dLLMs show that DMark achieves 92.0-99.5% detection rates at 1% false positive rate while maintaining text quality, compared to only 49.6-71.2% for naive adaptations of existing methods. DMark also demonstrates robustness against text manipulations, establishing that effective watermarking is feasible for non-autoregressive language models.
title DMark: Order-Agnostic Watermarking for Diffusion Large Language Models
topic Machine Learning
Artificial Intelligence
Cryptography and Security
url https://arxiv.org/abs/2510.02902