Title: An efficient natural language watermarking method via robust-word substitution

Authors: Lingyun Xiang; Chenling Zhang; Minghao Huang; Jinghan Zhou

Addresses: School of Computer Science and Technology, Changsha University of Science and Technology, Changsha, 410114, China ' School of Computer Science and Technology, Changsha University of Science and Technology, Changsha, 410114, China ' School of Computer Science and Technology, Changsha University of Science and Technology, Changsha, 410114, China ' School of Computer Science and Technology, Changsha University of Science and Technology, Changsha, 410114, China

Abstract: Current modification-based natural language watermarking methods often face limitations in extraction efficiency and resilience against text editing attacks. To tackle these issues, this paper proposes a Robust Word Substitution Watermarking (RWSW) method, which enhances robustness while preserving the original text's fluency and quality. RWSW embeds watermarks at the sentence level by substituting robust words, drawing on complex sentence simplification insights. A robust word recogniser, trained via an encoder-decoder framework, identifies words unchanged in normal editing. For optimal substitutions, a BiLSTM network and Transformer predict suitable replacements, ensuring natural integration. Watermark info is redundantly embedded by replacing multiple sentence-level robust words with the same bit, boosting resistance to perturbations. Extraction decodes values and uses majority voting for high-accuracy recovery. Experiments show RWSW generates high-quality watermarked texts, with superior robustness over existing methods and high extraction accuracy under common editing attacks.

Keywords: natural language watermarking; robust words; word prediction model; word substitution.

DOI: 10.1504/IJAACS.2025.150829

International Journal of Autonomous and Adaptive Communications Systems, 2025 Vol.18 No.6, pp.509 - 526

Received: 23 Mar 2025
Accepted: 29 Apr 2025
Published online: 23 Dec 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article