网络与信息安全学报2026，Vol.12Issue(4)：50-61,12.DOI:10.11959/j.issn.2096-109x.AQ26094
基于StyleID的神经风格迁移水印算法
Neural style transfer watermarking algorithm based on StyleID
摘要
Abstract
Neural style transfer has been widely used in digital creation,cultural heritage restoration,and other fields.However,the copyright attribution and traceability of generated content urgently need to be addressed.Exist-ing watermarking techniques suffer from problems such as weak fusion between style and watermark,insufficient robustness,and degraded visual quality,and are not well adapted to the generation characteristics of latent diffusion models.To solve these issues,a neural style transfer watermarking algorithm based on StyleID for copyright protec-tion and traceability of style-transferred generated content was proposed.The algorithm dynamically captures the key fusion regions of content and style features through an attention mechanism,generates a fused latent vector af-ter style injection,and completes watermark embedding at the decoder stage of the latent diffusion model,realizing deep binding between the watermark and the style transfer process.Experimental results show that the proposed al-gorithm maintains high visual quality of style-transferred images,while the extraction accuracy of the watermark against common attacks such as cropping,blurring,and JPEG compression is above 98%.It can effectively support copyright protection and traceability of style-transferred generated content,providing technical support for the le-gitimate application of generative AI content.关键词
风格迁移/数字水印/潜在扩散模型/内容溯源/版权保护Key words
style transfer/digital watermarking/latent diffusion models/content tracing/copyright protection分类
信息技术与安全科学引用本文复制引用
王建行,狄富强,朱率率,刘浪..基于StyleID的神经风格迁移水印算法[J].网络与信息安全学报,2026,12(4):50-61,12.基金项目
国家自然科学基金资助项目(No.62102451)The National Natural Science Foundation of China(No.62102451) （No.62102451）
