图像纹理引导的迭代水印模型  

Image texture-guided iterative watermarking model

作　　者：武鑫婷 黄樱 牛保宁[1] 关虎[3] 兰方鹏[1] 刘杰[2,3] WU Xinting;HUANG Ying;NIU Baoning;GUAN Hu;LAN Fangpeng;LIU Jie(School of Computer Science and Technology(Big Data College),Taiyuan University of Technology,Jinzhong 030699,China;School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100089,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]太原理工大学计算机科学与技术学院(大数据学院),山西晋中030699 [2]北京邮电大学人工智能学院,北京100089 [3]中国科学院自动化研究所,北京100190

出　　处：《西安电子科技大学学报》2024年第5期110-121,共12页Journal of Xidian University

基　　金：国家重点研发计划(2021YFF0900100);山西省重点研发计划(202302010101004);山西省重点研发计划(202102010101004);山西省基础研究计划(202203021222093);山西省基础研究计划(202203021212282);山西省研究生教育创新计划(2024KY193)。

摘　　要：近年来,深度神经网络被成功应用于数字水印领域。常见深度学习水印模型由嵌入水印的编码器、模拟攻击的噪声层、提取水印的解码器三部分构成。常用的端到端训练方式要求参与训练的攻击必须可导,此要求限制了水印模型在面对真实不可导攻击时的鲁棒性能力。此外,在一幅完整图像中,通常包含平滑纹理和粗糙纹理,然而目前的水印模型却很少直接利用其纹理信息实现水印的嵌入过程。针对上述问题,提出了一种图像纹理引导的迭代水印模型。引入图像纹理注意模块,使模型能够根据图像纹理的粗糙程度来引导水印的嵌入过程,提升水印的不可见性。采用两阶段迭代的训练方式实现对不可导攻击的学习。第一阶段,结合图像纹理注意模块,进行无攻击的端到端编-解码器联合训练,最大限度地保证水印的不可见性;第二阶段,进行不可导攻击参与的解码器独立训练,通过对任意真实攻击分布的学习,构建具有强鲁棒性的解码器。两阶段训练迭代进行,最终通过编、解码器协同训练的方式,达到模型的全局最优状态,实现水印不可见性和鲁棒性较好的平衡。实验结果表明,该模型在不可见性和鲁棒性上均优于主流的深度学习水印模型。Deep neural networks have been successfully used in the field of digital watermarking in recent years.An encoder that embeds the watermark,a noise layer that simulates the attacks,and a decoder that retrieves the watermark make up a typical deep learning watermarking model.The commonly used end-to-end training approach requires that the attacks involved in training should be conductible,and this requirement limits the robustness capability of watermarking models in the face of real differentiable attacks.In addition,in a complete image,which usually contains both smooth and rough textures,current watermarking models seldom utilize the texture information directly to realize the embedding process of the watermark.To address the above problem,this paper proposes an iterative watermarking model guided by the image texture.An image texture attention module is introduced so that the model can guide the embedding process of the watermark,and according to the roughness of the image texture it can improve the imperceptibility of the watermark.To make the learning of unguided attacks,a two-stage iterative training approach is adopted.In order to optimize the imperceptibility of the model,the first stage involves conducting joint end-to-end encoder-decoder training without attacks,along with the inclusion of the image texture attention module.In the second stage,the decoder involved in differentiable attacks is trained independently,and the decoder with strong robustness is built by learning from any real attack distribution.Through cooperative training of encoders and decoders,which realizes a superior balance of watermarking imperceptibility and robustness,the two-stage training is eventually carried out repeatedly to the model′s global optimum.According to experimental findings,the proposed approach outperforms popular deep learning watermarking techniques in terms of both imperceptibility and robustness.

关 键 词：深度学习 图像水印 纹理 鲁棒性 不可见性 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]