基于密集连接和特征消冗网络的零水印方法  

A zero-watermarking method based on dense connection and redundant feature elimination network

作　　者：何灵强 骆挺[1,2] 李黎[3] 何周燕 徐海勇[1] HE Lingqiang;LUO Ting;LI Li;HE Zhouyan;XU Haiyong(Faculty of Information Science and Engineering,Ningbo University,Ningbo,Zhejiang 31521l,China;College of Science and Technology,Ningbo University,Ningbo,Zhejiang 315000,China;Hangzhou Dianzi University,Zhejiang 310018,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211 [2]宁波大学科学技术学院,浙江宁波315000 [3]杭州电子科技大学,浙江杭州310000

出　　处：《光电子．激光》2023年第5期543-553,共11页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61971247,62171243,61501270);浙江省自然科学基金(LY22F020020,LQ23F010011);宁波市自然科学基金(2021J134,2022J136,2022J066);浙江省教育厅科研科研项目(Y202248989)资助项目。

摘　　要：针对鲁棒水印不可见性和鲁棒性的矛盾,提出了一种基于密集连接和特征消冗网络(dense connection and redundant feature elimination network,DCRFENet)的零水印方法。首先,为了抵抗不同图像攻击,设计密集连接模块,即从不同卷积层提取浅层和深层图像的鲁棒特征。同时,为了增强零水印的唯一性,结合特征间权重学习与特征内权重学习设计特征消冗模块,从而消除冗余特征以及增强图像的有效特征。其次,融合有效特征与鲁棒特征,生成图像特征图,并进行抗攻击训练。最后,基于训练的DCRFENet,将特征图进行分块,比较分块均值与块内每一特征值的大小构造零水印。实验结果表明,在CIFAR10、COCO、VOC数据集上抵抗单一攻击的平均比特误差率(bit error rate,BER)均低于0.03。此外,与现有方法相比,提出的零水印方法对训练的攻击、非训练的攻击以及混合攻击均具有较好的鲁棒性。To address the contradiction between watermarking invisibility and robustness,a zero-watermarking method based on dense connection and redundant feature elimination network(DCRFENet)is proposed.Firstly,in order to resist different image attacks,the dense connection module is designed to extract shallow and deep layer image robust features from different convolution layers.Meanwhile,to enhance the uniqueness of zero-watermarking,the redundant feature elimination module is presented to eliminate redundant information and enhance image valid features by learning weight of inter-feature and intra-feature.Secondly,valid features and robust features are fused to generate the final image feature map,which is used for robustness training.Finally,based on training of DCRFENet,the image feature map is divided into blocks,and zero watermark is obtained by comparing each feature value within the block with its average value.The experimental results show that the average bit error rate(BER)is lower than 0.03 for CIFAR10,COCO and VOC datasets.Moreover,compared with the existing methods,the proposed zero-watermarking model is robuster to trained attacks,untrained attacks and hybrid attacks.

关 键 词：零水印 唯一性 鲁棒性 密集连接 特征消冗 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]