首页|结合可逆神经网络和逆梯度注意力的抗屏摄攻击水印方法

结合可逆神经网络和逆梯度注意力的抗屏摄攻击水印方法

扫码查看
随着智能设备的普及,数字媒体内容的传播和分享变得更加便捷,人们可以通过手机拍摄屏幕等简单方式轻松获取未经授权的信息,导致屏幕拍摄传播成为版权侵权的热点问题.为此,该文针对屏幕盗摄版权保护任务提出一种端到端的基于可逆神经网络和逆梯度注意力的抗屏摄攻击图像水印框架,实现屏幕盗摄场景下版权维护的目标.该文将水印的嵌入和提取视为相互关联的逆问题,利用可逆神经网络实现编解码网络的一体化,有助于减少信息传递损失.进一步地,通过引入逆梯度注意模块,捕捉载体图像中鲁棒性强且视觉质量高的像素值,并将水印信息嵌入到载体图像中不易被察觉和破坏的区域,保证水印的不可见性和模型的鲁棒性.最后,通过可学习感知图像块相似度(LPIPS)损失函数优化模型参数,指导模型最小化水印图像感知差异.实验结果表明,所提方法在鲁棒性和水印图像视觉质量上优于目前同类的基于深度学习的抗屏摄攻击水印方法.
Screen-Shooting Resilient Watermarking Scheme Combining Invertible Neural Network and Inverse Gradient Attention
With the growing use of smart devices,the ease of sharing digital media content has been enhanced.Concerns have been raised about unauthorized access,particularly via screen shooting.In this paper,a novel end-to-end watermarking framework is proposed,employing invertible neural networks and inverse gradient attention,to tackle the copyright infringement challenges related to screen content leakage.A single invertible neural network is employed by the proposed method for watermark embedding and extraction,ensuring information integrity during network propagation.Additionally,robustness and visual quality are enhanced by an inverse gradient attention module,which emphasizes pixel values and embeds the watermark in imperceptible areas for better invisibility and model resilience.Model parameters are optimized using the Learnable Perceptual Image Patch Similarity(LPIPS)loss function,minimizing perception differences in watermarked images.The superiority of this approach over existing learning-based screen-shooting resilient watermarking methods in terms of robustness and visual quality is demonstrated by experimental results.

Digital watermarkingInvertible neural networkInverse gradient attentionScreen shooting

李谢华、娄芹、杨俊雪、廖鑫

展开 >

湖南大学信息科学与工程学院 长沙 410082

数字水印 可逆神经网络 逆梯度注意力 屏幕拍摄

国家自然科学基金国家自然科学基金湖南省自然科学基金湖南省杰出青年科学基金长沙市科技重大专项

U22A2030619721422021JJ301402024JJ2025kh2205033

2024

电子与信息学报
中国科学院电子学研究所 国家自然科学基金委员会信息科学部

电子与信息学报

CSTPCD北大核心
影响因子:1.302
ISSN:1009-5896
年,卷(期):2024.46(7)
  • 2