基于自适应掩码和生成式修复的图像隐私保护技术  

Image privacy protection based on adaptive masking and generative restoration

作　　者：方世超 滕旭阳 王子南 陈晗 仇兆炀 毕美华[1] FANG Shichao;TENG Xuyang;WANG Zinan;CHEN Han;QIU Zhaoyang;BI Meihua(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310010,Zhejiang,China;Heilongjiang Jiongyuan Technology Co.,Ltd.,Harbin 150000,Heilongjiang,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310010 [2]黑龙江炅源科技有限公司,黑龙江哈尔滨150000

出　　处：《山东大学学报（工学版）》2024年第5期111-121,共11页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(No.61906055);浙江省自然科学基金资助项目(LQ19F020009)。

摘　　要：针对现有图像保护技术中全图加密增加计算成本和区域遮挡无法判定多目标等问题,提出基于自适应掩码和生成式修复的图像保护框架。该框架采用Score-CAM(class activation mapping)技术自适应判别图像的核心区域,准确生成多目标核心区域掩膜;采用遮挡方法保护图像隐私来降低计算开销;引入区域感知的CAM损失函数,确保修复图像重点区域的一致性。将有遮挡的图像送入修复网络进行训练,对训练好的网络参数进行椭圆加密;在发送阶段将掩码图像和密钥分开发送,接收端通过密钥解密,Shift-Net网络载入参数对掩码图像进行准确修复。在ImageNet数据集中的试验表明,CAM损失函数的修复模型使得生成图像的结构相似性指标提高了0.2%、学习感知图像块相似度降低了0.2%。本研究在接收端自适应对图像重点区域进行掩码,使得识别模型失效进而保护图像隐私。In order to solve the problems of increasing the computational cost of full-image encryption and the inability of region oc-clusion to determine multiple targets in the existing image protection technology,an image protection framework based on the adaptive mask and generative inpainting was proposed.The framework used Score-CAM(class activation mapping)to adaptively discriminate the core region of the image and accurately generate the multi-target core region mask.The occlusion method was used to protect image privacy to reduce computational overhead.The region-aware CAM loss function was introduced to ensure the consistency of the key ar-eas of the repaired image.The occluded images were sent to the repair network for training,and the trained network parameters were elliptically encrypted.The mask image and the key were sent separately at the sending stage,and decrypted by the key at the receiving end.Then parameters in the Shift-Net were loaded to repair the mask image accurately.The experimental results on the ImageNet data-set showed that the restoration model with CAM loss function improved the structural similarity of the generated image by 0.2%,and reduced the learned perceptual image patch similarity by 0.2%.This study adaptively masked the key areas of the image at the receiving end,rendering the recognition model ineffective and thus protecting the privacy of the image.

关 键 词：自适应掩码 生成式修复 区域感知 类激活映射 图像隐私保护 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]