基于对抗样本的深度学习图像压缩感知方法

来源:用户上传      作者:王继良 周四望 金灿灿

　　摘要：压缩感知是研究数据采样压缩与重构的信号处理新理论，近年来研究人员将深度学习运用到图像压缩感知算法中，显著提高了图像重构质量.然而，图像信息常与隐私关联，高质量的重构图像在方便人们观赏的同时，带来了隐私保护的问题.本文基于深度学习理论，提出一种对抗的图像压缩感知方法，该方法将压缩理论和对抗样本技术统一于同一个压缩感知算法，通过设计损失函数，联合重构误差和分类误差来训练压缩感知深度神经网络，使得压缩感知重构样本同时也是一个对抗样本.因此，重构图像在保证重构质量的同时，也能对抗图像分类算法，降低其识别率，达到保护图像隐私的效果.在Cifar-10和MNIST图像集上进行的实验结果表明，和已有的压缩感知方法相比，我们提出的对抗压缩感知方法以损失仅10%的图像重构质量为代价，使得图像分类精度下降了74%，获得了很好的对抗性能.
　　关键词：对抗样本;深度学习;图像;压缩感知
　　中图分类号：TP391文献标志码：A
　　Method of Deep Learning Image Compressed Sensing Based on Adversarial Samples
　　WANG Jiliang ZHOU Siwang17，JIN Cancan1
　　（1. College of Computer Science and Electronic Engineering，Hunan University，Changsha 41008 China;
　　2. Changsha Environmental Protection College，Changsha 41000 China）
　　Abstract ：Compressed sensing is a new signal processing theory focusing on data sampling compression and reconstruction. In recent years，researchers have applied deep learning to image compressed sensing algorithms，which significantly improves the quality of the recovered images. However，images are often associated with personal privacy，and high-quality recovered images often bring privacy protection problems while facilitating peoplers viewing. Based on deep neural network，this paper proposes an image compressed sensing algorithm with adversarial learning. This method integrates data compression and adversary sample technique into the compressed sensing algorithm. By training the neural network with a loss function combining reconstruction loss and classification loss，the output samples，i. e.，the recovered images，become adversarial samples. The recovered images with our proposed algorithm can then be adversarial to image classifications algorithms，decreasing their recognition rate and achieving the performance of protecting image privacy while guaranteeing a reasonable image quality. Experimental results on Cifar-10 and MNIST show that，compared with the existing compressed sensing methods，the proposed adversarial algorithm achieves excellent adversarial performance，as the classification accuracy is decreased by 74% at the cost of 10% loss of image reconstruction quality.
　　Key words：adversarial sample;deep learning;image;compressed sensing
　　核醺兄是研究数据采样压缩与重构的信号处理新理论[1-3].压缩感知理论突破了奈奎斯特采样定理的限制，能降低图像获取成本、节省图像的存储空间和传输开销，在图像处理领域已经取得了成功应用.迄今为止，已有多种图像压缩感知算法被提出，目标是获得更高的图像重构质量.经典的图像压缩感知重构算法包括基于消息传递AMP框架的算法[4-5]、应用于二进制图像的压缩感知算法[6]、自适应压缩感知算法[7].我们则提出了基于分块的图像压缩感知算法[8，9].压缩感知理论有着严谨、完备的数学基础，但图像重构算法复杂度高，运行时间长.
　　受深度学习研究进展的鼓舞，近年来研究人员开始探索基于深度神经网络的图像压缩感知算法[10-12].深度学习压缩感知利用深度神经网络的学习能力，在有标签的训练集中学习从原始输入样本到重构样本的映射，实现压缩感知重构.ReconNet是较早提出的压缩感知深度网络模型[13]，文献[14]对此网络模型做了改进，通过联合学习测量进程和重构进程来优化压缩感知测量矩阵，在低采样率下有更好的重构性能.受分块压缩感知算法的启发，文献[15-16]提出CSNet网络结构，图像压缩采用分块方法，但用一个深度网络实现整体图像重构，从而提高了图像重构质量.我们对CSNet做了深入研究，根据图像各块的重要性自适应分配采样率，进一步提高了CSNet的重构效果[17].和传统压缩感知方法相比，深度学习算法有显著更快的重构速度，在低采样率时有更好的图像重构效果.

nlc202208291748

转载注明来源:https://www.xzbu.com/4/view-15438677.htm