一种新的卷积神经网络图像隐写分析模型
A New Image Steganalysis Model Based on Convolutional Neural Network
刘首岳 1段学明 1张猛 1张春英2
作者信息
- 1. 华北理工大学理学院,河北 唐山 063210
- 2. 华北理工大学理学院,河北 唐山 063210;河北省数据科学与应用重点实验室,河北 唐山 063210
- 折叠
摘要
针对现有卷积神经网络模型在图像隐写分析领域提取特征不充分、检测准确率不高的问题,提出一种融合转置卷积与普通卷积的图像隐写分析神经网络模型TCIS(Transposed Convolution-Convolutional Neural Network Image Steganalysis),包括四大模块:一是预处理模块,使用30 个高通滤波器,从多个尺度提取图像噪声的残差信息,减少图像内容的影响;二是转置卷积模块,对特征图进行上采样,放大隐写特征;三是普通卷积模块,由卷积层、BN层和激活函数组成,卷积层包括 5 个,最后一层使用全局卷积的方式精简识别特征;四是分类模块,通过全连接层和Softmax层判断图像是否隐写.实验结果表明,相比于典型卷积神经网络图像隐写分析模型,TCIS模型在嵌入率 0.4bpp情况下使用S-UNIWARD和HUGO算法的隐写分析准确率分别提升了2.94%～25.24%和 3.92%～21.64%.
Abstract
Aiming at the problems of insufficient feature extraction and low detection accuracy in the field of image steganalysis by existing convolutional neural network models,an image steganalysis neural network model TCIS(Transposed Convolution Convolution-Convolutional Neural Network Image Steganalysis)is proposed,which combines transposed convolution and ordinary convolution.It includes four modules:one is the preprocessing module,which u-ses 30 high-pass filters to extract the residual information of image noise from multiple scales to reduce the influence of image content.The second is the transposed convolution module,which upsamples the feature map and amplifies the steganographic features.The third is the ordinary convolution module,which consists of a convolution layer,a BN layer and an activation function.There are five convolution layers,and the last layer uses the global convolution meth-od to simplify the recognition features.The fourth is the classification module,which judges whether the image is steg-anographic through the fully connected layer and the Softmax layer.The experimental results show that compared with the typical image steganalysis models based on convolutional neural network,the steganalysis accuracy of the TCIS model using the S-UNIWARD and HUGO algorithms is improved by 2.94%～25.24%and 3.92%～21.64%respec-tively at the embedding rate is 0.4bpp.
关键词
隐写分析/转置卷积/卷积神经网络/图像隐写Key words
Steganalysis/Transposed convolution/Convolutional neural network/Image steganography引用本文复制引用
基金项目
河北省专业硕士教学案例库建设项目(KCJSZ2022073)
河北省研究生课程思政示范课程建设(YKCSZ2021091)
出版年
2024
NETL
NSTL
万方数据