针对图像隐写分析的卷积神经网络结构设计

Structure design of convolution neural network for image steganography

为提高图像隐写分析的检测效果,构建了一个基于卷积神经网络的图像隐写分析模型(steganalysis convolutional neural network,SCNN)进行图像隐写分析,使用2层3×3的卷积层代替1层5×5的卷积层,在减少网络参数,提高模型训练效率的同时,可以提取到图像更加抽象的特征;使用ReLU激活函数代替TanH激活函数,提高了隐写分析效果;取消池化层,减少了嵌入信息的损失。实验结果表明,相比空域富模型+集成分类器的隐写分析方法和5层的CNN隐写分析模型,SCNN模型将隐写分析检测率分别提高了4.6%和3.89%,并提高了隐写分析模型的泛化能力。

To improve the effectiveness of image steganalysis,a steganalysis convolutional neural network(SCNN)model is applied to image steganalysis.The model uses two layers of 3×3 convolution layer instead of a 5×5 convolution layer,which reduces the network parameters,improves the training efficiency of the model,and extracts the more abstract features of the image.ReLU activation function is used instead of the TanH activation function to improve the effect of steganography.The pooling layer is canceled to reduce the loss of embedded information.The experimental results show that compared with the steganography analysis method and the 5 level CNN steganalysis model,our model improves the detection rate of steganography by 4.6%and 3.89%,and improves the generalization ability of the steganalysis model.