Reserch On Spatial Image Steganalysis Based On Prior High-frequency Knowledge And Attention

With the continuous enhancement of the performance of adaptive image steganography algorithms,traditional image steganalysis methods encounter bottleneck due to their own limitations.Deep learning is introduced into the field of image steganalysis,which further improves the detection performance.However,compared with image classification tasks in other fields,the detection performance of deep learning-based image steganalysis methods still have room for improvement.In addition,the embedded signal is weak at low payload,and it is extremely easy to be covered by image content,which makes steganalysis difficult.Therefore,how to solve the problem of high error detection rate at low payload is worthy of attention.This thesis focuses on the above-mentioned problems and conducts further research on the deep learning-based image steganalysis methods.The main contributions of this thesis can be summarized in the following two aspects:(1)Aiming at the embedding strategy of the image steganography methods,this thesis attempts to apply attention mechanism to image steganography algorithms.In this thesis,a new deep learning-based image steganalysis method ResNet-CBAM,which is combined with CBAM(Convolutional Block Attention Module)and is improved based on ResNet(Residual Network)is proposed.CBAM is a plug-and-play lightweight module that can be plugged in any network.CBAM forces the network focuses on the main part of the image,emphasizes meaningful features along the channel axis and spatial axis,suppresses the meaningless features,and further improves the expression ability of embedding signal,which is conducive to steganalysis at low payload.The experimental results show that the detection performance of ResNet-CBAM surpasses the traditional image steganalysis method max SRMd2(Maximum Spatial Rich Models diagonal 2)and the deep learning-based image steganalysis method SRNet(Steganalysis Residual Network)for the mainstream BOSSBase v1.01(Break Our Steganographic System data Base)and BOWS-2(Break Our Watermarking System)databases,and it is confirmed that plugging CBAM into the deep learning-based image steganalysis method can enhance the detection performance.The attention mechanism shows its development prospects in the field of image steganalysis for the first time.(2)By analyzing the characteristics of the image steganography algorithms,the optimization principles of the image steganalysis methods are summarized in this thesis.A universal spatial image steganalysis method,which is named as He-Net is proposed in this thesis according to the optimization principles.He-Net is designed based on the bottleneck architecture and CBAM,uses a high-pass filter bank and truncated linear unit for preprocessing to enhance the signal-to-noise ratio,and uses global covariance pooling layer to replace the global average pooling layer,which can retain the channel correlation of the feature map and improves the expressive ability of the embedding signal.The experimental results show that the universal spatial image steganalysis method He-Net,which is designed according to the optimization principles,further reduces the error rate of image steganalysis for the mainstream BOSSBase v1.01 and BOWS-2 databases.He-Net has achieved the best the detection performance when detect three types of stego images of WOW(Wavelet Obtained Weights),S-UNIWARD(Spatial-UNIversal WAvelet Relative Distortion)and Mi POD(Minimizing the Performance of Optimal Detector)at 0.1,0.2,0.3,0.4 and0.5bpp(bits per pixel),and,especially,the effect improve obviously at low payload of 0.1bpp.