| In the field of steganography algorithms,most scholars are more concerned about the stages of sending and receiving in the steganography of digital images,that is,how to enhance the imperceptibility of secret information,but ignore the possible occurrence of secret images during transmission.The effect of distortion on whether the receiver can correctly extract secret information.Whether the receiver can correctly extract the secret information under distortion transformation is determined by the robustness of the encrypted image.In the field of steganalysis,neural network-based methods generally have higher accuracy than traditional methods of "feature extraction +classifier",but neural network methods rely heavily on the size of the input image,and classification accuracy depends on the image Steganography algorithm.If the steganographic images of different steganography algorithms used in the training phase are input in the test,the classification accuracy of the neural network cannot reach the expected.In view of the above problems in the study of steganography algorithm and steganography analysis,on the basis of analyzing the JPEG compression algorithm and the common scaling principle,this paper proposes an anti-compression spatial image steganography method based on loss difference,based on directional cubic convolution interpolation An anti-scaling steganography method in the frequency domain,and a steganalysis method for deep residual network based on heterogeneous kernels is proposed.The main work of this article is as follows:1.Propose an anti-compression spatial image steganography method based on loss difference.Firstly,analyze the changes of airspace image after compression and the effect on steganography.Methods HUGO,WOW and SUNIWARD distortion functions are taken as examples to calculate the embedding cost,combined with STC encoding,after the secret information isembedded in the dense image,the change of the dense image pixel before and after lossy compression occurs in the conversion process,Write the loss in the compression process to the encrypted image in advance to generate a new encrypted image,to ensure that the airspace steganographic image can still correctly extract secret information after compression.Experiments show that the method guarantees the ability of anti-steganography analysis.At the same time,after the compression conversion of the encrypted image,the receiver can still extract the secret information with 90% accuracy under the predicted compression quality.2.Propose a frequency domain anti-scaling steganography method based on directional cubic convolution interpolation.First,analyze the effect of frequency domain image after zooming on steganography.Methods Taking UED and J-UNIWARD algorithms as examples,the distortion function is combined with STC encoding to embed secret information,then the edge directions of the interpolation points in four directions are calculated,and the pixel points are interpolated to ensure that the frequency domain steganographic image is scaled Has the ability to extract secret information correctly.Experiments show that,while guaranteeing the ability to resist steganography analysis,the encrypted image can control the error rate of the secret information extracted by the receiver to less than 30% after scaling transformation.3.Proposed and steganalysis method of deep residual network based on heterogeneous kernel.First introduce the rich model,deep residual network and heterogeneous kernel.This steganalysis method uses a rich model to generate a feature matrix before inputting the dense image into the network,and then uses the feature matrix as the actual input of the heterogeneous kernel depth residual network.It ensures that the steganographic image is always input to the network in a certain size,and the generalization ability of the neural network is improved.Experiments show that the method has higher accuracy of steganography analysis in both spatial and frequency domain steganography,and low networkparameters can be quickly trained.The cross-experiment of multiple steganography algorithms shows that the method has strong generalization ability. |
