Deep learning color image steganography detection method based on surfacelet filtering preprocessing

Abstract

The invention provides a deep learning color image steganography detection method based on surfacelet filtering preprocessing. The method comprises the following steps of: 1, filtering a target color image by utilizing Surfacelet filtering so as to extract a steganography residual signal; step 2, constructing a steganography detection model: constructing a steganography detection network, optimizing parameters of the steganography detection network by adopting steganography residual signals of a training set, and taking the steganography detection network with the optimal parameters as the steganography detection model; and 3, inputting the steganography residual signal into a steganography detection model to obtain a detection result.

Description

Technical Field

[0001] The present invention relates to the technical field of digital image steganography detection, and in particular to a deep learning color image steganography detection method based on surfacelet filter preprocessing. Background Art

[0002] With the widespread popularity of the Internet, information and communication technology has gradually become one of the ways of communication between governments, enterprises and individuals, but the transmission security issues it faces have also gradually become prominent. Although existing encryption technology can convert original information into meaningless ciphertext, the characteristics of garbled code can easily attract people's attention, allowing relevant personnel to intercept and decipher it. Therefore, in order to ensure that the security of information transmission is further improved, it is also necessary to be able to hide the behavior of data transmission.

[0003] At present, digital steganography is a common technology for concealing data transmission. One of the important components of digital steganography is image steganography, which is to hide secret information in pictures and transmit them through pictures to achieve the effect of secret communication; image steganography detection uses the characteristics of carrier data changes caused by information embedding to perform statistical analysis on the image so that the results can be extracted and detected, and can also attack and damage the original carrier data to obtain secret information.

[0004] Image steganography technology is constantly developing, and color images are slowly beginning to replace simple grayscale images as the carrier data of steganography. Therefore, the steganography detection technology of color images is becoming increasingly important. The improvement of steganography detection technology is conducive to preventing the abuse of steganography technology and effectively preventing terrorist organizations from using it to transmit illegal information and conduct illegal transactions. However, there are still few steganography detection methods for color images, and the relevant steganography detection methods only extract features from a single channel or color space, without fully considering the impact of the correlation between different channels in the image on the hidden information, resulting in low detection accuracy at low embedding rates. Summary of the invention

[0005] In order to improve the accuracy of steganalysis detection for color images, the present invention provides a deep learning color image steganalysis detection method based on surfacelet filtering preprocessing.

[0006] The deep learning color image steganography detection method based on surfacelet filter preprocessing provided by the present invention comprises:

[0007] Step 1: Filter the target color image using Surfacelet filtering to extract the steganalysis residual signal;

[0008] Step 2: Building a steganalysis detection model, including: constructing a steganalysis detection network, optimizing the parameters of the steganalysis detection network using the steganalysis residual signal of the training set, and using the steganalysis detection network with the optimal parameters as the steganalysis detection model;

[0009] Step 3: Input the steganalysis residual signal into the steganalysis detection model to obtain the detection result.

[0010] Furthermore, step 1 specifically includes:

[0011] The target color image of size m×n×l is expanded three-dimensionally to a three-dimensional image of size m1×n1×l1; where m, n and l represent the number of rows, columns and channels of the target color image, respectively, and l=3; m1, n1 and l1 represent the number of rows, columns and channels of the three-dimensional image, respectively, and m1 <m,n1<n,l1> l;

[0012] Import l1 matrices of size m1×n1 into the Surfdec() function for surfacelet transform decomposition to obtain l1 transformation coefficient matrices of size m1×n1;

[0013] Perform hard thresholding on l1 transform coefficient matrices of size m1×n1;

[0014] Perform surfacelet inverse transformation on the l1 transformation coefficient matrices of size m1×n1 after hard threshold processing and reconstruct the image to obtain the filtered image;

[0015] The pixel values ​​in the RGB three channels of the filtered image are subtracted from the pixel values ​​in the RGB three channels of the original target color image to obtain the stego residual signal.

[0016] Further, the target color image of size m×n×l is three-dimensionally expanded to a three-dimensional image of size m1×n1×l1, specifically including:

[0017] The pixels in the three RGB channels of the target color image are downsampled from m×n pixels to m1×n1 pixels;

[0018] In the channel dimension, the m1×n1 pixels corresponding to the three RGB channels are copied until the number of channels is expanded from l to l1, and finally a three-dimensional image of m1×n1×l1 is obtained.

[0019] Furthermore, ResNet101 is adopted as the steganalysis detection network.

[0020] Beneficial effects of the present invention:

[0021] The present invention analyzes the color image data set, performs surfacelet filtering transformation on the color image according to the characteristics of steganography, performs hard threshold processing in the transform domain, and extracts the residual image. Then, a color image steganalysis detection model is constructed based on the residual image, which can effectively improve the accuracy of color image steganalysis detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic flow chart of a deep learning color image steganography detection method based on surfacelet filter preprocessing provided in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of three-dimensional expansion of a color image provided by an embodiment of the present invention;

[0024] Figure 3 A comparison diagram of residual extraction results provided by an embodiment of the present invention;

[0025] Figure 4 A diagram of the ResNet101 network structure provided in an embodiment of the present invention;

[0026] Figure 5 Feature graphs obtained at different feature extraction stages in the ResNet101 network provided in an embodiment of the present invention;

[0027] Figure 6 The confusion matrix and ROC curve of the ResNet101 model provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, an embodiment of the present invention provides a deep learning color image steganography detection method based on surfacelet filtering preprocessing, comprising the following steps:

[0030] S101: filtering the target color image using surfacelet filtering to extract a steganographic residual signal;

[0031] Specifically, it mainly includes the following sub-steps:

[0032] S1011: Expand the target color image of size m×n×l in three dimensions to a three-dimensional image of size m1×n1×l1; wherein m, n and l represent the number of rows, columns and channels of the target color image, respectively, and l=3; m1, n1 and l1 represent the number of rows, columns and channels of the three-dimensional image, respectively, and m1 <m,n1<n,l1> l; in this embodiment, m1=n1=l1=256.

[0033] As an implementable method, the embodiment of the present invention adopts the cat() function in Matlab to realize the three-dimensional expansion of the original target color image. The specific process includes: downsampling the pixels in the three RGB channels of the target color image from m×n pixels to m1×n1 pixels; copying the m1×n1 pixels corresponding to the three RGB channels in the channel dimension until the number of channels is expanded from l to l1, and finally obtaining a three-dimensional image of m1×n1×l1. Figure 2 Schematic diagram of three-dimensional expansion of color images.

[0034] It can be understood that the number of channels of the original target color image is 3. If the subsequent surfacelet transformation is to be performed on the target color image, the target color image needs to be expanded first.

[0035] S1012: importing l1 matrices of size m1×n1 into the surfdec() function for surfacelet transformation decomposition, and obtaining l1 transformation coefficient matrices of size m1×n1;

[0036] S1013: performing hard threshold processing on l1 transform coefficient matrices of size m1×n1;

[0037] It should be noted that when the embedding rate is 0.2 bit / pixel, the parameter σ used for hard threshold processing is set to 1, and the processing effect is optimal, and the processed coefficients are obtained.

[0038] S1014: performing surfacelet inverse transformation on the l1 transformation coefficient matrices of size m1×n1 after hard threshold processing and performing image reconstruction to obtain a filtered image;

[0039] S1015: Subtract the pixel values ​​in the RGB three channels of the filtered image from the pixel values ​​in the RGB three channels of the original target color image to obtain a steganographic residual signal. Figure 3 Schematic diagram of the steganalysis residual signal.

[0040] It can be understood that the relevant operations used for the surfacelet transformation in this embodiment can all be implemented using relevant functions or modules in the SurfBox library.

[0041] S102: Building a steganalysis detection model, including: constructing a steganalysis detection network, optimizing the parameters of the steganalysis detection network using the steganalysis residual signal of the training set, and using the steganalysis detection network with the optimal parameters as the steganalysis detection model;

[0042] It can be understood that the process of obtaining the steganalysis residual signal of the training set is as shown in the above step S101, which will not be repeated here.

[0043] Specifically, in this embodiment, ResNet101 is used as the steganalysis detection network for feature extraction and detection; the network structure diagram is shown in Figure 4 As shown, it includes the following five parts: conv1, block1, block2, block3 and block4. Among them, block1, block2, block3 and block4 are 4 residual blocks. The network processes the input stego residual signal as follows:

[0044] First, in the conv1 block, 64 convolution kernels of size 7×7 are used to perform convolution with a step size of 2 on the input stego residual signal, and the resulting feature map is as follows: Figure 5 As shown in (a); a pooling layer of size 3×3 is then used to perform maximum pooling with a step size of 2 on the feature map to obtain an output of 64 channels;

[0045] Then, in block1, the output data of conv1 block is first convolved with 64 convolution kernels of size 1×1, then the input data is convolved with 64 convolution kernels of size 3×3, and finally the input data is convolved with 256 convolution kernels of size 1×1; this block1 is repeated 3 times, and the feature map obtained is as follows Figure 5 (b) as shown.

[0046] Next, the other three blocks (block2, block3, and block4) undergo similar operations; among them, block2 is repeated 4 times, and the resulting feature map is as follows Figure 5 (c) shows that block3 is repeated 23 times, and the feature map obtained is as follows Figure 5 (d) shows that block4 is repeated three times, and the feature map obtained is as follows: Figure 5 As shown in (e), it can be seen that a total of 3+4+23+3=33 residual blocks have been passed, each residual block has 3 layers, so there are 33×3=99 layers.

[0047] Finally, the output data of the 99th layer is input into a fully connected layer (including the average pooling layer and the Softmax layer) for carrier image and dense image classification.

[0048] S103: Input the steganalysis residual signal of the target color image into a steganalysis detection model to obtain a detection result.

[0049] In order to verify the effectiveness of the scheme of the present invention, the present invention uses the J-UNIWARD steganography algorithm to embed secret information in each channel, divides the obtained secret images into training sets and test sets, trains and tests the ResNet101 model, and finally obtains the confusion matrix and ROC curve, as shown in Figure 6 As shown; through the confusion matrix, it can be calculated that the accuracy of the ResNet101 model is: 99.4%, and the ROC value is: 0.999.

[0050] It can be seen from the experimental results that the steganalysis detection model provided by the present invention has achieved good detection results in the classification of residual images.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning color image steganography detection method based on surfacelet filter preprocessing, characterized in that: include: Step 1: Filter the target color image using Surfacelet filtering to extract the steganalysis residual signal; Step 2: Building a steganalysis detection model, including: constructing a steganalysis detection network, optimizing the parameters of the steganalysis detection network using the steganalysis residual signal of the training set, and using the steganalysis detection network with the optimal parameters as the steganalysis detection model; Step 3: Input the steganalysis residual signal into the steganalysis detection model to obtain the detection result.

2. According to claim 1, the deep learning color image steganography detection method based on surfacelet filter preprocessing is characterized in that: Step 1 specifically includes: The target color image of size m×n×l is expanded three-dimensionally to a three-dimensional image of size m1×n1×l1; where m, n and l represent the number of rows, columns and channels of the target color image, respectively, and l=3; m1, n1 and l1 represent the number of rows, columns and channels of the three-dimensional image, respectively, and m1 <m,n1<n,l1> l; Import l1 matrices of size m1×n1 into the Surfdec() function for surfacelet transform decomposition to obtain l1 transformation coefficient matrices of size m1×n1; Perform hard thresholding on l1 transform coefficient matrices of size m1×n1; Perform surfacelet inverse transformation on the l1 transformation coefficient matrices of size m1×n1 after hard threshold processing and reconstruct the image to obtain the filtered image; The pixel values ​​in the RGB three channels of the filtered image are subtracted from the pixel values ​​in the RGB three channels of the original target color image to obtain the stego residual signal.

3. The deep learning color image steganography detection method based on surfacelet filter preprocessing according to claim 2 is characterized in that: The target color image of size m×n×l is three-dimensionally expanded to a three-dimensional image of size m1×n1×l1, specifically including: The pixels in the three RGB channels of the target color image are downsampled from m×n pixels to m1×n1 pixels; In the channel dimension, the m1×n1 pixels corresponding to the three RGB channels are copied until the number of channels is expanded from l to l1, and finally a three-dimensional image of m1×n1×l1 is obtained.

4. The deep learning color image steganography detection method based on surfacelet filter preprocessing according to claim 1 is characterized in that: ResNet101 is used as the steganalysis detection network.

Images