A spatial adaptive image steganalysis security detection method based on T-CNN

By introducing transposed convolution and high-pass filters into convolutional neural networks, image noise residual information is extracted, and combined with multi-layer convolutional layers and classification modules, the problem of poor image steganography analysis detection effect in the prior art is solved. , realizes efficient steganographic feature extraction and recognition, and improves network information security.

CN114240723BActive Publication Date: 2024-12-06NORTH CHINA UNIVERSITY OF SCIENCE AND TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111555265.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-17
Publication Date
2024-12-06
Estimated Expiration
2041-12-17

AI Technical Summary

Technical Problem

The detection effect of existing image steganography analysis methods has not yet reached the ideal level, and it is difficult to effectively identify the illegal use of image steganography, posing a potential threat to national security.

Method used

A T-CNN network structure is proposed, which fuses transposed convolution into a convolution neural network, extracts image noise residual information through a high-pass filter, combines transposed convolution and multi-layer convolution layer, and uses ReLU activation function and Softmax classification module realizes adaptive image steganography analysis in the airspace.

Benefits of technology

It significantly improves the detection accuracy of image steganography analysis, can effectively identify the use of image steganography, reduces threats to national interests and security, and improves network information security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114240723B_ABST
    Figure CN114240723B_ABST
Patent Text Reader

Abstract

The present invention discloses a spatial domain adaptive image steganalysis security detection method based on T-CNN, comprising the following steps: determining the size of the original image in the data set: adjusting it to 128×128 pixels, randomly selecting 30% of the processed images as carrier images in the training set, and 10% as carrier images in the test set; using a steganalysis algorithm to steganographically perform steganalysis on the randomly selected carrier images; inputting the T-CNN network structure for preprocessing, extracting residual information of image noise, and highlighting noise; performing a transposed convolution operation to double the size of the feature map; performing a convolution operation; classifying the image to detect whether the image is a stego image. The method can effectively extract image steganalysis features, realize efficient detection of spatial domain adaptive image steganalysis, reduce the damage caused when steganography is used by lawbreakers for illegal activities, and better maintain the information security of images in the network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image steganalysis in cyberspace security, and in particular to a spatial domain adaptive image steganalysis security detection method that fuses transposed convolution into a convolutional neural network. Background Art

[0002] Image steganography is a technology that embeds secret information into a carrier image so that the two communicating parties can communicate covertly without being discovered by other users. It plays an important role in network information security, military field and other aspects, and is of great significance for encrypted information transmission and maintaining data security storage. At the same time, image steganography can also be used by terrorists and criminals to spread messages, bringing potential serious harm to the country and society. How to effectively supervise the use of image steganography technology and prevent illegal applications has become a matter of close concern to national security and other departments. Image steganalysis is an attack on steganography and is a key technology to solve the problem of illegal use of image steganography.

[0003] As steganography continues to improve, it becomes increasingly difficult to correctly detect images that carry secrets. Steganography can easily be exploited by people with ill intentions. They can divide the secret information that needs to be transmitted into multiple pieces of data, and then hide these pieces of data in multiple images. By reducing the amount of information embedded in each image, they can evade detection by steganalysis methods. At the receiving end, they only need to extract the secret data from these images and then merge them to obtain the secret information. Therefore, the detection effect of existing steganalysis methods needs to be further improved.

[0004] In recent years, many researchers have used convolutional neural networks to study image steganalysis methods. Through a large amount of data training, feature expressions can be automatically learned in the data, thereby improving the accuracy of image steganalysis, which has very important practical significance for combating terrorist actions, preventing the leakage of commercial secrets and other cyberspace security aspects. Qian et al. proposed a five-layer convolutional neural network structure. The preprocessing layer of the structure uses a high-pass filter. According to the characteristics of steganalysis noise, Gaussian activation function and average pooling layer are used in the structure. The detection accuracy of the experimental results is only 3%~4% lower than that of the traditional image steganalysis rich model. Liu Mingming et al. reduced the number of convolutional layers, increased the number of convolution kernels and disabled the pooling layer. The experimental results showed that the detection results of the network structure were relatively ideal when the embedding rate was 0.4bpp and 0.1bpp. While transposed convolution has been widely used in target detection, natural language processing, image segmentation and face recognition, it is still rarely used in the field of image steganalysis. Integrating transposed convolution into convolutional neural networks for image steganalysis can effectively improve the detection accuracy of image steganalysis, which is conducive to accurately identifying whether the image steganography is used by some lawless elements or illegal organizations, thereby avoiding direct or indirect damage to national interests and posing a huge threat to national security. Summary of the invention

[0005] The purpose of this invention is to apply the concept of deep learning to the field of image steganalysis, utilize transposed convolution to fuse into convolutional neural network, and propose an improved network structure T-CNN, which is called Transpose-Convolutional Neural Network, for spatial adaptive image steganalysis security detection, improve the detection effect of image steganalysis, reduce the use of image steganography by criminals, damage national interests and threaten national security, and maintain the information security of images in the network.

[0006] In order to solve the problem that the detection effect of the existing steganalysis method needs to be further improved, the present invention provides a spatial domain adaptive image steganalysis security detection method based on T-CNN, wherein the method comprises the following steps:

[0007] Step 1: Determine the size of the original image in the dataset. If the image size is larger than 128×128 pixels, crop it to 128×128 pixels. If the image size is smaller than 128×128 pixels, enlarge the image size to 128×128 pixels.

[0008] Step 2: Randomly select 30% of the processed images as carrier images in the training set, and randomly select 10% as carrier images in the test set;

[0009] Step 3: Use the steganography algorithm to steganographically record the randomly selected carrier image, and obtain the secret images in the training set and the secret images in the test set respectively;

[0010] Step 4: Input the training set images into the T-CNN network structure for preprocessing, extract the residual information of image noise, and highlight the noise. The preprocessing layer contains 30 high-pass filters, whose weights are fixed, the learning rate is set to 0, and they do not participate in the learning of the network. The TLU activation function is used after the preprocessing layer to retain the positive and negative residual information and the high-frequency noise introduced by image steganography.

[0011] Step 5: Perform a transposed convolution operation on the preprocessed feature map to amplify the image feature information: transposed convolution → batch normalization → ReLU activation function;

[0012] Step 6: Perform convolution operation on the feature map obtained after transposed convolution to extract image features: convolution layer → batch normalization → ReLU activation function, a total of five convolution layers, the last layer is the global convolution layer, and the convolution kernel size used is equal to the size of the feature map input to this layer;

[0013] Step 7: The feature map obtained after the convolution operation enters the classification module: two fully connected layers and one Softmax layer. The ReLU activation function is used between the two fully connected layers to further increase the nonlinear expression ability of the model. Finally, the Softmax layer outputs the training detection result of whether the image is a steganographic image.

[0014] Step8: Input the test set image into the T-CNN network structure and process it from Step4 to Step7 to verify whether it is the detection result of the steganographic image.

[0015] Furthermore, the steganographic algorithms in Step 3 are S-UNIWARD and HUGO.

[0016] Furthermore, in the preprocessing operation in Step 4, a convolution layer with 30 high-pass filters is used, one of which has a convolution kernel size of 3 × 3 and the other 29 have a convolution kernel size of 5 × 5. All high-pass filters are used after standardization.

[0017] Furthermore, the convolution kernel size of the transposed convolution operation in Step 5 is 3×3, and the step size is 2. After the transposed convolution, the size of the feature map is doubled to achieve the purpose of amplifying the steganographic noise, making the steganographic features easier to extract.

[0018] Furthermore, the convolution module in Step 6 fully extracts image features. The first convolution layer uses 64 3×3 convolution kernels and uses a step size of 2 to replace the pooling layer. Then, 32 1×1 convolution kernels are used to reduce the number of channels. The parameters of the third convolution layer are consistent with those of the first layer. The fourth convolution layer contains 64 1×1 convolution kernels. The last layer is a global convolution layer, which uses 128 64×64 convolution kernels, which is equal to the size of the feature map input to this layer. The output size of this layer is 1×1.

[0019] Furthermore, the number of neurons in the two fully connected layers in Step 7 is 1000 and 2 respectively, and the outputs of multiple neurons are mapped to the interval [0, 1] by the Softmax function, and whether it is a detection result of a stego image is determined according to the size of the output value.

[0020] The beneficial effects of the present invention are as follows: by inputting the image into the T-CNN network structure for preprocessing, the residual information of the image noise is extracted, the noise is highlighted, and the TLU activation function is used after the preprocessing layer to retain the positive and negative residual information and the high-frequency noise introduced by the image steganography; after the transposed convolution operation, the size of the feature map is doubled to achieve the purpose of amplifying the steganographic noise, so that the steganographic features are easier to extract; the convolution module operation is performed to fully extract the image features; the ReLU activation function is used after the transposed convolution layer and the five convolution layers to solve the gradient explosion and gradient disappearance problems, and accelerate the convergence speed of the network structure; finally, the image is classified through the fully connected layer and the Softmax layer. The method can effectively extract the image steganographic features, realize the efficient detection of spatial domain adaptive image steganalysis, and accurately identify whether the image steganography is used by some lawless elements or illegal organizations, so as to avoid directly or indirectly damaging the interests of the country and posing a huge threat to national security. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The overall structure diagram of the present invention;

[0022] Figure 2 Overall method steps diagram;

[0023] Figure 3 Example of a secret image;

[0024] Figure 4 TLU activation function diagram;

[0025] Figure 5 Transposed convolution operation process diagram;

[0026] Figure 6 ReLU activation function diagram;

[0027] Figure 7Detailed structure diagram of transposed convolution and convolution module;

[0028] Figure 8 Comparison curve of accuracy with and without transposed convolution structure;

[0029] Fig. 9 Comparison chart of using S-UNIWARD steganography algorithm and other network structures;

[0030] Fig.10 Comparison chart of HUGO steganography algorithm and other network structures. DETAILED DESCRIPTION

[0031] The present invention introduces transposed convolution into convolutional neural network and proposes a new network structure T-CNN for spatial domain adaptive image steganalysis security detection. Figure 1 As shown in the figure. First, the image is preprocessed through a high-pass filter to extract the residual information of the image noise and highlight the noise; then the transposed convolution operation is performed to fully analyze the steganographic features; then the magnified image features are fully extracted through five convolutional layers; finally, the image is classified through the fully connected layer and the softmax layer. The image steganalysis accuracy of the image using the spatial domain adaptive steganography algorithm of this network structure has achieved excellent detection results.

[0032] In general, the image steganalysis security detection process of the present invention can be used as follows: Figure 2 As shown:

[0033] 1. Obtain a dataset for image steganalysis

[0034] Determine the size of the original image in the data set: if the image size is larger than 128×128 pixels, crop it to 128×128 pixels; if the image size is smaller than 128×128 pixels, enlarge the image size to 128×128 pixels. Randomly select 30% of the processed images as carrier images in the training set, and randomly select 10% as carrier images in the test set. The original images are steganographically processed using the spatial domain adaptive steganography algorithm at embedding rates of 0.1, 0.2, 0.3, and 0.4bpp to obtain the steganographic images. The original images and the steganographic images together constitute the training set and the test set. Examples of steganographic images are as follows: Figure 3 shown.

[0035] Preprocessing images

[0036] The noise of image steganography is mainly concentrated in the high-frequency part. The high-pass filter processing can extract the edges and details of the image, especially the part where the pixels change sharply. The preprocessed image is more conducive to steganalysis. The preprocessing layer contains a convolution layer composed of 30 high-pass filters, one of which has a convolution kernel size of 3×3 and the other 29 have a convolution kernel size of 5×5. There are seven types of high-pass filters, as shown below:

[0037]

[0038] Among them, "1st", "2nd" and "3rd" can obtain 8, 4 and 8 filters respectively after a 45-degree rotation operation, and "EDGE3×3" and "EDGE5×5" can obtain 4 filters respectively after a 90-degree rotation operation. By filling 0, the filter sizes of "1st", "2nd", "3rd", "EDGE3×3" and "EDGE5×5" are unified to 5×5. The learning rate of the preprocessing layer is set to 0 and does not participate in the learning of the network. The use of multiple high-pass filters can effectively extract noise residuals, which is beneficial for the convolution layer to extract image steganographic features, and the model is easier to converge during the training process. Since the preprocessing layer mainly extracts the residual information of image noise, and these residual information are positive and negative, the information they contain is meaningful for steganalysis. Considering this problem, the TLU function is selected as the activation function after the preprocessing layer, and its function curve is as follows Figure 4 shown.

[0039] Amplifying stegographic features

[0040] The relationship between transposed convolution and ordinary convolution is only the transformation relationship in size before and after convolution. The transposed convolution operation will not expand the perception field of view. Suppose a 5×5 pixel image is used as input, where the convolution kernel size is 3×3, the boundary expansion is 1, and the step size is 2. Ordinary convolution will output a 3×3 pixel image. Now take the 3×3 pixel image as input, perform a transposed convolution operation, expand the image size by zero padding, set the step size to 1, and traverse the output image size to 5×5 pixels. The process of transposed convolution operation is as follows: Figure 5 shown.

[0041] After image preprocessing, transposed convolution is used to amplify the steganographic features. The transposed convolution layer uses a convolution kernel size of 3×3 and a step size of 2. After transposed convolution, the input feature map size will be doubled. After that, the batch normalization layer is connected to realize normalization processing, and the numerical distribution is changed to a normal distribution with a mean of 0 and a variance of 1, which speeds up the convergence of the network. The transposed convolution layer uses the ReLU function as the activation function to introduce nonlinear factors into the network and improve the fitting ability of the convolutional neural network. Its curve is as follows: Figure 6 shown.

[0042] Feature extraction and recognition image classification

[0043] After the steganographic features are enlarged by transposed convolution, the image is convolved to extract image features. The convolution module contains 5 ordinary convolution layers. The first convolution layer uses 64 3×3 convolution kernels and uses a step size of 2 to replace the use of the pooling layer to prevent the loss of steganographic features and suppress overfitting. Then, 32 1×1 convolution kernels are used to reduce the number of channels, effectively reduce the number of parameters, save the GPU computing cost, and speed up the training of the model. The parameters of the third convolution layer are the same as those of the first layer. The fourth convolution layer contains 64 1×1 convolution kernels, which are mainly used to integrate cross-channel features and provide cross-channel feature extraction information interaction. The last layer is a global convolution layer, which is used to simplify features and improve detection performance. This layer uses 128 64×64 convolution kernels, which is equal to the size of the feature map input to this layer. Batch normalization layer processing is introduced after each convolution layer in the convolution module to effectively solve the problem of changes in data distribution. After batch normalization, the activation function uses the ReLU function. The detailed structure of the transposed convolution and convolution module is as follows Figure 7 shown.

[0044] After extracting image features, the image is classified to determine whether it is a confidential image. The classification module acts like a classifier, and contains two fully connected layers and one Softmax layer. The number of neurons in the two fully connected layers is 1000 and 2 respectively. The classification module finally uses the Softmax function to map the output of multiple neurons to the [0,1] interval, and determines the final classification result based on the size of the output value. The ReLU activation function is still used between the two fully connected layers to further increase the nonlinear expression ability of the network structure.

[0045] In order to illustrate the effectiveness of the above-mentioned solution implemented by the present invention, an experimental description is given below.

[0046] Experiment on the effectiveness of transposed convolution fusion into convolutional neural network

[0047] In order to prove that using transposed convolution after the preprocessing layer can help the network perform image steganalysis better, a comparative experiment was conducted on the use of the transposed convolution structure. The network layer where the transposed convolution module is located is removed, and the rest of the structure remains unchanged, forming structure A; the transposed convolution layer is replaced with a normal convolution layer, and the convolution layer parameters are set to 32 convolution kernels of size 3×3, and the step size is set to 1, forming structure B; in addition, the step size is set to 2, forming structure C. The setting of structure B and structure C is to reduce the interference of the number of network layers. The experimental accuracy comparison curve is shown in the figure below. Figure 8 As shown. Figure 8It can be concluded that the accuracy of structure A drops significantly after removing the transposed convolution layer, indicating that the transposed convolution layer has a positive impact on the model performance. The setting of different step sizes for structures B and C is to make the experiment more rigorous and prevent accuracy loss caused by inappropriate step sizes. The number of network layers of structures B and C is the same as that of T-CNN, but the accuracy is still not as good as T-CNN, which proves that the excellent performance of T-CNN is indeed contributed by the transposed convolution, not by the deeper network structure.

[0048] Comparison experiments with other models

[0049] In order to evaluate the effectiveness of the present invention as a whole, the method was compared with the existing steganalysis network structure. At the same time, considering the multi-faceted evaluation of the performance of the method, data sets with various embedding rates and two steganography algorithms were selected for the experiment. The embedding rates were selected as 0.1, 0.2, 0.3 and 0.4bpp, and the steganography algorithms were selected as S-UNIWARD and HUGO. The accuracy comparison results of this method with the other three network structures Xu-Net, Xiang's model and S-CNN are shown in the figure below. Fig. 9 and 10 As shown in the figure, it can be seen that the steganalysis accuracy of T-CNN is the best. This excellent performance is not only reflected in the case of slightly higher embedding rate, but also in the case of low embedding rate data, T-CNN can still show good results. This can fully illustrate that if steganography is used by criminals, even by reducing the amount of embedded information in each image, the method of the present invention can also have excellent steganalysis detection performance.

[0050] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A spatial domain adaptive image steganalysis security detection method based on T-CNN, the method comprising the following steps: Step 1: Determine the size of the original image in the dataset. If the image size is larger than 128×128 pixels, crop it to 128×128 pixels. If the image size is smaller than 128×128 pixels, enlarge the image size to 128×128 pixels. Step 2: Randomly select 30% of the processed images as carrier images in the training set, and randomly select 10% as carrier images in the test set; Step 3: Use the steganography algorithm to steganographically record the randomly selected carrier image, and obtain the secret images in the training set and the secret images in the test set respectively; Step 4: Input the training set images into the T-CNN network structure for preprocessing, extract the residual information of image noise, and highlight the noise; the preprocessing layer contains 30 high-pass filters, whose weights are fixed and the learning rate is set to 0, and they do not participate in the learning of the network; the TLU activation function is used after the preprocessing layer to retain the positive and negative residual information and the high-frequency noise introduced by image steganography; Step 5: Perform a transposed convolution operation on the preprocessed feature map to amplify the image feature information: transposed convolution → batch normalization → ReLU activation function; Step 6: Perform convolution operation on the feature map obtained after transposed convolution to extract image features: convolution layer → batch normalization → ReLU activation function, a total of five convolution layers, the last layer is the global convolution layer, and the convolution kernel size used is equal to the size of the feature map input to this layer; Step 7: The feature map obtained after the convolution operation enters the classification module: two fully connected layers and one Softmax layer. The ReLU activation function is used between the two fully connected layers to further increase the nonlinear expression ability of the model. Finally, the Softmax layer outputs the training detection result of whether the image is a stego image; Step8: Input the test set image into the T-CNN network structure and process it from Step4 to Step7 to verify whether it is the detection result of the steganographic image.

2. According to claim 1, a spatial domain adaptive image steganalysis security detection method based on T-CNN is characterized in that: The steganographic algorithms in Step 3 are S-UNIWARD and HUGO.

3. According to claim 1, a spatial domain adaptive image steganalysis security detection method based on T-CNN is characterized in that: In the preprocessing operation in Step 4, a convolution layer with 30 high-pass filters is used, one of which has a convolution kernel size of 3×3 and the other 29 have a convolution kernel size of 5×5. All high-pass filters are used after standardization.

4. According to claim 1, a spatial domain adaptive image steganalysis security detection method based on T-CNN is characterized in that: The convolution kernel size of the transposed convolution operation in Step 5 is 3×3, and the step size is 2. After the transposed convolution, the size of the feature map is doubled to achieve the purpose of amplifying the steganographic noise, making the steganographic features easier to extract.

5. According to the T-CNN-based spatial adaptive image steganalysis security detection method of claim 1, it is characterized in that: The convolution module in Step 6 fully extracts image features. The first convolution layer uses 64 3×3 convolution kernels and uses a step size of 2 to replace the pooling layer. Then, 32 1×1 convolution kernels are used to reduce the number of channels. The parameters of the third convolution layer are consistent with the first layer. The fourth convolution layer contains 64 1×1 convolution kernels. The last layer is a global convolution layer, which uses 128 64×64 convolution kernels, which is equal to the size of the feature map input to this layer. The output size of this layer is 1×1.

6. According to the T-CNN-based spatial adaptive image steganalysis security detection method of claim 1, it is characterized in that: The number of neurons in the two fully connected layers in Step 7 is 1000 and 2 respectively, and the outputs of multiple neurons are mapped to Within the interval, whether it is a detection result of a stego image is determined based on the size of the output value.