Image steganography method based on high-frequency distribution mapping depth reversible network

Abstract

The invention relates to an image steganography method based on a high-frequency distribution mapping depth reversible network, and the method comprises the steps: designing a reversible steganography module and a high-frequency distribution mapping module based on an affine coupling structure, and constructing a depth reversible steganography network; combining the forward hidden network and the reverse recovery network, constructing a corresponding loss function, and training an image steganography process, an image recovery process and a high-frequency distribution mapping process for the deep reversible steganography network to obtain a trained deep reversible steganography network; and on the basis of the trained deep reversible steganography network, image steganography, high-frequency distribution mapping and image recovery processing are sequentially carried out on the current carrier image and the secret image, and a secret-carrying image and a recovered image are respectively obtained. Compared with the prior art, the method has the advantages that the high-frequency distribution mapping module is designed to improve the recovery quality of the secret image in the reversible neural network, global end-to-end image steganography is realized, and the image steganography safety, the recovery image quality and the steganalysis resistance can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an image steganography method based on a high-frequency distribution mapping deep reversible network. Background Art

[0002] In recent years, with the advancement of computing power and artificial intelligence, deep learning-based image steganography has become mainstream. Image steganography, where images are hidden within images, aims to conceal a secret image within a carrier image, achieving both undetectability and recoverability. This means that a good image steganography scheme should not only achieve high-quality image hiding but also high-quality image recovery. However, improving image recovery quality remains a significant challenge. Furthermore, most existing image steganography schemes often result in noticeable outline artifacts and color distortion, making the stored image easily detectable by steganalysis tools.

[0003] Research has shown that reversible neural networks treat the image steganography and recovery processes as a pair of reversible processes using a mutually inverse encoder-decoder pair. Furthermore, the forward hiding process and the reverse recovery process share all learnable parameters, significantly reducing the number of model parameters. Therefore, introducing reversible neural networks into the field of image steganography not only allows the image to be hidden as secret information within another carrier image, effectively generating a secret image, but also leverages the symmetrical structure of the reversible neural network to recover the hidden secret image with high quality. This tightly coupled structure significantly enhances the effectiveness of image steganography.

[0004] However, due to the strict symmetry of reversible neural networks, the forward hidden network outputs both the secret image and the lost information. Assuming the true receiver can obtain the lost information, they can achieve perfect image recovery. However, in reality, the lost information is unknown, and the receiver often uses the received secret image and randomly sampled auxiliary variables for recovery. Therefore, how to map the lost information from the received secret image is a pressing problem. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an image steganography method based on a high-frequency distribution mapping deep reversible network, which can model and map the secret image and lost information, and effectively improve the image steganography ability and image recovery ability.

[0006] The purpose of the present invention can be achieved by the following technical solution: an image steganography method based on a high-frequency distribution mapping deep reversible network, comprising the following steps:

[0007] S1. Based on the affine coupling structure, a reversible steganography module and a high-frequency distribution mapping module are designed to construct a deep reversible steganography network.

[0008] S2. Combining the forward hidden network and the reverse recovery network, and constructing the corresponding loss function, the deep reversible steganography network is trained to perform image steganography, image recovery, and high-frequency distribution mapping processes to obtain the trained deep reversible steganography network;

[0009] S3. Based on the trained deep reversible steganography network, image steganography, high-frequency distribution mapping, and image recovery processing are performed on the current carrier image and secret image in sequence to obtain the carrier image and restored image respectively.

[0010] Furthermore, the step S1 includes the following steps:

[0011] S11. Construct a reversible 1×1 convolutional layer to perform channel permutation strategy on the input carrier image and secret image;

[0012] S12, realizing affine transformation of the reversible steganography module and the high-frequency distribution mapping module by using additive coupling and multiplicative coupling structures;

[0013] S13. Design nonlinear functions based on the DenseNet structure and integrate the spatial attention mechanism into the DenseNet structure.

[0014] Furthermore, the forward process of the reversible steganography module is as follows:

[0015] h1=x1·exp(α(s(x2)))+s(x2)

[0016] h2=x2·exp(α(s(h1)))+s(h1)

[0017] Where x1 and x2 are the features of the carrier image and secret image, respectively; α() is the activation function; s() is a nonlinear function constructed by a neural network; h1 and h2 are the features of the secret image and the features of the lost information, respectively, which will be used as the input of the subsequent high-frequency distribution mapping module for modeling;

[0018] The forward process of the high frequency distribution mapping module is as follows:

[0019] t=h2·exp(α(s(h1)))+s(h1)

[0020] The high-frequency distribution mapping module learns and models h1 and h2 through an affine coupling structure to obtain t;

[0021] The reverse process of the high frequency distribution mapping module is as follows:

[0022] hf =(zs(y1))÷exp(α(s(y1)))

[0023] Among them, y1 and z are the input ends of the high-frequency distribution mapping module, representing the characteristics of the encrypted image and auxiliary variables respectively, α() is the activation function, s() is a nonlinear function constructed by a neural network, and the high-frequency distribution mapping module maps the auxiliary variable z to h f , so that h f It not only conforms to the Gaussian distribution, but also contains effective information that helps with recovery;

[0024] The reverse process of the reversible steganography module is as follows:

[0025] x2′=(h f -s(y1))÷exp(α(s(y1)))

[0026] x1′=(y1-s(x2′))÷exp(α(s(x2′)))

[0027] Among them, x′2 and x′1 are the recovered carrier image and secret image.

[0028] Furthermore, the specific process of step S11 is as follows:

[0029] S111. Superimpose the secret image and the carrier image into a feature map of size (h, w, c), where h, w, and c represent the length and width and the number of channels of the feature map, respectively.

[0030] S112, using LR decomposition to generate a rotation matrix as a custom weight matrix W;

[0031] S113: Use a 1×1 filter and a weight matrix W to perform a convolution operation with a step size of 1 on the feature map to obtain a mixed feature after channel permutation.

[0032] Furthermore, the specific process of step S13 is as follows:

[0033] S131. Construct a DenseNet network consisting of five convolutional layers and initialize the weights of the five convolutional layers to 0 using Xavier's initialization scheme.

[0034] S132, use a 3×3 filter to convolve the input features, and use the LeakyReLU activation function after each convolution layer;

[0035] S133, after the fifth convolution layer, use exp() to get the final result;

[0036] S134. Insert a spatial attention module between the fourth and fifth layers in the DenseNet network;

[0037] S135, perform maximum pooling and average pooling on the features processed by the DenseNet network, and superimpose the results of the two along the channel axis;

[0038] S136, using a 7×7 filter to perform a convolution operation on the superposition result, and then using a sigmoid function for activation to obtain a spatial feature map;

[0039] S137. The final output is the product of the input feature and the spatial feature map.

[0040] Furthermore, the image steganography process in step S2 includes:

[0041] Using Discrete Wavelet Transform (DWT), the wavelet subbands of the carrier image and the secret image are used as inputs to the forward hidden network.

[0042] Image steganography is completed through the reversible steganography module to obtain the secret image and the information lost in the steganography process. The high-frequency distribution mapping module is used to model the secret image and the information lost in the steganography process. Finally, the inverse transform of the discrete wavelet transform (IWT) is used to restore the secret image and complete the image steganography process.

[0043] Furthermore, the image restoration process in step S2 includes:

[0044] Using discrete wavelet transform (DWT), the wavelet subband of the input secret image and auxiliary variables is used as the input of the inverse recovery network. The high-frequency distribution mapping module maps out the lost information according to the input secret image and auxiliary variables, and completes the image steganography through the reversible steganography module. Finally, the inverse transform (IWT) of the discrete wavelet transform is used for restoration to obtain the recovered secret image and secret image, completing the image recovery process.

[0045] Furthermore, the auxiliary variables are specifically sampled from a Gaussian distribution: Among them, z is an auxiliary variable used to replace the lost information in the steganographic process, μ0 is the mathematical expectation of the normal distribution, is the variance of the normal distribution.

[0046] Furthermore, the loss function in step S2 includes a steganographic process loss function, a steganographic process wavelet loss function and a restoration process loss function.

[0047] Furthermore, the loss function of the steganography process is specifically:

[0048]

[0049] Where N represents the number of training samples, MSE represents the mean square error, Represent the carrier image and the generated encrypted image respectively;

[0050] The wavelet loss function of the steganography process is specifically:

[0051]

[0052] Among them, DWT() LL Represents the low-frequency component of discrete wavelet transform;

[0053] The recovery process loss function is specifically:

[0054]

[0055] in, They represent the secret image and the secret image finally recovered from the secret image.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] The present invention designs a reversible steganography module and a high-frequency distribution mapping module based on an affine coupling structure to construct a deep reversible steganography network. This network is then combined with a forward hidden network and a reverse recovery network, and a corresponding loss function is constructed. The deep reversible steganography network is trained through image steganography, image recovery, and high-frequency distribution mapping to obtain a trained deep reversible steganography network. This utilizes the high-frequency distribution mapping module to model and map the secret image and lost information, effectively improving the quality of secret image recovery. This method also achieves global, end-to-end image steganography, with high image steganography and image recovery capabilities, while ensuring high security and resistance to steganalysis. Furthermore, the proposed method exhibits excellent scalability, enabling the use of a cascaded structure to implement a larger number of image steganography applications and apply them to more complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the method flow of the present invention;

[0059] Figure 2 A schematic diagram of the overall framework of a multi-image steganography method based on a high-frequency distribution mapping deep reversible network provided in an embodiment;

[0060] Figure 3 Schematic diagram of the cascade structure in the embodiment;

[0061] Figure 4 Schematic diagram of the DenseNet network and spatial attention module in the embodiment;

[0062] Figure 5 Schematic diagram of the residual before and after image steganography in the embodiment;

[0063] Figure 6 Schematic diagram of the comparison of high-frequency components of the encrypted image in the embodiment;

[0064] Figure 7 Schematic diagram of comparison of high-frequency components of the restored secret image 1 in the embodiment;

[0065] Figure 8 Schematic diagram of comparison of high-frequency components of the restored secret image 2 in the embodiment. DETAILED DESCRIPTION

[0066] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] Example

[0068] like Figure 1 As shown, an image steganography method based on a high-frequency distribution mapping deep reversible network includes the following steps:

[0069] S1. Based on the affine coupling structure, a reversible steganography module and a high-frequency distribution mapping module are designed to construct a deep reversible steganography network.

[0070] S2. Combining the forward hidden network and the reverse recovery network, and constructing the corresponding loss function, the deep reversible steganography network is trained to perform image steganography, image recovery, and high-frequency distribution mapping processes to obtain the trained deep reversible steganography network;

[0071] S3. Based on the trained deep reversible steganography network, image steganography, high-frequency distribution mapping, and image recovery processing are performed on the current carrier image and secret image in sequence to obtain the carrier image and restored image respectively.

[0072] This embodiment applies the above technical solution to build a multi-image steganography scene. Figure 2 The application framework shown in the figure mainly includes the following processes:

[0073] (a) Based on the affine coupling structure, a reversible steganography module and a high-frequency distribution mapping module are built to realize a deep reversible steganography network;

[0074] (b) Using discrete wavelet transform (DWT), the wavelet subbands of the carrier image and the secret image are used as the input of the forward hidden network, and the wavelet subbands of the input secret image and auxiliary variables are used as the input of the reverse recovery network;

[0075] (c) Image steganography is completed through the reversible steganography module, and the secret image and the information lost during the steganography process are obtained. The high-frequency distribution mapping module is used to model the two. Finally, the inverse discrete wavelet transform (IWT) is used to restore the secret image, completing the steganography process of the deep reversible steganography network.

[0076] (d) Based on cascade structure (such as Figure 3 As shown), the generated secret image and the new secret image are used as new inputs, and step (c) is repeated until all the secret images are hidden;

[0077] (e) The high-frequency distribution mapping module maps out the lost information based on the input secret image and auxiliary variables, and completes the image steganography through the reversible steganography module. Finally, the inverse discrete wavelet transform (IWT) is used to restore the image to obtain the recovered secret image and the secret image, completing the recovery process of the deep reversible steganography network.

[0078] (f) Based on the cascade structure, the generated secret image and the new auxiliary variable are used as new inputs, and step (e) is repeated until the secret image is completely recovered;

[0079] (g) In addition, a loss function is constructed to train the steganography process, the recovery process and the high-frequency distribution mapping process of the deep reversible steganography network, wherein the steganography process, the recovery process and the high-frequency distribution mapping process share a deep reversible steganography network.

[0080] The specific process of step (a) is:

[0081] (a1) Construct a reversible 1×1 convolutional layer to perform channel permutation strategy on the input carrier image and secret image:

[0082] First, the secret image and the carrier image are superimposed into a feature map of size (h, w, c), where h, w, and c represent the length, width, and number of channels of the feature map, respectively.

[0083] Then use LR decomposition to generate the rotation matrix as the custom weight matrix W;

[0084] Then, a 1×1 filter and weight matrix W are used to perform a convolution operation on the feature map with a step size of 1 to obtain the mixed features after channel permutation;

[0085] (a2) Affine transformation of the reversible steganography module and the high-frequency distribution mapping module is realized by additive coupling and multiplicative coupling structures;

[0086] For the reversible steganography module, the forward process is as follows:

[0087] h1=x1·exp(α(s(x2)))+s(x2)

[0088] h2=x2·exp(α(s(h1)))+s(h1)

[0089] Among them, x1 and x2 are the features of the carrier image and the secret image respectively, α() is the activation function, s() is a nonlinear function constructed by a neural network, h1 and h2 are the features of the secret image and the features of the lost information respectively, which will be modeled as the input of the subsequent high-frequency distribution mapping module.

[0090] The forward process of the high frequency distribution mapping module is as follows:

[0091] t=h2·exp(α(s(h1)))+s(h1)

[0092] The high-frequency distribution mapping module learns and models h1 and h2 through an affine coupling structure to obtain t.

[0093] The reverse process of the high frequency distribution mapping module is as follows:

[0094] h f =(zs(y1))÷exp(α(s(y1)))

[0095] Among them, y1 and z are the input ends of the high-frequency distribution mapping module, representing the characteristics of the encrypted image and auxiliary variables respectively, α() is the activation function, s() is a nonlinear function constructed by a neural network, and the high-frequency distribution mapping module maps the auxiliary variable z to h f , so that h f It not only conforms to the Gaussian distribution, but also contains effective information that is helpful for recovery.

[0096] The reverse process of the reversible steganography module is as follows:

[0097] x2′=(h f -s(y1))÷exp(α(s(y1)))

[0098] x1′=(y1-s(x2′))÷exp(α(s(x2′)))

[0099] Among them, x′2 and x′1 are the recovered carrier image and secret image.

[0100] (a3) Figure 4 As shown in the figure, s() is designed according to the DenseNet structure, and the spatial attention mechanism is integrated into the DenseNet structure:

[0101] First, a DenseNet network consisting of five convolutional layers is constructed, and the weights of the five convolutional layers are initialized to 0 using Xavier's initialization scheme;

[0102] Use a 3×3 filter to convolve the input features, and use the LeakyReLU activation function after each convolution layer;

[0103] After the fifth layer of convolution, exp() is used to process the final result;

[0104] Insert a spatial attention module between the fourth and fifth layers in the DenseNet network;

[0105] Perform maximum pooling and average pooling on the features processed by the DenseNet network, and superimpose the results of the two along the channel axis;

[0106] A 7×7 filter is used to perform a convolution operation on the superposition result, and then a sigmoid function is used for activation to obtain a spatial feature map;

[0107] The final output is the product of the input feature and the spatial feature map.

[0108] In this embodiment, the parameters of the DenseNet network structure combined with the spatial attention module are shown in Table 1. The network consists of five convolutional layers and one spatial attention layer. The convolutional kernel size of the convolutional layer is 3×3, the stride is 2, and Leaky ReLU is used as the activation function. The number of convolutional kernels in the first four layers is 32, which is used to extract rich image features. Dense connections are used between convolutional layers. Each layer uses the output of all previous layers as input and passes its own output to all subsequent layers. The spatial attention module is introduced between the fourth and fifth convolutional layers to further extract image texture features. The output size of the fifth layer is the same as the input size of the first layer.

[0109] Table 1

[0110]

[0111] The specific process of step (b) is:

[0112] (b1) Decompose the image into low-frequency components and high-frequency components through DWT, and extract the features of the image at different frequencies;

[0113] (b2) During the recovery process, auxiliary variables are used to replace the lost information in the steganographic process. The auxiliary variables are sampled from a Gaussian distribution:

[0114] The specific process of step (g) is:

[0115] (g1) Loss function of the steganographic process:

[0116]

[0117] Where N represents the number of training samples, MSE represents the mean square error, Represent the carrier image and the generated encrypted image respectively;

[0118] (g2) Wavelet loss function of the steganographic process:

[0119]

[0120] Among them, DWT() LL Represents the low-frequency component of the discrete wavelet transform.

[0121] (g3) Recovery process loss function

[0122]

[0123] in, They represent the secret image and the secret image finally recovered from the secret image.

[0124] In order to verify the effectiveness of this solution, this embodiment uses a traditional technical solution and the method of the present invention for comparative testing, and compares the test results by means of scientific demonstration to verify the real effect of this solution.

[0125] Experimental platform: DELL W-2102 workstation; CPU: Intel i7-9700k; GPU: RTX 2080SUPER; RAM: 16GB; Operating system: Windows 10; Deep learning framework: PyTorch 1.12.1-gpu, CUDA 11.3;

[0126] The experiments used the DIVerse 2K (DIV2K), COCO, and ImageNet datasets. DIV2K contains 800 training images, 100 validation images, and 100 test images, for a total of 1,000 images at 2K resolution. We selected 800 DIV2K training images as the training set, cropped to a resolution of 128×128. We used 100 DIV2K test images at a resolution of 1024×1024 for testing. The COCO dataset contains over 330,000 images, covering most everyday scenes. This example selected 5,000 images from COCO 2017 at a resolution of 256×256 as the test set. ImageNet is an extremely large-scale visual database used for research on visual object recognition software. We selected 50,000 images from ImageNet at a resolution of 256×256 for testing. During training, due to the varying sizes of images in the dataset, we used the torchvision tool to resize them to a uniform 128×128 size using a center crop.

[0127] The experimental settings are as follows: Adam optimizer is used and the learning rate is set to 10 -4.5 , weight_decay is set to 0.98 and updated every 20 rounds; the total training rounds are 3000 times, and the batch size is set to 3, corresponding to the carrier image, secret image 1 and secret image 2 respectively; according to the above settings, the reversible multi-image steganography model based on high-frequency distribution mapping optimizes the loss function through continuous learning and training, so that the forward network can realize image hiding, and the reverse network can realize image recovery; the final model parameters are obtained when the parameters in the model remain stable or reach the given maximum number of iterations. Figure 5 、 Figure 6 、 Figure 7 and Figure 8 The visual presentation effects of this method are demonstrated in turn. To test the steganographic security of this method, this example uses the rich model-based SRM steganalysis algorithm to evaluate its effectiveness. It also uses the deep learning-based model SRNet to test the number of training samples that need to be leaked for SRNet to distinguish between the carrier image and the secret image. The test results are shown in Tables 2 and 3, respectively.

[0128] Table 2. SRM detection error rate for different steganography schemes

[0129]

[0130] Table 3 uses SRNet to test "how many training samples need to be leaked before SRNet can detect anomalies"

[0131]

[0132]

[0133] From the test results, it can be seen that by comparing the rich model-based steganalyzer with the deep learning-based steganalyzer, the security performance of this method is obviously improved compared with the iSCMIS method and the DeepMIH method, which fully demonstrates that the method of the present invention has a strong anti-steganalysis ability.

[0134] In summary, this scheme uses the high-frequency distribution mapping module to model and map the secret image and lost information, which can effectively improve the recovery quality of the secret image; this scheme realizes global end-to-end image steganography, has high image steganography and image recovery capabilities, and at the same time ensures high security and anti-steganography analysis capabilities; this scheme also has good scalability and can realize a large number of image steganography with the help of a cascade structure and be applied to more complex scenarios.

Claims

1. An image steganography method based on high-frequency distribution mapping deep reversible network, characterized in that: The following steps are involved: S1. Based on the affine coupling structure, a reversible steganography module and a high-frequency distribution mapping module are designed to construct a deep reversible steganography network. S2. Combining the forward hidden network and the reverse recovery network, and constructing the corresponding loss function, the deep reversible steganography network is trained for the image steganography process, the image recovery process and the high-frequency distribution mapping process to obtain the trained deep reversible steganography network; S3. Based on the trained deep reversible steganographic network, image steganography, high-frequency distribution mapping and image recovery processing are performed in sequence for the current carrier image and the secret image to obtain the carrier image and the restored image respectively.

2. According to claim 1, an image steganography method based on high-frequency distribution mapping deep reversible network is characterized in that: The step S1 comprises the following steps: S11, construct a reversible 1×1 convolutional layer to perform channel permutation strategy on the input carrier image and secret image; S12, realizing the affine transformation of the reversible steganography module and the high-frequency distribution mapping module by means of additive coupling and multiplicative coupling structures; S13. Design nonlinear functions based on the DenseNet structure and integrate the spatial attention mechanism into the DenseNet structure.

3. According to claim 2, an image steganography method based on high-frequency distribution mapping deep reversible network is characterized in that: The forward process of the reversible steganography module is as follows: h1=x1·exp(α(s(x2)))+s(x2) h2=x2·exp(α(s(h1)))+s(h1) Among them, x1 and x2 are the features of the carrier image and the secret image respectively, α() is the activation function, s() is a nonlinear function constructed by a neural network, h1 and h2 are the features of the secret image and the features of the lost information respectively, which will be used as the input of the subsequent high-frequency distribution mapping module for modeling; The forward process of the high frequency distribution mapping module is as follows: t=h2·exp(α(s(h1)))+s(h1) The high-frequency distribution mapping module learns and models h1 and h2 through an affine coupling structure to obtain t; The reverse process of the high frequency distribution mapping module is as follows: h f =(zs(y1))÷exp(α(s(y1))) Among them, y1 and z are the input ends of the high-frequency distribution mapping module, representing the characteristics of the dense image and the auxiliary variable respectively, α() is the activation function, s() is a nonlinear function constructed by a neural network, and the high-frequency distribution mapping module maps the auxiliary variable z to h g , so that h f It not only conforms to the Gaussian distribution, but also contains effective information that is helpful for recovery; The reverse process of the reversible steganography module is as follows: x2′=(h f -s(y1))÷exp(α(s(y1))) x1′=(y1-s(x2′))÷exp(α(s(x2′))) Among them, x′2 and x′1 are the restored carrier image and secret image.

4. According to claim 2, the image steganography method based on high-frequency distribution mapping deep reversible network is characterized in that: The specific process of step S11 is as follows: S111, superimposing the secret image and the carrier image into a feature map of size (h, w, c), where h, w, and c represent the length and width and the number of channels of the feature map respectively; S112, using LR decomposition to generate a rotation matrix as a custom weight matrix W; S113, using a 1×1 filter and a weight matrix W to perform a convolution operation with a step size of 1 on the feature map to obtain a mixed feature after channel permutation.

5. According to claim 4, the image steganography method based on high frequency distribution mapping deep reversible network is characterized in that: The specific process of step S13 is as follows: S131. Construct a DenseNet network consisting of five convolutional layers, and use Xavier's initialization scheme to initialize the weights of the five convolutional layers to 0; S132, use a 3×3 filter to convolve the input features, and use the Leaky ReLU activation function after each convolution; S133, after the fifth convolution, use exp() to get the final result; S134, insert a spatial attention module between the fourth and fifth layers in the DenseNet network; S135, performing maximum pooling and average pooling on the features processed by the DenseNet network, and superimposing the results of the two along the channel axis; S136, using a 7×7 filter to perform a convolution operation on the superposition result, and then using a sigmoid function to activate it to obtain a spatial feature map; S137. The final output is the product of the input feature and the spatial feature map.

6. According to claim 1, the image steganography method based on high frequency distribution mapping deep reversible network is characterized in that: The image steganography process in step S2 includes: Using discrete wavelet transform DWT, the wavelet subbands of the carrier image and the secret image are used as the input of the forward hidden network; The image steganography is completed through the reversible steganography module to obtain the secret image and the information lost in the steganography process, and the high-frequency distribution mapping module is used to model the secret image and the information lost in the steganography process. Finally, the inverse transform IWT of the discrete wavelet transform is used to restore them to obtain the secret image and complete the image steganography process.

7. According to claim 6, the image steganography method based on high frequency distribution mapping deep reversible network is characterized in that: The image restoration process in step S2 includes: Using discrete wavelet transform DWT, the wavelet subband of the input secret image and auxiliary variables is used as the input of the reverse recovery network. The high-frequency distribution mapping module maps the lost information according to the input secret image and auxiliary variables, and completes the image steganography through the reversible steganography module. Finally, the inverse transform IWT of the discrete wavelet transform is used for restoration to obtain the recovered secret image and secret image, completing the image recovery process.

8. According to claim 7, the image steganography method based on high frequency distribution mapping deep reversible network is characterized in that: The auxiliary variables are specifically sampled from a Gaussian distribution: Among them, z is an auxiliary variable used to replace the lost information in the steganographic process, μ0 is the mathematical expectation of the normal distribution, is the variance of the normal distribution.

9. The image steganography method based on high frequency distribution mapping deep reversible network according to claim 1 is characterized in that: The loss function in step S2 includes a steganographic process loss function, a steganographic process wavelet loss function and a recovery process loss function.

10. The image steganography method based on high frequency distribution mapping deep reversible network according to claim 9 is characterized in that: The loss function of the steganalysis process is specifically: Where N represents the number of training samples, MSE represents the mean square error, Represent the carrier image and the generated encrypted image respectively; The wavelet loss function of the steganalysis process is specifically: Among them, DWT() LL Represents the low-frequency component of discrete wavelet transform; The recovery process loss function is specifically: in, They represent the secret image and the secret image finally recovered from the secret image respectively.

Images