An Airspace Image Steganography Method and System Based on Minimizing Texture Disturbance

Abstract

The present invention discloses a spatial domain image steganography method based on texture perturbation minimization, which relates to the field of secure communication technology. The designed neighboring embedding pixel enhancement module can consider the mutual influence between neighboring embedding pixels, making the embedding directions in the same direction more concentrated, effectively reducing the perturbation to the image to improve security, and generating a collaborative adversarial loss using comprehensive information such as the obtained collaboration value, probability map, and gradient information. During the iterative update process of the generator set and discriminator set, the obtained loss function can be used to synchronously modify adjacent ones, thereby improving the security performance. A texture consistency loss is proposed, which integrates pixel-level and image-level texture perturbations to limit the allocation of embedding costs within high-texture regions. Based on generating asymmetric embedding costs, satisfactory security performance is achieved by adaptively minimizing the texture perturbation brought by information embedding.

Description

Technical Field

[0001] The present invention relates to the technical field of secure communication, and specifically to a spatial domain image steganography method and system based on minimizing texture perturbation. Background Art

[0002] Image steganography is an algorithm that imperceptibly embeds secret information into an image. Recently, steganography methods based on generative adversarial networks have received more attention due to their ability to automatically generate embedding probabilities and excellent security. However, in this aspect of research, the mutual influence between adjacent embedding points and the feature perturbation of the embedding on the texture region have been ignored, resulting in these algorithms being more easily detected by deep steganography analysis models that use texture features and representation learning for steganography analysis, inevitably limiting the security of existing algorithms. Summary of the Invention

[0003] To solve the deficiencies mentioned in the above background art, the purpose of the present invention is to provide a spatial domain image steganography method and system based on minimizing texture perturbation.

[0004] In a first aspect, the purpose of the present invention can be achieved through the following technical solutions: A spatial domain image steganography method based on minimizing texture perturbation, the method comprising the following steps:

[0005] Receive a grayscale carrier image and secret information, input the grayscale carrier image and secret information into a pre-established embedding probability generation network model, output an embedding modification probability map, input the embedding modification probability map into an embedding simulator to obtain an embedding modification map, and then output a simulated stego image;

[0006] Input the grayscale carrier image and the simulated stego image into a pre-established discriminator network model, output an image discrimination result, and calculate an adaptive binary cross-entropy loss function and pixel-level gradient feedback information according to the discrimination result;

[0007] Input the embedding modification map into a pre-constructed adjacent embedding pixel enhancement module, output a cooperation value matrix, combine the pixel-level gradient feedback information, the cooperation value matrix, and the embedding modification probability map to obtain a cooperation adversarial loss function, and obtain an entropy constraint loss function according to the embedding modification probability map and a pre-set embedding rate;

[0008] Obtain an image-level texture consistency loss function and a pixel-level texture consistency loss function according to the texture features of the grayscale carrier image and the simulated stego image, and combine the image-level texture consistency loss function and the pixel-level texture consistency loss function with the cooperation value matrix and the embedding modification probability map to obtain a texture consistency loss function;

[0009] According to the collaborative adversarial loss function, the texture consistency loss function, and the entropy constraint loss function, the total loss function of the generator group is obtained. The adaptive binary cross-entropy loss function is used as the loss function of the discriminator group. The total loss function of the generator group and the loss function of the discriminator group are alternately updated. With the goal of minimizing the total loss function of the generator group and the loss function of the discriminator group, the embedding probability map of image steganography is output.

[0010] Combined with the first aspect, in some implementations of the first aspect, the method further includes: The pre-established embedding probability generation network model includes: a plurality of convolutional operation groups and a plurality of deconvolutional operation groups. The plurality of convolutional operation groups include: the first convolutional operation group, the second convolutional operation group, the third convolutional operation group, the fourth convolutional operation group, the fifth convolutional operation group, the sixth convolutional operation group, the seventh convolutional operation group, the eighth convolutional operation group, the ninth convolutional operation group. The plurality of deconvolutional operation groups include the first deconvolutional operation group, the second deconvolutional operation group, the third deconvolutional operation group, the fourth deconvolutional operation group, the fifth deconvolutional operation group, the sixth deconvolutional operation group, the seventh deconvolutional operation group, the eighth deconvolutional operation group;

[0011] The output of the first convolutional operation group and the output of the first deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0012] The output of the second convolutional operation group and the output of the second deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0013] The output of the third convolutional operation group and the output of the third deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0014] The output of the fourth convolutional operation group and the output of the fourth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0015] The output of the fifth convolutional operation group and the output of the fifth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0016] The output of the sixth convolutional operation group and the output of the sixth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0017] The output of the seventh convolutional operation group and the output of the seventh deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0018] Among them, each of the multiple convolution operation groups includes a convolution layer, an activation layer, a batch normalization layer, a convolution layer, an activation layer, and a batch normalization layer arranged in sequence; each of the multiple deconvolution operation groups includes a deconvolution layer, an activation layer, a batch normalization layer, a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence;

[0019] The output of the generation network is an embedded change probability map.

[0020] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the pre-established discriminator network model is a steganalysis network model combination based on SiaStegNet and SRNet, and the adjacent embedding pixel enhancement module includes a weight unit and a collaborative value matrix generation process.

[0021] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the weight unit W is expressed as:

[0022]

[0023] Where is the value of the weight unit at index index;

[0024] The collaborative value matrix generation process is expressed as:

[0025]

[0026] Where, is the value of the collaborative value matrix at index, is the value of the embedding perturbation map at index, u is the horizontal control parameter of the weight unit, and v is the vertical control parameter of the weight unit.

[0027] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the texture consistency loss function is as follows:

[0028]

[0029]

[0030] Where, , are the texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. , are the pixel-level texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. is the image-level texture consistency loss function. is a weight hyperparameter used to control the range of the loss function.

[0031] Combined with the first aspect, in some implementations of the first aspect, the method further includes: The pixel-level texture loss is expressed as:

[0032]

[0033]

[0034]

[0035]

[0036] and respectively represent the carrier image and the stego image The high-frequency features extracted by 30 high-pass filter kernels, represents the pixel-level texture difference, represents the k-th filter kernel among 30 high-pass filter kernels, is an indicator function that indicates the part greater than 0 in, , , respectively represent the cooperation value matrix, the positive embedding probability, and the negative embedding probability at the index value, represents the gradient calculated by the steganalysis group. The calculation process is:

[0037]

[0038] where, represents the binary cross-entropy loss corresponding to the steganalysis with the strongest current discrimination ability. The input is and , representing the carrier image and the simulated embedded image respectively, represents the th discriminator's discrimination result for the input , and respectively represent the true values of the cover image and the simulated stego image, then represents the differential operation.

[0039] Combined with the first aspect, in some implementations of the first aspect, the method further includes: The image-level texture loss is expressed as:

[0040]

[0041]

[0042]

[0043] Among them, , respectively represent the image-level texture features of the carrier image and the stego image, represents the value of this feature at index . represents taking the L1-norm of represents the texture feature of the carrier image at index . represents the texture feature of the carrier image at index . represents the index of the channel.

[0044] The input of the pixel-level texture loss is the carrier image and the stego image , and the output is the pixel-level texture loss.

[0045] Combined with the first aspect, in some implementations of the first aspect, the method further includes: the adaptive binary cross-entropy loss function is as follows:

[0046]

[0047] Among them, the inputs are and , which respectively represent the carrier image and the simulated embedding image, represents the discrimination result of the th discriminator for the input , and respectively represent the true values of the cover image and the simulated stego image, represents the differential operation.

[0048] The collaborative adversarial loss function is as follows:

[0049]

[0050] Among them, is an indicator function that indicates the part greater than 0 in , , , respectively represent the collaboration value matrix, the positive embedding probability, and the negative embedding probability at​​ The value of the index represents the calculated gradient of the steganalysis group and respectively represent the height and length of the image;

[0051] The loss function of the generator group is as follows:

[0052]

[0053] where , , are parameters for controlling the weights of each part , respectively represent the total loss functions of the positive and negative embedding probability generators , respectively represent the texture consistency loss functions corresponding to the positive and negative embedding probability generators represents a preset entropy constraint loss function, and the calculation is as follows:

[0054]

[0055] where represents the modification direction represents the embedding rate.

[0056] In a second aspect, to achieve the above object, the present invention discloses a spatial domain image steganography system based on minimizing texture perturbation, including:

[0057] A simulated stego module, configured to receive a grayscale carrier image and a secret message, input the grayscale carrier image and the secret message into a pre-established embedding probability generation network model, output an embedding change probability map, input the embedding change probability map into an embedding simulator, obtain an embedding change map, and then output a simulated stego image;

[0058] An image discrimination module, configured to input the grayscale carrier image and the simulated stego image into a pre-established discriminator network model, output an image discrimination result, and calculate an adaptive binary cross-entropy loss function and pixel-level gradient feedback information according to the discrimination result;

[0059] A cooperative adversarial module, configured to input the embedding change map into a pre-constructed neighboring embedding pixel enhancement module, output a cooperation value matrix, combine the pixel-level gradient feedback information, the cooperation value matrix, and the embedding change probability map to obtain a cooperative adversarial loss function, and obtain an entropy constraint loss function according to the embedding change probability map and a preset embedding rate;

[0060] The texture loss module is used to obtain the image-level texture consistency loss function and the pixel-level texture consistency loss function based on the texture features of the grayscale carrier image and the simulated stego image, and combine the image-level texture consistency loss function and the pixel-level texture consistency loss function with the collaboration value matrix and the embedding modification probability map to obtain the texture consistency loss function;

[0061] The image steganography module is used to obtain the total loss function of the generator group according to the collaborative adversarial loss function, the texture consistency loss function and the entropy constraint loss function, use the adaptive binary cross-entropy loss function as the loss function of the discriminator group, and alternately update the total loss function of the generator group and the loss function of the discriminator group. With the goal of minimizing the total loss function of the generator group and the loss function of the discriminator group, the embedding probability map of image steganography is output.

[0062] Combined with the second aspect, in some implementation manners of the second aspect, the system further includes: The embedding probability generation network model pre-established by the simulated stego module includes: a plurality of convolutional operation groups and a plurality of deconvolutional operation groups. The plurality of convolutional operation groups include: the first convolutional operation group, the second convolutional operation group, the third convolutional operation group, the fourth convolutional operation group, the fifth convolutional operation group, the sixth convolutional operation group, the seventh convolutional operation group, the eighth convolutional operation group, the ninth convolutional operation group. The plurality of deconvolutional operation groups include the first deconvolutional operation group, the second deconvolutional operation group, the third deconvolutional operation group, the fourth deconvolutional operation group, the fifth deconvolutional operation group, the sixth deconvolutional operation group, the seventh deconvolutional operation group, the eighth deconvolutional operation group;

[0063] The output of the first convolutional operation group and the output of the first deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0064] The output of the second convolutional operation group and the output of the second deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0065] The output of the third convolutional operation group and the output of the third deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0066] The output of the fourth convolutional operation group and the output of the fourth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0067] The output of the fifth convolutional operation group and the output of the fifth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0068] The output of the sixth convolution operation group and the output of the sixth deconvolution operation group are concatenated in channels and then input into the first deconvolution operation group to form a skip structure;

[0069] The output of the seventh convolution operation group and the output of the seventh deconvolution operation group are concatenated in channels and then input into the first deconvolution operation group to form a skip structure;

[0070] Among them, each of the multiple convolution operation groups includes a convolution layer, an activation layer, a batch normalization layer, a convolution layer, an activation layer, and a batch normalization layer arranged in sequence; each of the multiple deconvolution operation groups includes a deconvolution layer, an activation layer, a batch normalization layer, a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence;

[0071] The output of the generation network is an embedded modification probability map;

[0072] The discriminator network model pre-established by the image discrimination module is a steganalysis network model combination based on SiaStegNet and SRNet, and the adjacent embedding pixel enhancement module includes a weight unit and a collaborative value matrix generation process;

[0073] Among them, the weight unit of the image discrimination module is expressed as:

[0074]

[0075] Among them is the value of the weight unit at index the index.

[0076] The collaborative value matrix generation process is expressed as:

[0077]

[0078] Among them, is the value of the collaborative value matrix at the index, is the value of the embedded perturbation map at the index, u is the horizontal control parameter of the weight unit, and v is the vertical control parameter of the weight unit;

[0079] The texture consistency loss function of the texture loss module is as follows:

[0080]

[0081]

[0082] Among them, , are the texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator respectively. , The pixel-level texture consistency loss functions corresponding to the forward embedding probability generator and the negative embedding probability generator, respectively. The image-level texture consistency loss function. Is a weight hyperparameter used to control the range of the loss function.

[0083] The pixel-level texture loss is expressed as:

[0084]

[0085]

[0086]

[0087]

[0088] and respectively represent the high-frequency features extracted by 30 high-pass filter kernels from the cover image and the stego image The pixel-level texture difference is represented by where represents the k-th filter kernel among 30 high-pass filter kernels, is an indicator function that indicates the part greater than 0 in , , respectively represent the values of the cooperation value matrix, the forward embedding probability, and the negative embedding probability at the index, represents the gradient calculated by the steganalysis group, and the calculation process is:

[0089]

[0090] where represents the binary cross-entropy loss corresponding to the steganalysis detector with the strongest current discrimination ability, and the inputs are and , representing the cover image and the simulated embedded image respectively, represents the discrimination result of the th discriminator for the input , and represent the true values of the cover image and the simulated stego image respectively, represents the differential operation;

[0091] The image-level texture loss is expressed as:

[0092]

[0093]

[0094]

[0095] Among them, , respectively represent the image-level texture features of the carrier image and the stego image, represents the value of this feature at index . represents taking the L1-norm of . represents the texture feature of the carrier image at index . represents the texture feature of the carrier image at index . represents the index of the channel.

[0096] The input of the pixel-level texture loss is the carrier image and the stego image , and the output is the pixel-level texture loss;

[0097] The image steganography module adaptive binary cross-entropy loss function is as follows:

[0098]

[0099] Among them, the input is and , which respectively represent the carrier image and the simulated embedded image, represents the discrimination result of the th discriminator for the input , and respectively represent the true values of the cover image and the simulated stego image, represents the differential operation.

[0100] The described collaborative adversarial loss function is as follows:

[0101]

[0102]

[0103] Among them, is an indicator function that indicates the part greater than 0 in , , , ​​respectively represent the collaboration value matrix, the positive embedding probability, and the negative embedding probability at the value of the index, represents the gradient calculated by the steganalysis group, and respectively represent the height and length of the image;

[0104] The loss function of the generator group is as follows:

[0105]

[0106] Among them, , , are parameters for controlling the weights of each part, , respectively represent the total loss functions of the positive and negative embedding probability generators, , respectively represent the texture consistency loss functions corresponding to the positive and negative embedding probability generators, represents the pre-set entropy constraint loss function, which is calculated as follows:

[0107]

[0108] Among them represents the modification direction, represents the embedding rate.

[0109] Advantages of the present invention:

[0110] The designed adjacent embedding pixel enhancement module of the present invention can consider the mutual influence between adjacent embedding pixels, making the embedding directions in the same direction more concentrated, effectively reducing the perturbation to the image to improve security, and generating a collaborative adversarial loss using the obtained collaboration value, probability map, gradient information and other comprehensive information. During the iterative update process of the generator set and the discriminator set, the obtained loss function can be used to synchronize adjacent modifications, thereby improving the security performance, and a texture consistency loss is proposed. It integrates pixel-level and image-level texture perturbations to limit the distribution of embedding costs within high-texture regions, and based on generating asymmetric embedding costs, achieves satisfactory security performance by adaptively minimizing the texture perturbations brought by information embedding. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts;

[0112] Figure 1 It is a schematic flowchart of the method of the present invention;

[0113] Figure 2 It is a schematic diagram of the adjacent embedding pixel enhancement module of the present invention;

[0114] Figure 3 It is a schematic diagram of the texture consistency loss module of the present invention;

[0115] Figure 4 It is a schematic diagram of the system structure of the present invention. Specific implementation manners

[0116] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0117] Embodiment 1:

[0118] Next, the relevant terms involved in the embodiments of the present application will be introduced:

[0119] ‌Image steganography is a technique for hiding information in images, aiming to protect the security of information transmission and storage. It is different from cryptography because steganography hides the process of information transmission rather than the content of the information itself. Image steganography can be divided into two categories: ‌robust steganography and ‌fragile steganography, as well as ‌reversible steganography and irreversible steganography. In addition, there are also traditional steganography and ‌deep learning steganography.

[0120] The loss function or cost function is a function that maps the values of random events or their related random variables to non-negative real numbers to represent the "risk" or "loss" of the random event. In applications, the loss function is usually associated with the learning criterion and the optimization problem, that is, the model is solved and evaluated by minimizing the loss function.

[0121] As Figure 1 shown, a spatial domain image steganography method based on minimizing texture perturbation, the method includes the following steps:

[0122] S101: Receive a grayscale carrier image and secret information, input the grayscale carrier image and secret information into a pre-established embedding probability generation network model, output an embedding modification probability map, input the embedding modification probability map into an embedding simulator to obtain an embedding modification map, and then output a simulated stego image;

[0123] Among them, the pre-established embedding probability generation network model includes: a plurality of convolutional operation groups and a plurality of deconvolutional operation groups. The plurality of convolutional operation groups include: a first convolutional operation group, a second convolutional operation group, a third convolutional operation group, a fourth convolutional operation group, a fifth convolutional operation group, a sixth convolutional operation group, a seventh convolutional operation group, an eighth convolutional operation group, a ninth convolutional operation group. The plurality of deconvolutional operation groups include a first deconvolutional operation group, a second deconvolutional operation group, a third deconvolutional operation group, a fourth deconvolutional operation group, a fifth deconvolutional operation group, a sixth deconvolutional operation group, a seventh deconvolutional operation group, an eighth deconvolutional operation group;

[0124] The output of the first convolutional operation group and the output of the first deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0125] The output of the second convolutional operation group and the output of the second deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0126] The output of the third convolutional operation group and the output of the third deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0127] The output of the fourth convolutional operation group and the output of the fourth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0128] The output of the fifth convolutional operation group and the output of the fifth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0129] The output of the sixth convolutional operation group and the output of the sixth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0130] The output of the seventh convolutional operation group and the output of the seventh deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0131] Among them, each of the plurality of convolutional operation groups includes a convolutional layer, an activation layer, a batch normalization layer, a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; each of the plurality of deconvolutional operation groups includes a deconvolutional layer, an activation layer, a batch normalization layer, a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence;

[0132] The output of the generation network is an embedded change probability map.

[0133] S102: Input the grayscale carrier image and the simulated stego image into a pre-established discriminator network model, output the image discrimination result, and calculate the adaptive binary cross-entropy loss function and pixel-level gradient feedback information according to the discrimination result;

[0134] The pre-established discriminator network model is a steganalysis network model combination based on SiaStegNet and SRNet. The adjacent embedding pixel enhancement module includes a weight unit and a collaborative value matrix generation process;

[0135] Weight unit It is expressed as:

[0136]

[0137] Where is the value of the weight unit at index at the index.

[0138] As Figure 2 shown, the collaborative value matrix generation process is expressed as:

[0139]

[0140] Where, is the value of the collaborative value matrix at index, is the value of the embedding perturbation map at index, u is the horizontal control parameter of the weight unit, and v is the vertical control parameter of the weight unit;

[0141] S103: Input the embedding modification map into a pre-constructed adjacent embedding pixel enhancement module, output the collaborative value matrix, combine the pixel-level gradient feedback information, the collaborative value matrix, and the embedding modification probability map to obtain the collaborative adversarial loss function, and obtain the entropy constraint loss function according to the embedding modification probability map combined with a preset steganography rate;

[0142] S104: As Figure 3 shown, obtain the image-level texture consistency loss function and the pixel-level texture consistency loss function according to the texture features of the grayscale carrier image and the simulated stego image, and combine the image-level texture consistency loss function and the pixel-level texture consistency loss function with the collaborative value matrix and the embedding modification probability map to obtain the texture consistency loss function;

[0143] The texture consistency loss function is as follows:

[0144]

[0145]

[0146] Where, , They are the texture consistency loss functions corresponding to the forward embedding probability generator and the negative embedding probability generator respectively. , They are the pixel-level texture consistency loss functions corresponding to the forward embedding probability generator and the negative embedding probability generator respectively. It is the image-level texture consistency loss function. is a weight hyperparameter used to control the range of the loss function.

[0147] The pixel-level texture loss is expressed as:

[0148]

[0149]

[0150]

[0151]

[0152] and represent the carrier image and the stego image The high-frequency features extracted by 30 high-pass filter kernels, represents the pixel-level texture difference, represents the k-th filter kernel among 30 high-pass filter kernels, is an indicator function that indicates the part greater than 0 in , , represent the collaborative value matrix, the forward embedding probability, and the negative embedding probability at the index value, represents the gradient calculated by the steganalysis group. The calculation process is:

[0153]

[0154] Among them, represents the binary cross-entropy loss corresponding to the steganalysis with the strongest current discrimination ability. The input is and , representing the carrier image and the simulated embedded image respectively, represents the th discriminator's discrimination result for the input , and represent the true values of the cover image and the simulated stego image respectively, represents the differential operation;

[0155] The pixel-level texture loss is expressed as:

[0156]

[0157]

[0158]

[0159] where , represent the image-level texture features of the cover image and the stego image respectively, represents the value of this feature at index . represents taking the L1-norm of . represents the texture feature of the cover image at index . . represents the texture feature of the cover image at index . . represents the index of the channel.

[0160] The input of the pixel-level texture loss is the cover image and the stego image , and the output is the pixel-level texture loss.

[0161] S105: Obtain the total loss function of the generator group according to the collaborative adversarial loss function and the texture consistency loss function, use the adaptive binary cross-entropy loss function as the loss function of the discriminator group, alternately update the total loss function of the generator group and the loss function of the discriminator group, take minimizing the total loss function of the generator group and the loss function of the discriminator group as the optimization goal, and output the image steganography result.

[0162] In actual embedding, convert the generated probability map into an embedding cost through a conversion formula, and use a pre-prepared steganographic encoding such as STC to implement the embedding and extraction of secret information. The above-mentioned conversion formula is as follows:

[0163]

[0164] The adaptive binary cross-entropy loss function is as follows:

[0165]

[0166] where the inputs are and , representing the cover image and the simulated embedding image respectively Indicates the discriminant result of the th discriminator for the input and represent the true values of the cover image and the simulated stego image respectively, then represents the differential operation.

[0167] The collaborative adversarial loss function is as follows:

[0168]

[0169]

[0170] Among them, is an indicator function that points out the part greater than 0 in , , , respectively represent the collaborative value matrix, the positive embedding probability, and the negative embedding probability at the index value, represents the gradient calculated by the steganalysis group. and represent the height and length of the image respectively.

[0171] The loss function of the generator group is as follows:

[0172]

[0173] Among them, , , are parameters for controlling the weights of each part. , respectively represent the total loss functions of the positive and negative embedding probability generators, , respectively represent the texture consistency loss functions corresponding to the positive and negative embedding probability generators. represents the pre-set entropy constraint loss function, and its calculation method is as follows:

[0174]

[0175] Among them represents the possible modification direction, represents the embedding rate, and its unit is bpp (bit per pixel).

[0176] The detection error rate of the steganalysis is used to evaluate the performance of the steganography algorithm , where represents the false detection rate, Represents the false alarm rate.

[0177] Experiments and comparisons were carried out from 0.1 bpp to 0.4 bpp.

[0178] On the BossBase dataset, the results are as shown in Table 1 below:

[0179] Table 1

[0180]

[0181] On the Bows2 dataset, the results are as shown in Table 2 below:

[0182] Table 2

[0183]

[0184] Experiments on multiple datasets show that the algorithm has high and outstanding security performance.

[0185] Example 2: Second aspect, as Figure 4 shown, to achieve the above object, the present invention discloses a spatial domain image steganography system based on texture perturbation minimization, including:

[0186] A simulated stego module 11, configured to receive a grayscale carrier image and secret information, input the grayscale carrier image and the secret information into a pre-established embedding probability generation network model, output an embedding change probability map, input the embedding change probability map into an embedding simulator to obtain an embedding change map, and then output a simulated stego image;

[0187] An image discrimination module 12, configured to input the grayscale carrier image and the simulated stego image into a pre-established discriminator network model, output an image discrimination result, and calculate an adaptive binary cross-entropy loss function and pixel-level gradient feedback information according to the discrimination result;

[0188] A collaborative adversarial module 13, configured to input the embedding change map into a pre-constructed adjacent embedding pixel enhancement module, output a collaborative value matrix, combine the pixel-level gradient feedback information, the collaborative value matrix, and the embedding change probability map to obtain a collaborative adversarial loss function, and obtain an entropy constraint loss function according to the embedding change probability map in combination with a pre-set embedding rate;

[0189] A texture loss module 14, configured to obtain an image-level texture consistency loss function and a pixel-level texture consistency loss function according to the texture features of the grayscale carrier image and the simulated stego image, and combine the image-level texture consistency loss function and the pixel-level texture consistency loss function with the collaborative value matrix and the embedding change probability map to obtain a texture consistency loss function;

[0190] The image steganography module 15 is used to obtain the total loss function of the generator group according to the collaborative adversarial loss function, the texture consistency loss function, and the entropy constraint loss function, use the adaptive binary cross-entropy loss function as the loss function of the discriminator group, and alternately update the total loss function of the generator group and the loss function of the discriminator group. With the goal of minimizing the total loss function of the generator group and the loss function of the discriminator group, an embedded probability map of image steganography is output.

[0191] Combined with the second aspect, in some implementation manners of the second aspect, the system further includes: The embedding probability generation network model pre-established by the simulated stego module 11 includes: a plurality of convolutional operation groups and a plurality of deconvolutional operation groups. The plurality of convolutional operation groups include: a first convolutional operation group, a second convolutional operation group, a third convolutional operation group, a fourth convolutional operation group, a fifth convolutional operation group, a sixth convolutional operation group, a seventh convolutional operation group, an eighth convolutional operation group, a ninth convolutional operation group. The plurality of deconvolutional operation groups include a first deconvolutional operation group, a second deconvolutional operation group, a third deconvolutional operation group, a fourth deconvolutional operation group, a fifth deconvolutional operation group, a sixth deconvolutional operation group, a seventh deconvolutional operation group, an eighth deconvolutional operation group;

[0192] The output of the first convolutional operation group and the output of the first deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0193] The output of the second convolutional operation group and the output of the second deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0194] The output of the third convolutional operation group and the output of the third deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0195] The output of the fourth convolutional operation group and the output of the fourth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0196] The output of the fifth convolutional operation group and the output of the fifth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0197] The output of the sixth convolutional operation group and the output of the sixth deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0198] The output of the seventh convolutional operation group and the output of the seventh deconvolutional operation group are concatenated on the channel and then input into the first deconvolutional operation group to form a skip structure;

[0199] Among them, each of the multiple convolution operation groups includes a convolution layer, an activation layer, a batch normalization layer, a convolution layer, an activation layer, and a batch normalization layer arranged in sequence; each of the multiple deconvolution operation groups includes a deconvolution layer, an activation layer, a batch normalization layer, a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence;

[0200] The output of the generation network is an embedded modification probability map;

[0201] The discriminator network model pre-established by the image discrimination module 12 is a steganalysis network model combination based on SiaStegNet and SRNet. The adjacent embedding pixel enhancement module includes a weight unit and a collaborative value matrix generation process;

[0202] Among them, the weight unit of the image discrimination module 12 is expressed as:

[0203]

[0204] where is the value of the weight unit at index at the index.

[0205] The collaborative value matrix generation process is expressed as:

[0206]

[0207] where is the value of the collaborative value matrix at the index, is the value of the embedded perturbation map at the index, u is the horizontal control parameter of the weight unit, and v is the vertical control parameter of the weight unit;

[0208] The texture consistency loss function of the texture loss module 15 is as follows:

[0209]

[0210]

[0211] where , are the texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. , are the pixel-level texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. is the image-level texture consistency loss function. is the weight hyperparameter used to control the range of the loss function.

[0212] The pixel-level texture loss is expressed as:

[0213]

[0214]

[0215]

[0216]

[0217] and represent the carrier image and the stego image The high-frequency features extracted by 30 high-pass filter kernels, represent the texture difference at the pixel level, represents the k-th filter kernel among 30 high-pass filter kernels, is an indicator function that indicates the part greater than 0 in , , represent the collaboration value matrix, the positive embedding probability, and the negative embedding probability at the index value, represents the gradient calculated by the steganalysis group. The calculation process is as follows:

[0218]

[0219] Among them, represents the binary cross-entropy loss corresponding to the steganalysis detector with the strongest current discrimination ability. The inputs are and , representing the carrier image and the simulated embedded image respectively, represents the th discriminator's discrimination result for the input , and represent the true values of the cover image and the simulated stego image respectively, represents the differential operation;

[0220] The image-level texture loss is expressed as:

[0221]

[0222]

[0223]

[0224] Among them, , represent the image-level texture features of the carrier image and the stego image respectively, represents the value of this feature at the index . represents taking the L1-norm of represents the texture feature of the carrier image . at the index . represents the texture feature of the carrier image . at the index . represents the index of the channel.

[0225] The input of the pixel-level texture loss is the carrier image and the stego image , and the output is the pixel-level texture loss;

[0226] The adaptive binary cross-entropy loss function of the image steganography module is as follows:

[0227]

[0228] where the inputs are and , representing the carrier image and the simulated embedded image respectively, represents the discrimination result of the -th discriminator for the input , and represent the true values of the cover image and the simulated stego image respectively, represents the differential operation.

[0229] The described collaborative adversarial loss function is as follows:

[0230]

[0231]

[0232] where is an indicator function that indicates the part greater than 0 in , , , represent the collaborative value matrix, the positive embedding probability, and the negative embedding probability at the index of respectively, represents the gradient calculated by the steganalysis group. and represent the height and length of the image respectively.

[0233] The loss function of the generator group described above is as follows:

[0234]

[0235] Wherein, , , is a parameter for controlling the weights of each part. , respectively represent the total loss functions of the forward and backward embedding probability generators, , respectively represent the texture consistency loss functions corresponding to the forward and backward embedding probability generators. represents a preset entropy constraint loss function, and its calculation method is as follows:

[0236]

[0237] Wherein represents a possible modification direction, represents the embedding density, and its unit is bpp (bit per pixel, bits per pixel).

[0238] Based on the same inventive concept, the present invention further provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is configured to execute the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is used to implement one or more instructions. Specifically, it is used to load and execute one or more instructions in the computer storage medium to implement the above method.

[0239] It should be further noted that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the above-mentioned method. The storage medium may be any combination of one or more computer-readable media. The computer-readable media may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electro-magnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device.

[0240] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0241] The above shows and describes the basic principles, main features, and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and the descriptions in the specification only illustrate the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure will have various changes and improvements, and these changes and improvements all fall within the scope of the present disclosure claimed.

Claims

1. A spatial domain image steganography method based on texture perturbation minimization, characterized in that: The method comprises the following steps: Receiving a grayscale carrier image and secret information, inputting the grayscale carrier image and secret information into a pre-established embedding probability generation network model, outputting an embedding change probability map, inputting the embedding change probability map into an embedding simulator to obtain an embedding change map, and then outputting a simulated secret image; The grayscale carrier image and the simulated encrypted image are input into the pre-established discriminator network model, and the image discrimination result is output. The adaptive binary cross entropy loss function and pixel-level gradient feedback information are calculated based on the discrimination result. The embedding change map is input into the pre-built neighboring embedding pixel enhancement module, and the collaborative value matrix is ​​output. The collaborative adversarial loss function is obtained by combining the pixel-level gradient feedback information, the collaborative value matrix and the embedding change probability map. The entropy constraint loss function is obtained according to the embedding change probability map combined with the pre-set embedding density ratio. According to the texture features of the grayscale carrier image and the simulated dense image, an image-level texture consistency loss function and a pixel-level texture consistency loss function are obtained, and the image-level texture consistency loss function and the pixel-level texture consistency loss function are combined with the collaborative value matrix and the embedding change probability map to obtain a texture consistency loss function; According to the collaborative adversarial loss function, texture consistency loss function and entropy constraint loss function, the total loss function of the generator group is obtained. The adaptive binary cross entropy loss function is used as the loss function of the discriminator group. The total loss function of the generator group and the loss function of the discriminator group are updated alternately. The optimization goal is to minimize the total loss function of the generator group and the loss function of the discriminator group, and the output is the embedding probability map of image steganography.

2. According to claim 1, a spatial domain image steganography method based on texture perturbation minimization is characterized in that: The pre-established embedding probability generation network model includes: a plurality of convolution operation groups and a plurality of deconvolution operation groups, the plurality of convolution operation groups include: a first convolution operation group, a second convolution operation group, a third convolution operation group, a fourth convolution operation group, a fifth convolution operation group, a sixth convolution operation group, a seventh convolution operation group, an eighth convolution operation group, and a ninth convolution operation group, the plurality of deconvolution operation groups include a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group, a fourth deconvolution operation group, a fifth deconvolution operation group, a sixth deconvolution operation group, a seventh deconvolution operation group, and an eighth deconvolution operation group; The output of the first convolution operation group and the output of the first deconvolution operation group are spliced ​​on the channel and then input into the deconvolution first operation group to form a jump structure; The output of the second convolution operation group and the output of the second deconvolution operation group are spliced ​​on the channel and then input into the deconvolution first operation group to form a jump structure; The output of the third convolution operation group and the output of the third deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the fourth convolution operation group and the output of the fourth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the fifth convolution operation group and the output of the fifth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the sixth convolution operation group and the output of the sixth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the seventh convolution operation group and the output of the seventh deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; Among them, multiple convolution operation groups all include convolution layer, activation layer, batch normalization layer, convolution layer, activation layer, batch normalization layer arranged in sequence; multiple deconvolution operation groups all include deconvolution layer, activation layer, batch normalization layer, deconvolution layer, activation layer, batch normalization layer arranged in sequence; The output of the embedding probability generation network model is an embedding change probability map.

3. According to the spatial domain image steganography method based on texture perturbation minimization according to claim 1, it is characterized in that: The pre-established discriminator network model is a combination of steganalysis network models based on SiaStegNet and SRNet, and the neighboring embedded pixel enhancement module includes a weight unit.

4. According to claim 3, a spatial domain image steganography method based on texture perturbation minimization is characterized in that: The weight unit W is expressed as: where w i,j is the value of the weight unit at index (i, j); The process of generating the collaboration value matrix is ​​expressed as: Among them, c i,j is the value of the collaboration value matrix at the (i,j) index, m i,j is the value of the embedded perturbation map at index (i, j), u is the lateral control parameter of the weight unit, and v is the longitudinal control parameter of the weight unit.

5. The spatial domain image steganography method based on texture perturbation minimization according to claim 1, characterized in that: The texture consistency loss function is as follows: in, are the texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. are the pixel-level texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. is the image-level texture consistency loss function, and α is the weight parameter.

6. The spatial domain image steganography method based on texture perturbation minimization according to claim 1, characterized in that: The pixel-level texture consistency loss function is expressed as: T x and T y They represent the high-frequency features of the carrier image X and the encrypted image Y extracted by 30 high-pass filter kernels, respectively. x -T y |Indicates pixel-level texture differences, HPF k represents the kth filter kernel among the 30 high-pass filter kernels. I(·) is an indicator function that indicates the part of · that is greater than 0. c i,j , Respectively represent the collaborative value matrix, the positive embedding probability, the negative embedding probability at the (i, j) index, and g i,j represents the gradient calculated by the steganalyzer group. The calculation process is: Among them, L w represents the binary cross entropy loss corresponding to the most discriminative steganalyzer at present. The input is X and Y, which represent the carrier image and the secret image respectively. D i (·) represents the discrimination result of the i-th discriminator on the input (·), z0 and z1 represent the true values ​​of the cover image and the simulated encrypted image, respectively. It means the differentiation operation.

7. The spatial domain image steganography method based on texture perturbation minimization according to claim 1, characterized in that: The image-level texture consistency loss function is expressed as: Among them, E X , E Y denote the image-level texture features of the carrier image and the encrypted image, respectively, and e i,j,k represents the value of the feature at index (i, j, k), ||·||1 represents the L1-norm for ·, Represents the texture feature T of the carrier image X x The value at index (k,i,l), Represents the texture feature T of the encrypted image Y Y The value at index (k,i,l), where k represents the index of the channel; The input of the pixel-level texture consistency loss function is the carrier image X and the encrypted image Y, and the output is the pixel-level texture consistency loss function.

8. The spatial domain image steganography method based on texture perturbation minimization according to claim 1, characterized in that: The adaptive binary cross entropy loss function l d as follows: The input is X and Y, representing the carrier image and the secret image respectively, D i (·) represents the discrimination result of the i-th discriminator on the input (·), z0 and z1 represent the true values ​​of the cover image and the simulated encrypted image respectively; The collaborative adversarial loss function is as follows: Among them, I(·) is an indicator function, indicating the part of · that is greater than 0, c i,j , Respectively represent the collaborative value matrix, the positive embedding probability, the negative embedding probability at the (i, j) index, and g i,j represents the gradient calculated by the steganalyzer group, H and W represent the height and length of the image respectively; The loss function of the generator group is as follows: Among them, β, χ, δ are parameters that control the weights of each part. Represent the total loss function of the positive and negative embedding probability generators, respectively. Represent the texture consistency loss function corresponding to the positive and negative embedding probability generators, respectively, l e represents the pre-set entropy constraint loss function, which is calculated as follows: Where m∈{-1,0,1} represents the modification direction and R represents the embedding rate.

9. A spatial domain image steganography system based on texture perturbation minimization, characterized in that: include: A simulated secret-containing module is used to receive a grayscale carrier image and secret information, input the grayscale carrier image and secret information into a pre-established embedding probability generation network model, output an embedding change probability map, input the embedding change probability map into an embedding simulator to obtain an embedding change map, and then output a simulated secret-containing image; The image discrimination module is used to input the grayscale carrier image and the simulated secret image into the pre-established discriminator network model, output the image discrimination result, and calculate the adaptive binary cross entropy loss function and pixel-level gradient feedback information according to the discrimination result; The collaborative adversarial module is used to input the embedding change map into the pre-built neighboring embedding pixel enhancement module, and output the collaborative value matrix. The collaborative adversarial loss function is obtained by combining the pixel-level gradient feedback information, the collaborative value matrix and the embedding change probability map. The entropy constraint loss function is obtained according to the embedding change probability map combined with the pre-set embedding density ratio. A texture loss module is used to obtain an image-level texture consistency loss function and a pixel-level texture consistency loss function according to the texture features of the grayscale carrier image and the simulated dense image, and to obtain a texture consistency loss function by combining the image-level texture consistency loss function and the pixel-level texture consistency loss function with the collaboration value matrix and the embedding change probability map; The image steganography module is used to obtain the total loss function of the generator group based on the collaborative adversarial loss function, the texture consistency loss function and the entropy constraint loss function, and use the adaptive binary cross entropy loss function as the loss function of the discriminator group. The total loss function of the generator group and the loss function of the discriminator group are alternately updated to minimize the total loss function of the generator group and the loss function of the discriminator group as the optimization goal, and the embedded probability map of the image steganography is output.

10. The spatial domain image steganography system based on texture perturbation minimization according to claim 9, characterized in that: The pre-established embedding probability generation network model of the simulated secret module includes: a plurality of convolution operation groups and a plurality of deconvolution operation groups, the plurality of convolution operation groups include: a first convolution operation group, a second convolution operation group, a third convolution operation group, a fourth convolution operation group, a fifth convolution operation group, a sixth convolution operation group, a seventh convolution operation group, an eighth convolution operation group, and a ninth convolution operation group, the plurality of deconvolution operation groups include a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group, a fourth deconvolution operation group, a fifth deconvolution operation group, a sixth deconvolution operation group, a seventh deconvolution operation group, and an eighth deconvolution operation group; The output of the first convolution operation group and the output of the first deconvolution operation group are spliced ​​on the channel and then input into the deconvolution first operation group to form a jump structure; The output of the second convolution operation group and the output of the second deconvolution operation group are spliced ​​on the channel and then input into the deconvolution first operation group to form a jump structure; The output of the third convolution operation group and the output of the third deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the fourth convolution operation group and the output of the fourth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the fifth convolution operation group and the output of the fifth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the sixth convolution operation group and the output of the sixth deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; The output of the seventh convolution operation group and the output of the seventh deconvolution operation group are spliced ​​on the channel and then input into the first deconvolution operation group to form a jump structure; Among them, multiple convolution operation groups all include convolution layer, activation layer, batch normalization layer, convolution layer, activation layer, batch normalization layer arranged in sequence; multiple deconvolution operation groups all include deconvolution layer, activation layer, batch normalization layer, deconvolution layer, activation layer, batch normalization layer arranged in sequence; The output of the embedding probability generation network model is an embedding change probability map; The discriminator network model pre-established in the image discrimination module is a combination of steganalysis network models composed of SiaStegNet and SRNet, and the neighboring embedded pixel enhancement module includes a weight unit; Among them, the weight unit W of the image discrimination module is expressed as: where w i,j is the value of the weight unit at index (i, j); The process of generating the collaboration value matrix is ​​expressed as: Among them, c i,j is the value of the collaboration value matrix at the (i,j) index, m i,j is the value of the embedded perturbation graph at index (i, j), u is the lateral control parameter of the weight unit, and v is the longitudinal control parameter of the weight unit; The texture consistency loss function of the texture loss module is as follows: in, are the texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. are the pixel-level texture consistency loss functions corresponding to the positive embedding probability generator and the negative embedding probability generator, respectively. is the image-level texture consistency loss function, α is the weight parameter; The pixel-level texture consistency loss function is expressed as: T x and T y They represent the high-frequency features of the carrier image X and the encrypted image Y extracted by 30 high-pass filter kernels, respectively. x -T y |Indicates pixel-level texture differences, HPF k represents the kth filter kernel among the 30 high-pass filter kernels. I(·) is an indicator function that indicates the part of · that is greater than 0. c i,j , Respectively represent the collaborative value matrix, the positive embedding probability, the negative embedding probability at the (i, j) index, and g i,j represents the gradient calculated by the steganalyzer group. The calculation process is: Among them, L w represents the binary cross entropy loss corresponding to the most discriminative steganalyzer at present. The input is X and Y, which represent the carrier image and the secret image respectively. D i (·) represents the discrimination result of the i-th discriminator on the input (·), z0 and z1 represent the true values ​​of the cover image and the simulated encrypted image, respectively. It means the differential operation; The image-level texture consistency loss function is expressed as: Among them, E X , E Y denote the image-level texture features of the carrier image and the encrypted image, respectively, and e i,j,k represents the value of the feature at index (i, j, k), ||·||1 represents the L1-norm for ·, Represents the texture feature T of the carrier image X x The value at index (k,i,l), Represents the texture feature T of the encrypted image Y Y The value at index (k,i,l), where k represents the index of the channel; The input of the pixel-level texture consistency loss function is the carrier image X and the encrypted image Y, and the output is the pixel-level texture consistency loss function; Image steganography module adaptive binary cross entropy loss function l d as follows: The input is X and Y, representing the carrier image and the secret image respectively, D i (·) represents the discrimination result of the i-th discriminator on the input (·), z0 and z1 represent the true values ​​of the cover image and the simulated encrypted image respectively; The collaborative adversarial loss function is as follows: Among them, I(·) is an indicator function, indicating the part of · that is greater than 0, c i,j , Respectively represent the collaborative value matrix, the positive embedding probability, the negative embedding probability at the (i, j) index, and g i,j represents the gradient calculated by the steganalyzer group, H and W represent the height and length of the image respectively; The loss function of the generator group is as follows: Among them, β, χ, δ are parameters that control the weights of each part. Represent the total loss function of the positive and negative embedding probability generators, respectively. Represent the texture consistency loss function corresponding to the positive and negative embedding probability generators, respectively, l e represents the pre-set entropy constraint loss function, which is calculated as follows: Where m∈{-1,0,1} represents the modification direction and R represents the embedding rate.

Images