A method and device for training a steganalysis network that combines source supervision and target unsupervised information

Abstract

The present invention provides a method and device for training a steganalysis network by combining source supervision and target unsupervised information. The steps of the method include: inputting a test data set into the steganalysis network to output the category of each target domain image in the test data set; obtaining the parameter matrix corresponding to each source domain image in a preset training data set through a feature extraction layer, calculating the source domain feature mean matrix corresponding to each category based on the parameter matrix of the source domain image, and calculating the target domain feature mean matrix corresponding to each category based on the test data set; calculating the inter-domain average difference corresponding to each category based on the source domain feature mean matrix and the target domain feature mean matrix of each category, and calculating the value of the cross-entropy loss function; calculating the individual prediction entropy and the average prediction entropy based on the probability parameters of the target domain feature image belonging to two categories, and calculating the total prediction entropy; calculating the value of the loss function based on the value of the cross-entropy loss function and the total prediction entropy, and training the steganalysis network.

Description

Technical Field

[0001] The present invention relates to the technical field of steganalysis, and in particular to a method and device for training a steganalysis network that combines source supervision and target unsupervised information. Background Art

[0002] Steganography is a covert communication technology that embeds data into a multimedia carrier in an imperceptible manner to hide the existence of the embedded data. Conversely, steganalysis is a technology aimed at revealing the existence of embedded data in steganographically altered multimedia. The most commonly used carrier is digital images.

[0003] So far, existing steganalysis algorithms can be divided into two categories, namely steganalysis algorithms based on handcrafted features and steganalysis algorithms based on deep learning. Previous algorithms essentially perform feature extraction and classification separately, and rely heavily on expert experience during the feature extraction process. In contrast, the latter has the ability to optimize feature extraction and classification under a unified framework and can automatically learn the feature representation of images.

[0004] In a laboratory environment, both steganalysis algorithms based on handcrafted features and steganalysis algorithms based on deep learning are based on the assumption that the labeled images in the source domain and the unlabeled images in the target domain follow the same distribution. However, in the real world, the target images to be detected are diverse and even unknown. More specifically, the distribution of target domain images usually does not match or is completely different from the distribution of source domain images used for training. In this case, when the classifier learned on the source domain is used for the target domain, the detection accuracy will be significantly reduced. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and device for training a steganalysis network that combines source supervision and target unsupervised information to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a method for training a steganalysis network that combines source supervision and target unsupervised information. The steganalysis network includes a feature extraction layer and a classification layer. The feature extraction layer is used to extract an input image into a parameter matrix, and the classification layer is used to determine the category of the input image. The category is a stego image or a carrier image. The steps of the method include:

[0007] Input a test data set into the steganalysis network, and output whether each target domain image in the test data set is a stego image or a carrier image;

[0008] Input the preset training dataset into the feature extraction layer to obtain the parameter matrix of each source domain image in the corresponding training dataset. Calculate the source domain feature mean matrix for each category based on the parameter matrix of each source domain image, and calculate the target domain feature mean matrix for each category based on the parameter matrix of each target domain image in the test dataset;

[0009] Calculate the inter-domain average difference for each category based on the source domain feature mean matrix and the target domain feature mean matrix for each category, and calculate the cross-entropy loss function value based on the inter-domain average difference;

[0010] Calculate the individual prediction entropy and the average prediction entropy based on the probability parameters of each target domain feature image belonging to two categories calculated by the classification layer, and calculate the total prediction entropy based on the individual prediction entropy and the average prediction entropy;

[0011] Calculate the loss function value based on the cross-entropy loss function value and the total prediction entropy, and train the steganalysis network based on the loss function value.

[0012] Adopting the above scheme, in this paper, by integrating the supervised information of the source domain and the unsupervised information of the target domain into the training process of the classifier, through the cross-entropy loss function value calculated based on the inter-domain average difference and the total prediction entropy calculated through the individual prediction entropy and the average prediction entropy, it alleviates the significant reduction in detection accuracy caused by carrier source mismatch (CSM). This scheme can transfer the knowledge learned from the source enhanced features to the target domain, thereby enhancing the classification ability. Different weights are assigned to the individual prediction entropy and the average prediction entropy during the calculation of the total prediction entropy, and by fusing the unsupervised information of the target domain, the detection accuracy is further improved.

[0013] In some embodiments of the present invention, the step of calculating the cross-entropy loss function value based on the inter-domain average difference includes:

[0014] Construct a weight matrix based on the weight parameters of the classification layer, where the weight matrix includes a first weight parameter matrix for calculating the stego image category and a second weight parameter matrix for calculating the carrier image category;

[0015] Calculate the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix;

[0016] Calculate the second covariance parameter based on the first covariance parameter;

[0017] Calculate the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference.

[0018] In some embodiments of the present invention, in the step of calculating the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix, calculate the first covariance parameter according to the following formula:

[0019]

[0020] where, σ c represents the first covariance parameter corresponding to category c, λ represents a preset scale factor parameter, W represents a weight matrix, and W c represents the first weight parameter matrix or the second weight parameter matrix corresponding to category c, and W T represents the transpose matrix of the weight matrix, represents the average value of the covariance matrix corresponding to the parameter matrix of category c in the test dataset.

[0021] In some embodiments of the present invention, in the step of calculating the second covariance parameter based on the first covariance parameter, the second covariance parameter is calculated according to the following formula:

[0022]

[0023] where, d c represents the second covariance parameter corresponding to category c, and σ c represents the first covariance parameter corresponding to category c.

[0024] In some embodiments of the present invention, in the step of calculating the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference, the cross-entropy loss function value is calculated according to the following formula:

[0025]

[0026] where, L FM represents the cross-entropy loss function value, n s represents the number of images in the training dataset, k represents any image in the test dataset, D represents the category number, c represents any category, e is the natural constant, W T represents the transpose matrix of the weight matrix, and W c represents the first weight parameter matrix or the second weight parameter matrix corresponding to category c, Δμ c represents the inter-domain average difference corresponding to category c, λ represents a preset scale factor parameter, b represents the bias parameter vector constructed by the bias parameters of the classification layer, and b c represents the first bias parameter vector or the second bias parameter vector corresponding to category c, f k represents the parameter matrix corresponding to image k in the test dataset, and d c represents the second covariance parameter corresponding to category c.

[0027] In some embodiments of the present invention, the parameter matrix of each source domain image is a two-dimensional matrix corresponding to two categories. In the step of calculating the source domain feature mean matrix corresponding to each category based on the parameter matrix of each source domain image, the two-dimensional matrices of the two dimensions of the parameter matrix of each training data set and test data set are respectively split. The matrices corresponding to one category in all parameter matrices are constructed into a multi-dimensional matrix, the average value of the parameters at the corresponding positions in the multi-dimensional matrix is calculated, the multi-dimensional matrix is converted into a one-dimensional matrix, the one-dimensional matrices obtained by converting the two categories of the training data set are output as the source domain feature mean matrices corresponding to the two categories, and the one-dimensional matrices obtained by converting the two categories of the test data set are output as the target domain feature mean matrices corresponding to the two categories.

[0028] In some embodiments of the present invention, in the step of calculating the inter-domain average difference corresponding to each category based on the source domain feature mean matrix and the target domain feature mean matrix of each category, the inter-domain average difference is calculated based on the following formula:

[0029]

[0030] where, Δμ c represents the inter-domain average difference, represents the target domain feature mean matrix corresponding to category c in the test data set, represents the target domain feature mean matrix corresponding to category c in the test data set.

[0031] In some embodiments of the present invention, in the step of calculating the individual prediction entropy and the average prediction entropy based on the probability parameters of each target domain feature image belonging to two categories calculated by the classification layer, the individual prediction entropy and the average prediction entropy are calculated based on the following formula:

[0032]

[0033] where, L EI represents the individual prediction entropy, n t represents the number of images in the test data set, j represents any image in the test data set, c represents the category. If c = 0, it represents the carrier image category. If c = 1, it represents the stego image category, represents the probability parameter corresponding to category c of the j-th image output by the classification layer;

[0034]

[0035] where, L EA represents the average prediction entropy, represents the average value of the probability parameters of the images in the test data set corresponding to category c.

[0036] In some embodiments of the present invention, in the step of calculating the total prediction entropy based on the individual prediction entropy and the average prediction entropy, the total prediction entropy is calculated according to the following formula:

[0037] L EIA = α LEI - β LEA ;

[0038] Wherein, L EIA represents the total prediction entropy, L EA represents the average prediction entropy, L EI represents the individual prediction entropy, and α and β are the weight parameters corresponding to L EI and L EA respectively.

[0039] In some embodiments of the present invention, in the step of calculating the loss function value based on the cross-entropy loss function value and the total prediction entropy, the loss function value is calculated according to the following formula:

[0040] L FME = L FM + L EIA ;

[0041] Wherein, L FME represents the loss function value, L EIA represents the total prediction entropy, and L FM represents the cross-entropy loss function value.

[0042] The additional advantages, objects, and features of the present invention will be partially described below, and will become partially apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objects and other advantages of the present invention can be pointed out and obtained in the specification and the accompanying drawings.

[0043] Those skilled in the art will understand that the objects and advantages achievable by the present invention are not limited to the above specifically described, and the above and other objects achievable by the present invention will be more clearly understood according to the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of the present invention, and constitute a part of this application, and do not limit the present invention.

[0045] Figure 1 It is a schematic diagram of an embodiment of a steganography analysis network training method combining source supervision and target unsupervised information of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0047] Herein, it should also be noted that in order to avoid obscuring the present invention due to unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less related to the present invention are omitted.

[0048] It should be emphasized that the term "comprising / including" when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0049] Herein, it should also be noted that if not otherwise specified, the term "connection" herein can refer not only to direct connection, but also to indirect connection with an intermediate.

[0050] In the following, embodiments of the present invention will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0051] As Figure 1 shown, an aspect of the present invention provides a method for training a steganalysis network that combines source supervision and target unsupervised information. The steganalysis network includes a feature extraction layer and a classification layer. The feature extraction layer is used to extract an input image into a parameter matrix, and the classification layer is used to determine the category of the input image. The category is a stego image or a carrier image. The steps of the method include:

[0052] In some embodiments of the present invention, the feature extraction layer includes 5 convolutional layers connected in sequence and 1 fully connected layer, the classification layer is a fully connected layer, and the input of the feature extraction layer is an image of 512*512.

[0053] In the specific implementation process, the steganalysis network of this solution is similar to the yedroudj-net structure. The difference is that this solution uses two fully connected layers, and the input image is modified from an image of 256*256 of yedroudj-net to an image of 512*512.

[0054] In the specific implementation process, the classification layer outputs the probability parameters of the input image belonging to a stego image and a carrier image, and takes the category with the larger probability parameter as the category of the input image.

[0055] Step S100, input the test data set into the steganalysis network, and output whether each target domain image in the test data set is a stego image or a carrier image;

[0056] In the specific implementation process, the test data set includes multiple target domain images.

[0057] In the specific implementation process, before the step of inputting the test data set into the steganalysis network in step S100 and outputting whether each target domain image in the test data set is a stego image or a carrier image, the steganalysis network is pre-trained. The pre-training step is to input the source domain images in the preset training data set into the steganalysis network to output classification results, calculate the binary cross-entropy loss function through the classification results and the true classification labels, and pre-train the steganalysis network through the binary cross-entropy loss function.

[0058] Step S200: Input the preset training data set into the feature extraction layer to obtain the parameter matrix corresponding to each source domain image in the training data set. Calculate the source domain feature mean matrix corresponding to each category based on the parameter matrix of each source domain image, and calculate the target domain feature mean matrix corresponding to each category based on the parameter matrix of each target domain image in the test data set.

[0059] In the specific implementation process, the training data set includes multiple source domain images, and each source domain image is marked with the category label of the source domain image, and the category label indicates whether the source domain image belongs to a stego image or a carrier image.

[0060] In the specific implementation process, the fully connected layer of the feature extraction layer outputs the parameter matrix corresponding to the image, and the parameter matrix is two 512-length parameter sets corresponding to two categories respectively.

[0061] In the specific implementation process, the source domain feature mean matrix is to list the parameter sets of one category. If there are 32 source domain images, the listed matrix is a 512*32 matrix, and the average value of 32 parameters in each column is calculated to obtain a new 512-length parameter set, which is the source domain feature mean matrix.

[0062] Step S300: Calculate the inter-domain average difference corresponding to each category based on the source domain feature mean matrix and the target domain feature mean matrix corresponding to each category, and calculate the value of the cross-entropy loss function based on the inter-domain average difference.

[0063] Step S400: Calculate the individual prediction entropy and the average prediction entropy based on the probability parameters that each target domain feature image belongs to two categories calculated by the classification layer, and calculate the total prediction entropy based on the individual prediction entropy and the average prediction entropy.

[0064] In the specific implementation process, the classification layer receives the parameter matrices of two categories corresponding to each image output by the feature extraction layer, and outputs the probability parameters of each image corresponding to the category based on each parameter matrix.

[0065] Step S500, calculate the loss function value based on the cross-entropy loss function value and the total prediction entropy, and train the steganalysis network based on the loss function value.

[0066] Adopting the above scheme, in this paper, the supervised information of the source domain and the unsupervised information of the target domain are integrated into the training process of the classifier. By using the cross-entropy loss function value calculated based on the inter-domain average difference and the total prediction entropy calculated by the individual prediction entropy and the average prediction entropy, it is established to alleviate the significant reduction of the detection accuracy caused by the carrier source mismatch (CSM). This scheme can transfer the knowledge learned from the source enhanced features to the target domain, thereby enhancing the classification ability. Different weights are assigned to the individual prediction entropy and the average prediction entropy in the process of calculating the total prediction entropy. By fusing the unsupervised information of the target domain, the detection accuracy is further improved.

[0067] In some embodiments of the present invention, the step of calculating the cross-entropy loss function value based on the inter-domain average difference includes:

[0068] Construct a weight matrix based on the weight parameters of the classification layer. The weight matrix includes a first weight parameter matrix for calculating the stego-image category and a second weight parameter matrix for calculating the carrier-image category;

[0069] Calculate the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix;

[0070] Calculate the second covariance parameter based on the first covariance parameter;

[0071] Calculate the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference.

[0072] In the specific implementation process, the parameters in the weight matrix are all from the weight parameters of the classification layer.

[0073] Adopting the above scheme, this scheme can accurately calculate the cross-entropy loss function value and improve the network training efficiency.

[0074] In some embodiments of the present invention, in the step of calculating the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix, the first covariance parameter is calculated according to the following formula:

[0075]

[0076] where σ c represents the first covariance parameter corresponding to the c category, λ represents a preset scale factor parameter, W represents the weight matrix, and W c represents the first weight parameter matrix or the second weight parameter matrix corresponding to the c category, and WT represents the transpose matrix of the weight matrix, represents the average value of the covariance matrix corresponding to the parameter matrix of class c in the test dataset.

[0077] In some embodiments of the present invention, in the step of calculating the second covariance parameter based on the first covariance parameter, the second covariance parameter is calculated according to the following formula:

[0078]

[0079] where d c represents the second covariance parameter corresponding to class c, and σ c represents the first covariance parameter corresponding to class c.

[0080] In some embodiments of the present invention, in the step of calculating the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference, the cross-entropy loss function value is calculated according to the following formula:

[0081]

[0082] where L FM represents the cross-entropy loss function value, n s represents the number of images in the training dataset, k represents any image in the test dataset, D represents the class number, c represents any class, e is the natural constant, W T represents the transpose matrix of the weight matrix, W c represents the first weight parameter matrix or the second weight parameter matrix corresponding to class c, Δμ c represents the inter-domain average difference corresponding to class c, λ represents the preset scale factor parameter, b represents the bias parameter vector constructed by the bias parameters of the classification layer, b c represents class c, f k represents the parameter matrix corresponding to image k in the test dataset, d c represents the second covariance parameter corresponding to class c.

[0083] The first bias parameter vector is a matrix composed of the bias parameters of the corresponding stego-image class, and the second bias parameter vector is a matrix composed of the bias parameters of the corresponding cover-image class.

[0084] In the implementation process of this solution, during the training process, by minimizing the upper bound of the cross-entropy loss function value, the classifier adaptation from the source domain to the target domain is realized. Therefore, the source feature enhancement process is implicitly implemented, and the computational cost can be ignored.

[0085] Adopting the above solution, according to the characteristics of the steganographic signal, this solution models the distribution of the target domain features, significantly enhances the classifier's ability by enhancing the source domain features in the target domain, and derives the upper bound of the cross-entropy loss of the implicitly enhanced source features. This enables the classifier trained on the source enhanced features to generalize well on the target domain without incurring additional computational costs.

[0086] In some embodiments of the present invention, the parameter matrix of each source domain image is a two-dimensional matrix corresponding to two categories. In the step of calculating the source domain feature mean matrix corresponding to each category based on the parameter matrix of each source domain image, the two-dimensional matrices of the two dimensions of the parameter matrix of the training data set and the test data set are respectively split, the matrices corresponding to one category in all parameter matrices are constructed into a multi-dimensional matrix, the average value of the parameters at the corresponding positions in the multi-dimensional matrix is calculated, the multi-dimensional matrix is converted into a one-dimensional matrix, the one-dimensional matrices obtained by converting the two categories of the training data set are output as the source domain feature mean matrices corresponding to the two categories, and the one-dimensional matrices obtained by converting the two categories of the test data set are output as the target domain feature mean matrices corresponding to the two categories.

[0087] In some embodiments of the present invention, in the step of calculating the inter-domain average difference corresponding to each category based on the source domain feature mean matrix and the target domain feature mean matrix of each category, the inter-domain average difference is calculated based on the following formula:

[0088]

[0089] where, Δμ c represents the inter-domain average difference, represents the target domain feature mean matrix corresponding to category c in the test data set, represents the target domain feature mean matrix corresponding to category c in the test data set.

[0090] In some embodiments of the present invention, in the step of calculating the individual prediction entropy and the average prediction entropy based on the probability parameters of each target domain feature image belonging to two categories calculated by the classification layer, the individual prediction entropy and the average prediction entropy are calculated based on the following formula:

[0091]

[0092] where, L EI represents the individual prediction entropy, n t represents the number of images in the test data set, j represents any image in the test data set, c represents the category, if c = 0, it represents the carrier image category, if c = 1, it represents the stego image category, represents the probability parameter of the j-th image corresponding to category c output by the classification layer;

[0093]

[0094] Among them, L EA represents the average prediction entropy, which represents the average value of the probability parameters of the images in the test dataset for the c category.

[0095] In some embodiments of the present invention, in the step of calculating the total prediction entropy based on the individual prediction entropy and the average prediction entropy, the total prediction entropy is calculated according to the following formula:

[0096] L EIA = α LEI - β LEA ;

[0097] Among them, L EIA represents the total prediction entropy, L EA represents the average prediction entropy, L EI represents the individual prediction entropy, and α and β are the weight parameters corresponding to L EI and L EA respectively.

[0098] Adopting the above scheme, in order to balance the relationship between EI and EA, EIA assigns different weights to them according to their respective characteristics, so as to provide a more meaningful transformation direction for enhancing source features, and obtains a trained source enhancement classifier that is most adapted to the target domain, thereby alleviating the performance degradation brought by CSM.

[0099] In some embodiments of the present invention, in the step of calculating the loss function value based on the cross-entropy loss function value and the total prediction entropy, the loss function value is calculated according to the following formula:

[0100] L FME = L FM + L EIA ;

[0101] Among them, L FME represents the loss function value, L EIA represents the total prediction entropy, and L FM represents the cross-entropy loss function value.

[0102] In the specific implementation process, by minimizing the overall loss, this scheme fully considers the supervised information from the source domain and the unsupervised information from the target domain.

[0103] Deep learning-based networks have the advantage of capturing transferable features; however, as the network depth increases, the transferability of features drops sharply. For the shallow convolutional layers with strong feature transferability, the corresponding pre-trained parameters obtained by minimizing the source domain cross-entropy loss can be directly used in the target domain, and these parameters remain unchanged during the training process. However, due to the poor transferability of the parameter features of the deep fully-connected layers, they cannot be directly used in the target domain, so they will be randomly initialized and updated during the training process of this solution.

[0104] At the beginning of the training process, since the pseudo-labels of the target domain images cannot be accurately predicted, the enhancement direction is inaccurate. As the training process progresses, the pseudo-labels become more and more accurate. To weaken the influence of the early training process and at the same time enhance the influence of the subsequent training process on the transformation direction, a scale factor λ is introduced to control the enhancement intensity, and the calculation formula is as follows:

[0105]

[0106] where t and T c represent the current iteration number and the maximum iteration number respectively. As the training progresses, λ varies from 0 to λ0, and λ0 is the preset upper limit of variation.

[0107] In the prior art, the cover-source mismatch (CSM) poses a huge challenge to the application of steganalysis in the real world because its occurrence means a large domain difference between the source domain and the target domain, which will inevitably lead to a significant decrease in the detection accuracy. To effectively reduce the huge domain difference between the source and the target, this paper proposes a deep steganalysis network (FME-Net) that combines the supervised information of the source domain and the unsupervised information of the target domain to improve the performance of the classifier in the target domain. FME-Net models the target feature distribution as a multivariable Laplace distribution (MLD) according to the statistical characteristics of the steganographic signal, and enhances the source features towards the target domain along the meaningful transformation direction sampled from the MLD, thereby improving the classification ability. To train the classifier by combining the unsupervised information from the target domain, FME-Net treats the two differently according to the different characteristics of minimizing the entropy of individual predictions and maximizing the entropy of average predictions, and assigns different weights. Experimental results also show that compared with several advanced methods, FME-Net has an advantage in alleviating the decline in the detection accuracy of CSM.

[0108] The beneficial effects of this solution include:

[0109] 1. Under the premise of considering the statistical characteristics of steganographic signals, the FAS+MLD method is proposed to guide the source augmented features (SAFs) closer to the target domain. Therefore, FAS+MLD generates rich and meaningful cross-domain transformation directions, effectively making the SAFs closer to the target domain features, thus greatly enhancing the generalization ability of the classifier and further alleviating the impact of CSM on the detection accuracy.

[0110] 2. FAS+MLD only considers the supervised information from the source domain and ignores the unsupervised information from the target domain. Therefore, EIA that considers the unsupervised information of the target domain is combined with FAS+MLD to train the classifier. Different from treating EI and EA without discrimination, EIA can handle them differently according to their respective characteristics and assign different weights to them, which will inevitably make the classifier obtain satisfactory generalization.

[0111] 3. In CSM, when the source domain and the target domain are very different, it is very difficult to directly transfer the knowledge learned from the source domain to the target domain. Therefore, constructing an intermediate domain to effectively reduce the huge domain difference between the source domain and the target domain is an attractive option. In this paper, the intermediate domain is constructed by enhancing the source features in the direction of the target domain feature distribution. Therefore, modeling the target domain feature distribution is crucial for constructing an intermediate domain close to the target domain. To sum up, the more accurate the modeling of the target domain feature distribution is, the closer the intermediate domain is to the target domain, the better the classifier trained on the intermediate domain performs in the target domain, and the less the accuracy degradation caused by CSM.

[0112] Experimental Example

[0113] We conduct several experiments to demonstrate the advantages of our scheme in alleviating the accuracy degradation caused by CSM. Its performance is compared with some recent works, namely J-Net. Three publicly available image datasets are used in the experiments, including ImageNe, MIRFlickr-25K, and BOSSBase v1.01. BOSSBase itself contains 10,000 grayscale images in PGM format with a size of 512×512. We randomly select 10,000 color images from ImageNet and MIRFlickr-25K to generate the corresponding mini-I and mini-M respectively. For mini-I or mini-M, first use the imresize function with default settings in MATLAB to change its image size to 512×512, and then convert it to an 8-bit grayscale image in PGM format.

[0114] We randomly divided mini-I, mini-M, and BOSSBase into training sets and test sets according to the ratio of 8:2. In fact, for BOSSBase, in addition to the random division method, it can also be divided into training sets and test sets in order according to the ratio of 8:2. Let the data sets obtained by the two methods be denoted as BBR and BBS respectively. The test set of BBS consists of 2000 pictures taken from the same camera, while the test set of BBR is randomly selected. Therefore, it consists of images from 7 different cameras. Therefore, the test set distributions of BBS and BBR are different and can be regarded as different domains. During the pre-training process, both the training set and the test set are regarded as the source domain. During the training process, the training set is used as the source domain and the test set is used as the target domain.

[0115] For a given steganographic algorithm (i.e., WOW or S-UNIWARD) and embedding rate (0.2 bpp or 0.4 bpp), we construct a training set and a test set for each image data set.

[0116] To reduce the computational cost of training, Yedroudj-Net is selected as the backbone of FME-Net (i.e., the feature extractor of FME-Net). Yedroudj-Net was originally designed for images of size 256×256, rather than the 512×512 size images used in our experiments. Therefore, we slightly adjusted the convolutional layers of these two backbones to be applicable to 512×512 size images. Specifically, the stride of the 4th convolutional layer of Yedroudj-Net is changed from 1 to 2 as in J-Net. Similar to J-Net, FME-Net uses two FC layers in the experiment and takes the output of the 1st FC layer as the input of FAS+MLD. In addition, λ0 is set to 0.25 in the experiment.

[0117] During the entire training process, the optimizer and learning rate settings of FME-Net are consistent with those of the backbone. In addition, the batch size is set to 32 and the number of epochs is set to 100. All networks are implemented on a GPU with Pytorch 1.11.0 and RTX3090Ti.

[0118] The main task of this scheme is to solve these four cases where the performance drops severely. In this section, we compare the detection accuracies of this scheme with those of three networks, J-Net, TSA, and MEDM, in four cases.

[0119] As can be seen from Table 1, when the embedding rate is 0.4 bpp and the network performance under two steganography algorithms, WOW and S-UNIWARD, is evaluated, the detection accuracy of this scheme is 5.45%-18.33% higher than that of the backbone network used (i.e., Yedroudj-Net), while the detection accuracy of J-Net, the first and only one that proposed to solve the CSM problem, is only 2.60%-9.15% higher than that of Yedroudj-Net. In short, in terms of detection accuracy, this scheme is significantly better than J-Net.

[0120] As can be seen from Table 2, when the embedding rate is 0.2 bpp and the network performance under two steganography algorithms, WOW and S-UNIWARD, is evaluated, the detection accuracy of this scheme is 1.25%-6.27% higher than that of Yedroudj-Net, and the detection accuracy of J-Net is 1.02%-4.38% higher than that of Yedroudj-Net. As can be seen from Table 5, this scheme is significantly better than J-Net.

[0121] To further prove the effectiveness of this scheme, taking mini-I→BBS as an example, the target feature distributions separately extracted by this scheme and J-Net are visualized using t-SNE, where the cover images and stego-images are given different colors. For J-Net, most of the cover and stego features are mixed together, which means its weakness in class recognition. Compared with J-Net, the results of FME-Net in extracting cover features and stego features in the target domain are consistent with expectations, with less overlapping regions, indicating its higher intra-class compactness and larger inter-class distance. Therefore, compared with J-Net, this scheme can further extract discriminative features, which is consistent with the conclusions in Table 1 and Table 2. That is to say, this scheme can effectively alleviate the problem of the decline in detection accuracy caused by CSM.

[0122] Performance of this scheme when the backbone network is Yedroudj-Net and the embedding rate is 0.4 bpp

[0123]

[0124] Performance of this scheme when the backbone network is Yedroudj-Net and the embedding rate is 0.2 bpp

[0125]

[0126] Experiments prove that by integrating the supervised information of the source domain and the unsupervised information of the target domain into the training process of the classifier, this solution alleviates the significant reduction in detection accuracy caused by CSM. This solution enhances the source features towards the target domain along the meaningful transformation direction sampled by MLD, transfers the knowledge learned from the source-enhanced features to the target domain, thereby enhancing the classification ability. In addition, this solution uses EIA to process EI and EA differently and assigns different weights, further improving the detection accuracy by fusing the unsupervised information of the target domain. The experimental results also verify that FME-Net is significantly superior to J-Net, TSA, and MEDM in terms of detection accuracy.

[0127] The present invention also provides a steganography analysis network training device that combines source supervision and target unsupervised information. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps achieved by the method described above.

[0128] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps achieved by the aforementioned railway engineering ECM management method based on metadata. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0129] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement it in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted via a data signal carried in a carrier wave on a transmission medium or a communication link.

[0130] It should be clear that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0131] In the present invention, features described and / or illustrated for one embodiment can be used in the same or similar manner in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0132] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for training a steganalysis network that combines source supervision and target unsupervised information, characterized in that The steganalysis network includes a feature extraction layer and a classification layer. The feature extraction layer is used to extract an input image into a parameter matrix, and the classification layer is used to determine the category of the input image. The category is a stego image or a carrier image. The steps of the method include: Input the test data set into the steganalysis network, and output whether each target domain image in the test data set is a stego image or a carrier image; Input the preset training data set into the feature extraction layer to obtain the parameter matrix of each source domain image in the corresponding training data set. Calculate the source domain feature mean matrix of each category based on the parameter matrix of each source domain image. Split the parameter matrices of two dimensions of each parameter matrix in the training data set and the test data set respectively. Construct the matrices corresponding to one category in all parameter matrices into a multi-dimensional matrix. Calculate the average value of the parameters at the corresponding positions in the multi-dimensional matrix. Convert the multi-dimensional matrix into a one-dimensional matrix. Output the one-dimensional matrices obtained by converting the two categories in the training data set as the source domain feature mean matrices corresponding to the two categories, and output the one-dimensional matrices obtained by converting the two categories in the test data set as the target domain feature mean matrices corresponding to the two categories. Calculate the target domain feature mean matrix of each category based on the parameter matrix of each target domain image in the test data set; Calculate the inter-domain average difference of each category based on the source domain feature mean matrix and the target domain feature mean matrix of each category. Calculate the cross-entropy loss function value based on the inter-domain average difference. Construct a weight matrix based on the weight parameters of the classification layer. The weight matrix includes a first weight parameter matrix for calculating the stego image category and a second weight parameter matrix for calculating the carrier image category; Calculate the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix; Calculate the second covariance parameter based on the first covariance parameter; Calculate the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference; Calculate the individual prediction entropy and the average prediction entropy based on the probability parameters that each target domain feature image obtained by the classification layer belongs to the two categories, and calculate the total prediction entropy based on the individual prediction entropy and the average prediction entropy; Calculate the loss function value based on the cross-entropy loss function value and the total prediction entropy, and train the steganalysis network based on the loss function value.

2. The method according to claim 1, characterized in that, In the step of calculating the first covariance parameter based on the weight matrix, the first weight parameter matrix, and the second weight parameter matrix, calculate the first covariance parameter according to the following formula: ； Among them, represents the first covariance parameter corresponding to category c, represents a preset scale factor parameter, represents a weight matrix, represents the first weight parameter matrix or the second weight parameter matrix corresponding to category c, represents the transpose matrix of the weight matrix, represents the average value of the covariance matrix corresponding to the parameter matrix of category c in the test dataset.

3. The method according to claim 1, wherein In the step of calculating the second covariance parameter based on the first covariance parameter, calculate the second covariance parameter according to the following formula: Among them, represents the second covariance parameter corresponding to category c, represents the first covariance parameter corresponding to category c.

4. The method according to claim 1, wherein In the step of calculating the cross-entropy loss function value based on the second covariance parameter and the inter-domain average difference, calculate the cross-entropy loss function value according to the following formula: Among them, represents the value of the cross-entropy loss function, represents the number of images in the training dataset, k represents any one image in the test dataset, D represents the class number, c represents any class, and e is the natural constant. represents the transpose matrix of the weight matrix, represents the first weight parameter matrix or the second weight parameter matrix corresponding to class c, represents the inter-domain average difference corresponding to class c, represents the preset scale factor parameter, represents the bias parameter vector constructed by the bias parameters of the classification layer, represents class c, represents the parameter matrix corresponding to image k in the test dataset, represents the second covariance parameter corresponding to class c.

5. The method according to claim 1, wherein In the step of calculating the inter-domain average difference of each category based on the source domain feature mean matrix and the target domain feature mean matrix of each category, calculate the inter-domain average difference according to the following formula: Among them, represents the inter-domain average difference, represents the mean matrix of the target domain features corresponding to category c in the test dataset, represents the mean matrix of the target domain features corresponding to category c in the test dataset.

6. The method according to claim 1, characterized in that, In the step of calculating the individual prediction entropy and the average prediction entropy based on the probability parameters that each target domain feature image obtained by the classification layer belongs to the two categories, calculate the individual prediction entropy and the average prediction entropy according to the following formula: Among them, represents the individual prediction entropy, represents the number of images in the test dataset, j represents any image in the test dataset, c represents the category. If c = 0, it represents the carrier image category. If c = 1, it represents the stego-image category, represents the probability parameter of the j-th image corresponding to the c-th category output by the classification layer; Among them, represents the average predicted entropy, represents the average of the probability parameters of the images in the test dataset for the c category.

7. The method according to claim 1, characterized in that In the step of calculating the loss function value based on the cross-entropy loss function value and the total prediction entropy, the loss function value is calculated according to the following formula: Among them, represents the loss function value, represents the total prediction entropy, represents the cross-entropy loss function value.

8. A steganalysis network training device that combines source supervision and target unsupervised information, characterized in that The device includes a computer device, the computer device includes a processor and a memory, computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the method according to any one of claims 1-7.

Images