An image steganalysis method, model and system

Abstract

The present invention discloses an image steganalysis method, model and system, wherein the image steganalysis method based on feature enhancement transfer and channel attention comprises: constructing a feature enhancement transfer module and a channel attention downsampling module; using the feature enhancement transfer module and the channel attention downsampling module to form an image steganalysis model; constructing an image steganalysis data set, wherein the image steganalysis data set comprises an image steganalysis training set and an image steganalysis test set; using the image steganalysis training set to train the image steganalysis model; when the image steganalysis model training is completed, using the image test set to evaluate the image steganalysis model; when the image steganalysis model is evaluated, applying the image steganalysis model to analyze the image steganalysis information. The technical solution of the present invention can solve the problem that the weak steganalysis signal in the prior art is prone to feature disappearance.

Description

Technical Field

[0001] The present invention relates to the technical field of image steganalysis, and in particular to an image steganalysis method, model and system based on feature enhancement transfer and channel attention. Background Art

[0002] With the rapid development of Internet technology and the increasing maturity of digital multimedia technology, people pay more and more attention to information security issues in the communication process. Therefore, as a means to ensure information security, image steganography algorithms have developed rapidly. Image steganography algorithms can embed image hidden information into complex areas in the image to avoid visible artifacts and changes in statistical properties. Most grayscale image steganography algorithms can treat each color band as a grayscale image and embed the secret bits independently into the grayscale image. Therefore, image steganography technology can be used to enhance information security.

[0003] Although image steganography can enhance information security, it is a double-edged sword. Many dangerous elements can use image steganography to hide image steganographic information that endangers national security in various carriers and publish them on websites. If not screened, the image steganographic information written by image steganography will cause incalculable losses. Because such adaptive image steganography algorithms are difficult to detect, it is even more difficult to manually detect these visually invisible weak information for the massive image data on the Internet. Therefore, efficient and accurate automatic steganalysis methods have important research significance and application value.

[0004] As an adversarial technology for image steganography algorithms, image steganalysis technology can determine whether a given image contains image steganographic information, which has important research significance for ensuring information security. Most image steganalysis methods use convolutional neural networks (CNNs) to directly learn a large number of features from the input image and classify the features; then calculate the loss and calculate the gradient through the back-propagation algorithm, and update the network parameters according to the obtained gradient to obtain better classification performance and achieve fast automatic detection. Specifically, most of these methods use convolution kernels to preprocess the image to be detected; then a series of convolution and downsampling operations are used to further abstract and transform the preprocessed features so that the obtained features can be better identified. At present, the most advanced image steganalysis methods are constructed using feature-based steganalysis and machine learning, and share the same processing flow, namely: noise residual calculation, feature construction, and binary classification. For the above methods, the residual helps to improve the signal-to-noise ratio of the stegographic signal to extract more accurate features; however, the most advanced feature set is a combination of coexisting residuals of different filters, the so-called rich model. Rich models are usually high-dimensional, such as 30,000 or more dimensions. From the perspective of steganalysis, in order to obtain a more complete description of the original image, high-dimensional representation is an inevitable trend. High-dimensional representation will make the features obtained by steganalysis more and more complex, and these features may deviate from the actual features. At the same time, obtaining high-dimensional features means using more convolutions and pooling to obtain a deeper receptive field. However, convolution and pooling themselves have the characteristics of suppressing weak signals, which will cause the already weak steganalysis signals to disappear easily as the network deepens. That is, the combination of simple convolution and fully connected layers is difficult to learn the long-range dependency between deep features and shallow features. Under this premise, the already weak steganalysis signals obtained by downsampling in the shallow network are more likely to disappear, and using only deep features cannot achieve ideal detection results. Summary of the invention

[0005] The present invention provides an image steganalysis method, model and system based on feature enhancement transfer and channel attention, aiming to solve the problem that the combination of simple convolution and fully connected layers in the prior art is difficult to learn the long-range dependency relationship between deep features and shallow features. Under this premise, the shallow network downsampling to obtain the already weak steganalysis signal is more likely to cause feature disappearance.

[0006] To achieve the above object, according to a first aspect of the present invention, the present invention proposes an image steganalysis method based on feature enhancement transfer and channel attention, comprising:

[0007] Construct a feature enhancement transfer module and a channel attention downsampling module, wherein the feature enhancement transfer module is used to perform feature enhancement transfer operations on the image, and the channel attention downsampling module is used to perform channel attention weighting and downsampling operations on the image;

[0008] Use the feature enhancement transfer module and the channel attention downsampling module to build an image steganalysis model;

[0009] Constructing an image steganography dataset, wherein the image steganography dataset includes an image steganography training set and an image steganography test set;

[0010] Use the image steganalysis training set to train the image steganalysis model;

[0011] When the image steganalysis model training is completed, the image steganalysis model is evaluated using the image test set;

[0012] When the image steganalysis model evaluation passes, the image steganalysis model is applied to analyze the image steganalysis information.

[0013] Preferably, in the above-mentioned image steganalysis method, the steps of constructing a feature enhancement transfer module and a channel attention downsampling module include:

[0014] Use residual multi-convolution blocks and convolution short sides in parallel to build the first feature transfer module;

[0015] Use the first feature transfer module to splice with a directly connected branch to obtain a feature splicing structure;

[0016] Use the feature splicing structure to connect in series with the second feature transfer module to obtain a feature enhancement transfer module;

[0017] as well as,

[0018] Construct a channel attention module, and connect the channel attention module in series with the pixel-by-pixel downsampling module and the third feature transfer module in sequence to construct a channel attention downsampling module.

[0019] Preferably, in the above-mentioned image steganalysis method, the step of establishing an image steganalysis model comprises:

[0020] Using multiple SRM cores, a preprocessing module is constructed;

[0021] Use the adaptive average pooling layer, the filtering layer and the fully connected layer in series to construct the steganalysis classification head;

[0022] The image steganalysis model is constructed by using the preprocessing module, feature enhancement transfer module, channel attention downsampling module and steganalysis classification head in sequence.

[0023] Preferably, in the above-mentioned image steganalysis method, the step of constructing an image steganalysis data set comprises:

[0024] Acquire an original image set, wherein the original image set includes a plurality of original images;

[0025] Embed the image steganographic information into each original image to generate a steganographic image corresponding to the original image;

[0026] The original image and the stego image are taken as the same group of images to form an image steganographic dataset;

[0027] According to the predetermined division ratio, the image steganography dataset is divided into an image steganography training set and an image steganography test set.

[0028] Preferably, in the above-mentioned image steganalysis method, the step of training the image steganalysis model using the image steganalysis training set comprises:

[0029] The preprocessing module uses the SRM kernel as the convolution kernel to perform convolution operations on the images of the image steganography training set to obtain the preprocessed feature map;

[0030] In the feature enhancement transfer module, a first feature transfer module is used to perform feature transfer on the preprocessed feature map to obtain a first feature map; the first feature map is concatenated with the preprocessed feature map to obtain a second feature map; a second feature transfer module is used to perform feature transfer operation on the second feature map to obtain an output feature map;

[0031] In the channel attention downsampling module, the output feature map is sequentially subjected to channel attention weighting using the channel attention module, pixel downsampling using the pixel-by-pixel downsampling module, and feature transfer operation using the third feature transfer module to obtain the downsampled features after channel attention weighting;

[0032] The steganalysis classification head is used to classify the channel-weighted downsampled features to obtain the image classification results.

[0033] Preferably, in the above-mentioned image steganalysis method, the step of using the first feature transfer module to perform feature transfer on the preprocessed feature map to obtain the first feature map includes:

[0034] Perform multi-convolution operation on the preprocessed feature map using multiple convolution pairs of the residual multi-convolution block and a residual connection to obtain a first convolution feature sub-map;

[0035] Use the convolution short edge to perform a convolution operation on the preprocessed feature map to obtain the second convolution feature sub-map;

[0036] The first convolution feature subgraph is concatenated with the second convolution feature subgraph, and a convolution operation is performed to obtain a first feature graph.

[0037] Preferably, the above-mentioned image steganalysis method, the step of using the image training set to train the image steganalysis model comprises:

[0038] Input the image training set into the image steganalysis model, and set the training and optimization parameters of the image steganalysis model;

[0039] The multi-step decay algorithm and online data augmentation technology are used to tune the training and optimization parameters of the image steganalysis model.

[0040] Preferably, in the above-mentioned image steganalysis method, the step of evaluating the image steganalysis model using the image test set comprises:

[0041] Input the image test set into the image steganalysis model;

[0042] Use the true and false positive values ​​and true and false negative values ​​measured by the image steganalysis model in the image test set to calculate the precision and recall of the image steganalysis model respectively;

[0043] Using precision and recall, calculate the model metric score for the image steganalysis model;

[0044] Select the model weight corresponding to the optimal value of the model indicator score and set it as the inference weight of the image steganalysis model;

[0045] Use the image steganalysis model after setting the inference weight to analyze the image steganalysis information.

[0046] According to a second aspect of the present invention, the present invention provides an image steganalysis model based on feature enhancement transfer and channel attention, comprising:

[0047] Preprocessing module, feature enhancement transfer module, channel attention downsampling module and steganalysis classification head; among them,

[0048] The preprocessing module includes convolution kernels composed of multiple SRM matrices;

[0049] The feature enhancement transfer module includes a first feature transfer module, a direct branch, and a second feature transfer module, wherein the first feature transfer module includes a residual multi-convolution block and a convolution short edge; the first feature transfer module and the direct branch form a feature splicing structure, and the feature splicing structure is connected in series with the second feature transfer module;

[0050] The channel attention downsampling module includes a channel attention module, an alternate pixel downsampling module and a third feature transfer module connected in series in sequence;

[0051] The steganalysis classification head consists of an adaptive average pooling layer, a filtering layer, and a fully connected layer.

[0052] According to a third party of the present invention, the present invention provides an image steganalysis system based on feature enhancement transfer and channel attention, comprising:

[0053] A memory, a processor, and an image steganalysis program stored in the memory and executable on the processor. When the image steganalysis program is executed by the processor, the steps of the image steganalysis method described in any one of the above technical solutions are implemented.

[0054] In summary, the image steganalysis scheme based on feature enhancement transfer and channel attention provided by the above technical solution of the present invention, by constructing a feature enhancement transfer module and a channel attention downsampling module, the feature enhancement transfer module can enhance the features of the image and pass the enhanced features as much as possible to reduce the loss of features; in addition, the channel attention downsampling module is adopted, by weighting the features with channel attention, and then sampling, the feature diversity can be further enhanced and the feature information of the image can be retained. After the image steganalysis model is constructed by the above feature enhancement transfer module and the channel attention downsampling module, the image steganalysis training set in the image steganalysis data set is input into the image steganalysis model, and the image steganalysis model is trained to obtain the strength of the enhanced weak feature signal, and the shallow layer features are transferred to the deep layer, and the information loss is avoided as much as possible during the downsampling process by weighting the attention. Then the image steganalysis model is evaluated by the image test set, the high performance of the image steganalysis model is maintained, and the weak steganographic features are extracted and processed, thereby improving the analysis ability of the image steganalysis model. In summary, the above method can solve the problem in the prior art that the features themselves are relatively weak and the neural network is prone to feature disappearance. The combination of simple convolution and fully connected layers is difficult to learn the long-range dependency relationship between deep features and shallow features, and weak steganographic signals are prone to feature disappearance. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0056] Figure 1 It is a structural schematic diagram of an image steganalysis model based on feature enhancement transfer and channel attention provided by an embodiment of the present invention;

[0057] Figure 2 yes Figure 1 A schematic diagram of the structure of a feature transfer model provided by the illustrated embodiment;

[0058] Figure 3 yes Figure 1 A schematic diagram of the structure of a channel attention downsampling module provided in the illustrated embodiment;

[0059] Figure 4 It is a flowchart of an image steganalysis method based on feature enhancement transfer and channel attention provided by an embodiment of the present invention;

[0060] Figure 5 yes Figure 4 A schematic flow chart of a method for constructing a feature enhancement transfer module and a channel attention downsampling module provided in the illustrated embodiment;

[0061] Figure 6 yes Figure 4 A schematic diagram of a flow chart of a method for constructing an image steganalysis model provided by the illustrated embodiment;

[0062] Figure 7 yes Figure 4 A schematic diagram of a flow chart of a method for constructing an image steganographic dataset provided by the illustrated embodiment;

[0063] Figure 8 yes Figure 4 A schematic flow chart of a first method for training an image steganalysis model provided in the illustrated embodiment;

[0064] Fig. 9 yes Figure 8 A schematic flow chart of a method for constructing a first characteristic map provided in the illustrated embodiment;

[0065] Fig.10 yes Figure 4 A schematic flow chart of a second method for training an image steganalysis model provided in the illustrated embodiment;

[0066] Fig.11 yes Figure 4 A schematic flow chart of an evaluation method of an image steganalysis model provided by the illustrated embodiment;

[0067] Fig.12 It is a structural schematic diagram of an image steganalysis system based on feature enhancement transfer and channel attention provided by an embodiment of the present invention.

[0068] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0069] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0070] The main technical problems solved by the embodiments of the present invention are:

[0071] Most of the existing image steganalysis methods are built using feature-based steganalysis and machine learning, and share the same processing flow. The most advanced feature set available is the result of the coexistence of different filter residuals, i.e., a rich model. High-dimensional rich models extract a large number of complex features, including key steganalysis features and a large number of redundant features. However, the steganalysis features that help improve detection performance are weak. Since the combination of simple convolution and fully connected layers is difficult to learn the long-range dependency between deep features and shallow features, as the number of neural network layers continues to increase, the weak steganalysis features extracted by such shallow networks are more and more likely to disappear during the neural network transmission process. The loss of such key features will lead to a decline in network detection performance.

[0072] In order to solve the above problems, that is, to obtain excellent classification performance, the design of the neural network structure must consider two key factors: 1. How to design the structure to more effectively help shallow features to be transferred to deep layers; 2. How to downsample to avoid information loss as much as possible. The following embodiment of the present invention provides an image steganalysis scheme based on feature enhancement transfer and channel attention. By using feature enhancement transfer and channel attention weighted downsampling, shallow features are transferred to deep layers and the purpose of reducing information loss is achieved.

[0073] To achieve this, see Figure 1 , Figure 1 A structural diagram of an image steganalysis model based on feature enhancement transfer and channel attention provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the image steganalysis model includes:

[0074] Preprocessing module SRM, feature enhancement transfer module FETB, channel attention downsampling module CAD and steganalysis classification head Cls Head; Among them,

[0075] The preprocessing module SRM includes a convolution kernel composed of multiple SRM matrices. The spatial rich model steganalysis SRM matrix is ​​a steganalysis matrix commonly used in the steganalysis module. In the specific implementation of the embodiment of the present application, a convolution operation with 30 SRM matrices as convolution kernels is used; it should be noted that the convolution kernel parameters of the convolution layer will be initialized to 30 SRM matrices before training, but the parameters of the convolution kernel are learnable during training; this means that its parameter values ​​will change as the back-propagation parameters are updated.

[0076] The feature enhancement transfer module FETB includes a first feature transfer module FTB, a direct connection branch and a second feature transfer module FTB. Figure 2, the first feature transfer module FTB includes a residual multi-convolution block and a convolution short edge Conv1×1; the first feature transfer module FTB and the direct branch form a feature splicing structure ( Figure 2 The feature splicing structure is connected in series with the second feature transfer module; see Figure 1 and Figure 2 The feature enhancement transfer module FETB is composed of two FTB modules. After the preprocessing module processes and outputs the initial feature map b1, it is convolved through the first feature transfer module to enhance certain features in the initial feature map to obtain b1_1. In addition, the initial feature map b1 is directly output through the direct branch, and b1_1 and b1 are directly concatenated through the above-mentioned feature concat structure to obtain b2. b2 is then convolved through the second feature transfer module to enhance the features of b2. In this way, the entire feature enhancement transfer module can output the final feature map b3.

[0077] Among them, Figure 2 As shown in , each feature transfer module FTB includes a residual multi-convolution block and a convolution short side; the residual multi-convolution block and the convolution short side are connected in parallel and spliced, and then undergo a convolution to form the feature transfer module. Figure 2 As shown in the figure, the residual multi-convolution block includes a 1×1 convolution, a 3×3 convolution and a 1×1 convolution. Specifically, a feature map is input, assuming that the feature map is a1. After a1 is input into the FTB module, a1 undergoes a 1×1 convolution to obtain a1_1, and then a1_1 undergoes a 1×1 convolution and a 3×3 convolution to obtain a1_2; a1_1 and a1_2 are added to obtain a1_3; for another parallel short edge, the short edge includes a 1×1 convolution, and the original input a1 undergoes a 1×1 convolution to obtain a2; finally, a2 and a1_3 are concatenated through a concat operation and then a 1×1 convolution is performed to obtain the FTB output result a3; it should be noted that each convolution layer used here (except the convolution used by the first layer of SRM) has a BN layer and a Relu layer after the convolution layer.

[0078] In addition, if Figure 1 As shown, the channel attention downsampling module includes a channel attention module, an alternate pixel downsampling module and a third feature transfer module which are connected in series in sequence.

[0079] Among them, Figure 3As shown in the figure, the channel attention module includes a mean function and a 1×1 convolution kernel sigmoid layer. Specifically, taking the input feature map c1 as an example, the mean value of c1 is calculated, and then a layer of 1×1 convolution and a sigmoid layer is used to obtain the attention c1_1 of different channels; then c1 is multiplied by c1_1 and the result is passed through a BN layer. The channel attention module can obtain the attention of different channels, and multiply the attention of the channel with the original input feature map, and input the result into a BN layer to obtain a new feature map. Through the weighting of channel attention, the feature information of the highlighted image can be improved and the diversity of features can be enhanced. In addition, by sampling through the pixel downsampling module, the feature map output by the channel attention module can be sampled pixel by pixel in the resolution dimension, so as to achieve expansion in the channel dimension and increase the number of sampled pixels. The sampled feature map is then processed by the third feature transfer module, which can smooth the roughness of the original feature map and improve the accuracy of the feature.

[0080] The steganalysis classification head includes an adaptive average pooling layer, a filtering layer, and a fully connected layer. Figure 2 As shown in FIG. 1 , the steganalysis classification head has a specific structure including an adaptive average pooling layer, a filtering Dropout layer, and a fully connected FC layer. The above three layers are combined in sequence. The retention ratio of the Dropout layer is 0.1.

[0081] For the entire image steganalysis model, the first module of data input is the preprocessing module composed of SRM; then it passes through two consecutive feature enhancement transfer modules FETB and two consecutive channel attention downsampling modules CAD, and finally the extracted features are classified by the steganalysis classification head Cls Head; the classification result is whether the detected image has image steganalysis information. If there is, the model outputs 1, otherwise it outputs 0.

[0082] See also Figure 4 , Figure 4 is a flow chart of an image steganalysis method based on feature enhancement transfer and channel attention provided by an embodiment of the present invention, such as Figure 4 As shown, the image steganalysis method includes the following steps:

[0083] S110: Construct a feature enhancement transfer module and a channel attention downsampling module, wherein the feature enhancement transfer module is used to perform feature enhancement transfer operations on the image, and the channel attention downsampling module is used to perform channel attention weighting and downsampling operations on the image. The feature enhancement transfer module can use the residual convolutional network to transfer features of the original image, and splice the feature image after feature transfer with the original feature image to further enhance the features, and then use the channel attention downsampling module to perform channel attention weighting processing on the image to further increase the diversity of features. In addition, the channel attention downsampling module performs a downsampling operation after channel attention weighting, and superimposes the sampled features on the channel vector dimension, thereby reducing the disappearance of features.

[0084] Specifically, as a preferred embodiment, Figure 5 As shown, in the technical solution provided in the embodiment of the present application, the steps of constructing the feature enhancement transfer module and the channel attention downsampling block module include:

[0085] S111: Use the residual multi-convolution block and the convolution short side in parallel to build the first feature transfer module. Figure 2 As shown in the structure of the feature transfer module, a residual multi-convolution block and the short side of the convolution are connected in parallel and spliced, and then a convolution is performed to form the feature transfer module. Figure 2 As shown in the figure, the residual multi-convolution block includes a 1×1 convolution, a 3×3 convolution and a 1×1 convolution. Specifically, a feature map is input, assuming that the feature map is a1. After a1 is input into the FTB module, a1 undergoes a 1×1 convolution to obtain a1_1, and then a1_1 undergoes a 1×1 convolution and a 3×3 convolution to obtain a1_2; a1_1 and a1_2 are added to obtain a1_3; for the other parallel short edge, the original input a1 undergoes a 1×1 convolution to obtain a2; finally, a2 and a1_3 are concatenated through a concat operation and then a 1×1 convolution is performed to obtain the FTB output result a3; it should be noted that each convolution layer used here (except the convolution used by the first layer of SRM) has a BN layer and a ReLU layer after the convolution layer.

[0086] S112: Use the first feature transfer module to splice with a directly connected branch to obtain a feature splicing structure.

[0087] Through this feature splicing structure, the feature map obtained by convolution of the first feature transfer module can be spliced ​​with the initial feature map obtained directly from the preprocessing module by the direct branch. The spliced ​​feature map can retain the features of the original image on the basis of enhancing some features, retain the original weak image steganographic information, and reduce the disappearance of features.

[0088] S113: Connecting the feature splicing structure in series with the second feature transfer module to obtain a feature enhancement transfer module. The feature splicing structure is connected in series with the second feature transfer module, so that the second feature transfer module can be used to further enhance some features and enhance the smoothness of the features.

[0089] S114: Construct a channel attention module, and connect the channel attention module in series with the pixel-by-pixel downsampling module and the third feature transfer module in sequence to construct a channel attention downsampling module.

[0090] See also Figure 3 For a channel attention downsampling module CAD, the channel attention module used includes a mean function, a 1×1 convolution and a sigmoid layer. In this way, for the input feature map, assuming that the feature map is c1, the mean function can be used to calculate the average of c1, and the above 1×1 convolution and a sigmoid layer are used to obtain the attention c1_1 of different channels; then the input feature map c1 is multiplied by the channel attention c1_1, and the result is passed through a BN layer to obtain the feature map c2.

[0091] In addition, for the downsampling layer IS features in the channel attention downsampling module CAD, the specific process is as follows: for the output c2 of the channel attention module CA, every other pixel is sampled in the resolution dimension, so that the length and width resolution of the entire feature map are halved to H / 2 and W / 2 respectively, and the number of channels becomes 4C, and the dimension of the output feature map c3 of IS is (4C, H / 2, W / 2). Through the above process, the downsampled features can be stacked on the channel attention dimension, thereby obtaining a large number of features and reducing feature loss. In addition, the third feature transfer module is connected in series to reduce the roughness of features and retain important features.

[0092] like Figure 4 As shown in Figure 2, after constructing the feature enhancement transfer module and the channel attention downsampling module, Figure 4 The image steganalysis method shown also includes:

[0093] S120: Use the feature enhancement transfer module and the channel attention downsampling module to build an image steganalysis model.

[0094] Specifically, as a preferred embodiment, Figure 6 As shown, in the image steganalysis method provided in the embodiment of the present application, the step of establishing the image steganalysis model includes:

[0095] S121: Use multiple SRM cores as convolution kernels to construct a preprocessing module. The preprocessing module of the present application uses multiple SRM cores. In the specific implementation, the convolution operation using 30 SRM matrices as convolution kernels is implemented. In addition, it should be noted that the convolution kernel parameters of the convolution layer will be initialized to 30 SRM matrices before training; however, during training, the parameters of the convolution kernel are learnable; this means that its parameter values ​​will change as the backpropagation parameters are updated.

[0096] S122: Use the adaptive average pooling layer, the filter layer and the fully connected layer to be connected in series in sequence to construct a steganalysis classification head. For the steganalysis classification head in the image steganalysis model network, its specific structure includes an adaptive average pooling Adaptive Avg Pooling layer, a filter Dropout layer and a fully connected FC layer; by sequentially combining the above three-layer network, a steganalysis classification head can be obtained, and the classification result of the steganalysis classification head is whether the detected image has image steganalysis information. If there is image steganalysis information, 1 is output, and if there is no image steganalysis information, 0 is output; wherein the retention ratio of the filter layer Dropout is 0.1.

[0097] S123: Using the preprocessing module, feature enhancement transfer module, channel attention downsampling module and steganalysis classification head in sequence, an image steganalysis model is constructed.

[0098] For the entire image steganalysis model, the first module for data input is SRM; followed by two consecutive FETBs and two consecutive CADs, and finally the features extracted by the above modules are classified by the steganalysis classification head Cls Head. In summary, the image steganalysis method based on feature enhancement transfer and channel attention designed in the embodiment of the present application describes a new image steganalysis model based on feature enhancement transfer and channel attention. Because the image steganalysis model can realize feature enhancement transfer and can realize channel attention weighting and downsampling, compared with the most advanced image steganalysis detection algorithm, the model has obvious advantages in terms of parameter quantity, computational complexity, detection accuracy and other aspects.

[0099] After building the image steganalysis model, Figure 4 The image steganalysis method provided by the illustrated embodiment further includes the following steps:

[0100] S130: constructing an image steganography dataset, wherein the image steganography dataset includes an image steganography training set and an image steganography test set.

[0101] In the embodiment of the present application, the current image steganography algorithm, such as S-Uniward, WOW, HILL and other algorithms, is used to embed the image steganographic information into the image of the original data set, thereby constructing an image steganographic data set. After the image steganographic data set is constructed, the image steganographic data set is classified according to a certain ratio to obtain an image steganographic training set for training the image steganalysis model and an image steganographic test set for testing.

[0102] Since the network model based on deep learning needs to learn features from data samples, it is necessary to create positive and negative samples for network learning (pictures with secret information and original pictures). Usually, there is no image steganographic information (i.e. secret information) hidden in the picture, so we need to create learning samples for the network model, that is, manually embed the secret information into normal pictures.

[0103] Specifically, as a preferred embodiment, Figure 7 As shown, in the embodiment of the present application, the step of constructing the image steganographic data set includes:

[0104] S131: Acquire an original image set, wherein the original image set includes a plurality of original images. The original image set is a collection of images that do not contain image steganographic information, and each original image in the original image set does not contain image steganographic information.

[0105] S132: Embed the image stego information into each original image respectively to generate a stego image corresponding to the original image. In the embodiment of the present application, the image stego information is embedded into the original image by an image steganography algorithm. The image steganography algorithms used in the embodiment of the present application include S-Uniward, WOW and HILL algorithms. When the image stego information is embedded using the steganalysis algorithm, the embedding rate is set to p, and the image stego information is embedded at the embedding rate p for each picture. It should also be noted that when the image stego information is embedded in the original image, the amount of embedded information is expressed by the embedding rate. The larger the embedding rate, the more information is embedded, and in this case, the easier it is to detect that the image contains image stego information. The smaller the embedding rate, the less information is embedded, and in this case, the more difficult it is to detect that the image contains image stego information. For the convenience of the experiment, the embodiment of the present application uniformly uses the embedding rate p to embed the image stego information into the image. For the convenience of the experiment, the embedded image stego information is usually random noise; since the embedded image stego information can be any content, the embodiment of the present application uses random noise as the embedded information. The random noise here refers to a set of random numbers with no regular pattern, such as [1,4,5,2,6,9,14,....].

[0106] S133: The original image and the stego image are taken as the same group of images to form an image steganalysis data set. In addition, it is assumed here that the original image before embedding is called cover, and the image after embedding the image stego information is called steg; there is no difference between the images before and after embedding under human observation; the difference is that the images before and after steganography (the images before and after embedding the secret information) may have one or two pixel differences in some pixels; a cover and the steg generated by this cover are together called a pair and are classified into the same group.

[0107] After the image steganographic information is embedded, the size of the entire image steganographic dataset is twice the original size (cover and steg)

[0108] S134: Dividing the image steganography data set into an image steganography training set and an image steganography test set according to a predetermined division ratio.

[0109] The image steganography dataset with embedded information is divided into image steganography training set and image steganography test set in proportion. Before division, the order of the dataset needs to be randomly shuffled; the purpose of random shuffling here is to make the sample distribution more uniform; the division ratio is set to r = 6:4. Here r is the size ratio of the training set to the test set, and the specific value of r can be adjusted appropriately according to different needs. In addition, it should be noted that the shuffling here is for pairs, that is, after dividing the training set and the test set, the two images (steg and cover) belonging to the same pair will be divided into the same dataset (training set or test set).

[0110] After constructing the image steganography dataset, Figure 4 The image steganalysis method provided by the illustrated embodiment further includes the following steps:

[0111] S140: training an image steganalysis model using the image steganalysis training set.

[0112] The image steganalysis model training and optimization provided in the embodiments of the present application are as follows:

[0113] When training the image steganalysis model, set the training cycle epoch to 500, the batch size to 16, and the learning rate to 0.001 (different data sets may require corresponding adjustments); use CrossEntropyLoss as the loss function. In addition, use Adamax as the optimizer, set its eps to 1e-8, and weight_decay to 1e-4. Among them, CrossEntropyLoss is as follows: CrossEntropyLoss = H(p,q) = -∑p(x)logq(x)+(1-p(x)log(1-q(x))).

[0114] During training, multi-step decay is used as the learning decay algorithm, and its milestones are set to [300, 400] and gamma is set to 0.1; the learning rate is updated after each epoch.

[0115] No normalization is performed on the data during training; online data augmentation is performed on the training images using random rotation and random flipping to improve the generalization ability of the model.

[0116] As a preferred embodiment, the specific content of using the image steganography training set to train the image steganalysis model is as follows: Figure 8 As shown, the steps of training the image steganalysis model using the image steganalysis training set include:

[0117] S141: The preprocessing module uses the SRM kernel as a convolution kernel to perform a convolution operation on the image of the image steganography training set to obtain a preprocessed feature map. In the embodiment of the present application, the SRM kernel is used as the preprocessing module, and the convolution operation using 30 SRM matrices as the convolution kernel is implemented.

[0118] S142: In the feature enhancement transfer module, use the first feature transfer module to perform feature transfer on the preprocessed feature map to obtain a first feature map; concatenate the first feature map with the preprocessed feature map to obtain a second feature map; use the second feature transfer module to perform feature transfer operation on the second feature map to obtain an output feature map. For details, see Figure 1 and Figure 2 The feature enhancement transfer module FETB is composed of two FTB modules, so that after the preprocessing module processes to obtain the preprocessed feature map b1, it is convolved through the first feature transfer module to enhance certain features in the initial feature map to obtain the first feature map b1_1. In addition, the preprocessed feature map b1 is directly obtained through the direct branch. The first feature map b1_1 and the preprocessed feature map b1 are directly concatenated with the preprocessed feature map b1 through the above-mentioned feature concat structure to obtain the second feature map b2. The second feature map b2 is then convolved through the second feature transfer module to enhance the features of b2. In this way, the entire feature enhancement transfer module can output the final output feature map b3.

[0119] S143: In the channel attention downsampling module, the output feature map is sequentially subjected to channel attention weighting using the channel attention module, pixel downsampling using the pixel-by-pixel downsampling module, and feature transfer operation using the third feature transfer module to obtain the downsampled features after channel attention weighting.

[0120] Combination Figure 3As shown in the structure of the channel attention downsampling module, the channel attention downsampling CAD module can be specifically divided into three parts: channel attention (CA), every other pixel downsampling (IS) and FTB; its specific structure is as follows Figure 3 As shown: For the CA module used in the CAD module, the specific process is as follows:

[0121] For the input feature map c1, firstly, the mean value of c1 is calculated by mean, and then the attention c1_1 to different channels is obtained through a 1×1 convolution and a sigmoid layer; then c1 is multiplied by c1_1 and the result is passed through a BN layer to obtain the feature map c2. For the downsampling layer IS feature in CAD, the specific process is as follows: for the CA module output c2, every other pixel is sampled in the resolution dimension, so that the length and width resolution of the entire feature map are halved to H / 2 and W / 2 respectively, and the number of channels becomes 4C, and the dimension of the IS output feature map c3 is (4C, H / 2, W / 2). After the CAD module performs channel attention weighting and every other pixel downsampling, the features need to be further processed through the FTB module.

[0122] Among them, the calculation method of channel attention weighting is as follows:

[0123] sigmoid(conv(mean(f))).

[0124] S144: Use the steganalysis classification head to classify the image according to the channel weighted sampling features to obtain the image type. The steganalysis classification head has a specific structure including an Adaptive Avg Pooling layer, a Dropout layer, and an FC layer; the three layers are combined in sequence; the retention ratio of the Dropout layer is 0.1. The classification result is whether the detected image has image steganalysis information. If yes, the model outputs 1; if no, the model outputs 0.

[0125] In addition, as a preferred embodiment, Fig. 9 As shown, the step of using the first feature transfer module to transfer features of the preprocessed feature map to obtain the first feature map includes:

[0126] S1421: Use multiple convolutions of the residual multi-convolution block and a residual connection to perform a multi-convolution operation on the preprocessed feature map to obtain a first convolution feature sub-map.

[0127] S1422: Perform a convolution operation on the preprocessed feature map using the convolution short edge to obtain a second convolution feature sub-map.

[0128] S1423: Concatenate the first convolution feature subgraph and the second convolution feature subgraph, and perform a convolution operation to obtain a first feature graph.

[0129] Specific as Figure 2 As shown in the figure, the residual multi-convolution block includes a 1×1 convolution, a 3×3 convolution and a 1×1 convolution. Specifically, a feature map is input, assuming that the feature map is a1. After a1 is input into the FTB module, a1 undergoes a 1×1 convolution to obtain a1_1, and then a1_1 undergoes a 1×1 convolution and a 3×3 convolution to obtain a1_2; a1_1 and a1_2 are added to obtain the first convolution feature submap a1_3; for another parallel short edge, the original input feature map a1 undergoes a 1×1 convolution to obtain the second convolution feature submap a2; finally, the second convolution feature submap a2 and the first convolution feature submap a1_3 are concatenated through a concat operation, and then a 1×1 convolution is performed to obtain the output result of FTB, that is, the first feature map a3; it should be noted that each convolution layer used here (except the convolution used by the first layer of SRM) has a BN layer and a ReLU layer after the convolution layer.

[0130] In addition, as a preferred embodiment, Fig.10 As shown, the steps of using the image training set to train the image steganalysis model in the above embodiment include:

[0131] S145: Input the image training set into the image steganalysis model, and set the training and optimization parameters of the image steganalysis model.

[0132] S146: Use the multi-step decay algorithm and online data augmentation technology to tune the training and optimization parameters of the image steganalysis model.

[0133] In the technical solution provided in the embodiment of the present application, when training the image steganalysis model, the training cycle epoch is set to 500, the batch size is 16, and the learning rate is 0.001 (different data sets may require corresponding adjustments); CrossEntropyLoss is used as the loss function; Adamax is used as the optimizer, and its eps is set to 1e-8 and weight_decay is set to 1e-4. During training, multi-step decay is used as the learning decay algorithm, and its milestones are set to [300,400] and gamma is 0.1; the learning rate size is updated after each epoch. No normalization is performed on the data during training; random rotation and random flipping are used to perform online data enhancement on the training images.

[0134] After training the image steganalysis model, Figure 4 The image steganalysis method provided by the illustrated embodiment further includes the following steps:

[0135] S150: When the image steganalysis model training is completed, the image steganalysis model is evaluated using the image test set.

[0136] As a preferred embodiment, Fig.11 As shown, the steps of evaluating the image steganalysis model using the image test set specifically include:

[0137] S151: Input the image test set into the image steganalysis model.

[0138] S152: Use the true and false positive values ​​and true and false negative values ​​measured by the image steganalysis model on the image test set to calculate the precision and recall rate of the image steganalysis model respectively.

[0139] S153: Calculate the model metric score of the image steganalysis model using precision and recall.

[0140] S154: Select a model weight corresponding to the optimal value of the model indicator score and set it as the inference weight of the image steganalysis model.

[0141] S155: Analyze the image steganalysis information using the image steganalysis model after setting the inference weight.

[0142] The model performance verification is as follows: after the model training is completed, the performance of the model needs to be evaluated, and the F1 Score score is calculated for the inference results and labels on the test set; the performance of the model is judged according to the size of the model indicator score - the F1Score score; the larger the F1 Score, the higher the recognition rate of the model, and vice versa, the lower the F1 Score, the higher the recognition rate of the model, and ... F1 Score of the model, and the lower the F1 Score of the model, the higher the F1 Score of the model, and the lower the F1 Score of the model, the higher the F1 Score of the model, and the lower the F1 Score of the model, the higher the F1 Score of the model, and the lower the F1 Score of the model, the higher the F1 Score of the model, and the lower the F1 Score of the model, the higher the F1 Score

[0143] in,

[0144] The accuracy of the image steganalysis model is calculated using the true and false positive values ​​and the true and false negative values. The calculation formula is as follows:

[0145] The calculation formula for calculating the recall rate of the image steganalysis model using true and false positive values ​​and true and false negative values ​​is as follows:

[0146] The formula for calculating the model index score of the image steganalysis model using precision and recall is as follows:

[0147] After evaluating the performance of the model on the test set, the weight with the highest F1Score is selected as the weight for the final application of reasoning. That is, when the F1Score is the highest, it is determined that the values ​​of the training parameters of the model are optimal, and the training parameter values ​​at this time are maintained, and the image steganalysis model is applied to analyze the image steganalysis information.

[0148] When the image steganalysis model passes the evaluation, the image steganalysis model is applied to analyze the image steganalysis information. When the image steganalysis model is applied to analyze the image steganalysis information, the above-mentioned inference weights are used, that is, the various parameters when the model maintains the optimal value of the above-mentioned model indicator score, and the image steganalysis model at this time is used to analyze the image steganalysis information.

[0149] In summary, the image steganalysis scheme based on feature enhancement transfer and channel attention provided by the embodiment of the present invention, by constructing a feature enhancement transfer module and a channel attention downsampling module, the feature enhancement transfer module can enhance the features of the image and pass the enhanced features as far as possible to reduce the loss of features; in addition, the channel attention downsampling module is adopted, by weighting the features with channel attention, and then sampling, the feature diversity can be further enhanced and the feature information of the image can be retained. After the image steganalysis model is constructed by the above-mentioned feature enhancement transfer module and the channel attention downsampling module, the image steganalysis training set in the image steganalysis data set is input into the image steganalysis model, and the image steganalysis model is trained to obtain the strength of the enhanced weak feature signal, and the shallow layer features are passed to the deep layer, and the information loss is avoided as much as possible during the downsampling process by weighting the attention. Then the image steganalysis model is evaluated by the image test set, the high performance of the image steganalysis model is maintained, and the weak steganographic features are extracted and processed, thereby improving the analysis ability of the image steganalysis model. In summary, the above method can solve the problem in the prior art that the features themselves are relatively weak, the neural network is prone to feature disappearance, the simple combination of convolution kernels and fully connected layers is difficult to learn the long-range dependency relationship between deep features and shallow features, and the weak steganographic signal is prone to feature disappearance.

[0150] Based on the same concept of the above method embodiment, the embodiment of the present invention also proposes an image steganalysis model and system based on feature enhancement transfer and channel attention, which are used to implement the above method of the present invention. Since the principle and method of solving the problem in the embodiment of the system are similar, it at least has all the beneficial effects brought by the technical solutions of the above embodiments, which will not be repeated here one by one.

[0151] See also Fig.12 , Fig.12 A structural diagram of an image steganalysis system based on feature enhancement transfer and channel attention provided by an embodiment of the present invention. Fig.12As shown, the image steganalysis system includes:

[0152] A communication line 1002, a memory 1004, a processor 1001, and an image steganalysis program stored in the memory 1004 and executable on the processor 1001. When the image steganalysis program is executed by the processor, the steps of the image steganalysis method provided in any of the above embodiments are implemented.

[0153] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0157] It should be noted that in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claim. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0158] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0159] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An image steganalysis method based on feature enhancement transfer and channel attention, characterized in that: include: Constructing a feature enhancement transfer module and a channel attention downsampling module, wherein the feature enhancement transfer module is used to perform a feature enhancement transfer operation on the image, and the channel attention downsampling module is used to perform a channel attention weighting and downsampling operation on the image; Using the feature enhancement transfer module and the channel attention downsampling module, an image steganalysis model is constructed; Constructing an image steganography dataset, wherein the image steganography dataset includes an image steganography training set and an image steganography test set; Using the image steganography training set to train the image steganalysis model; When the image steganalysis model training is completed, the image steganalysis model is evaluated using the image test set; When the image steganalysis model passes the evaluation, applying the image steganalysis model to analyze the image steganalysis information; The step of constructing the feature enhancement transfer module and the channel attention downsampling module includes: Use residual multi-convolution blocks and convolution short sides in parallel to build the first feature transfer module; Use the first feature transfer module to splice with a directly connected branch to obtain a feature splicing structure; Use the feature splicing structure to connect in series with a second feature transfer module to obtain the feature enhancement transfer module; as well as, Constructing a channel attention module, connecting the channel attention module in series with the every-pixel downsampling module and the third feature transfer module in sequence to construct the channel attention downsampling module; The step of building an image steganalysis model includes: Use multiple SRM kernels as convolution kernels to construct a preprocessing module; Use the adaptive average pooling layer, the filtering layer and the fully connected layer in series to construct the steganalysis classification head; The image steganalysis model is constructed by sequentially using the preprocessing module, the feature enhancement transfer module, the channel attention downsampling module and the steganalysis classification head.

2. The image steganalysis method according to claim 1, characterized in that: The step of constructing an image steganographic data set includes: Acquire an original image set, wherein the original image set includes a plurality of original images; Embedding the image steganographic information into each of the original images to generate a steganographic image corresponding to the original image; The original image and the steganographic image are taken as the same group of images to form the image steganographic data set; According to a predetermined division ratio, the image steganography data set is divided into an image steganography training set and an image steganography test set.

3. The image steganalysis method according to claim 2, characterized in that: The step of using the image steganography training set to train the image steganalysis model comprises: The preprocessing module uses the SRM kernel as a convolution kernel to perform a convolution operation on the image of the image steganography training set to obtain a preprocessed feature map; In the feature enhancement transfer module, the first feature transfer module is used to perform feature transfer on the preprocessed feature map to obtain a first feature map; the first feature map is spliced ​​with the preprocessed feature map to obtain a second feature map; and the second feature transfer module is used to perform feature transfer operation on the second feature map to obtain an output feature map; In the channel attention downsampling module, the output feature map is sequentially subjected to channel attention weighting using the channel attention module, pixel downsampling using the pixel-by-pixel downsampling module, and feature transfer operation using the third feature transfer module to obtain the downsampling feature after channel attention weighting; The steganalysis classification head is used to perform image classification on the channel-weighted down-sampled features to obtain an image classification result.

4. The image steganalysis method according to claim 3, characterized in that: The step of using a first feature transfer module to perform feature transfer on the preprocessed feature map to obtain a first feature map includes: Performing a multi-convolution operation on the preprocessed feature map using multiple convolutions of the residual multi-convolution block and a residual connection to obtain a first convolution feature sub-map; Using the convolution short edge to perform a convolution operation on the preprocessed feature map to obtain a second convolution feature sub-map; The first convolution feature subgraph is concatenated with the second convolution feature subgraph, and a convolution operation is performed to obtain the first feature graph.

5. The image steganalysis method according to claim 1, characterized in that: The step of using the image training set to train the image steganalysis model comprises: Inputting the image training set into the image steganalysis model, and setting training and optimization parameters of the image steganalysis model; The training and optimization parameters of the image steganalysis model are tuned using a multi-step decay algorithm and online data enhancement technology.

6. The image steganalysis method according to claim 1, characterized in that: The step of evaluating the image steganalysis model using an image test set comprises: Inputting the image test set into the image steganalysis model; Using the true and false positive values ​​and true and false negative values ​​measured by the image steganalysis model on the image test set, respectively calculating the precision and recall rate of the image steganalysis model; Using the precision and recall, calculating a model index score of the image steganalysis model; Selecting a model weight corresponding to the optimal value of the model indicator score and setting it as the inference weight of the image steganalysis model; The image steganalysis model after setting the inference weight is used to analyze the image steganalysis information.

7. An image steganalysis model based on feature enhancement transfer and channel attention, characterized in that: The method for image steganalysis based on feature enhancement transfer and channel attention as claimed in any one of claims 1 to 6 is used to implement the processing, including: Preprocessing module, feature enhancement transfer module, channel attention downsampling module and steganalysis classification head; among them, The preprocessing module includes a convolution kernel composed of multiple SRM matrices; The feature enhancement transfer module includes a first feature transfer module, a direct branch, and a second feature transfer module, wherein the first feature transfer module includes a residual multi-convolution block and a convolution short edge; the first feature transfer module and the direct branch form a feature splicing structure, and the feature splicing structure is connected in series with the second feature transfer module; The channel attention downsampling module includes a channel attention module, an alternate pixel downsampling module and a third feature transfer module connected in series in sequence; The steganalysis classification head includes an adaptive average pooling layer, a filtering layer and a fully connected layer.

8. An image steganalysis system based on feature enhancement transfer and channel attention, characterized in that: include: A memory, a processor, and an image steganalysis program stored in the memory and executable on the processor, wherein when the image steganalysis program is executed by the processor, the steps of the image steganalysis method according to any one of claims 1 to 6 are implemented.

Images