An image spatial domain steganalysis system and method
By utilizing the feature extraction and feature analysis units of the CDNet model, and employing the SRM high-pass filter and residual channel attention block structure, combined with composite downsampling blocks and global covariance pooling layers, the problem of insufficient attention to stegographic regions and loss of detail information in existing models during steganalysis is solved, thereby improving detection accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUXI UNIV
- Filing Date
- 2026-04-01
- Publication Date
- 2026-06-23
Smart Images
Figure CN122265756A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image steganalysis technology, specifically to an image spatial domain steganalysis system and method. Background Technology
[0002] Image steganography is a technique that embeds secret information into digital images to achieve the purpose of covert information transmission without arousing suspicion from third parties. As a countermeasure to steganography, image steganography analysis aims to detect whether the input image contains secret information, and focuses on capturing the subtle perturbations caused by the steganography process.
[0003] Traditional image steganalysis methods typically consist of two parts: feature extraction and binary classification. Feature construction often relies heavily on human experience, and adjusting feature parameters is time-consuming and labor-intensive. Existing technologies use convolutional neural networks with powerful feature extraction capabilities to solve image steganalysis tasks. The entire process is automated, integrating feature extraction and classification into an end-to-end architecture. A series of spatial domain steganalysis models have been proposed. In terms of architecture design, some classic convolutional neural network ideas, such as residual networks, deep separable networks, and Siamese networks, are gradually incorporated. In terms of training strategies, stepwise transfer learning and data augmentation strategies are used during training to accelerate model convergence and prevent overfitting. However, this image steganalysis method clearly has at least the following shortcomings: 1. When extracting noise residuals, existing spatial domain steganalysis models mostly use multiple stacked convolutional layers, allocating the same attention to smooth areas and complex texture edges that may contain hidden information, lacking a focused attention on steganalytic regions.
[0004] 2. Most existing steganalysis analyzers often use a combination of ordinary convolution and average pooling during the downsampling stage, which can easily lose detailed information, thus limiting the performance of the model to some extent. Summary of the Invention
[0005] To address the aforementioned technical shortcomings, the present invention aims to provide an image spatial domain steganalysis system and method.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In the first aspect, the present invention provides an image spatial domain steganalysis system, including: a CDNet model module.
[0007] The CDNet model module includes a feature extraction unit, a feature analysis unit, and a classification unit.
[0008] The feature extraction unit first inputs the original image into the CDNet model, preprocesses the input image using an SRM high-pass filter, and then constructs a residual channel attention block structure by embedding a lightweight channel attention mechanism into the residual structure to enhance the steganalytic signal. The process is as follows: In the preprocessing layer, 30 basic high-pass filters commonly used in steganalysis from the SRM rich model are used, and their parameters are fixed during training. After obtaining the noise residual through preprocessing, it is enhanced. The feature enhancement employs the idea of a residual network, embedding the channel attention mechanism into the residual network. Within the difference block, a residual channel attention block is formed. The backbone network of this residual channel attention block uses a common lightweight structure of 1×1 convolutional layer-3×3 convolutional layer-1×1 convolutional layer, with each convolutional layer followed by a BN layer and a SiLU function. The number of convolutional kernels used in the three convolutional layers is always 30. Local spatial features of the image are extracted through the 3×3 convolutional layer, and the extracted local spatial features are embedded into the channel attention module. In the channel attention module, the global information of the image is first integrated through global average pooling, and then processed by two 1×1 convolutional layers and activation functions to obtain attention weights.
[0009] The feature analysis unit is used to construct four stacked composite downsampling blocks. Each composite downsampling block is composed of a normal convolution, a detail enhancement convolution, and an average pooling layer. The detail enhancement convolution consists of five convolutions deployed in parallel, namely a normal convolution, a central difference convolution (CDC), an angular difference convolution (ADC), a horizontal difference convolution (HDC), and a vertical difference convolution (VDC). By utilizing the linear property of the convolutional layer, these five parallel convolutions are simplified into a single convolutional layer.
[0010] The classification unit is used to first convert the feature map obtained by the feature analysis module into a feature vector through a global covariance pooling layer, and then output the classification result through a fully connected layer and a softmax function.
[0011] Secondly, the present invention provides an image spatial domain steganalysis method system, comprising: S1, constructing a dataset: firstly, dividing the dataset into a training set, a validation set and a test set according to a preset ratio; secondly, using a spatial domain adaptive steganalysis algorithm to embed secret information into each original image according to a preset embedding rate to obtain the steganalysis image of each original image; and finally, binding each original image and its steganalysis image into pairs and inputting them into the CDNet model.
[0012] S2. Optimize CDNet model parameters: During the training of the CDNet model for image spatial domain steganalysis, the stochastic gradient descent algorithm is used to optimize the model parameters, and its accuracy on the validation set is evaluated after each round of training.
[0013] S3. Calculate the accuracy: After training is completed, the final model obtained from the training is used for detection on the test set, and the overall detection accuracy of the model is calculated.
[0014] The beneficial effects of this invention are as follows: 1. This invention provides an image spatial domain steganalysis system and method. The feature extraction unit preprocesses the input image through an SRM high-pass filter and constructs a residual channel attention block structure by embedding a lightweight channel attention mechanism into the residual structure. The four stacked composite downsampling blocks of the feature analysis unit are composed of ordinary convolution, detail enhancement convolution, and average pooling layers. The classification unit converts the feature map obtained by the previous module into a feature vector through a global covariance pooling layer, and outputs the classification result through a fully connected layer and a softmax function. This effectively enhances the noise residual features extracted by the preprocessing layer, effectively reduces the loss of steganalytic signals during downsampling, and ensures detection accuracy.
[0015] 2. In the feature extraction unit of this invention, the original image is first input into the CDNet model, and the input image is preprocessed by an SRM high-pass filter. Then, a residual channel attention block structure is constructed by embedding a lightweight channel attention mechanism into the residual structure to enhance the stegographic signal. The backbone network of the residual channel attention block uses a lightweight structure of 1×1 convolutional layer-3×3 convolutional layer-1×1 convolutional layer, and each convolutional layer is equipped with a BN layer and a SiLU function. The local spatial features of the image are extracted by the 3×3 convolutional layer, and the extracted local spatial features are embedded into the channel attention module. In the channel attention module, the global information of the image is first integrated by global average pooling, and then processed by two 1×1 convolutional layers and activation functions to obtain attention weights, which effectively enhances the noise residual features extracted by the preprocessing layer.
[0016] 3. The present invention constructs four stacked composite downsampling blocks in the feature analysis unit, and each composite downsampling block is composed of ordinary convolution, detail enhancement convolution and average pooling layer. By integrating ordinary convolution, detail enhancement convolution and average pooling layer and applying them to the feature dimensionality reduction process, the loss of stegographic signal during downsampling is effectively reduced. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a schematic diagram of the system structure connection of the present invention.
[0019] Figure 2 The network architecture diagram of the image spatial domain steganalysis model provided by this invention.
[0020] Figure 3 This diagram illustrates the structure of the channel attention mechanism in this invention.
[0021] Figure 4 This is a structural diagram of the residual channel attention module in this invention.
[0022] Figure 5 This diagram illustrates the structure of the composite downsampling block in this invention.
[0023] Figure 6 This represents a composition diagram of the detail-enhancing convolution in this invention.
[0024] Figure 7 The diagram illustrates the derivation of the central difference convolution in this invention.
[0025] Figure 8 This diagram illustrates the derivation process of horizontal difference convolution in this invention.
[0026] Figure 9 This diagram illustrates the derivation process of vertical difference convolution in this invention.
[0027] Figure 10 This diagram illustrates the derivation process of the angular difference convolution in this invention.
[0028] Figure 11 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Please see Figure 1 and Figure 2 As shown, the present invention provides an image spatial domain steganalysis system, including: a CDNet model module.
[0031] The CDNet model module includes a feature extraction unit, a feature analysis unit, and a classification unit.
[0032] Please see Figure 3As shown, the feature extraction unit first inputs the original image into the CDNet model, preprocesses the input image using an SRM high-pass filter, and then constructs a residual channel attention block structure by embedding a lightweight channel attention mechanism into the residual structure to enhance the steganalytic signal. The process is as follows: In the preprocessing layer, 30 basic high-pass filters commonly used in steganalysis from the SRM rich model are used, and their parameters are fixed during training. After obtaining the noise residual through preprocessing, it is enhanced. The feature enhancement employs the idea of a residual network, embedding the channel attention mechanism... The residual channel attention block is formed in the residual block. The backbone network of the residual channel attention block uses a common lightweight structure of 1×1 convolutional layer-3×3 convolutional layer-1×1 convolutional layer. Each convolutional layer is equipped with a BN layer and a SiLU function. The number of convolutional kernels used in the three convolutional layers is always 30. The local spatial features of the image are extracted through the 3×3 convolutional layer and embedded into the channel attention module. In the channel attention module, the global information of the image is first integrated through global average pooling, and then processed by two 1×1 convolutional layers and activation functions to obtain the attention weights.
[0033] It should be noted that the feature extraction unit implementation is as follows: In the feature extraction unit, the preprocessing layer uses 30 basic high-pass filters commonly used in steganalysis from SRM rich models, and their parameters are fixed during training. After obtaining the noise residuals through preprocessing, this method does not directly perform feature aggregation, but rather enhances them to improve the signal-to-noise ratio. To prevent the steganalytic signal from attenuating as the network deepens, this method adopts the idea of residual networks during feature enhancement. Simultaneously, to guide the model to quickly focus on the steganalytic region, a channel attention mechanism is embedded into the residual block, forming a Residual Channel Attention Block (RCAB). Specific implementation details are as follows... Figure 4 As shown, the backbone network uses a common lightweight structure of 1×1 convolutional layers - 3×3 convolutional layers - 1×1 convolutional layers, with each convolutional layer followed by a BN layer and a SiLU function. The number of kernels used in the three convolutional layers is always 30, which is completely consistent with the number of channels in the feature map output during the preprocessing stage. Therefore, the input and output can be directly connected. After extracting local spatial features in the middle 3×3 convolutional layer, a channel attention module is embedded, which integrates global information through global average pooling. Then, it is processed by two 1×1 convolutional layers and an activation function to obtain attention weights. The feature map output from the 3×3 convolutional layer is multiplied with the obtained weights to complete the recalibration of the feature map, enhancing useful features while suppressing useless features. After feature enhancement processing by 3 RCABs, 30 feature maps of size 256×256 are obtained.
[0034] Please see Figure 5 and Figure 6 As shown, the feature analysis unit is used to construct four stacked composite downsampling blocks. Each composite downsampling block is composed of a normal convolution, a detail enhancement convolution, and an average pooling layer. The detail enhancement convolution consists of five convolutions deployed in parallel, namely, a normal convolution, a center difference convolution (CDC), an angle difference convolution (ADC), a horizontal difference convolution (HDC), and a vertical difference convolution (VDC). By utilizing the linear property of the convolutional layer, these five parallel convolutions are simplified into one convolutional layer.
[0035] It should be noted that the feature analysis unit implementation includes four stacked composite downsampling blocks (CDBs) to refine and analyze the features obtained from the feature extraction module. Considering that the combination of convolutional layers and average pooling commonly used in spatial steganalysis during the downsampling stage easily loses detail information, and that detail-enhancing convolution excels at capturing texture details, this invention introduces detail-enhancing convolution into the feature analysis unit based on conventional downsampling methods. The detail-enhancing convolution mainly comprises five parallel convolutions: ordinary convolution, center difference convolution, angular difference convolution, horizontal difference convolution, and vertical difference convolution. The difference between ordinary convolution and differential convolution lies in the fact that ordinary convolution directly performs convolution operations between the feature map and kernel weights, while differential convolution first calculates the difference between adjacent pixels and then multiplies it with the kernel weights to obtain the feature map, effectively capturing texture, edge, and other features in the image. During model training, since the pixel values of the input feature map are fixed, the purpose of differential convolution is mainly achieved by adjusting the kernel weights. First is the Center Difference Convolution (CDC). The weight adjustment in CDC mainly aims to achieve the goal of differencing each neighboring pixel with the center pixel of the feature map. The derivation process is as follows: Figure 7 As shown. Regarding Horizontal Differential Convolution (HDC), its weight adjustment mainly aims to achieve the goal of differencing the selected pixel pairs in the image along the horizontal direction in order to calculate the gradient in the horizontal direction. The derivation process is as follows: Figure 8 As shown. Regarding Vertical Differential Convolution (VDC), its weight adjustment mainly aims to achieve the goal of differencing the selected pixel pairs in the image along the vertical direction in order to calculate the gradient in the vertical direction. The derivation process is as follows: Figure 9 As shown in the figure. Regarding the angular difference convolution ADC, the adjustment of its weights mainly aims to achieve the goal of differencing adjacent pixels around the center pixel according to the direction indicated in the figure, in order to capture edge detail information. The derivation process is as follows: Figure 10 As shown, the kernel weights of the learned parallel convolutions are reparameterized to reduce the number of parameters, accelerating the training and testing process. Leveraging the linear nature of convolutional layers, the five parallel convolutions of the kernel convolution are simplified into a single ordinary convolutional layer.
[0036] The detailed derivation of these differential convolutions reveals that, compared to ordinary convolutions, different differential convolutions calculate gradients in different directions. This means the model can extract multi-scale features, enhancing its sensitivity to steganalytic details. Applying this to the downsampling stage of a steganalysis model, the collaboration of ordinary convolutions and different differential convolutions preserves more steganalytic details, reducing the loss of steganalytic signals in conventional downsampling operations. In the Composite Downsampling Block (CDB), a standard 3×3 convolution is first used to extract basic image features, providing a foundation for subsequent feature processing. Then, a detail-enhancing convolution is used to extract multi-scale feature information, with parallel processing of the ordinary convolution and four differential convolutions to obtain more texture details. Finally, an average pooling layer further aggregates and refines the detail-enhanced features to obtain a more representative feature representation. The three components in this composite downsampling block work together to solve the problem of lost detail information in conventional downsampling methods, improving the model's ability to capture weak steganalytic signals. To compensate for the loss of spatial information, while reducing the feature map size, this invention increases the number of convolutional kernels to ensure information capacity in the downsampling stage. Therefore, the number of convolutional kernels is increased in the first and third CDBs. After feature aggregation processing through four cascaded CDBs, 120 feature maps of size 16×16 are obtained.
[0037] In a specific embodiment, the five parallel convolutions are simplified into a single convolutional layer, and the specific expression of this convolutional layer is as follows: In the formula Represents detail difference convolution. The convolution kernel representing a regular convolution. The kernel representing the central difference convolution (CDC) is... The convolution kernel representing horizontal difference convolution (HDC) The convolution kernel representing the angle difference convolution ADC, The convolution kernel representing the vertical difference convolution (VDC) Represents the transformed convolutional kernel. These represent the numbers for ordinary convolution, central difference convolution (CDC), angular difference convolution (ADC), horizontal difference convolution (HDC), and vertical difference convolution (VDC). Representative and Number The details of the convolution operation performed by the convolution kernel are enhanced by the input feature map of the convolution.
[0038] The classification unit is used to first convert the feature map obtained by the feature analysis module into a feature vector through a global covariance pooling layer, and then output the classification result through a fully connected layer and a softmax function.
[0039] It should be noted that in the implementation of the classification unit: global average pooling only calculates the statistical information of each channel in the feature map, ignoring the relationships between channels and losing some detailed feature information. Global covariance pooling, on the other hand, is equivalent to second-order pooling, which can retain more in-depth information about the image. Existing research has shown that using global covariance pooling instead of global average pooling can improve the performance of CNN-based image steganalysis. Therefore, in the classification unit, global covariance pooling is used to convert the 120 feature maps into... The system generates a 3D feature vector, which is then fed into a fully connected layer. The Softmax function is used to calculate the probability distribution, and the prediction results are output.
[0040] Please see Figure 11 As shown, the present invention provides an image spatial domain steganalysis method, including: S1, constructing a dataset: firstly, the dataset is divided into a training set, a validation set and a test set according to a preset ratio; secondly, a spatial domain adaptive steganography algorithm is used to embed secret information into each original image according to a preset embedding rate to obtain the steganalysis image of each original image; finally, each original image and its steganalysis image are bound together as a pair and input into the CDNet model.
[0041] It should be noted that the dataset is divided into training set, validation set, and test set in a ratio of 4:1:5.
[0042] In a specific embodiment, the construction of the dataset is carried out as follows: S11, the size of each original image is uniformly adjusted to 256×256 using the imresize function in Maltab.
[0043] S12. Construct a steganalysis dataset using the S-UNIWARD, WOW, and HILL spatial domain adaptive steganography methods.
[0044] S13. Expand the steganography dataset using data augmentation techniques involving random mirroring and rotation.
[0045] It should be noted that the authoritative benchmark datasets BOSSBase and BOWS2 in the field of steganalysis were used, and two different configurations of training sets were prepared during the experiment.
[0046] (1) Training set BOSS: The 10,000 images in the BOSSBase dataset are divided into training set, validation set and test set in a ratio of 4:1:5. The training set contains 4,000 images, the validation set contains 1,000 images and the test set contains 5,000 images.
[0047] (2) Training set BOSS+BOWS2: In order to enhance the dataset and improve the detection performance of the model, the training set was expanded. Based on the 4,000 images in the training set in (1), 10,000 images from BOWS2 were added, resulting in 14,000 images as the training set. The validation set and test set are the same as in (1).
[0048] Considering hardware resource limitations, the image size was adjusted to 256×256 using the `imresize` function in Matlab. In experiments, three widely used spatial adaptive steganography methods, S-UNIWARD, WOW, and HILL, were used to evaluate the performance of the proposed steganalysis method. All steganalysis embedding algorithms generated steganalytic images using Matlab code at embedding rates ranging from 0.1 to 0.4 bpp.
[0049] S2. Optimize CDNet model parameters: During the training of the CDNet model for image spatial domain steganalysis, the stochastic gradient descent algorithm is used to optimize the model parameters, and its accuracy on the validation set is evaluated after each round of training.
[0050] In a specific embodiment, the optimization of CDNet model parameters is carried out as follows: S21, the weights of the CDNet model are updated using the SGD optimizer.
[0051] S22. Initialize and regularize the CDNet model.
[0052] It should be noted that the initialization and regularization settings of the CDNet model are the same as those of LWENet, and L2 regularization is disabled for the bias terms of each convolutional layer and fully connected layer.
[0053] S23. Train the CDNet model. The CDNet model is trained for a total of 250 epochs. When training the 120th and 210th epochs, the learning rate is set to 0.1 times the initial learning rate.
[0054] It should be noted that when the training set is BOSS+BOWS2, only 70 epochs are needed for transfer learning. The initial learning rate is set to 0.001, and then divided by 10 at the 30th and 55th epochs respectively.
[0055] S24. During the training process, the embedding rate of each image in the steganalysis dataset is obtained and compared with the preset embedding rate threshold. If the embedding rate of a certain image is less than the preset embedding rate threshold, the transfer learning strategy is adopted for that image.
[0056] It should be noted that the preset embedding rate threshold is set by the staff. For images with an embedding rate lower than the preset embedding rate threshold, a transfer learning strategy is adopted, that is, the weights obtained from training on high embedding rate images are used to initialize the model, and fine-tuning is performed during training so that the model converges as quickly as possible on low embedding rate images.
[0057] Among them, a high embedding rate image refers to an image whose embedding rate is greater than or equal to a preset embedding rate threshold.
[0058] In the above, the initial learning rate of the CDNet model is set to 0.01, the momentum value is set to 0.9, the weight decay value is set to 0.0005, and the batch size is set to 32.
[0059] S3. Calculate the accuracy: After training is completed, the final model obtained from the training is used for detection on the test set, and the overall detection accuracy of the model is calculated.
[0060] In a specific embodiment, the accuracy calculation process is as follows: when the CDNet model is trained in each epoch, the current CDNet model is saved and its performance is evaluated on the validation set. The CDNet models saved in the last 40 epochs are obtained, and the CDNet model with the highest accuracy on the validation set classification is selected as the final CDNet model. At the same time, the performance of the CDNet model is tested through the test set, and the accuracy is calculated.
[0061] It should be noted that when the training set is BOSS+BOWS2, the model with the highest accuracy in the last 15 epochs of the validation set is selected as the final model during transfer learning.
[0062] Test Case of this Invention: The experiments in this test case were run on Nvidia K40c GPUs, comparing CDNet with four mainstream spatial steganalysis models proposed in recent years, including SRNet, LWENet, FPNet, and CovNet. The experiments also tested three widely used spatial adaptive steganography algorithms: S-UNIWARD, WOW, and HILL. To make the experimental results more convincing, all steganalysis models involved in the comparison experiments used identical training, validation, and test sets. To demonstrate the model's performance on different datasets, the experiments were conducted on both the BOSS and BOSS+BOWS2 datasets.
[0063] Table 1 shows the detection accuracy of each steganalysis model when the training set is BOSS.
[0064] Table 1:
[0065] Table 2 shows the detection accuracy of each steganalysis model when the training set is BOSS+BOWS2.
[0066] Table 2:
[0067] Experimental results show that the proposed CDNet has the best detection performance.
[0068] In this embodiment of the invention, the feature extraction unit preprocesses the input image using an SRM high-pass filter and constructs a residual channel attention block structure by embedding a lightweight channel attention mechanism into the residual structure. The four stacked composite downsampling blocks of the feature analysis unit are composed of ordinary convolution, detail enhancement convolution, and average pooling layers. The classification unit converts the feature map obtained by the previous modules into a feature vector through a global covariance pooling layer, and outputs the classification result through a fully connected layer and a softmax function. This effectively enhances the noise residual features extracted by the preprocessing layer, effectively reduces the loss of stegographic signals during downsampling, and ensures detection accuracy.
[0069] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.
[0070] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.
Claims
1. An image spatial domain steganalysis system, characterized in that, Includes the following modules: The CDNet model module includes a feature extraction unit, a feature analysis unit, and a classification unit: The feature extraction unit first inputs the original image into the CDNet model, preprocesses the input image using an SRM high-pass filter, and then constructs a residual channel attention block structure by embedding a lightweight channel attention mechanism into the residual structure to enhance the steganalytic signal. The process is as follows: In the preprocessing layer, 30 basic high-pass filters commonly used in steganalysis from the SRM rich model are used, and their parameters are fixed during training. After obtaining the noise residual through preprocessing, it is enhanced. The feature enhancement employs the idea of a residual network, embedding the channel attention mechanism into the residual network. Within the difference block, a residual channel attention block is formed. The backbone network of this residual channel attention block uses a common lightweight structure of 1×1 convolutional layer-3×3 convolutional layer-1×1 convolutional layer, with each convolutional layer followed by a BN layer and a SiLU function. The number of convolutional kernels used in the three convolutional layers is always 30. Local spatial features of the image are extracted through the 3×3 convolutional layer, and the extracted local spatial features are embedded into the channel attention module. In the channel attention module, the global information of the image is first integrated through global average pooling, and then processed by two 1×1 convolutional layers and activation functions to obtain attention weights. The feature analysis unit is used to construct four stacked composite downsampling blocks. Each composite downsampling block is composed of a normal convolution, a detail enhancement convolution, and an average pooling layer. The detail enhancement convolution consists of five convolutions deployed in parallel, namely a normal convolution, a center difference convolution (CDC), an angle difference convolution (ADC), a horizontal difference convolution (HDC), and a vertical difference convolution (VDC). By utilizing the linear property of the convolutional layer, these five parallel convolutions are simplified into one convolutional layer. The classification unit is used to first convert the feature map obtained by the feature analysis module into a feature vector through a global covariance pooling layer, and then output the classification result through a fully connected layer and a softmax function.
2. The image spatial domain steganalysis system according to claim 1, characterized in that, The five parallel convolutions are simplified into a single convolutional layer, and the specific expression for this convolutional layer is as follows: In the formula Represents detail difference convolution. The convolution kernel representing a regular convolution. The kernel representing the central difference convolution (CDC) is... The convolution kernel representing horizontal difference convolution (HDC) The convolution kernel representing the angle difference convolution ADC, The convolution kernel representing the vertical difference convolution (VDC) Represents the transformed convolutional kernel. These represent the numbers for ordinary convolution, central difference convolution (CDC), angular difference convolution (ADC), horizontal difference convolution (HDC), and vertical difference convolution (VDC). Representative and Number The details of the convolution operation are enhanced by the input feature map of the convolution.
3. An image spatial domain steganalysis method utilizing the image spatial domain steganalysis system according to any one of claims 1-2, characterized in that, include: S1. Constructing the dataset: First, the dataset is divided into training set, validation set and test set according to a preset ratio. Second, the spatial adaptive steganography algorithm is used to embed secret information into each original image according to a preset embedding rate to obtain the steganographic image of each original image. Finally, each original image and its steganographic image are paired and input into the CDNet model. S2. Optimize CDNet model parameters: During the training of the CDNet model for image spatial domain steganalysis, the stochastic gradient descent algorithm is used to optimize the model parameters, and its accuracy on the validation set is evaluated after each round of training. S3. Calculate the accuracy: After training is completed, the final model obtained from the training is used for detection on the test set, and the overall detection accuracy of the model is calculated.
4. The image spatial domain steganalysis method according to claim 3, characterized in that, The specific process for constructing the dataset is as follows: S11. Use the imresize function in Maltab to uniformly adjust the size of each original image to 256×256. S12. Construct a steganalysis dataset using the S-UNIWARD, WOW, and HILL spatial domain adaptive steganography methods; S13. Expand the steganography dataset using data augmentation techniques involving random mirroring and rotation.
5. The image spatial domain steganalysis method according to claim 3, characterized in that, The specific process for optimizing the CDNet model parameters is as follows: S21. Weights updated using the SGD optimizer; S22. Initialize and regularize the CDNet model; S23. Train the CDNet model. The CDNet model is trained for a total of 250 epochs. When training the 120th and 210th epochs, the learning rate is set to 0.1 times the initial learning rate. S24. During the training process, the embedding rate of each image in the steganalysis dataset is obtained and compared with the preset embedding rate threshold. If the embedding rate of a certain image is less than the preset embedding rate threshold, the transfer learning strategy is adopted for that image.
6. The image spatial domain steganalysis method according to claim 5, characterized in that, The initial learning rate of the CDNet model was set to 0.01, the momentum value to 0.9, the weight decay value to 0.0005, and the batch size to 32.
7. The image spatial domain steganalysis method according to claim 3, characterized in that, The accuracy calculation process is as follows: When the CDNet model is trained in each epoch, the current CDNet model is saved and its performance is evaluated on the validation set. The CDNet models saved in the last 40 epochs are obtained, and the CDNet model with the highest accuracy on the validation set is selected as the final CDNet model. At the same time, the performance of the CDNet model is tested on the test set, and the accuracy is calculated.