An image steganalysis method based on multi-scale residual network

Abstract

The application discloses a steganalysis method based on a multi-scale residual network, and relates to the technical field of image steganalysis, and solves the technical problem that in the prior art, a single-channel group is difficult to adapt to different image natural characteristics and steganographic feature distribution differences, weak feature signals are easily lost, and detection effect is limited. By introducing a same-layer multi-scale receptive field mechanism and a same-layer mixed feature channel group mechanism, a multi-scale receptive field network model is constructed. The multi-scale receptive field network model selects multiple different scales of receptive fields in the same layer, generates a same-layer mixed feature channel group, and is subjected to interval mixed identification. Different feature scale channel groups are fused, and MRNet is optimized, so that steganographic features are extracted and detected by using multi-range receptive fields. The problem of feature loss of weak feature signals when steganographic image information is input into a neural network and in the process of transmission of weak features from a low layer to a high layer of the network is solved.

Description

Technical Field

[0001] This invention relates to the field of image steganalysis technology, specifically to an image steganalysis method based on multi-scale residual networks. Background Technology

[0002] With the continuous development of steganography technology, more and more steganography tools have appeared on the internet, which has also brought convenience to criminals. A large amount of image data exists online, and a considerable amount of steganographic information also exists. From the perspective of preventing information leakage and blocking the transmission of dangerous information, it is necessary to identify whether there is steganographic information in images, that is, to conduct image steganography analysis.

[0003] Steganalysis generally consists of two steps: steganalytic feature extraction and steganalytic feature classification. Early steganalysis largely relied on manual feature selection based on experience, followed by classification using classifiers such as Support Vector Machines. The effectiveness of steganalysis depended primarily on the quality of feature selection. As steganalysis methods have become increasingly diverse and complex, manual feature selection has become increasingly difficult. Steganalysis algorithms rely heavily on manual feature selection, requiring significant experience and manpower. Therefore, it is necessary to find an effective method for faster and more convenient feature selection.

[0004] Compared to classic image steganalysis algorithms, deep learning has been widely applied to the field in recent years due to its lower reliance on human experience and broader adaptability. Deep neural network models integrate preprocessing, feature extraction, and classification, achieving a unified approach to image steganalysis. However, in image steganalysis, current deep neural networks use single-channel groups for feature extraction. While increasing the number of convolutional kernels can improve detection accuracy to some extent, once a certain limit is reached, further increases in kernel size offer no significant improvement. Analyzing the natural information of images reveals differences in texture regions, density, and other natural characteristics among different image types, resulting in varying distributions of implicit steganalytic features. Using the same receptive field size for all images makes it difficult to accurately and comprehensively extract information, easily leading to the loss of weak feature signals during steganalysis.

[0005] Therefore, when using deep neural network algorithms for steganalysis, it is necessary to combine the regional differences in image steganography and make more targeted designs to achieve better steganalysis results. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides a multi-scale receptive field network model that introduces a mechanism of multiple receptive fields of different scales within the same layer and a mechanism of mixed feature channel groups within the same layer. By utilizing the collaborative extraction of receptive fields across multiple ranges and fusing channel groups of different feature scales, this invention solves the problem of loss of weak steganalytic features during network transmission, thereby achieving efficient and accurate image steganalysis based on a multi-scale residual network.

[0007] To achieve the above objectives, the present invention provides an image steganalysis method based on multi-scale residual networks. This method, based on a steganalysis residual network (SRNet), constructs a multi-scale receptive field network model (MRNet) by introducing multiple receptive field mechanisms of different scales within the same layer and a mixed feature channel group mechanism within the same layer. The specific steps include the following:

[0008] Step 1: Prepare the dataset and perform data preprocessing;

[0009] Step 2, multi-scale receptive field feature extraction, is used for natural information suppression and initial feature extraction of the image; by selecting multiple receptive fields of different scales in the same layer, a multi-scale receptive field network based on dual convolutional layers is constructed, and the stegographic association features of each region of the preprocessed data image in Step 1 are extracted using the perception range of convolutional kernels of different scales.

[0010] Step 3, multi-channel group feature fusion, is used to suppress the extraction of natural features and steganalytic features; combined with dual-channel group convolutional network and feature map group fusion, the initial features extracted by the same layer mixed feature channel group in step 2 are further extracted and fused.

[0011] Step 4, feature channel group dimensionality reduction; used to reduce the dimensionality of the steganalytic features output in step 3 by constructing a dimensionality-reduced feature channel group, and outputting the dimensionality-reduced image steganalytic features.

[0012] Step 5, Feature Classification and Result Output: This step is used to classify and ultimately determine the steganographic features of the dimensionality-reduced image extracted in Step 4. By constructing a feature classification channel group, the final image steganographic detection result is output.

[0013] Preferably, the MRNet network selects multiple receptive fields of different scales in the same layer to generate a mixed feature channel group in the same layer, and then performs interval mixing recognition to fuse the channel groups of different feature scales. By optimizing MRNet, it is possible to use multiple receptive fields to jointly extract and detect image steganalysis features.

[0014] The receptive field mechanism of different scales constructs a multi-scale receptive field network based on dual convolutional layers, and uses different sizes of receptive range to extract stegographic features from various regions of the image. The same-layer hybrid feature channel group mechanism introduces parallel channel groups between different layers of the neural network to handle the process of weak stegographic features being transmitted from the lower layers to the higher layers of the network.

[0015] Preferably, step 1, preparing the dataset and data preprocessing specifically includes:

[0016] Prepare a base image dataset for training and testing. Based on the base image dataset, generate sub-datasets for model training and testing respectively, and divide the sub-datasets into training set, validation set and test set according to a preset ratio.

[0017] Preferably, in step 2, multi-scale receptive field feature extraction, as the first part of the MRNet network, involves selecting multiple receptive fields of different scales in the same layer to generate a mixed feature channel group in the same layer. Specifically, this includes constructing a receptive field network of different scales based on dual convolutional layers, including convolutional channel group A1, convolutional channel group A2, and convolutional channel group A3, and using different sizes of receptive range to extract stegographic association features for each region of the preprocessed data image in step 1.

[0018] Preferably, the construction of a multi-scale receptive field network based on dual convolutional layers using convolutional kernels of different scales is as follows: the first convolutional layer of convolutional channel group A1 uses a 3×3 convolutional kernel, and the second convolutional layer uses a 3×3 convolutional kernel;

[0019] The first convolutional layer of convolutional channel group A2 uses a 5×5 convolutional kernel, and the second convolutional layer uses a 3×3 convolutional kernel;

[0020] The first convolutional layer of convolutional channel group A3 uses a 7×7 convolutional kernel, and the second convolutional layer uses a 3×3 convolutional kernel;

[0021] The outputs of the three parallel convolutional channel groups A1, A2, and A3 are concatenated to form 22 output feature channels, which serve as the initial output features of the first part of the MRNet network.

[0022] Preferably, in step 3, the multi-channel group feature fusion, as the second part of the MRNet network, combines a dual-channel group convolutional network and feature map group fusion to suppress the extraction of natural features and steganalytic features. Specifically, it includes: constructing two parallel convolutional channel groups B1 and B2 to deepen the feature extraction and fusion of the initial features extracted by the feature channel groups mixed in the same layer in step 2; convolutional channel group B1 is processed through 5 S1 structure residual layers to output steganalytic features of 22 feature channels; convolutional channel group B2 is processed through 3 S2 structure residual layers to output steganalytic features of 22 feature channels.

[0023] The feature channels output by convolutional channel group B1 and the feature channels output by convolutional channel group B2 are fused along the channel dimension to form a total of 44 fused feature channels, which serve as the steganalytic feature output channels for the second part of the MRNet network.

[0024] Preferably, the S1 structure residual layer in the convolutional channel group B1 includes: a first convolutional layer, a BN layer, a ReLU activation function, a second convolutional layer, and a BN layer connected in sequence, and the output is the result of summing the output of the BN layer after the second convolutional layer and the initial input mapping branch information of the S1 structure residual layer; each convolutional layer in the S1 structure residual layer uses a 3×3 convolutional kernel;

[0025] The S2 structure residual layer in the convolutional channel group B2 includes: a first convolutional layer, a BN layer, a ReLU activation function, a second convolutional layer, and a BN layer connected in sequence. The output is the sum of the output of the BN layer after the second convolutional layer and the initial input mapping branch information of the S2 structure residual layer. Each convolutional layer in the S2 structure residual layer uses a 5×5 convolutional kernel.

[0026] Preferably, in step 4, the feature channel group dimensionality reduction is used as the third part of the MRNet network. By constructing the dimensionality reduction feature channel group C1, the steganographic features output in step 3 are dimensionality reduced. Specifically, the dimensionality reduction feature channel group C1 adopts a single channel and is processed through four S3 structure residual layers in sequence. After each processing, the feature map size is halved, and the dimensionality-reduced image steganographic features are output.

[0027] The main branch of the S3 structure residual layer sequentially includes a first convolutional layer, a BN layer, a ReLU activation function layer, a second convolutional layer, and another BN layer, followed by an average pooling layer with a stride of 2. All convolutional layers use 3×3 convolutional kernels. The shortcut branch of the S3 structure residual layer sets a 1×1 convolutional layer with a stride of 2 to match the channel size of the input feature mapping with the output of the main branch.

[0028] The output of the S3 structure residual layer is the result of element-wise summation of the output of the main branch after average pooling and the output of the shortcut branch after 1×1 convolution.

[0029] Preferably, in step 5, feature classification and result output are the fourth part of the MRNet network. By constructing feature classification channel group D1, it is used to classify and ultimately determine the steganalytic features of the dimensionality-reduced image output in step 4. Specifically, this includes:

[0030] The network layers of the D1 channel group consist of: a first convolutional layer, a batch normalization (BN) layer, a ReLU activation function layer, a second convolutional layer, and another BN layer, followed by a global average pooling layer, a fully connected layer, and a softmax layer. The global average pooling layer converts 512 feature maps into 512-dimensional feature vectors. These feature vectors are then processed by the fully connected layer and subjected to binary classification by the softmax layer, outputting the final image steganalysis detection result.

[0031] Preferably, the method further includes step 6, configuring multi-scale residual network parameters for training and testing the MRNet network model. This is achieved by setting parameters such as the input image specifications, the padding mode of each convolutional layer, the activation function type, the convolutional kernel initialization method, the number of training steps, the learning rate, and the optimizer type of the MRNet network model.

[0032] The image steganalysis method based on multi-scale residual networks provided by this invention has the following beneficial effects:

[0033] This invention presents an image steganalysis method based on multi-scale residual networks. By introducing a receptive field mechanism with multiple receptive fields of different scales within the same layer and a mixed feature channel group mechanism within the same layer, a multi-scale residual network model is constructed to achieve high-precision detection of steganalyst images. The receptive field mechanism with multiple receptive fields of different scales within the same layer utilizes the synergistic effect of the perceptual range at different scales to extract steganalyst-related features from various regions of the image. The receptive fields of different scales can cover regional information of different ranges in the image, specifically capturing steganalyst features with different distributions. This reduces the loss of weak feature signals when steganalyst image information is input into the neural network from the source, providing a more complete initial feature foundation for subsequent feature processing.

[0034] Meanwhile, for features extracted from multi-scale receptive fields, parallel-designed convolutional dual-channel groups are used to deepen and fuse the initial features. Different channel groups use residual layers with specific structures to process features, forming a fully optimized architecture from feature extraction and deepening to classification. The same-layer hybrid feature channel group mechanism can effectively preserve the transmission of weak steganalytic features from low to high layers between network layers, avoiding feature loss caused by information compression or transformation during feature transmission in a single channel group. This achieves efficient detection of steganalytic images while reducing the loss of steganalytic features throughout the extraction, transmission, and fusion processes, solving the problem of feature loss during the transmission of weak steganalytic features from low to high layers of the network, and further improving the accuracy of image steganalysis. Attached Figure Description

[0035] Figure 1 This is a flowchart of the image steganalysis method based on multi-scale residual networks in this invention;

[0036] Figure 2This is a structural framework diagram of the multi-scale residual network model in this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0038] like Figure 1 As shown, the image steganalysis method based on multi-scale residual networks of the present invention is based on steganalysis residual network (SRNet). It constructs a multi-scale residual network model (MRNet) by introducing a receptive field mechanism of multiple receptive fields of different scales in the same layer and a feature channel group mechanism of the same layer. The MRNet network selects receptive fields of multiple different scales in the same layer to generate a feature channel group of the same layer. After interval mixing recognition, the channel groups of different feature scales are fused. By optimizing MRNet, it is possible to use multiple receptive fields to jointly extract and detect image steganalytic features. The receptive field mechanism of different scales is constructed by constructing a multi-scale receptive field network based on double convolutional layers. It uses different sizes of receptive range to extract steganalytic association features of various regions of the image. The feature channel group mechanism of the same layer is to handle the process of weak steganalytic features being transmitted from the lower layers to the higher layers of the network by introducing parallel channel groups between the layers of the neural network.

[0039] The specific steps are as follows:

[0040] Step 1: Dataset Preparation and Preprocessing; Using the BOSSbase V1.01 and BOWS2 datasets as the base image datasets, 76 sets of training and testing datasets, totaling 760,000 images, were prepared. Corresponding subsets for model training and testing were generated, and these subsets were divided into training, validation, and testing sets according to a preset ratio. Specifically, the following steps were included:

[0041] (1) Experimental Dataset Preparation: The experimental dataset was generated from two standard datasets. One dataset was based on the BOSSbase V1.01 dataset, which contained 10,000 grayscale images of 512×512 pixels. The BOSSbase V1.01 dataset was used to generate the training, validation, and test datasets. The other dataset was based on BOWS2. This dataset contained 10,000 grayscale images of 512×512 pixels. The BOWS2 dataset was mainly used to generate the training dataset. The experimental dataset was mainly divided into two parts: the spatial domain dataset and the JPEG domain dataset.

[0042] (2) Spatial Domain Dataset Preprocessing: The spatial domain dataset was converted from the BOSSbase V1.01 dataset and the BOWS2 dataset into a 256×256 pixel PGM image set of 20,000 images (10,000 from the BOSSbase V1.01 dataset and 10,000 from the BOWS2 dataset) using the Resize method. The PGM image set used the S-UNIWARD, HILL, and WOW steganography algorithms, with each algorithm using 5 embedding rates: 0.1 bpp, 0.2 bpp, 0.3 bpp, 0.4 bpp, and 0.5 bpp. Therefore, the spatial domain dataset contains a total of 320,000 images.

[0043] (3) JPEG Domain Dataset Preprocessing: The JPEG domain dataset converts the BOSSbase V1.01 and BOWS2 datasets into 256×256 pixel JPEG image sets using the Resize method. Two image sets with quality factors of QF=75 and QF=95 are generated for each dataset. The JPEG image sets are processed using the J-UNIWARD and UED-JC steganography algorithms. Each steganography algorithm uses five steganography embedding rates: 0.1 bpnzac, 0.2 bpnzac, 0.3 bpnzac, 0.4 bpnzac, and 0.5 bpnzac. Therefore, the JPEG domain dataset contains a total of 440,000 images.

[0044] (4) Dataset Grouping: The entire dataset contains 760,000 images, divided into 76 subsets, numbered 1-76, with 10,000 images in each subset. BOSSbase and BOWS2 each consist of 38 subsets. In each subset of BOSSbase V1.01, 4,000 images are used for training, 1,000 for validation, and 5,000 for testing. All subsets of BOWS2 are used for training. In the subsets of BOSSbase V1.01, the training set accounts for 40%, the validation set for 10%, and the test set for 50%. In the subsets of BOWS2, the training set accounts for 100%.

[0045] Step 2, multi-scale receptive field feature extraction, such as Figure 2As shown, the first part of the MRNet network is used for natural information suppression and initial feature extraction. Multiple receptive fields of different scales are selected within the same layer to generate mixed feature channel groups. Using convolutional kernels of different scales, a multi-scale receptive field network based on dual convolutional layers is constructed. Steganographic association features are extracted from various regions of the preprocessed image data from step 1 using different receptive ranges. Using different receptive fields to extract image features of different ranges can adapt to more image texture characteristics and generate image feature channel groups of different ranges. This part constructs three convolutional channel groups A1, A2, and A3 based on dual convolutional layers. The specific construction steps are as follows:

[0046] (1) Selection of kernel sizes for three parallel channel groups: As the first part of MRNet, the kernel sizes of the first convolutional layers of the three channel groups A1, A2, and A3 are 3×3, 5×5, and 7×7, respectively. Due to channel limitations, the three channel groups in the first part of MRNet were studied independently. Using the controlled variable method, the detection accuracy was highest when the number of network convolutional kernels was 64. Therefore, the network ultimately selected 64 kernels as the first-level convolutional kernels of convolutional channel group A1. Similarly, the dimensions of convolutional channel groups A2 and A3 were 16 and 8, respectively. After testing, it can be found that there is a certain upper limit to increasing the number of channels while keeping the main body of the model unchanged. The upper limit of channel capacity is different at different scales.

[0047] (2) Ablation of three parallel channel groups: Ablation test is performed on the combination of the three channel groups in step (1) above. When the number of channels is the same, the mixed use of convolutional channel groups A1, A2 and A3 helps to improve the detection effect. Among them, convolutional channel group A1 plays a leading role in steganalysis, convolutional channel group A2 and convolutional channel group A3 help to improve the accuracy, while further increasing the number of channel groups (such as 9×9 convolutional kernels) does almost no help to improve the accuracy of steganalysis.

[0048] (3) Calculate the number of layers in the three-channel group. The number of layers in the three-channel group A1, A2 and A3 was studied and tested. When the channel group passes through only 1 convolutional layer, the detection accuracy is poor. When the channel group passes through 2 convolutional layers, the detection accuracy is relatively ideal. When it passes through 3 or more convolutional layers, the accuracy begins to decrease.

[0049] (4) Construct convolutional channel groups A1, A2 and A3. Among them, there are no pooling layers in the three channel groups, because pooling layers reduce the natural characteristics of the image itself and also greatly weaken the stegographic signal. Convolutional channel group A1 sequentially passes through the first convolutional layer (3×3 kernel size, 64 kernels), a batch normalization (BN) layer, a ReLU activation function, the second convolutional layer (3×3 kernel size, 16 kernels), a BN layer, and a ReLU activation function, then outputs the result of convolutional channel group A1. Convolutional channel group A2 sequentially passes through the first convolutional layer (5×5 kernel size, 16 kernels), a BN layer, a ReLU activation function, the second convolutional layer (3×3 kernel size, 4 kernels), a BN layer, and a ReLU activation function, then outputs the result of convolutional channel group A2. Convolutional channel group A3 passes through the first convolutional layer (7×7 kernel size, 8 kernels), a BN layer, a ReLU activation function, the second convolutional layer (3×3 kernel size, 2 kernels), a BN layer, and a ReLU activation function, then outputs the result of convolutional channel group A3. Finally, the results of the three channel groups are concatenated, resulting in a total of 22 feature channels, which serve as the final output feature channels of the first part of MRNet.

[0050] Step 3, as follows Figure 2 As shown, multi-channel group feature fusion, as the second part of the MRNet network, is used to suppress the extraction of natural and steganalytic features. Combining dual-channel group convolutional network and feature map group fusion, two parallel convolutional channel groups B1 and B2 are constructed to deepen feature extraction and fusion of the initial features extracted in step 2. The specific construction steps are as follows:

[0051] (1) Select the kernel size of the dual-channel group. The kernel sizes of the convolutional channel group B1 and the convolutional channel group B2 are 3×3 and 5×5, respectively. After testing, it was found that adding a 7×7 channel group did not significantly improve the detection accuracy.

[0052] (2) Dual-channel group structure test. When the first part of MRNet uses a single-channel group, the second part can also improve the detection accuracy. The number of residual layers passed through the convolutional channel group B1 and B2 and the detection results show that the convolutional channel group B2 needs to pass through fewer residual layers than the convolutional channel group B1.

[0053] (3) Calculate the depth of the dual-channel groups. This step uses the S-UNIWARD and J-UNIWARD algorithms to train and test the network, and improves the network structure based on the results. Here, S-UNIWARD is a spatial domain algorithm, and J-UNIWARD is a JPEG domain algorithm, which can enhance the network's adaptability in different domains. The HILL, WOW, and UED-JC algorithms are used for MRNet structure detection. In the spatial domain, the steganography embedding rate of S-UNIWARD, HILL, and WOW is 0.4 bpp. In the JPEG domain, J-UNIWARD and UED-JC have a QF=75 and a steganography embedding rate of 0.4 bpnzac. Experimental results show that too many or too few channel group layers are not good. Too many feature channel groups will lead to inter-group feature interference, while too few feature groups will reduce the effect of feature extraction in the next layer.

[0054] (4) Construct the S1 structure residual layer. The S1 structure residual layer consists of a first convolutional layer, a BN layer and a ReLU activation function, a second convolutional layer, a BN layer, and finally summed with the mapping branch information of the initial input. Each convolutional layer uses a 3×3 convolutional kernel.

[0055] (5) Construct the S2 structure residual layer. The S2 structure residual layer consists of a first convolutional layer, a BN layer and a ReLU activation function, a second convolutional layer, a BN layer, and finally summed with the mapping branch information of the initial input. Each convolutional layer uses a 5×5 convolutional kernel.

[0056] (6) Construct the second part of MRNet, namely convolutional channel group B1 and convolutional channel group B2. Convolutional channel group B1 goes through 5 S1 structure residual layers, and the convolutional layers in each residual layer use 3×3 convolutional kernels, with a channel dimension of 22 feature channels; convolutional channel group B2 goes through 3 S2 structure residual layers, and the convolutional layers in each residual layer use 5×5 convolutional kernels, with a channel dimension of 22 feature channels.

[0057] (7) Convolutional channel group B1 and convolutional channel group B2 are connected in a dual-channel group. The results of the two channel groups are combined to obtain a total of 44 feature channels, which are used as the final output feature channels of the second part of MRNet. Since all convolutional layers in the first two parts use the "SAME" mode, the feature map size remains unchanged at 256×256.

[0058] Step 4, Feature channel group dimensionality reduction: such as Figure 2 As shown, as the third part of the MRNet network, it is used to perform dimensionality reduction processing on the steganalytic features output in step 3; by constructing the dimensionality reduction feature channel group C1, the dimensionality reduction image steganalytic features are extracted.

[0059] The dimensionality reduction feature channel group C1 uses a single channel. The main branch of the S3 structure residual layer sequentially includes a first convolutional layer, a BN layer, a ReLU activation function layer, a second convolutional layer, and another BN layer, followed by an average pooling layer with a stride of 2. Due to the characteristics of image steganography, average pooling preserves steganalytic features better than max pooling. All convolutional layers use 3×3 convolutional kernels. The shortcut branch of the S3 structure residual layer has a 1×1 convolutional layer with a stride of 2, which is used to match the channel size of the input feature mapping with the output of the main branch. The output of the S3 structure residual layer is the result of element-wise summation of the output of the main branch after average pooling and the output of the shortcut branch after 1×1 convolution.

[0060] The dimensionality reduction feature channel group C1 passes through four S3 structure residual layers. Each time it passes through an S3 structure residual layer, the feature map size is halved. Therefore, after four S3 structure residual layers, the feature map group sizes are 128×128, 64×64, 32×32, and 16×16, respectively, and the feature channel depths are 64, 128, 256, and 512, respectively.

[0061] Step 5, as follows Figure 2 As shown, feature classification and result output, as the fourth part of the MRNet network, is used to classify and finally determine the steganatical features of the image after dimensionality reduction output in step 4. By constructing feature classification channel group D1 and setting a threshold, the output results are classified into two categories: result 1 is true and result 0 is false, and the final image steganalysis detection result is output.

[0062] The network layers of feature classification channel group D1 are as follows: first convolutional layer, BN layer, ReLU activation function layer, second convolutional layer, BN layer, followed by global average pooling layer, fully connected layer and softmax layer. Among them, the global average pooling layer converts 512 feature maps into 512-dimensional feature vectors. After the feature vectors are processed by the fully connected layer, they are classified by the softmax layer to output the final steganalysis detection result.

[0063] Step 6: Configure multi-scale residual network parameters for training and testing the MRNet network model. This involves setting parameters such as input image specifications, convolutional layer padding mode, activation function type, convolutional kernel initialization method, training steps, learning rate, and optimizer type to enable model training and testing. The MRNet network model training and testing parameter configurations are as follows:

[0064] The multi-scale residual network model takes a 256×256 grayscale image as input. All convolutional and pooling layers in the network use a "SAME" padding mode, which keeps the feature map size of each channel constant or halved. No pooling layers are used in the first and second parts of the model. Experiments show that adding pooling layers in the first two parts reduces detection accuracy. Pooling layers reduce the size of the feature maps and also weaken the steganalytic signal strength. In the second and third parts of the model, the deepest channel group has 18 convolutional layers. The network uses a shortcut structure to effectively prevent gradient vanishing. The model uses the ReLU activation function. Experiments show that the detection accuracy using ReLU is no lower than tanh and sigmoid, while ReLU is computationally faster.

[0065] For kernel initialization, constant, Gaussian, uniform, bilinear, and replace methods were tried in different channels. Gaussian yielded better accuracy, so the network used Gaussian to initialize the kernels for all channel groups. In experimental testing, the Gaussian function was used to initialize the weights with a mean of 0 and a standard deviation of 0.01. Training was performed for 500,000 steps each time, with the learning rate set to three parts: 0.001, 0.0001, and 0.00001. The corresponding boundaries were set to [300,000, 400,000]. The batch size was set to 16. The optimizer used was AdmaxOptimizer.

[0066] This invention presents an image steganalysis method based on multi-scale residual networks. It innovatively employs multi-scale receptive fields for image feature extraction, and for the first time attempts to construct a multi-scale receptive field network. By utilizing receptive ranges of different sizes, it collaboratively extracts steganalytic features from various regions of the image, comprehensively covering feature information from local to broader areas. This effectively reduces the risk of losing weak feature signals when steganalytic image information is input into the neural network, providing a more complete initial feature foundation for subsequent feature processing. Simultaneously, for features extracted by the multi-scale receptive field, the method introduces dual-channel groups for in-depth feature extraction and fusion. By setting parallel convolutional channel groups between layers of the neural network, a hybrid feature channel group mechanism is formed, allowing features of different dimensions to be transmitted independently and completely. This design specifically addresses the problem of weak steganalytic features being easily lost during transmission from lower to higher layers of the network, ensuring the continuity and recognizability of weak features.

[0067] The multi-channel residual neural network model finally constructed in this invention integrates the feature extraction capabilities of multi-scale channel groups with the feature fusion advantages of hybrid channel groups, significantly reducing the loss of steganalytic features throughout the feature extraction process. This greatly enhances the model's ability to capture steganalytic signals, thereby effectively improving the accuracy of image steganalysis.

[0068] Based on the above embodiments, the present invention continues to describe in detail the technical features involved therein and the functions and roles of these technical features in the present invention, so as to help those skilled in the art to fully understand the technical solution of the present invention and reproduce it.

[0069] Finally, although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-scale residual network based image steganalysis method, characterized in that, Based on steganalysis residual network (SRNet), a multi-scale receptive field network model (MRNet) is constructed by introducing a same-layer multi-scale receptive field mechanism and a same-layer mixed feature channel group mechanism. The MRNet network selects multiple different scales of receptive fields in the same layer, generates a same-layer mixed feature channel group, and performs interval mixing identification. The different feature scale channel groups are fused by optimizing the MRNet, so as to realize the extraction and detection of image steganographic features by using multi-range receptive fields. The different scale receptive field mechanism constructs a multi-scale receptive field network based on double convolution layers, and uses different size perception ranges to extract steganographic associated features of each region of the image. The same-layer mixed feature channel group mechanism introduces parallel channel groups between each layer of the neural network to process the transmission process of steganographic weak features from the low layer to the high layer of the network. The specific steps include the following: Step 1, preparing a data set and data preprocessing; Step 2, multi-scale receptive field feature extraction, for image natural information suppression and initial feature extraction; a multi-scale receptive field network based on double convolution layers is constructed by selecting multiple different scales of receptive fields in the same layer, and different size convolution kernels are used to extract steganographic associated features of each region of the data image after preprocessing in step 1; The multi-scale receptive field feature extraction, as the first part of the MRNet network, generates a same-layer mixed feature channel group by selecting multiple different scales of receptive fields in the same layer, which specifically includes: constructing a different scale receptive field network based on double convolution layers, including convolution channel group A1, convolution channel group A2 and convolution channel group A3, and using different size perception ranges to extract steganographic associated features of each region of the data image after preprocessing in step 1; A multi-scale receptive field network based on double convolution layers is constructed by using different scale convolution kernels, which specifically includes: the first convolution layer of the convolution channel group A1 uses a 3×3 convolution kernel, and the second convolution layer uses a 3×3 convolution kernel; The first convolution layer of the convolution channel group A2 uses a 5×5 convolution kernel, and the second convolution layer uses a 3×3 convolution kernel; The first convolution layer of the convolution channel group A3 uses a 7×7 convolution kernel, and the second convolution layer uses a 3×3 convolution kernel; The output results of the above three parallel convolution channel groups A1, A2 and A3 are connected and combined to form 22 output feature channels as the output initial features of the first part of the MRNet network; Step 3, multi-channel group feature fusion, for suppressing natural features and steganographic feature extraction; the initial features extracted by the same-layer mixed feature channel group in step 2 are deepened and fused by combining a double-channel group convolution network and a feature map group fusion; The multi-channel group feature fusion as the second part of the MRNet network combines a double-channel group convolutional network and a feature map group fusion, and is used for suppressing the extraction of natural features and steganographic features, and specifically includes: constructing two parallel convolutional channel groups B1 and B2, and deepening feature extraction and fusion on the initial features extracted from the mixed feature channel groups in the same layer in step 2; the convolutional channel group B1 is sequentially processed by five S1 structure residual layers, and outputs 22 steganographic features of the feature channels; the convolutional channel group B2 is sequentially processed by three S2 structure residual layers, and outputs 22 steganographic features of the feature channels; The S1 structure residual layer in the convolutional channel group B1 includes: a first convolutional layer, a BN layer, a ReLU activation function, a second convolutional layer, and a BN layer connected in sequence, and the output is the sum of the output of the second convolutional layer after the BN layer and the initial input mapping branch information of the S1 structure residual layer; each convolutional layer in the S1 structure residual layer uses a 3*3 convolutional kernel; The S2 structure residual layer in the convolutional channel group B2 includes: a first convolutional layer, a BN layer, a ReLU activation function, a second convolutional layer, and a BN layer connected in sequence, and the output is the sum of the output of the second convolutional layer after the BN layer and the initial input mapping branch information of the S2 structure residual layer; each convolutional layer in the S2 structure residual layer uses a 5*5 convolutional kernel; The feature channels output by the convolutional channel group B1 and the feature channels output by the convolutional channel group B2 are fused through the channel dimension to form a total of 44 fused features of the feature channels, which are used as the steganographic feature output channels of the second part of the MRNet network. Step 4, feature channel group dimension reduction; for reducing the steganographic features output in step 3, a dimension reduction feature channel group is constructed to extract the output image steganographic features after dimension reduction. Step 5, feature classification and result output; for classifying and finally determining the dimension-reduced image steganographic features extracted and output in step 4, a feature classification channel group is constructed to finally output the image steganographic detection result.

2. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that, In step 1, the data set and data preprocessing are prepared, which specifically includes: Prepare the basic image data set for training and testing, generate sub-data sets for model training and testing based on the basic image data set, and divide the sub-data sets into a training set, a validation set and a test set according to a preset ratio.

3. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that, In step 4, the feature channel group dimension reduction is used as the third part of the MRNet network, and a dimension reduction feature channel group C1 is constructed to reduce the steganographic features output in step 3, specifically: the dimension reduction feature channel group C1 uses a single channel, is sequentially processed by four S3 structure residual layers, and the feature map size is halved after each processing, and outputs the image steganographic features after dimension reduction; The main branch of the S3 structure residual layer comprises a first convolutional layer, a BN layer, a ReLU activation function layer, a second convolutional layer, a BN layer, and an average pooling layer with a step of 2 connected thereto, wherein the convolutional layers all adopt a 3*3 convolutional kernel; the shortcut branch of the S3 structure residual layer is provided with a 1*1 convolutional layer with a step of 2, which is used to match the channel size of the input feature mapping with the output of the main branch; The output of the S3 structure residual layer is the result of element-wise summation of the output of the main branch after average pooling and the output of the shortcut branch after 1*1 convolution.

4. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that, In step 5, feature classification and result output are taken as the fourth part of the MRNet network, a feature classification channel group D1 is constructed to classify and finally determine the stego feature of the reduced image output in step 4, and specifically includes: The network levels of the D1 channel group include, in sequence, a first convolutional layer, a BN layer, a ReLU activation function layer, a second convolutional layer, a BN layer, and then a global average pooling layer, a fully connected layer and a softmax layer connected in sequence; wherein the global average pooling layer converts 512 feature maps into a 512-dimensional feature vector, the feature vector is processed by the fully connected layer, and then the softmax layer is used for binary classification to output the final image steganography detection result.

5. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that, Step 6, multi-scale residual network parameter configuration, is further included for training and testing the MRNet network model, and each parameter of the input image specification, the padding mode of each convolutional layer, the activation function type, the convolutional kernel initialization method, the training step number, the learning rate and the optimizer type of the MRNet network model is set to realize the training and testing of the model.

Images