Image steganalysis method based on multi-scale residual network

Abstract

The invention discloses an image steganalysis method based on a multi-scale residual network, relates to the technical field of image steganalysis, and solves the problems that in existing image steganalysis, single-channel group extraction features are difficult to adapt to natural features and steganalysis feature distribution differences of different images, weak feature signals are easy to lose, and the image steganalysis accuracy is low. And the detection effect is limited. According to the method, a multi-scale receptive field network model is constructed by introducing a same-layer multiple different-scale receptive field mechanism and a same-layer mixed feature channel group mechanism, the multi-scale receptive field network model generates a same-layer mixed feature channel group by selecting multiple different-scale receptive fields on the same layer, and the multi-scale receptive field network model is subjected to interval mixed identification, so that the multi-scale receptive field network model is obtained. Different feature scale channel groups are fused, and MRNet is optimized, so that multi-range receptive field collaborative extraction and detection of steganography features are realized; the problem of feature loss of weak feature signals and weak features in the process of transmitting the weak features from a low layer to a high layer when steganography image information is input into a neural network is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image steganalysis, and in particular to an image steganalysis method based on a multi-scale residual network. Background Art

[0002] With the continuous development of steganography technology, more and more steganographic tools have appeared online, which has also brought convenience to criminals. A large amount of image data exists online, and a considerable amount of steganographic information is also present. 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 perform image steganalysis.

[0003] Steganalysis generally consists of two steps: stegographic feature extraction and classification. Early steganalysis relied heavily on manual feature selection, followed by classification using classifiers like support vector machines. The effectiveness of steganalysis hinged on the quality of feature selection. As steganography methods have become increasingly diverse and complex, manual feature selection has become increasingly difficult. Steganalysis algorithms rely on manual feature selection, requiring significant experience and manpower. Therefore, it became necessary to find an effective method for faster and more convenient feature selection.

[0004] Compared to classic image steganalysis algorithms, deep learning has been applied to image steganalysis in recent years due to its reduced reliance on human experience and wide adaptability. Deep neural network models combine preprocessing, feature extraction, and classification, achieving integrated image steganalysis. However, in image steganalysis, current deep neural networks use a single channel group for feature extraction. Increasing the number of convolution kernels within the network model can improve detection accuracy to a certain extent, but further increases in kernels have no significant effect after reaching a certain limit. Analyzing natural image information reveals that different image types exhibit varying natural characteristics, such as texture area and density, leading to varying distributions of underlying stegographic features. Using a single receptive field size for all images makes it difficult to accurately and comprehensively extract information from the images, and this can easily lead to the loss of weak feature signals during stegographic feature extraction.

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

[0006] In response to the shortcomings and deficiencies in the prior art, the present invention provides an image steganalysis method based on a multi-scale residual network, which constructs a multi-scale receptive field network model by introducing a receptive field mechanism of multiple scales at the same layer and a mixed feature channel group mechanism at the same layer. The method utilizes multi-range receptive fields for collaborative extraction and fuses channel groups of different feature scales to solve the problem of loss of weak steganographic features during network transmission, thereby achieving efficient and accurate analysis of image steganography.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: The image steganalysis method based on multi-scale residual network provided by the present invention is based on the steganalysis residual network (SRNet). By introducing the receptive field mechanism of multiple scales in the same layer and the mixed feature channel group mechanism in the same layer, a multi-scale receptive field network model (MRNet) is constructed. The specific steps include the following: Step 1: Prepare data set and preprocess data; Step 2: Multi-scale receptive field feature extraction, used for natural information suppression and initial feature extraction of the image. By selecting receptive fields of multiple scales in the same layer, a multi-scale receptive field network based on a double convolutional layer is constructed. The receptive range of convolution kernels of different scales is used to extract steganographic related features from various regions of the data image preprocessed in step 1. Step 3: Multi-channel group feature fusion is used to suppress the extraction of natural features and steganographic features. Combining the 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 subjected to in-depth feature extraction and fusion. Step 4, feature channel group dimensionality reduction; used to reduce the dimensionality of the steganographic features output in step 3 by constructing a dimensionality reduction feature channel group, and output the image steganographic features after dimensionality reduction; Step 5, feature classification and result output; used to classify and make final judgments on the dimensionality-reduced image steganographic features extracted and output in step 4, and finally output the image steganographic detection results by constructing a feature classification channel group.

[0008] Preferably, the MRNet network selects receptive fields of multiple scales at the same layer to generate a mixed feature channel group at the same layer, and fuses the channel groups of different feature scales through interval mixed recognition. By optimizing the MRNet, it is possible to realize the joint extraction and detection of image steganographic features using multi-range receptive fields. The receptive field mechanism of different scales constructs a multi-scale receptive field network based on double convolutional layers, and uses receptive ranges of different sizes to extract steganographic related features from various areas of the image. The same-layer mixed feature channel group mechanism introduces parallel channel groups between each layer of the neural network to handle the process of transferring weak steganographic features from the low layer to the high layer of the network.

[0009] Preferably, in step 1, preparing the data set and data preprocessing specifically include: Prepare a basic image dataset for training and testing, generate sub-datasets for model training and testing based on the basic image dataset, and divide the sub-datasets into training set, validation set and test set according to preset proportions.

[0010] Preferably, in step 2, multi-scale receptive field feature extraction, as the first part of the MRNet network, selects a plurality of receptive fields of different scales in the same layer to generate a mixed feature channel group in the same layer, specifically including: constructing a different-scale receptive field network based on a double convolutional layer, including convolution channel group A1, convolution channel group A2 and convolution channel group A3, and using different sizes of receptive ranges to extract steganographic correlation features from each area of ​​the data image preprocessed in step 1.

[0011] Preferably, using convolution kernels of different scales to construct a multi-scale receptive field network based on a double convolution layer is as follows: the first convolution layer of the convolution channel group A1 adopts a 3×3 convolution kernel, and the second convolution layer adopts a 3×3 convolution kernel; The first convolution layer of 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 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 three parallel convolution channel groups A1, A2, and A3 are connected to form 22 output feature channels, which serve as the output initial features of the first part of the MRNet network.

[0012] Preferably, in step 3, multi-channel group feature fusion is used as the second part of the MRNet network, combined with the dual-channel group convolutional network and the feature map group fusion, to suppress the extraction of natural features and stego features, specifically including: constructing two parallel convolution channel groups B1 and convolution channel group B2, and deepening feature extraction and fusion of the initial features extracted by the same-layer mixed feature channel group in step 2; the convolution channel group B1 is sequentially processed by the S1 structure residual layer 5 times to output the stego features of 22 feature channels; the convolution channel group B2 is sequentially processed by the S2 structure residual layer 3 times to output the stego features of 22 feature channels; The feature channels output by convolution channel group B1 and the feature channels output by convolution channel group B2 are fused through the channel dimension to form a fusion feature of a total of 44 feature channels, which serves as the steganalysis feature output channel of the second part of the MRNet network.

[0013] Preferably, the S1 structure residual layer in the convolution channel group B1 includes: a first convolution layer, a BN layer, a ReLU activation function, a second convolution layer, and a BN layer connected in sequence, and the output is the sum of the output of the BN layer after the second convolution layer and the initial input mapping branch information of the S1 structure residual layer; each convolution layer in the S1 structure residual layer uses a 3×3 convolution kernel; The S2 structure residual layer in the convolution channel group B2 includes: the first convolution layer, BN layer, ReLU activation function, second convolution layer, and BN layer connected in sequence. The output is the sum of the output of the BN layer after the second convolution layer and the initial input mapping branch information of the S2 structure residual layer; each convolution layer in the S2 structure residual layer uses a 5×5 convolution kernel.

[0014] Preferably, in step 4, the feature channel group dimensionality reduction is used as the third part of the MRNet network. By constructing a dimensionality reduction feature channel group C1, the steganographic features output in step 3 are subjected to dimensionality reduction processing. Specifically, the dimensionality reduction feature channel group C1 adopts a single channel and is sequentially processed by four S3 structure residual layers. After each processing, the size of the feature map is halved, and the image steganographic features after dimensionality reduction are output; The main branch of the S3 residual layer sequentially includes the first convolutional layer, the BN layer, the ReLU activation function layer, the second convolutional layer, the BN layer, followed by an average pooling layer with a stride of 2. The convolutional layers all use 3×3 convolution kernels. The shortcut branch of the S3 residual layer sets a 1×1 convolutional layer with a stride of 2 to match the channel size of the input feature map with the output of the main branch. The output of the residual layer of the S3 structure is the element-by-element summation of the output of the main branch after average pooling and the output of the shortcut branch after 1×1 convolution.

[0015] Preferably, in step 5, feature classification and result output is the fourth part of the MRNet network, which is used to classify and make final judgments on the steganographic features of the image after dimensionality reduction output in step 4 by constructing a feature classification channel group D1, specifically including: The network layers of the D1 channel group include: the first convolutional layer, the BN layer, the ReLU activation function layer, the second convolutional layer, the BN layer, followed by the global average pooling layer, the fully connected layer and the softmax layer in sequence; among them, the global average pooling layer converts 512 feature maps into 512-dimensional feature vectors, which are processed by the fully connected layer and then binary classified by the softmax layer to output the final image steganalysis detection result.

[0016] Preferably, step 6 is also included, multi-scale residual network parameter configuration, which is used to train and test the MRNet network model, by setting the input image specifications, filling modes of each convolutional layer, activation function type, convolution kernel initialization method, training steps, learning rate and optimizer type of the MRNet network model to achieve model training and testing.

[0017] The image steganalysis method based on multi-scale residual network provided by the present invention has the following beneficial effects: The present invention's image steganalysis method based on a multi-scale residual network constructs a multi-scale residual network model by introducing a same-layer receptive field mechanism with multiple scales and a same-layer mixed feature channel group mechanism, achieving high-precision detection of steganographic images. This same-layer receptive field mechanism leverages the synergistic effects of the different scales of perception to extract steganalytic features from various image regions. Receptive fields of varying scales can cover information from different regions within the image, specifically capturing differently distributed steganographic features. This reduces the loss of weak feature signals when inputting steganographic image information into the neural network, providing a more complete initial feature foundation for subsequent feature processing.

[0018] At the same time, for features extracted from multi-scale receptive fields, the initial features are further extracted and fused through a parallel-designed convolutional dual-channel group. Different channel groups use residual layers with specific structures to process features, forming a full-process optimization architecture from feature extraction and deepening to classification. The same-layer mixed feature channel group mechanism can effectively preserve the transmission of weak steganographic features from low to high layers between network layers, avoiding feature loss caused by information compression or conversion during feature transmission in a single channel group. While achieving efficient detection of steganographic images, it also reduces the loss of steganographic features throughout the extraction, transmission, and fusion processes, solving the problem of feature loss during the transmission of weak steganographic features from low to high layers of the network, further improving the accuracy of image steganalysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of the image steganalysis method based on multi-scale residual network in the present invention; Figure 2 This is a structural framework diagram of the multi-scale residual network model in the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0021] like Figure 1As shown, the image steganalysis method based on a multi-scale residual network of the present invention is based on the steganalysis residual network (SRNet). By introducing a receptive field mechanism of multiple scales at the same layer and a mixed feature channel group mechanism at the same layer, a multi-scale residual network model (MRNet) is constructed. The MRNet network selects receptive fields of multiple scales at the same layer to generate a mixed feature channel group at the same layer. After interval mixed recognition, the channel groups of different feature scales are fused. By optimizing the MRNet, image steganalysis feature extraction and detection are achieved using multi-range receptive fields. The receptive field mechanism of different scales constructs a multi-scale receptive field network based on double convolutional layers, and uses receptive ranges of different sizes to extract steganalysis-related features from various regions of the image. The mixed feature channel group mechanism at the same layer introduces parallel channel groups between the layers of the neural network to process the process of transmitting weak steganalysis features from the lower layers to the higher layers of the network.

[0022] The specific steps are as follows: Step 1: Prepare the dataset and preprocess the data. Using the BOSSbase V1.01 dataset and the BOWS2 dataset as the basic image dataset, 76 sets of training and test datasets were prepared, totaling 760,000 images. Corresponding sub-datasets for model training and testing were generated, and the sub-datasets were divided into training set, validation set, and test set according to the preset ratio. The specific steps include: (1) Experimental dataset preparation: The experimental dataset is generated from two standard datasets. One of the experimental data is based on the BOSSbase V1.01 dataset, which contains 10,000 grayscale images of 512 × 512 pixels. The BOSSbase V1.01 dataset is used to generate the training dataset, validation dataset, and test dataset. The other dataset is based on BOWS2. This dataset contains 10,000 grayscale images of 512 × 512 pixels. The BOWS2 dataset is mainly used to generate the training dataset. The experimental dataset is mainly divided into two parts: the spatial domain dataset and the JPEG domain dataset.

[0023] (2) Spatial Dataset Data Preprocessing: The spatial dataset uses the Resize method to convert the BOSSbase V1.01 and BOWS2 datasets into a set of 20,000 PGM images of 256 × 256 pixels (10,000 from the BOSSbase V1.01 dataset and 10,000 from the BOWS2 dataset). The PGM image set uses the S-UNIWARD, HILL, and WOW steganographic algorithms. Each steganographic algorithm uses five embedding rates: 0.1 bpp, 0.2 bpp, 0.3 bpp, 0.4 bpp, and 0.5 bpp. Therefore, the spatial dataset contains a total of 320,000 images.

[0024] (3) Data preprocessing of the JPEG domain dataset: The JPEG domain dataset converts the BOSSbase V1.01 dataset and the BOWS2 dataset into 256×256 pixel JPEG image sets using the Resize method. Two image sets with quality factors QF=75 and QF=95 are generated for each dataset. The JPEG image sets are operated using the J-UNIWARD and UED-JC steganographic algorithms. Each steganographic algorithm uses five steganalysis 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.

[0025] (4) Dataset grouping number: The entire dataset has 760,000 images and consists of 76 sub-datasets, numbered 1-76, with 10,000 images in each sub-dataset. BOSSbase and BOWS2 each consist of 38 subsets. In each sub-dataset of BOSSbase V1.01, 4,000 images are used for training, 1,000 images are used for validation, and 5,000 images are used for testing. All sub-datasets of BOWS2 are used for training. In the sub-datasets of BOSSbase V1.01, the training set accounts for 40%, the validation set accounts for 10%, and the test set accounts for 50%. In the sub-datasets of BOWS2, the training set accounts for 100%.

[0026] Step 2, multi-scale receptive field feature extraction, such as Figure 2 As shown in the figure, as the first part of the MRNet network, it is used for natural information suppression and initial feature extraction of images; by selecting receptive fields of multiple different scales in the same layer, a mixed feature channel group of the same layer is generated, and convolution kernels of different scales are used to construct a multi-scale receptive field network based on a double convolution layer. The receptive ranges of different sizes are used to extract steganographic related features from various areas of the data image preprocessed in step 1. Using different receptive fields to extract image features of different ranges can adapt to the texture characteristics of more images and generate image feature channel groups of different ranges. This part constructs three convolution channel groups A1, convolution channel group A2, and convolution channel group A3 based on double convolution layers. The specific construction steps are as follows: (1) Select the convolution kernel sizes of the three parallel channel groups: As the first part of MRNet, the convolution kernel sizes of the first convolution layer of the three channel groups of convolution channel group A1, convolution channel group A2, and convolution channel group A3 are 3×3, 5×5, and 7×7, respectively. Due to channel limitations, the three channel groups of the first part of MRNet were independently studied. Using the control variable method, when the number of network convolution kernels was 64, the detection accuracy was the highest. Therefore, the network finally selected 64 kernels as the first-level convolution kernel of convolution channel group A1. Similarly, the dimensions of convolution channel groups A2 and A3 were 16 and 8, respectively. After testing, it was found that there is a certain upper limit to increasing the number of channels while keeping the main body of the model unchanged, and the upper limit of channel capacity is different at different scales.

[0027] (2) Ablation of three parallel channel groups: The combination of the three channel groups in the above step (1) is subjected to an ablation test. When the number of channels is the same, the mixed use of convolution channel groups A1, A2 and A3 helps to improve the detection effect. Among them, convolution channel group A1 plays a leading role in steganalysis, and convolution channel group A2 and convolution channel group A3 have certain help in improving the accuracy. However, continuing to increase the channel group (such as 9×9 convolution kernel) has almost no help in improving the accuracy of steganalysis.

[0028] (3) Calculate the number of layers for the three-channel group. The number of layers for the three-channel groups A1, A2, and A3 was studied and experimentally tested. When the channel group passed through only one convolutional layer, the detection accuracy was poor. When the channel group passed through two convolutional layers, the detection accuracy was ideal. When the channel group passed through three or more convolutional layers, the accuracy began to decline.

[0029] (4) Construct convolution channel groups A1, A2, and A3. There is no pooling layer in the three constructed channel groups because the pooling layer not only reduces the natural characteristics of the image itself, but also significantly weakens the steganographic signal. Convolutional channel group A1 passes through the first convolutional layer (3×3 kernel size, 64 kernels), a batch normalization layer, a ReLU activation function, a second convolutional layer (3×3 kernel size, 16 kernels), a batch normalization layer, and a ReLU activation function, and then outputs the convolutional channel group A1. Convolutional channel group A2 passes through the first convolutional layer (5×5 kernel size, 16 kernels), a batch normalization layer, a ReLU activation function, a second convolutional layer (3×3 kernel size, 4 kernels), a batch normalization layer, and a ReLU activation function, and then outputs the convolutional channel group A2. Convolutional channel group A3 passes through the first convolutional layer (7×7 kernel size, 8 kernels), a batch normalization layer, a ReLU activation function, a second convolutional layer (3×3 kernel size, 2 kernels), a batch normalization layer, a ReLU activation function, and then outputs the convolutional channel group A3. Finally, the results of the three channel groups are combined through channel group concatenation, totaling 22 feature channels, which serve as the final output feature channels of the first part of MRNet.

[0030] Step 3, such as Figure 2 As shown in the figure, multi-channel group feature fusion, as the second part of the MRNet network, is used to suppress the extraction of natural features and steganographic features; combined with the dual-channel group convolutional network and feature map group fusion, by constructing two parallel convolution channel groups B1 and B2, the initial features extracted in step 2 are deepened in feature extraction and fusion. The specific construction steps are as follows: (1) Select the convolution kernel size of the dual-channel group. The convolution kernel sizes of convolution channel group B1 and convolution channel group B2 are 3×3 and 5×5 respectively. After experimental testing, adding a 7×7 channel group does not significantly improve the detection accuracy.

[0031] (2) Dual-channel group structure test. When the first part of MRNet uses a single channel group, the second part can also improve detection accuracy. The number of residual layers passed through the convolutional channel groups B1 and B2 and the test results show that the convolutional channel group B2 requires fewer residual layers than the convolutional channel group B1.

[0032] (3) Calculate the depth of the dual-channel group. This step uses the S-UNIWARD algorithm and the J-UNIWARD algorithm to train and test the network, and improve the network structure based on the results. Here, S-UNIWARD is a spatial domain algorithm and J-UNIWARD is a JPEG domain algorithm. They can enhance the adaptability of the network in different fields. The HILL algorithm, WOW algorithm and UED-JC algorithm are used for MRNet structure detection. In the spatial domain, the steganographic 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 steganographic embedding rate of 0.4 bpnzac. The 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, and too few feature groups will reduce the effect of feature extraction in the next layer.

[0033] (4) Construct the S1 structure residual layer. The S1 structure residual layer consists of the first convolution layer, the BN layer and the ReLU activation function, the second convolution layer, the BN layer, and finally the summation of the mapping branch information with the initial input. Each convolution layer uses a 3×3 convolution kernel.

[0034] (5) Construct an S2 structure residual layer. The S2 structure residual layer consists of the first convolution layer, the BN layer and the ReLU activation function, the second convolution layer, the BN layer, and finally the summation of the mapping branch information with the initial input. Each convolution layer uses a 5×5 convolution kernel.

[0035] (6) Construct the convolution channel group B1 and convolution channel group B2 of the second part of MRNet. Among them, the convolution channel group B1 passes through the S1 structure residual layer 5 times, and the convolution layer in each residual layer uses a 3×3 convolution kernel, and the channel dimension is 22 feature channels; the convolution channel group B2 passes through the S2 structure residual layer 3 times, and the convolution layer in each residual layer uses a 5×5 convolution kernel, and the channel dimension is also 22 feature channels.

[0036] (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, totaling 44 feature channels, which serve 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 and is 256×256.

[0037] Step 4: Dimensionality reduction of feature channel groups: Figure 2 As shown in FIG, as the third part of the MRNet network, it is used to reduce the dimension of the steganographic features output in step 3; by constructing the dimensionality reduction feature channel group C1, the steganographic features of the image after the output dimensionality reduction are extracted.

[0038] The dimensionality reduction feature channel group C1 adopts a single channel. The main branch of the S3 structure residual layer sequentially includes the first convolution layer, the BN layer, the ReLU activation function layer, the second convolution layer, and the BN layer, followed by an average pooling layer with a stride of 2. Due to the characteristics of image steganography, average pooling can better preserve steganographic features than maximum pooling. Among them, the convolution layers all use 3×3 convolution kernels; the shortcut branch of the S3 structure residual layer sets a 1×1 convolution layer with a stride of 2 to match the channel size of the input feature map with the output of the main branch; the output of the S3 structure residual layer is the element-by-element summation of the output of the main branch after average pooling and the output of the shortcut branch after 1×1 convolution.

[0039] The dimensionality-reduced feature channel group C1 passes through four S3 residual layers, with the feature map size halved each time. Therefore, after four S3 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.

[0040] Step 5, such as Figure 2 As shown in the figure, feature classification and result output, as the fourth part of the MRNet network, is used to classify and make the final judgment on the image steganographic features after dimensionality reduction output in step 4. By constructing the feature classification channel group D1 and setting the threshold, the output result is classified into two categories, with result 1 being true and result 0 being false, and finally the image steganography detection result is output.

[0041] The network layers of the feature classification channel group D1 include: the first convolutional layer, the BN layer, the ReLU activation function layer, the second convolutional layer, the BN layer, followed by the global average pooling layer, the fully connected layer and the softmax layer in sequence; among them, the global average pooling layer converts 512 feature maps into 512-dimensional feature vectors, which are processed by the fully connected layer and then binary classified by the softmax layer to output the final steganalysis detection result.

[0042] Step 6, multi-scale residual network parameter configuration, is used to train and test the MRNet network model. By setting the input image specifications, filling mode of each convolution layer, activation function type, convolution kernel initialization method, number of training steps, learning rate and optimizer type of the MRNet network model, the training and testing of the model can be realized. The training and testing parameter configuration of the MRNet network model is as follows: The multi-scale residual network model takes a grayscale image with a size of 256×256 as input. All convolutional and pooling layers in the network use the "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 have shown that adding pooling layers to the first two parts reduces detection accuracy. While pooling layers reduce the size of feature maps, they also weaken the signal strength of steganographic features. 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 vanishing gradients. The model uses the ReLU activation function. Experimental testing has shown that the detection accuracy of the ReLU function is comparable to that of the tanh and sigmoid functions, while ReLU is computationally faster.

[0043] For convolution kernel initialization, we tried constant, Gaussian, uniform, bilinear, and replace methods across different channels. Gaussian achieved better accuracy, so the network used Gaussian initialization for convolution kernels across all channel groups. In experimental testing, the Gaussian function initialized weights with a mean of 0 and a standard deviation of 0.01. Training was conducted for 500,000 steps, with a learning rate set to three fractions: 0.001, 0.0001, and 0.00001. The corresponding bounds were set to [300,000, 400,000]. The batch size was set to 16. The optimizer used was AdmaxOptimizer.

[0044] The image steganalysis method based on a multi-scale residual network in the present invention innovatively uses a multi-scale receptive field for image feature extraction. It attempts to construct a multi-scale receptive field network for the first time, and uses perception ranges of different sizes to collaboratively extract steganographic related features from various regions of the image. It can comprehensively cover feature information from local to wider ranges, thereby effectively reducing the risk of losing weak feature signals when the steganographic image information is input into the neural network, and providing a more complete initial feature basis for subsequent feature processing. At the same time, for the features extracted by the multi-scale receptive field, the method introduces a dual-channel group for in-depth feature extraction and fusion. By setting parallel convolution channel groups between the layers of the neural network, a hybrid feature channel group mechanism is formed, which enables features of different dimensions to be transmitted independently and completely. This design specifically solves the problem that weak steganographic features are easily lost during the transmission process from the lower layer to the higher layer of the network, ensuring the coherence and recognizability of weak features.

[0045] The multi-channel group residual neural network model finally constructed by the present invention significantly reduces the loss of steganographic features in the entire process of feature extraction by integrating the feature extraction capabilities of the multi-scale channel group with the feature fusion advantages of the hybrid channel group, thereby greatly improving the model's ability to capture steganographic signals and effectively improving the accuracy of image steganalysis.

[0046] On the basis of the above embodiments, the present invention continues to describe in detail the technical features involved therein and the functions and roles played by the technical features in the present invention, so as to help technicians in this field fully understand the technical solutions of the present invention and reproduce them.

[0047] Finally, although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of 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 implementation methods that can be understood by those skilled in the art.

Claims

1. An image steganalysis method based on a multi-scale residual network, characterized in that: Based on the steganalysis residual network (SRNet), a multi-scale receptive field network model (MRNet) is constructed by introducing a receptive field mechanism of multiple scales at the same layer and a mixed feature channel group mechanism at the same layer. The specific steps include the following: Step 1: Prepare data set and preprocess data; Step 2: Multi-scale receptive field feature extraction, used for natural information suppression and initial feature extraction of the image. By selecting receptive fields of multiple scales in the same layer, a multi-scale receptive field network based on a double convolutional layer is constructed. The receptive range of convolution kernels of different sizes is used to extract steganographic related features from various regions of the data image preprocessed in step 1. Step 3: Multi-channel group feature fusion is used to suppress the extraction of natural features and steganographic features. Combining the 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 subjected to in-depth feature extraction and fusion. Step 4, feature channel group dimensionality reduction; used to reduce the dimensionality of the steganographic features output in step 3 by constructing a dimensionality reduction feature channel group, and extracting the steganographic features of the image after the dimensionality reduction; Step 5: feature classification and result output; It is used to classify and make final judgments on the dimensionality-reduced image steganographic features extracted and output in step 4, and finally output the image steganographic detection results by constructing a feature classification channel group.

2. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: The MRNet network selects receptive fields of multiple scales at the same layer to generate a mixed feature channel group at the same layer. After interval mixed recognition, the channel groups of different feature scales are fused. By optimizing MRNet, it is possible to extract and detect image steganographic features using multi-range receptive fields. The receptive field mechanism of different scales constructs a multi-scale receptive field network based on double convolutional layers, and uses receptive ranges of different sizes to extract steganographic related features from various areas of the image. The same-layer mixed feature channel group mechanism introduces parallel channel groups between each layer of the neural network to handle the process of transferring weak steganographic features from the low layer to the high layer of the network.

3. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: In step 1, data set preparation and data preprocessing specifically include: Prepare a basic image dataset for training and testing, generate sub-datasets for model training and testing based on the basic image dataset, and divide the sub-datasets into training set, validation set and test set according to preset proportions.

4. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: In step 2, multi-scale receptive field feature extraction, as the first part of the MRNet network, selects receptive fields of multiple scales in the same layer to generate a mixed feature channel group in the same layer. Specifically, it includes: constructing a different-scale receptive field network based on a double convolutional layer, including convolution channel group A1, convolution channel group A2 and convolution channel group A3, and using different sizes of receptive ranges to extract steganographic correlation features from various regions of the data image preprocessed in step 1.

5. The image steganalysis method based on multi-scale residual network according to claim 4, characterized in that: Using convolution kernels of different scales, a multi-scale receptive field network based on a double convolution layer is constructed as follows: the first convolution layer of the convolution channel group A1 adopts a 3×3 convolution kernel, and the second convolution layer adopts 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 adopts a 7×7 convolution kernel, and the second convolution layer adopts a 3×3 convolution kernel; The output results of the three parallel convolution channel groups A1, A2, and A3 are connected to form 22 output feature channels, which serve as the output initial features of the first part of the MRNet network.

6. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: In step 3, multi-channel group feature fusion is used as the second part of the MRNet network, combined with the dual-channel group convolutional network and feature map group fusion, to suppress the extraction of natural features and stego features. Specifically, it includes: constructing two parallel convolution channel groups B1 and convolution channel group B2, and deepening feature extraction and fusion of the initial features extracted by the same-layer mixed feature channel group in step 2; the convolution channel group B1 is sequentially processed by the S1 structure residual layer 5 times to output the stego features of 22 feature channels; the convolution channel group B2 is sequentially processed by the S2 structure residual layer 3 times to output the stego features of 22 feature channels; The feature channels output by the convolution channel group B1 and the feature channels output by the convolution channel group B2 are fused through the channel dimension to form a fusion feature of a total of 44 feature channels, which serves as the steganographic feature output channel of the second part of the MRNet network.

7. The image steganalysis method based on multi-scale residual network according to claim 6, characterized in that: The S1 structure residual layer in the convolution channel group B1 includes: a first convolution layer, a BN layer, a ReLU activation function, a second convolution layer, and a BN layer connected in sequence, and the output is the sum of the output of the BN layer after the second convolution layer and the initial input mapping branch information of the S1 structure residual layer; each convolution layer in the S1 structure residual layer uses a 3×3 convolution kernel; The S2 structure residual layer in the convolution channel group B2 includes: a first convolution layer, a BN layer, a ReLU activation function, a second convolution layer, and a BN layer connected in sequence, and the output is the sum of the output of the BN layer after the second convolution layer and the initial input mapping branch information of the S2 structure residual layer; each convolution layer in the S2 structure residual layer adopts a 5×5 convolution kernel.

8. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: In step 4, the feature channel group dimensionality reduction is used as the third part of the MRNet network. By constructing a dimensionality reduction feature channel group C1, the steganographic features output in step 3 are reduced in dimension. Specifically, the dimensionality reduction feature channel group C1 uses a single channel and is processed four times by the S3 structure residual layer in sequence. After each processing, the feature map size is halved, and the image steganographic features after dimensionality reduction are output. The main branch of the S3 structure residual layer sequentially includes the first convolutional layer, the BN layer, the ReLU activation function layer, the second convolutional layer, the BN layer, followed by an average pooling layer with a stride of 2, wherein the convolutional layers all use a 3×3 convolution kernel; the shortcut branch of the S3 structure residual layer is set to a 1×1 convolutional layer with a stride of 2 to match the channel size of the input feature map with the output of the main branch; The output of the residual layer of the S3 structure is the result of element-by-element summation of the output of the main branch after average pooling and the output of the shortcut branch after 1×1 convolution.

9. 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 is the fourth part of the MRNet network. By constructing the feature classification channel group D1, it is used to classify and make the final judgment on the steganographic features of the image after dimensionality reduction output in step 4. Specifically, it includes: The network layers of the D1 channel group include: the first convolutional layer, the BN layer, the ReLU activation function layer, the second convolutional layer, the BN layer, followed by the global average pooling layer, the fully connected layer and the softmax layer in sequence; among them, the global average pooling layer converts 512 feature maps into 512-dimensional feature vectors, which are processed by the fully connected layer and then binary classified by the softmax layer to output the final image steganalysis detection result.

10. The image steganalysis method based on multi-scale residual network according to claim 1, characterized in that: It also includes step 6, multi-scale residual network parameter configuration, which is used to train and test the MRNet network model, by setting the input image specifications, filling modes of each convolutional layer, activation function type, convolution kernel initialization method, training steps, learning rate and optimizer type of the MRNet network model to achieve model training and testing.

Images