General color image steganography detection method and system based on deep learning
By constructing a color image steganography detection model of twin networks and feature fusion layers, combining channel rearrangement and attention mechanisms, the problem of low color image detection accuracy is solved, and higher detection accuracy and generalization ability are achieved.
Patent Information
- Application Number
- CN202410984617.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-19
AI Technical Summary
The existing color image steganography detection methods have shortcomings in detection accuracy, especially the deep learning-based methods lack consideration of the correlation between the three channels, resulting in a decrease in detection accuracy.
Using a general color image steganography detection method based on deep learning, by constructing a model including a preprocessing layer, a feature extraction layer and a classifier, using a twin network and a feature fusion layer, combining channel rearrangement, filtering and attention mechanisms, Design an enhanced loss function to improve detection accuracy.
Significantly improves the accuracy and generalization ability of color image steganography detection. Compared with classic methods such as ConvNet and UCNet, it shows better detection effect, especially when dealing with airspace and JPEG adaptive steganography. ability to transform.
Smart Images

Figure CN118982450A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of digital steganography technology, and in particular to a general color image steganography detection method and system based on deep learning. Background Art
[0002] Digital steganography embeds secret information into digital carriers (such as images, videos, audio, web pages, etc.) and transmits it through public channels to achieve covert communication. These characteristics of steganography algorithms can also be easily used by criminals to transmit illegal information through steganography, thereby endangering national and social security. Steganography detection technology is a countermeasure technology for steganography technology, which is used to detect secret information in the carrier by analyzing the difference in image features. Current steganography detection methods are mostly aimed at grayscale analysis, and there is a lack of effective detection methods for color images that are more widely used and disseminated on the Internet. Therefore, conducting research on color image steganography detection technology to detect and discover illegal communications is of great significance for protecting network security.
[0003] The existing steganalysis detection technologies can be divided into detection technologies based on manual design and detection technologies based on deep learning. Among the steganalysis detection technologies based on manual design, there are steganalysis detection methods for low-dimensional features, such as SPA, RS, etc.; there are also steganalysis detection methods for high-dimensional features, such as the high-dimensional rich model feature SRM (Spatial Rich Model) proposed by Fridrich et al., the PSRM (Projection SpatialRich Model) steganalysis detection model proposed by Holub et al. on the basis of the SRM model, and the Adaptive SRM (Adaptive Spatial RichModel) steganalysis detection model proposed by Tang et al. The above methods are based on manual design features, and their feature extraction and classification are separate, which limits the improvement of steganalysis detection accuracy. The steganalysis detection method based on deep learning combines the feature extraction stage and the classification stage, and with its powerful automatic learning and fitting capabilities, it greatly improves the detection accuracy of the adaptive steganalysis algorithm. However, most of the current steganalysis detection methods based on deep learning are aimed at grayscale images, such as the ConvNet method (structure such as Figure 1 As shown in Figure 2, when detecting color images, the detection accuracy is reduced due to the lack of consideration of the correlation between the three channels of the color image. Among the color image steganography detection methods based on deep learning, such as the UCNet method proposed by Wei et al. (structure as Figure 2 As shown in the figure, the correlation between the three channels is not considered in the preprocessing and feature extraction stages, which leads to limited extracted residuals, thus limiting the improvement of the accuracy of color image steganalysis detection. Summary of the invention
[0004] In order to solve the problem of low accuracy of color image steganography detection, the present invention provides a general color image steganography detection method and system based on deep learning.
[0005] On the one hand, the present invention provides a general color image steganography detection method based on deep learning, comprising:
[0006] Step 1: Construct a color image steganography detection model, including a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps;
[0007] Step 2: construct a loss function, and train the color image steganography detection model based on the loss function;
[0008] Step 3: Input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
[0009] Furthermore, the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps, specifically including:
[0010] The input image is separated into three channels, and each channel feature map after separation is processed by SRM filtering and Gabor filtering. The three channel feature maps after filtering are spliced, and the channels of the spliced feature maps are rearranged.
[0011] Furthermore, the feature extraction layer adopts a twin network, and the twin network includes two feature extraction branches, each of which includes a convolution group, four first residual blocks, two second residual blocks, one third residual block and a global average pooling layer; correspondingly, the feature extraction process of the feature extraction layer includes:
[0012] The feature map after channel rearrangement is first processed by the convolution group, then processed by three of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the third residual blocks in sequence, and finally passed through a global average pooling layer to obtain an output feature map.
[0013] Furthermore, the convolution group includes two 3×3 convolution layers.
[0014] Furthermore, the first residual block includes two 3×3 convolutional layers, an SE attention mechanism layer and a skip connection; correspondingly, the feature map processing process of the first residual block includes:
[0015] Input Features Figure 1On the one hand, it passes through two 3×3 convolutional layers and the SE attention mechanism layer in sequence to obtain the first intermediate feature map; on the other hand, it is directly added to the first intermediate feature map through a jump connection to obtain the output feature map.
[0016] Furthermore, the second residual block includes two 3×3 convolutional layers, an Eca attention mechanism layer, a pixel rearrangement layer and a skip connection; correspondingly, the feature map processing process of the second residual block includes:
[0017] Input Features Figure 1 On the one hand, it passes through two 3×3 convolutional layers, the Eca attention mechanism layer and the pixel rearrangement layer in sequence to obtain the second intermediate feature map; on the other hand, it first passes through a 1×1 convolutional layer and then passes through a jump connection to add it to the second intermediate feature map to obtain the output feature map.
[0018] Furthermore, the third residual block includes two 3×3 convolutional layers, a channel attention layer, an SE attention mechanism layer and two skip connections; correspondingly, the feature map processing process of the third residual block includes:
[0019] The input feature map is divided into three branches. The first branch passes through a 3×3 convolution layer and a channel attention layer in sequence to obtain a third intermediate feature map. The second branch directly multiplies the third intermediate feature map through a jump connection, and then passes through a 3×3 convolution layer and an SE attention mechanism layer in sequence to obtain a fourth intermediate feature map. The third branch directly adds the fourth intermediate feature map through a jump connection to obtain an output feature map.
[0020] Furthermore, the feature fusion process of the feature fusion layer includes:
[0021] According to the feature map output by the global average pooling layer, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the first fused feature;
[0022] According to the feature map output by the last third residual block, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the second fused feature;
[0023] The first fusion feature and the second fusion feature are concatenated to obtain an output feature map.
[0024] Furthermore, the loss function is:
[0025] Loss(F)=L CEL (F)+L CTL (F a ,F b )+λ×L MC(F)
[0026] Among them, F represents the feature map output by the feature fusion layer, F a and F b Represents the feature map output by the two feature extraction branches in the twin network, L CEL represents the cross entropy loss function, L CTL represents the contrast loss function, L MC represents the cross-correlation loss function, L MC =MC1+MC2;
[0027] in, H and W represent the height and width of the feature map F, C represents the number of channels of the feature map F, C = c × ξ, c represents the number of channels of each class, ξ is a hyperparameter, represents the feature corresponding to the target category i, i = 0, 1, 2…c-1, y represents the category, y∈{0,1}; M i =diag(Mask i ), is a mask composed of 0 and 1, including 1 and 0, F i,j,k Represents the feature corresponding to the (j,k)th channel in the i-th category.
[0028] On the other hand, the present invention provides a general color image steganography detection system based on deep learning, comprising:
[0029] A model building module is used to build a color image steganography detection model, including a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps;
[0030] A training module, used for constructing a loss function, and training the color image steganography detection model based on the loss function;
[0031] The detection module is used to input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
[0032] Beneficial effects of the present invention:
[0033] (1) The present invention constructs a preprocessing layer for enhancing channel correlation in the color image steganography detection model. The preprocessing layer first filters the three separated channels and then uses channel rearrangement to enhance the network's ability to identify the correlation between channels, thereby solving the current problem of low detection accuracy due to the lack of consideration of the correlation between channels.
[0034] (2) The present invention constructs a feature extraction layer based on a twin network in a color image steganography detection model. The twin network includes two feature extraction branches, each of which includes multiple residual blocks, enriching the extraction of residual features, and fusing pixel rearrangement with the residual blocks to further enhance the expression capability of the three-channel features.
[0035] (3) The present invention also designs a weighted loss function, uses the cross entropy loss function to enhance the generalization ability of the network, and uses the contrast loss function to improve the network's ability to learn the steganographic features in the subgraph. At the same time, by calculating the maximum, minimum, average and variance of the features of the last two layers in the feature extraction layer, combined with the cross-correlation loss function, the role of inter-channel correlation in classification is fully utilized.
[0036] Using the ALASKA#2 universal dataset as the carrier image, a large number of experiments were conducted on the cryptanalysis images generated by spatial domain and JPEG adaptive steganography using the latest steganalysis network. The experimental results show that compared with the classic ConvNet steganalysis detection method and the latest UCNet universal color image detection method, the proposed scheme has good generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is a structural diagram of the existing ConvNet solution;
[0038] Figure 2 This is the structural diagram of the existing UCNet solution;
[0039] Figure 3 A structural diagram of the color image steganography detection model UCPNet provided in an embodiment of the present invention;
[0040] Figure 4 A structural diagram of a convolution group, a first residual block, a second residual block, and a third residual block provided in an embodiment of the present invention;
[0041] Figure 5 A structural diagram of a general color image steganography detection system based on deep learning provided by an embodiment of the present invention;
[0042] Figure 6 The training accuracy of the ConvNet, UCNet and UCPNet provided in the embodiment of the present invention for detecting the CMD-C-HILL algorithm is as follows: Figure 6 (a) and training Loss, such as Figure 6 (b);
[0043] Figure 7 The steganalysis detection effects of ConvNet, UCNet and UCPNet under algorithm mismatch provided in the embodiments of the present invention are shown. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] The automatic learning and classification capabilities of deep learning-based steganalysis detection methods have made their detection accuracy for secret images higher than that of traditional steganalysis detection methods based on artificial feature extraction, and have become a research hotspot in the field of information hiding in recent years. However, most of the existing deep learning-based steganalysis detection methods are targeted at grayscale images, and are difficult to work on steganalysis algorithms based on color images.
[0046] To this end, an embodiment of the present invention proposes a general color image steganography detection method based on deep learning, comprising the following steps:
[0047] S101: Build a color image steganography detection model;
[0048] Specifically, Figure 3 As shown, the color image steganography detection model constructed in the embodiment of the present invention includes a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then performs channel shuffle on the filtered three-channel feature maps;
[0049] As an implementable method, the preprocessing layer uses a spatial domain filter combined with a frequency domain Gabor filter to enhance the model's ability to recognize steganographic signals in spatial domain images and frequency domain images. At the same time, in addition to the inconsistency in the number of channels between color images and grayscale images, the most important thing is that color images need to consider the correlation between the three channels when detecting them, especially when detecting images using color image steganographic algorithms. This feature is even more obvious. Therefore, the embodiment of the present invention uses channel rearrangement after filtering to enhance the expression ability of features, which is also beneficial for the network to extract the correlation between channel residuals.
[0050] Specifically, the main steps of the preprocessing layer are as follows:
[0051] First, separate the three channels of the color image I.
[0052] [I r ,I g ,I b ]=ChannelSplit[I] (1)
[0053] Then filter each channel separately.
[0054] F r =SRM&Gab.(I r ) (2)
[0055] F g =SRM&Gab.(I g ) (3)
[0056] F b =SRM&Gab.(I b ) (4)
[0057] Finally, the filtered feature maps are concatenated, and channel rearrangement is used to enhance the correlation between the three channels.
[0058] F rgb =concate(F r ,F g ,F b ) (5)
[0059] F = ChannelShuffle(F rgb ) (6)
[0060] S102: constructing a loss function, and training the color image steganography detection model based on the loss function;
[0061] S103: Input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
[0062] The general color image steganography detection method based on deep learning provided by the embodiment of the present invention extracts the correlation between channels by using channel rearrangement, thereby enriching the residual features while reducing the feature loss caused by convolution, thereby improving the residual feature acquisition capability of the deep learning network.
[0063] Based on the above embodiments, in the embodiments of the present invention, Figure 3 As shown, the feature extraction layer adopts a twin network, and the twin network includes two feature extraction branches, each of which includes a convolution group, four first residual blocks (i.e., Block1 blocks), two second residual blocks (i.e., Block2 blocks), a third residual block (i.e., Block3 blocks) and a global average pooling layer; correspondingly, the feature extraction process of the feature extraction layer includes:
[0064] The feature map after channel rearrangement is first processed by the convolution group, then processed by three of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the third residual blocks in sequence, and finally passed through the global average pooling layer to obtain the output feature map. In deep learning, global average pooling can not only extract background information of the image, but also reduce parameters and prevent overfitting. Therefore, this embodiment adopts global average pooling.
[0065] As an implementation method, Figure 4 As shown in FIG. 1 , the convolution group includes two 3×3 convolution layers. Using two 3×3 convolutions can complete the conversion between channels and increase the rich residual features.
[0066] As an implementation method, Figure 4 As shown, the first residual block includes two 3×3 convolutional layers, an SE attention mechanism layer and a skip connection; correspondingly, the feature map processing process of the first residual block includes: input feature Figure 1 On the one hand, it passes through two 3×3 convolutional layers and the SE attention mechanism layer in sequence to obtain the first intermediate feature map; on the other hand, it is directly added to the first intermediate feature map through a jump connection to obtain the output feature map.
[0067] As an implementation method, Figure 4 As shown in FIG. 1 , the second residual block includes two 3×3 convolutional layers, an Eca attention mechanism layer, a pixel unshuffle layer and a jump connection; correspondingly, the feature map processing process of the second residual block includes: input feature Figure 1 On the one hand, it passes through two 3×3 convolutional layers, the Eca attention mechanism layer and the pixel rearrangement layer in sequence to obtain the second intermediate feature map; on the other hand, it first passes through a 1×1 convolutional layer and then passes through a jump connection to add it to the second intermediate feature map to obtain the output feature map.
[0068] Specifically, by using convolution layers on skip connections, the residual features are enriched. At the same time, in order to reduce the impact of image size on the number of parameters and the amount of computation, the step size of the image is generally set to 2. However, by setting the step size to half the original image size, the residual features are easily lost. To this end, this embodiment uses pixel rearrangement to reduce the size of the image to half of the original size, which can not only reduce the number of parameters and computation of the network, but also reduce the loss of residual features. The main steps of Block2 include: first, using 3×3 convolution to extract residual features, and then using the EcaLayer attention mechanism layer to enhance the network's ability to identify steganographic features; then, using pixel rearrangement to halve the image size. Finally, the features obtained by the skip connection are added to the features obtained by the pixel rearrangement.
[0069] As an implementation method, Figure 4 As shown, the third residual block includes two 3×3 convolutional layers, a channel attention layer, an SE attention mechanism layer and two jump connections; correspondingly, the feature map processing process of the third residual block includes: dividing the input feature map into three branches, the first branch passes through the 3×3 convolutional layer and the channel attention layer in sequence to obtain the third intermediate feature map, the second branch directly passes through the jump connection to multiply the third intermediate feature map, and then passes through the 3×3 convolutional layer and the SE attention mechanism layer in sequence to obtain the fourth intermediate feature map, and the third branch directly passes through the jump connection to add the fourth intermediate feature map to obtain the output feature map.
[0070] The three different residual blocks combine residual blocks, attention mechanisms, and pixel rearrangement. The organic combination of the three helps the network extract rich and effective residual features.
[0071] On the basis of the above embodiments, in an embodiment of the present invention, the feature fusion process of the feature fusion layer includes:
[0072] According to the feature map output by the global average pooling layer, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the first fused feature;
[0073] According to the feature map output by the last third residual block, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the second fused feature;
[0074] The first fusion feature and the second fusion feature are concatenated to obtain an output feature map.
[0075] Furthermore, the classifier uses a fully connected layer to classify the image, so this embodiment uses the fully connected layer as the last layer of the model.
[0076] For the color image steganography detection task, it is necessary not only to obtain effective residual features to improve classification efficiency, but also to learn the correlation between color image channels. In this way, it is not enough to use only one loss function. Therefore, this embodiment uses three loss functions to construct the total loss function. Among them, the cross entropy loss function (CEL) is used to enhance the ability of the network to learn steganographic features, the contrast loss function (CTL) is used to learn the features between the twin network subgraphs, and the mutual correlation loss function (MC) is calculated by the maximum value, minimum value, average value and variance of the last layer of the network, in order to enhance the ability of the network to learn the residual features between color image channels.
[0077] Specifically, the loss function used in this embodiment is as follows:
[0078] Loss(F)=L CEL (F)+L CTL (F a ,F b )+λ×L MC (F) (7)
[0079] Where F represents the features extracted after the input image passes through the convolutional network, that is, the feature map output by the feature fusion layer; F a and F b Represents the subgraph features of the twin network, that is, the feature graph output by the two feature extraction branches in the twin network; for an input image, C represents the number of channels, H and W represent the height and width of the feature map, where C = c × ξ, c represents the number of channels for each class, ξ is a hyperparameter.
[0080] For a given target category i, its features can be expressed as
[0081] F i ={F i×ξ+1 ,F i×ξ+2 ,...,F i×ξ+ξ} (8)
[0082] After obtaining the features of each category, the feature set F = {F0, F1, ..., F c-1} is sent to MC-loss to calculate the loss. MC-loss consists of MC1 and MC2. MC1 can pay more attention to specific categories and make the features of each channel more discriminative. MC2 means that the channels of the same category pay more attention to different areas of the image, which is similar to the spatial attention mechanism.
[0083] L MC =MC1+MC2
[0084] The MC1 loss function is defined as:
[0085]
[0086] y represents the category, y∈{0,1}, and g(.) can be defined as:
[0087]
[0088] is a mask composed of 0 and 1. 1, 0.
[0089] The MC2 loss function is defined as:
[0090]
[0091] Among them, F i,j,k Represents the feature corresponding to the (j,k)th channel in the i-th category.
[0092] The universal color image steganography detection scheme provided by the present invention provides a preprocessing layer that effectively captures the correlation between channels. The preprocessing layer extracts the correlation between channels by using channel shuffle and combines it with the residual by using pixel unshuffle. While enriching the residual features, it reduces the feature loss caused by convolution and improves the ability of deep learning network to obtain residual features.
[0093] Furthermore, the present invention also constructs a feature extraction layer based on the twin network, in which three residual blocks are constructed. The residual structure is used to enhance the network's ability to capture residual features, and the attention mechanism is combined to extract the correlation between channels.
[0094] Furthermore, the present invention also designs a weighted loss function that captures the correlation between channels. The generalization ability of the network is enhanced by using the cross entropy loss function, the residual features of the image are learned by using the contrast loss function, and the correlation between channels is learned by using the cross-correlation loss function. The three are organically combined to enhance the network's ability to learn the residual features between channels.
[0095] Corresponding to the above method, the embodiment of the present invention also provides a general color image steganography detection system based on deep learning, such as Figure 5 As shown, it includes a model building module, a training module and a detection module.
[0096] The model construction module is used to construct a color image steganography detection model, including a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps; the training module is used to construct a loss function, and train the color image steganography detection model based on the loss function; the detection module is used to input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
[0097] It should be noted that the color image steganography detection system provided in the embodiment of the present invention is to implement the above method. Its specific functions can be referred to the above method embodiments, which will not be described in detail here.
[0098] In order to verify the effectiveness of the solution of the present invention, the present invention also provides the following experimental data.
[0099] All experiments were conducted on the NVIDIA Telsa V100 graphics card. The data set used the ALASKA #2 public data set. The ALASKA #2 image library contains grayscale and color image libraries. This experiment used a color image library with a size of 256×256. At the same time, the image library was compressed to generate the ALASKA_75JPEG image library.
[0100] The UCPNet framework of the present invention is as follows Figure 3 The comparison schemes are ConvNet method and UCNet method; among them, the structure diagram of ConvNet method is shown in Figure 1 As shown in the figure, the preprocessing layer uses SRM high-pass filtering and TLU truncation function, the feature extraction layer uses 10 layers of 3×3 convolution layers and 3 layers of average pooling layers, and the classification stage uses global covariance features to obtain the second-order statistics of residual features, and combines the fully connected layer and softmax for classification. This method is only applicable to grayscale image steganalysis detection. When detecting color images, the detection accuracy will be reduced due to the lack of consideration of the correlation between the three channels of the color image. The structure of the UCNet method is shown in the figure. Figure 2 As shown in the figure, the network first has a preprocessing layer, then uses four different convolutional blocks to extract residual features, and finally uses a fully connected layer for classification. This method lacks consideration of the correlation between channels in the feature preprocessing layer and feature extraction, resulting in limited extracted steganalytic residual features, which limits the improvement of the accuracy of color image steganalytic detection.
[0101] (I) Network parameter settings
[0102] The learning rate (lr) initialization parameter used by the UCPNet of the present invention is 0.001, the learning rate is decreased at 80, 140 and 180 respectively, the minimum batch size is set to 30, the number of training epochs is set to 250, the Admax optimizer is used to solve the minimum value of the loss function, the weight decay is 0.0001, L2 regularization and a dropout layer are used to prevent overfitting, and the parameter of the dropout layer is 0.5. The present invention uses accuracy and AUC to evaluate the performance of UCPNet, and the formulas are as follows.
[0103]
[0104] Among them, P FA and P MD The horizontal and vertical axes of the ROC curve are the false positive rate (FPR), that is, P FAand True Positive Rate (TPR), which is 1-P MD .
[0105] Figure 6 The training accuracy, training loss and AUC of the proposed ConvNet, UCNet and UCPNet for the CMD-C-HILL steganography algorithm at an embedding rate of 0.4bpp are given. The dataset used here is ALASKA#2. Figure 6 (a) It can be seen that as the number of training rounds (Epoch) increases, the accuracy of the training set gradually improves. Figure 6 (b) It can be seen that as the number of training rounds increases, the loss of the training set gradually decreases. Figure 6 (a) and Figure 6 In (b), when Epoch is 80, the accuracy is greatly improved and the training loss is greatly reduced. The learning rate is 0.0001. When Epoch is 140, the accuracy of the training set is improved again and the loss is reduced. The learning rate is 0.00001. When Epoch is 180, the accuracy and loss of the training set are constant. The learning rate is 0.000001.
[0106] (II) Image library reconstruction
[0107] Dataset construction: The ALASKA#2 image library is steganographically processed using four classic adaptive steganographic algorithms: S-UNIWARD, HILL, CMD-C-SUNIWARD, and CMD-C-HILL, generating 20,000 images at four embedding rates: 0.1bpp, 0.2bpp, 0.3bpp, and 0.4bpp. The ALASKA_75 is steganographically processed using the J-UNIWARD adaptive steganographic algorithm, generating 20,000 images at four embedding rates: 0.1bpp, 0.2bpp, 0.3bpp, and 0.4bpp. The training set, validation set, and test set are set as follows:
[0108] The carrier image training set is 12000, the validation set is 2000, and the test set is 6000; the corresponding encrypted image training set is 12000, the validation set is 2000, and the test set is 6000.
[0109] (III) Generalizability of the detection method
[0110] The generalization of the steganalysis detection method refers to the detection effect of the steganalysis detection method on the steganalysis algorithm, which is affected by the inconsistency of the embedding rates of the training set and the test set. Figure 7 The detection effects of ConvNet, UCNet and UCPNet on the mismatch of two steganographic algorithms, CMD-C-SUNIWARD and CMD-C-HILL, are given.
[0111] from Figure 7 It can be seen that the algorithm mismatch has a great impact on the steganography detection effect, and the detection effect tends to decrease. For the detection of the same steganography algorithm for the training set and the test set, a good detection effect can be maintained. When the algorithm used in the training set and the test set is mismatched, the overall detection effect decreases. However, compared with the comparative method, the method proposed in the present invention still maintains a good detection effect. Specifically, when the training set is for the detection of CMD-C-SUNIWARD and the test set is the CMD-C-HILL steganography algorithm, compared with the ConvNet and UCNet methods, the UCPNet method proposed in the present invention improves the detection accuracy of the CMD-C-HILL steganography algorithm by 13.03% and 2.83%, respectively. Overall, the different algorithms used in the training set and the test set will affect the detection effect, but the UCPNet method proposed in the present invention still maintains a good detection advantage.
[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A general color image steganography detection method based on deep learning, characterized in that: include: Step 1: Construct a color image steganography detection model, including a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps; Step 2: construct a loss function, and train the color image steganography detection model based on the loss function; Step 3: Input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
2. The general color image steganography detection method based on deep learning according to claim 1 is characterized in that: The preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps, specifically including: The input image is separated into three channels, and each channel feature map after separation is processed by SRM filtering and Gabor filtering. The three channel feature maps after filtering are spliced, and the channels of the spliced feature maps are rearranged.
3. The general color image steganography detection method based on deep learning according to claim 1 is characterized in that: The feature extraction layer adopts a twin network, and the twin network includes two feature extraction branches, each of which includes a convolution group, four first residual blocks, two second residual blocks, one third residual block and a global average pooling layer; Correspondingly, the feature extraction process of the feature extraction layer includes: The feature map after channel rearrangement is first processed by the convolution group, then processed by three of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the first residual blocks in sequence, then processed by one of the second residual blocks and one of the third residual blocks in sequence, and finally passed through a global average pooling layer to obtain an output feature map.
4. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The convolution group includes two 3×3 convolutional layers.
5. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The first residual block includes two 3×3 convolutional layers, an SE attention mechanism layer and a skip connection; correspondingly, the feature map processing process of the first residual block includes: On the one hand, the input feature map passes through two 3×3 convolutional layers and the SE attention mechanism layer in sequence to obtain the first intermediate feature map; on the other hand, it is directly added to the first intermediate feature map through a jump connection to obtain the output feature map.
6. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The second residual block includes two 3×3 convolutional layers, an Eca attention mechanism layer, a pixel rearrangement layer and a skip connection; correspondingly, the feature map processing process of the second residual block includes: On the one hand, the input feature map passes through two 3×3 convolutional layers, the Eca attention mechanism layer and the pixel rearrangement layer in sequence to obtain the second intermediate feature map; on the other hand, it first passes through a 1×1 convolutional layer and then passes through a jump connection to be added to the second intermediate feature map to obtain the output feature map.
7. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The third residual block includes two 3×3 convolutional layers, a channel attention layer, an SE attention mechanism layer and two skip connections; correspondingly, the feature map processing process of the third residual block includes: The input feature map is divided into three branches. The first branch passes through a 3×3 convolution layer and a channel attention layer in sequence to obtain a third intermediate feature map. The second branch directly multiplies the third intermediate feature map through a jump connection, and then passes through a 3×3 convolution layer and an SE attention mechanism layer in sequence to obtain a fourth intermediate feature map. The third branch directly adds the fourth intermediate feature map through a jump connection to obtain an output feature map.
8. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The feature fusion process of the feature fusion layer includes: According to the feature map output by the global average pooling layer, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the first fused feature; According to the feature map output by the last third residual block, four statistical features, namely, maximum value, minimum value, average value and variance, are calculated, and the four statistical features are fused to obtain the second fused feature; The first fusion feature and the second fusion feature are concatenated to obtain an output feature map.
9. The general color image steganography detection method based on deep learning according to claim 3 is characterized in that: The loss function is: Loss(F)=L CEL (F)+L CTL (F a ,F b )+λ×L MC (F) Among them, F represents the feature map output by the feature fusion layer, F a and F b Represents the feature map output by the two feature extraction branches in the twin network, L CEL represents the cross entropy loss function, L CTL represents the contrast loss function, L MC represents the cross-correlation loss function, L MC =MC1+MC2; in, H and W represent the height and width of the feature map F, C represents the number of channels of the feature map F, C = c × ξ, c represents the number of channels of each class, ξ is a hyperparameter, represents the feature corresponding to the target category i, i = 0, 1, 2…c-1, y represents the category, y∈{0,1}; M i =diag(Mask i ), is a mask composed of 0 and 1, including 1 and 0, F i,j,k Indicates that in the i-th category, j,k ) channels corresponding to the features.
10. A general color image steganography detection system based on deep learning, characterized in that: include: A model building module is used to build a color image steganography detection model, including a preprocessing layer, a feature extraction layer, a feature fusion layer and a classifier connected in sequence; wherein the preprocessing layer filters the three-channel feature maps of the input image respectively, and then rearranges the channels of the filtered three-channel feature maps; A training module, used for constructing a loss function, and training the color image steganography detection model based on the loss function; The detection module is used to input the color image to be detected into the trained color image steganography detection model to obtain the detection result.
Citation Information
Cited By
Color image steganalysis model improvement method based on double contrast learning
CN120583188A
Image steganography method based on residual Haar wavelet transform and ECA attention mechanism
CN121792677A
