A jpeg steganalysis method and system

Abstract

The application discloses a kind of Jpeg steganalysis method and system, wherein method steps include: S1. a kind of network model DCANet for Jpeg steganalysis is constructed;S2. prepare data set, divide into training set, verification set and test set;S3. train DCANet, with the data with label is sent into model, with the way of gradient descent is adjusted parameter, so that loss function value is minimum;S4. using the network model trained to test the image to be measured, determine whether it contains secret information.The Jpeg steganalysis network proposed in the application combines the advantages of different networks, which can focus on the area with complex texture that may contain secret information, and achieves better detection results compared to typical Jpeg steganalysis methods.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and relates to an image steganalysis technology, in particular to a Jpeg steganalysis method and system. BACKGROUND

[0002] Image steganography hides secret information in images, and as the opposite of steganography, the purpose of image steganalysis is to analyze and judge whether secret information is contained in images. Jpeg images are the most prominent and widely used image file formats, so it is of great theoretical significance and practical value to study the Jpeg image steganalysis method.

[0003] Early Jpeg domain image steganography algorithms complete information embedding by replacing the least significant bits of DCT coefficients with secret information, mainly including nsF5, MB, etc. With the deepening understanding of adaptive coding, new adaptive steganography algorithms are constantly being proposed, such as the general wavelet distortion steganography algorithm J-UNIWARD, the consistent embedding distortion algorithm UED, and its corresponding improved UERD algorithm. This kind of Jpeg domain adaptive steganography algorithm can reduce the distortion caused by steganography embedding to a minimum, and the selected embedding secret information position is relatively concentrated in the texture complex area, making it difficult for steganalysis to conduct statistical modeling. In this regard, researchers have proposed a series of steganalysis models, such as DCTR, PHARM and GFR, which almost all have similar feature extraction steps, that is, first decompress the Jpeg image to the spatial domain, and then use a filter to extract noise residuals, and finally process the statistical features through a quantization stage. These traditional methods have achieved good detection results, but with the continuous development of adaptive steganography algorithms, the limitations of traditional methods are becoming more and more obvious: on the one hand, more complex image statistical characteristics need to be considered when designing features, which requires researchers to have a high degree of understanding of steganography technology; on the other hand, the adjustment of feature parameters requires a large amount of computing power, and the process is complicated and inefficient.

[0004] In recent years, deep learning technology has achieved great success in the field of computer vision, helping humans to solve many complex problems. Steganalysis is essentially a binary classification problem, and scholars at home and abroad have been trying to use deep learning to realize image steganalysis. Deep learning methods have strong feature learning and expression capabilities, allowing steganographers to focus on designing appropriate architectures without manually designing complex high-dimensional features. By combining multiple convolutions and activation functions, detailed features of images can be effectively extracted, and image classification can be completed under the same model. Through extensive data training, an end-to-end steganalyzer is constructed. Existing research shows that deep learning-based steganalyzers have significantly improved performance compared to steganalyzers based on manually extracted features. To improve the recognition accuracy of the model, existing Jpeg steganalysis networks use various techniques. In terms of network architecture design, some classic CNN network structures are combined, such as residual networks and dense networks. In terms of training strategies, transfer learning, data augmentation, pre-training, and other techniques are used. In addition, some models also introduce relevant domain knowledge, such as JPEG phase perception and high-pass filters. However, existing steganalysis models still have some shortcomings. On the one hand, almost all Jpeg steganalysis models use single-path structures, making it difficult to fully utilize the extracted noise residuals. On the other hand, most existing steganalyzers lack attention to areas that may contain secret information. SUMMARY

[0005] To solve the above problems, the present application provides a Jpeg steganalysis method and system. In the noise extraction stage, a dual-path network is used to explore hidden features in depth while preserving steganographic signals. In the noise aggregation stage, an improved coordinate attention mechanism is used to quickly locate areas that may contain secret information, extract more valuable features, and improve the performance of the model.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] A Jpeg steganalysis method, comprising the following steps:

[0008] S1, constructing a Jpeg steganalysis network DCANet based on a dual-path network and an improved coordinate attention, including noise extraction, noise aggregation, and classification modules; the noise extraction module is used to extract the noise residual component of the input image, and to explore hidden features in depth while preserving the noise residual component; the noise aggregation module aggregates and counts the noise features extracted previously; and the classification module determines whether the image contains secret information based on the aggregated features;

[0009] S2, preparing the data set, dividing the data set into training set, validation set and test set, using two Jpeg domain adaptive steganography algorithms J-UNIWARD and UERD to steganography the original data set;

[0010] S3, training of DCANet model: using random gradient descent optimizer Adamax to train the proposed model, detecting the validation set after the end of each epoch, selecting the model with the highest validation set accuracy as the final model applied to the test set;

[0011] S4, using the trained network model to test the image to be tested, and determining whether it contains secret information according to the predicted label result.

[0012] Further, the step S1 comprises the following sub-steps:

[0013] S11, the noise extraction module comprises a preprocessing layer and two double-path modules; the preprocessing layer uses four high-pass filters F KV , F P , F H and F V to extract noise residuals, wherein F KV is used to suppress the content of the image itself, F P is used as a point high-pass filter, and F KV is used in combination with F H to extract steganographic signals introduced by Jpeg steganography; F V and F C are second-order horizontal and vertical Gabor filters respectively, which accurately describe the texture and edge features of the image from different scales and directions; the parameters of the four filters are updated during the training process; then two double-path modules are used to realize the reuse and mining of noise features, and deeper features are extracted;

[0014] S12, the noise aggregation part includes an improved coordinate attention module and four residual blocks, first, the improved coordinate attention module is used to guide the network to focus on the area with complex texture and possible secret information, the residual connection is used to realize the fusion of low-level features and high-level features, then four residual blocks are used to aggregate the features and reduce the size of the feature map;

[0015] S13, the classification module uses global covariance pooling to convert the feature map into a feature vector, and then obtains the discrimination result through the fully connected layer and the softmax layer.

[0016] Further, the improved coordinate attention module realizes the following functions:

[0017] First, the input S=[s1, s2,..., s COne-dimensional average pooling operations are performed along the X-axis direction and the Y-axis direction, respectively, to complete the encoding of each channel and obtain the accurate position information in the space; the output of the cth channel at a height of h can be expressed as:

[0018]

[0019] where W represents the width of the feature map, s c (h, i) represents the feature value after average pooling of the cth channel at a height of h and a width of i; then, the obtained feature vector is first spliced along the channel dimension, and then subjected to 1x1 convolution, batch normalization processing and nonlinear operation in sequence, and the calculation formula is as follows:

[0020] f = δ (F1[z h ,z w ])

[0021] where [, ] represents the splicing operation of the feature vector, F1 is a convolution transformation function, and the role is to reduce the number of channels, δ is a nonlinear activation function, and the generated f is an intermediate vector containing horizontal and vertical direction information; then, f is divided into two independent feature tensors f h and f w along the horizontal and vertical directions, and is transformed to have the same number of channels as the input feature S through two 1x1 convolutions, and the calculation formula is as follows:

[0022] g h = Sigmoid (F h (f h ))

[0023] g w = Sigmoid (F w (f w ))

[0024] where F h and F w are convolution transformation functions, and g h and g w represent attention weights; finally, the input feature map S and the attention weight are multiplied correspondingly; a residual connection is introduced in the attention module, the input and the attention output are added to form a residual learning mode, and the high-level features and the low-level features are fused together without increasing excessive cost; the final output P of the attention module is expressed as:

[0025]

[0026] Further, the step S2 includes the following sub-steps:

[0027] S21, adjusting the picture size in the data set, dividing the data set into a training set, a validation set and a test set;

[0028] S22, using the imwrite function in Matlab to compress the image with two quality factors respectively, obtaining Jpeg images with different quality factors;

[0029] S23, using two Jpeg domain adaptive steganography methods J-UNIWARD and UERD to embed secret information with four different embedding rates respectively, obtaining eight steganography data sets.

[0030] Further, the step S3 comprises the following sub-steps:

[0031] S31, using the random gradient descent optimizer Adamax to train the proposed model;

[0032] S32, using a suitable batch number during the training process, and the images in each batch are data enhanced by random mirroring and 90-degree rotation, and the data set is randomly shuffled after each iteration to prevent the model from overfitting;

[0033] S33, migrating the model parameters trained under high embedding rate steganography images to low embedding rate images, and fine-tuning the model during the training process;

[0034] S34, training the network based on the above settings, using cross-entropy as the loss function, and evaluating the performance of the model by the false detection rate.

[0035] Further, the performance of the model is evaluated by the following formula:

[0036]

[0037] Wherein, P FA and P FN respectively represent the false alarm rate and the missed detection rate.

[0038] Further, the model is saved after each iteration, and finally the model with the highest classification accuracy on the validation set is selected as the network model trained in step S4 and applied to the test set for testing.

[0039] The application also provides a Jpeg steganalysis system, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, which realizes the above-mentioned Jpeg steganalysis method when loaded into the processor.

[0040] Compared with the prior art, the application has the following advantages and beneficial effects:

[0041] 1. The Jpeg steganalysis network provided by the present application combines the advantages of different networks, can make the model focus on the area with complex texture that may contain secret information, and achieves better detection results compared with typical Jpeg steganalysis methods.

[0042] 2. The noise extraction module of the present application does not directly enter the noise aggregation module after the preprocessing layer, but uses two double-path modules, combines the advantages of residual network and dense network, deeply explores the hidden features on the basis of preserving the steganographic signal, realizes the reuse and mining of noise residuals.

[0043] 3. The noise aggregation module of the present application proposes an improved coordinate attention module to guide the network to focus on the area with complex texture and extract more features conducive to steganalysis, and the improved residual connection combines high-level features and low-level features to further improve the learning ability of the attention module.

[0044] 4. The present application adopts two Jpeg domain adaptive steganographic algorithms J-UNIWARD and UERD, which have obvious improvement in detection accuracy and generalization ability compared with SRNet and DANet, and realize the balance between accuracy and training time. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The Jpeg steganalysis network framework structure diagram provided by the present application is shown.

[0046] Figure 2 The double-path module structure diagram in the noise extraction module is shown.

[0047] Figure 3 The improved coordinate attention module structure diagram is shown.

[0048] Figure 4 The residual module structure diagram is shown. DETAILED DESCRIPTION

[0049] The technical solutions provided by the present application will be described in detail below in combination with specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0050] The present application provides a Jpeg steganalysis method, comprising the following steps:

[0051] S1, a Jpeg steganalysis network DCANet based on double-path network and improved coordinate attention is constructed, including noise extraction, noise aggregation and classification modules; the noise extraction module is responsible for extracting the noise residual component of the input image, and further exploring the hidden features on the basis of preserving the noise residual component, the noise aggregation module aggregates and counts the noise features extracted before, and the classification module discriminates whether the image contains secret information according to the aggregated features. Specifically, the following sub-steps are included:

[0052] (1.1) As shown in Figure 1 , in the Jpeg steganalysis network, the noise extraction module includes a preprocessing layer and two double-path modules, corresponding to "Preprocessing", "DPB1" and "DPB2" in Figure 1 . Among them, the preprocessing layer uses four high-pass filters F KV , F P , F H and F V to extract noise residuals, wherein F KV is used to suppress the content of the image itself, F P is used as a point high-pass filter, and the combination of F KV can better extract steganographic signals introduced by Jpeg steganography; F H and F V are second-order horizontal and vertical Gabor filters respectively, which accurately describe the texture and edge features of the image from different scales and directions. The parameters of the four filters are updated during the training process.

[0053] Residual network can reuse image features and has low redundancy, but it is not good at mining new image features. The dense network uses the splicing method so that each layer of the network can accept the feature map of all previous layers. This mode can help the network to mine more new hidden features during training. In order to mine deeper hidden features on the basis of preserving the steganographic signals as much as possible, we use a double-path network module (DPB) after the preprocessing layer to further reuse and explore the noise residuals. The specific structure is as Figure 2As shown in FIG. 3, each DPB contains a bottleneck block which sequentially passes through a 1x1 convolution layer, a 3x3 convolution layer and a 1x1 convolution layer, where each convolution layer is preceded by a BN layer and a ReLU layer. The output of the bottleneck block is divided into two parts: 64 feature maps are added to the residual connection path, and the remaining 16 feature maps are connected to the dense connection path. In order to enhance the learning ability of each DPB, we use grouped convolution in the 3x3 convolution layer. In DPB-1, we first use 96 convolution kernels to increase the dimension of the extracted noise residual feature maps, and divide the resulting 96 feature maps into two parts, where 32 feature maps are used for dense connection, and the remaining 64 feature maps are used for residual connection. In DPB-2, we fuse the feature maps of the two paths by channel concatenation to obtain an output, which is input into the next module.

[0054] (1.2) Noise aggregation module The purpose of this module is to aggregate the statistical information of the noise features processed by the noise extraction module, while reducing the size of the feature maps. This module includes an improved coordinate attention module and four residual blocks, corresponding to Figure 1 ICA, ResBlock in FIG. 4. The Jpeg adaptive steganography algorithm mostly embeds secret messages in complex texture areas and edge-rich areas according to the content of the image to achieve higher security, so we add an improved coordinate attention module after the noise extraction module to guide the network to focus on these areas that may contain secret information and extract more useful features. The functions implemented by the improved coordinate attention module include coordinate information embedding, coordinate attention generation and added residual connection, and its design is as shown in Figure 3 FIG. 4. First, one-dimensional average pooling operations are performed on the input S = [s1, s2,..., s C ]∈R along the X-axis direction and the Y-axis direction, respectively, thereby completing the encoding of each channel and obtaining the precise position information in space. The output of the cth channel at height h can be represented as:

[0055]

[0056] where W represents the width of the feature map, s c (h, i) represents the feature value after average pooling at height h and width i of the cth channel. Subsequently, the obtained feature vector is first concatenated along the channel dimension, and then sequentially subjected to 1x1 convolution, batch normalization and nonlinear operation. The calculation formula is as follows:

[0057] f = δ (F1[z h ,z w ])

[0058] where [,] denotes the concatenation operation of feature vectors, F1is a convolutional transformation function, which is used to reduce the number of channels, and δ is a nonlinear activation function, and the generated f is an intermediate vector containing horizontal and vertical direction information. Then, f is divided into two independent feature tensors f h and f w along the horizontal and vertical directions, and is transformed to have the same number of channels as the input feature S through two 1x1 convolutions, and the calculation formula is as follows:

[0059] g h = Sigmoid(F h (f h ))

[0060] g w = Sigmoid(F w (f w ))

[0061] where F h and F w are convolutional transformation functions, and g h and g w represent attention weights. Finally, the input feature map S and the attention weight are multiplied correspondingly. In order to enhance the learning ability of the coordinate attention module for image detail features, we made certain improvements to the coordinate attention module, that is, we introduced a residual connection into the attention module, added the input and the attention output to form a residual learning mode, and fused the high-level features and the low-level features together without increasing too much cost. The final output P of the attention module is expressed as follows:

[0062]

[0063] Each residual block has a main layer and a skip connection structure, as shown in Figure 4 . The residual block is used to aggregate features and reduce the size of the feature map. The main layer is composed of a 3x3 convolution layer, a BN layer, a ReLU layer, another 3x3 convolution layer, and a BN layer in turn. Considering that the public average pooling weakens the steganographic signal to some extent, we realize the down-sampling of the feature map by adjusting the stride of the second 3x3 convolution layer to 2. The skip connection structure includes a 1x1 convolution layer and a BN layer, which mainly adjusts the dimension so that the input can be directly added to the output. After passing through 4 residual blocks, a feature map with a size of 16x16x256 is obtained.

[0064] The purpose of the classification module is to determine whether the image contains secret information according to the aggregated features. First, the global covariance pooling is used instead of the commonly used global average pooling to convert the 256 16x16 feature maps into The eigenvectors are then input into a fully connected layer. Finally, the probability distribution is calculated using a Softmax function and the prediction result is output.

[0065] S2, preparing a data set, and dividing the BOSSBase v1.01 data set into a training set, a validation set and a test set according to a ratio of 4:1:5, using two Jpeg domain adaptive steganography algorithms J-UNIWARD and UERD to perform steganography on the original data set; specifically including the following sub-steps:

[0066] (2.1) The data set used in the application is derived from BOSSbase v1.01, which contains 10,000 512x512 gray scale images. Considering the limitation of computing resources, we first adjust the image size to 256x256 through the "imrize" function in Matlab. 4000 images are randomly selected for training, 1000 for validation, and the remaining 5000 for testing.

[0067] (2.2) In order to generate Jpeg format images, the "imwrite" function in Matlab is used to compress the original images, with quality factors of 75 and 95, respectively, to obtain Jpeg images with different quality factors.

[0068] (2.3) For each quality factor, the two most advanced Jpeg adaptive steganography methods J-UNIWARD and UERD are used to evaluate the performance of the proposed steganalysis method. The two steganography algorithms are implemented using the Matlab version of the code, and 0.1, 0.2, 0.3 and 0.4 bpp of secret data are embedded, respectively, to obtain a total of 8 steganography data sets. In order to make the results more convincing, all network models included in the comparative experiment use the same training set, validation set and test set.

[0069] S3, training of DCANet model: the proposed model is trained using the random gradient descent optimizer Adamax, and after each epoch, the validation set is detected, and the model with the highest validation set accuracy is selected as the final model applied to the test set; specifically including the following sub-steps:

[0070] (3.1) The proposed model is trained using the random gradient descent optimizer Adamax, with a momentum value of 0.9 and a weight decay rate of 0.0002. The initial learning rate is set to 0.001, which decays to 0.0001 after 150 epochs of training, and the model is trained for a total of 200 epochs;

[0071] (3.2) Considering the limitation of GPU, the batch size used in the training process is 32 (16 cover / stego pairs). The images in each batch are augmented by random mirroring and 90-degree rotation, and the dataset is shuffled after each iteration to prevent the model from overfitting;

[0072] (3.3) In order to further improve the detection ability of the model for low embedding rate stego images, for the same type of steganography algorithm, we use the strategy of transfer learning, that is, we transfer the model parameters trained under high embedding rate stego images to low embedding rate images and fine-tune the model during the training process; for stego images with embedding rate of 0.4bpp and quality factor of 95, we use the model parameters trained by stego images with quality factor of 75 and embedding rate of 0.4bpp to initialize and fine-tune during the training process.

[0073] (3.4) Based on the above settings, the network is trained, and cross-entropy is used as the loss function, and the performance of the model is evaluated by the false detection rate. The mathematical expression is:

[0074]

[0075] Where P FA and P FN represent the false alarm rate and the missed detection rate, respectively.

[0076] S4, using the trained network model to test the images to be tested, and determining whether the secret information is contained in the images according to the predicted label result. As an improvement, the model is saved after each iteration, and the model with the highest classification accuracy on the validation set is finally selected to test the test set.

[0077] The application also provides a Jpeg steganalysis system, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the computer program realizes the above-mentioned Jpeg steganalysis method when loaded into the processor.

[0078] The technical means disclosed in the application scheme is not limited to the technical means disclosed in the above-mentioned embodiments, and also includes the technical solutions composed of any combination of the above technical features. It should be noted that for those skilled in the art, without departing from the principles of the application, some improvements and refinements can also be made, which are also considered within the protection scope of the application.

Claims

1. A method of Jpeg steganalysis, characterized by, Comprising the following steps: S1, constructing a Jpeg steganalysis network DCANet based on a double-path network and an improved coordinate attention, comprising a noise extraction module, a noise aggregation module and a classification module; The noise extraction module is used to extract the noise residual component of the input image, and the noise residual component is used to explore the hidden features in depth, the noise aggregation module is used to aggregate and count the noise features extracted in the foregoing, and the classification module is used to determine whether the image contains secret information according to the aggregated features; Comprising the following sub-steps: S11, the noise extraction module includes a preprocessing layer and two dual-path modules; the preprocessing layer uses four high-pass filters F. KV F P F H and F V Extract the noise residual, where F KV Used to suppress the content of the image itself, F P As a point high-pass filter, with F KV Combining these techniques can extract steganalytic signals introduced by JPEG steganography; F H and F V The four filters are second-order horizontal and vertical Gabor filters that accurately describe image texture and edge features from different scales and directions. The parameters of these four filters are updated during training. Subsequently, two dual-path modules are used to reuse and mine noise features, extracting deeper features. S12, the noise aggregation part comprises an improved coordinate attention module and four residual blocks, the improved coordinate attention module is used to guide the network to focus on the area with complex texture and possible secret information, the residual connection is used to realize the fusion of low-level features and high-level features, and then four residual blocks are used to aggregate the features and reduce the size of the feature map; The improved coordinate attention module realizes the following functions: First, one-dimensional average pooling operations are performed on the input S = [s1, s2, …, s C ]∈R along the X-axis direction and the Y-axis direction, respectively, to complete the encoding of each channel and obtain accurate position information in space; the output of the cth channel at a height of h can be expressed as: where W represents the width of the feature map, s c (h,i) represents the feature value after the average pooling of the cth channel at height h and width i; the obtained feature vector is then first spliced along the channel dimension, and then sequentially subjected to 1x1 convolution, batch normalization processing and non-linear operation, and the calculation formula is as follows: f = δ(F1[z h ,z w ]) where, [ ] represents the splicing operation of the feature vector, F1 is the convolution transformation function, which is used to reduce the number of channels, δ is a nonlinear activation function, and the generated f is an intermediate vector containing horizontal and vertical direction information; then, f is divided into two independent feature tensors f h and f w along the horizontal and vertical directions, and is transformed to have the same number of channels as the input feature S through two 1x1 convolutions, and the calculation formula is as follows: g h = Sigmoid(F h (f h )) g w = Sigmoid(F w (f w )) wherein F h and F w are convolutional transformation functions, g h and g w represent attention weights; finally, the input feature map S and the attention weights are multiplied correspondingly; a residual connection is introduced in the attention module, the input and the attention output are added to form a residual learning mode, and the high-level features and the low-level features are fused together without increasing excessive cost; the final output P of the attention module is expressed as follows: S13, the classification module converts the feature map into a feature vector by using global covariance pooling, and then obtains a discrimination result through a fully connected layer and a softmax layer; S2, preparing a data set, dividing the data set into a training set, a validation set and a test set, and using two Jpeg domain adaptive steganography algorithms J-UNIWARD and UERD to perform steganography on the original data set; S3, training of the DCANet model: using the random gradient descent optimizer Adamax to train the proposed model, detecting the validation set after the end of each epoch, and selecting the model with the highest validation set accuracy as the final model applied to the test set; S4, testing the to-be-tested image by using the trained network model, and determining whether the image contains secret information according to the predicted label result.

2. The method of Jpeg steganalysis according to claim 1, wherein, The step S2 comprises the following sub-steps: S21, adjusting the picture size in the data set, and dividing the data set into a training set, a validation set and a test set; S22, using the imwrite function in Matlab to compress the images respectively with two quality factors, and obtaining Jpeg images with different quality factors; S23, using two Jpeg domain adaptive steganography methods J-UNIWARD and UERD to embed secret information with four different embedding rates respectively, and obtaining eight steganography data sets.

3. The method of steganalysis of Jpegs as claimed in claim 1, wherein, The step S3 comprises the following sub-steps: S31, using the random gradient descent optimizer Adamax to train the proposed model; S32, using a suitable batch number during the training process, and performing data enhancement on the images in each batch by random mirroring and rotating 90 degrees, and randomly shuffling the data set after each iteration to prevent model overfitting; S33, migrating the model parameters trained under high embedding rate steganography images to images with low embedding rate, and fine-tuning the model during the training process; S34, training the network based on the above settings, using cross-entropy as the loss function, and evaluating the performance of the model by using the false detection rate.

4. The method of steganalysis of Jpegs according to claim 3, characterized in that, The performance of the model is evaluated by the following formula: where P FA and P FN represent the false alarm rate and the miss detection rate, respectively.

5. The method for Jpeg steganalysis of claim 1, wherein, The model is saved after each iteration, and the model with the highest classification accuracy on the validation set is selected as the trained network model in step S4 and applied to the test set for testing.

6. A Jpeg steganalysis system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program comprises computer program code configured to cause the processor to perform the steps of claim 1. The computer program, when loaded into the processor, implements the Jpeg steganalysis method as claimed in any one of claims 1-5.

Images