Image steganalysis method based on Siamese network contrastive learning

Abstract

This invention presents an image steganalysis method based on Siamese network contrastive learning. A steganalysis model is constructed using the SiaStegNet network as the basic framework. The model comprises two parallel sub-networks and a classifier. The image to be analyzed is divided into left and right sub-images, which are input into the two sub-networks. Each sub-network extracts feature vectors from its sub-image. These feature vectors are then classified by the classifier to determine whether the image to be analyzed contains steganalysis. The sub-networks include two stages: preprocessing and feature extraction. In the feature extraction stage, a residual attention transfer module is used to help transfer residual features completely from the shallow network to the deep network, assigning different channel weights to the residual features to encourage the network to focus more on the channels containing steganalysis information. During model training, prototype contrastive loss is used to assist cross-entropy loss in fitting the network, making steganalysis information easier to distinguish from other high-frequency information in the image, enhancing the network's representation ability, and improving the accuracy of detection and classification.

Description

Technical Field

[0001] This invention belongs to the field of image steganalysis technology in information security, specifically an image steganalysis method based on Siamese network contrastive learning. Background Technology

[0002] Steganography is a method of ensuring communication security by hiding information in multimedia media during data communication. Its purpose is to prevent information from being known to anyone other than the designated recipient. While steganography protects privacy and information security during multimedia transmission, it is also easily exploited. Image steganalysis aims to detect the presence of hidden information in images, and it is of great significance for communication security and maintaining social stability. In steganalysis tasks, images without hidden information are called cover images, and images containing hidden information are called stegog images. In traditional steganalysis methods, feature extraction and machine learning-based classification are separate, and these two steps cannot be optimized simultaneously, making it difficult to achieve a heterogeneous balance. Moreover, stegog features are generally calculated by professional researchers relying on their prior experience and continuous heuristic attempts, resulting in high complexity in stegog feature extraction.

[0003] In recent years, the rapid development of deep learning has provided new methods for steganalysis. More and more researchers are beginning to combine deep learning with steganalysis, gradually forming a three-step general model of preprocessing, feature extraction, and classification. In the preprocessing stage, stegographic signals are extracted through high-pass filtering; in the field of steganalysis, these stegographic signals are also called noise residuals or residual features. In the feature extraction stage, more abstract and representative features are generated. A classifier is used to determine whether the input image is a steganalyte. Steganalysis differs from conventional machine learning tasks. Conventional machine learning tasks tend to process the content of the image itself, while steganalysis aims to process the stegographic information in the high-frequency information of the steganalyte image, i.e., the stegographic signal. High-frequency information corresponds to the parts of the image that change drastically, including texture, semantic edges, and image noise. Therefore, the stegographic signal is difficult to distinguish from the high-frequency information of the image.

[0004] Currently, most deep learning-based steganalysis models use SRM to initialize high-pass filter kernels in the preprocessing stage, aiming to extract high-frequency information, including steganalytic signals, from images. However, in feature extraction and classifiers, they merely heuristically apply general components of deep learning without giving special attention to steganalytic signals. This leads to problems such as steganalytic signals being difficult to distinguish from high-frequency information and difficulties in extracting steganalytic signals, thus reducing the accuracy of detection and classification. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to propose an image steganalysis method based on Siamese network contrastive learning.

[0006] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0007] A steganalysis method based on Siamese network contrastive learning is proposed. The steganalysis model is constructed with the SiaStegNet network as the basic framework. The steganalysis model includes two parallel sub-networks and a classifier. The image to be analyzed is divided into two sub-images, left and right, and input into the two sub-networks. The sub-networks extract feature vectors from the sub-images. The feature vectors extracted by the two sub-networks are classified in the classifier to determine whether the image to be analyzed is a steganalysis image.

[0008] The sub-network comprises two stages: preprocessing and feature extraction. The feature extraction stage includes two sequential feature extraction modules and a global average pooling layer. Each feature extraction module includes a residual attention propagation module and a G-Block module. The residual attention propagation module is divided into two stages: channel weight learning and feature map downsampling. In the channel weight learning stage, the input feature map of the residual attention propagation module is multiplied by the input feature map of this module after global average pooling, convolution, and activation to obtain a weighted feature map. The feature map downsampling stage has two branches. The weighted feature map is passed through a point convolutional layer, a normalization layer, a depthwise convolutional layer, and a normalization layer, and then added to the feature map obtained after the weighted feature map has passed through a point convolutional layer and a normalization layer. The added feature map passes through an activation layer, then through a convolutional layer and a normalization layer to obtain the output feature map of the first branch. The weighted feature map passes through a point convolutional layer and a normalization layer to obtain the output feature map of the second branch. The output feature maps of the two branches are concatenated along the channel dimension and then passed through an activation layer to obtain the output feature map of the residual attention propagation module.

[0009] Furthermore, the classifier first performs element-wise operations on the two feature vectors to obtain four vectors: maximum value, minimum value, mean, and variance. These four vectors are then concatenated to obtain a fused feature vector. The fused feature vector is then passed through a fully connected layer and a classification layer to obtain the classification result.

[0010] Furthermore, the G-Block module includes two 3×3 convolutional layers. A normalization layer and an activation layer are added after the first convolutional layer, and a normalization layer is added after the second convolutional layer. The output of the second convolutional layer is added to the input of the G-Block module, and then passed through the activation layer to obtain the output of the G-Block module.

[0011] Furthermore, the method also includes model training. During model training, the classification loss is measured by cross-entropy loss, the prototype contrast loss is used to assist the cross-entropy loss in fitting the network, and the Euclidean distance is used to measure the similarity between two feature vectors.

[0012] The formula for calculating similarity loss is:

[0013]

[0014] Among them, L sml f represents the similarity loss. i and f j Let represent the feature vectors extracted by the two sub-networks, and y represent the label of the input image. Let m represent the L2 norm and m represent the margin.

[0015] The fused feature vector is passed through a multi-layer perceptron and an L2 regularization layer, mapping the fused feature vector to a spherical embedding space. The prototype contrast loss is then calculated in the spherical embedding space using the following formula:

[0016]

[0017] Among them, L spcl Let z represent the prototype contrastive loss, I represent the index of the sample / label pair, P(i) represent the index of all positive samples with the same label as the anchor sample, |P(i)| represent the cardinality of P(i), A(i) represent the index of all other samples except the anchor sample, and z i Represents the anchor sample, z p Represents anchor sample z i The corresponding positive samples, k represents the set of all class labels, O j This represents the prototype vector corresponding to the sample with label j. Represents the relationship with anchor sample z i Prototype features with the same labels;

[0018] In summary, the overall loss function is expressed as:

[0019] L = L CE +αL sml +βL spcl (6)

[0020] Among them, L CE This represents the cross-entropy loss; α and β are both weighting factors.

[0021] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0022] 1. Stegmatic information exists in the high-frequency information of an image, which in turn contains its own texture details, semantic edges, and other high spatial frequency components. Therefore, in steganalysis tasks, there is a problem of difficulty in extracting stegmatic information from complex high-frequency information. To solve this technical problem, this invention designs a residual attention transfer module. First, it helps the residual features (stegmatic information) to be completely transferred from the shallow network to the deep network. Second, it assigns different channel weights to the residual features, prompting the network to pay more attention to the channel where the stegmatic information is located, making the stegmatic information easier to distinguish and improving the network's representation ability.

[0023] 2. Because different steganalytes contain different content, and the location and quantity of stegographic information also vary, steganalysis is a classification task with significant intra-class differences. To better train the network, a loss function based on prototype contrastive learning was designed. Prototype contrastive loss utilizes label information to amplify inter-class differences in the embedding space, clustering features of the same class, thereby better guiding the network to focus on stegographic information and improving detection and classification accuracy. Prototype contrastive loss is combined with cross-entropy loss to distinguish stegographic signals from other high-frequency information in a steganalyte image. Furthermore, by incorporating prototype features into the loss function, the model's generalization performance is further improved, enabling the model to better handle stegographic signals from different steganalysis algorithms.

[0024] 3. Experiments were conducted on the BOSSbase fixed-size benchmark dataset and the ALASKA arbitrary-size benchmark dataset using three adaptive steganography algorithms at four different embedding rates. Experimental results show that, compared to the conventional steganalysis network SiaStegNet, the steganalysis model of this invention achieves varying degrees of improvement in classification accuracy and performance on each dataset. Therefore, the steganalysis model of this invention has better adaptability and performance. Attached Figure Description

[0025] Figure 1 Examples of carrier images, steganalog images, and residual images for the BOSSbase dataset;

[0026] Figure 2 Examples of carrier images, steg images, and residual images for the ALASKA dataset;

[0027] Figure 3 This is a structural diagram of the steganalysis model of the present invention;

[0028] Figure 4 Here is a structural diagram of the G-Block module;

[0029] Figure 5 This is a structural diagram of the RAP-Block module;

[0030] Figure 6This is a schematic diagram of the prototype contrast loss. Detailed Implementation

[0031] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific implementation methods, but this is not intended to limit the scope of protection of this application.

[0032] This invention relates to an image steganalysis method based on Siamese network contrastive learning (hereinafter referred to as the method, see [link]). Figures 1-6 The process includes the following steps:

[0033] Step 1: Create datasets, including both fixed-size and arbitrary-size datasets;

[0034] Fixed-size dataset: Using the original images from the BOSSbase v1.01 dataset as carrier images, four steganalysis images with payloads of 0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp were generated from the carrier images using the S-UNIWARD, HILL, and WOW steganography algorithms, respectively, resulting in a fixed-size dataset. The fixed-size dataset was then divided into training, validation, and test sets in a ratio of 6:1:3. The original images in the BOSSbase v1.01 dataset were 512×512 pixels, which were scaled to 256×256 pixels using bilinear interpolation.

[0035] Arbitrary Size Datasets: The original images from the ALASKA_512 and ALASKA_VAR datasets were used as carrier images. The original images in the ALASKA_512 dataset were 512×512 in size. The S-UNIWARD steganography algorithm was used to generate steganalysis images with a payload of 0.4 bpp from the carrier images of the ALASKA_512 dataset, and the resulting steganalysis images were divided into training and validation sets. When generating steganalysis images for the ALASKA_VAR dataset using the S-UNIWARD steganography algorithm, the payload was adjusted according to the square root law to achieve constant statistical detectability at different sizes. The ALASKA_VAR dataset contains images of 16 different sizes. Table 1 shows the payload and rate of change of images at different scales in the ALASKA_VAR dataset.

[0036] Table 1. Payload and rate of change of images at different scales in the ALASKA_VAR dataset.

[0037]

[0038] Step 2: Design a steganography analysis model;

[0039] In practical applications, the size of the images to be processed by steganalysis is unknown. In general vision domains, the size of the input image can usually be directly adjusted to meet the required specifications. However, for steganalysis tasks, directly adjusting the image size will affect the signal-to-noise ratio between the carrier image and the steganalytic information, reducing the detection performance of the steganalysis model. The goal of this application is to build a steganalysis model that can process images of arbitrary size. Therefore, the SiaStegNet (Siamese Steganalysis Network), which can process images of arbitrary size, is selected as the basic framework of the steganalysis model.

[0040] See Figure 3 The steganalysis model consists of two parallel sub-networks with shared structure, parameters, and weights, and a classifier. The input image is divided into two sub-images, left and right, which serve as the input to the two sub-networks. The feature vectors obtained from the two sub-networks are processed by the classifier to obtain the steganalysis results. Each sub-network includes two stages: preprocessing and feature extraction. The preprocessing stage is used to generate a noisy residual image that is highly correlated with the stegographic information, and the feature extraction stage is used to extract the feature vector of the noisy residual image.

[0041] In the preprocessing stage, a learnable high-pass filter is first used to initially extract noise residuals from the image. The weights of the high-pass filter are initialized by 30 5×5 filters from the SRM. Then, three consecutive unpooled convolutional blocks (G-Block modules) are used to enhance the extraction effect of noise residuals. The structure of the G-Block module is shown in [link to G-Block module]. Figure 4 The G-Block module consists of two convolutional layers with 30 channels and a kernel size of 3×3. After the first convolutional layer, a normalization layer (BN) and an activation layer (ReLU) are added. After the second convolutional layer, a normalization layer (BN) is added. The output of the second convolutional layer is added to the input of the G-Block module, and then passed through the activation layer to obtain the output of the G-Block module.

[0042] See Figure 3 The feature extraction stage includes two feature extraction modules. The feature vector generated by the global average pooling layer after the feature map obtained by the second feature extraction module is the output of the feature extraction stage, which is also the output of the sub-network. Each feature extraction module includes a Residual Attention Propagation Block (RAP-Block) and a G-Block module. The RAP-Block module helps the residual features to be passed from the shallow network to the deep network and makes the network give more attention to the channel where the steganographic information is located.

[0043] See Figure 5The RAP-Block module consists of two stages: channel weight learning and feature map downsampling. In the channel weight learning stage, the input feature map of the RAP-Block module first undergoes global average pooling (GAP) to obtain initial weights for each channel dimension. Then, it passes through a 1×1 one-dimensional convolutional layer for adaptive weight learning, and the weights are mapped to the interval [0,1] using the Sigmoid function. Finally, the learned weights are multiplied by the input feature map of the RAP-Block module to obtain a weighted feature map. The feature map downsampling stage has two branches. In the first branch, the weighted feature map passes through a point convolutional layer, a normalization layer, a depthwise convolutional layer, and another normalization layer. This result is then added to the feature map obtained after the weighted feature map has passed through a point convolutional layer with a stride of 2 and a normalization layer. The weighted feature map passes through an activation layer, then a 3×3 convolutional layer and a normalization layer to obtain the output feature map of the first branch. The depth convolutional layer has a 3×3 kernel and a stride of 2. In the second branch, the weighted feature map passes through a point convolutional layer with a stride of 2 and a normalization layer to obtain the output feature map of the second branch. The output feature maps of the two branches are concatenated along the channel dimension and then passed through an activation layer to obtain the output feature map of the RAP-Block module. In steganalysis tasks, steganalytic features are also called noise residual features because they are discontinuous, like image noise, and appear as isolated points. Residual features can be lost as the network deepens. The second branch helps shallow residual features to enter deeper networks more effectively, thus helping steganalytic information to be better transmitted to deeper networks.

[0044] The feature vectors obtained from the two sub-networks are fed into the classifier. First, element-wise operations are performed on the two feature vectors to obtain the maximum value vector, minimum value vector, mean vector, and variance vector. Then, these four vectors are concatenated in series to obtain the fused feature vector. The fused feature vector passes through a fully connected layer and a classification layer to obtain the classification result.

[0045] The formula for calculating the fused features is as follows:

[0046] Fusion(f i ,f j )=[m max ,m min ,m mean ,m var (1)

[0047] Where, m max m min m mean m var These represent the maximum, minimum, mean, and variance vectors, respectively.

[0048] Step 3: Use the training and validation sets obtained in Step 1 to train and validate the steganalysis model, calculate the model loss, and optimize the model through backpropagation;

[0049] For the training set in the ALASKA_512 dataset, after loading each training batch, the carrier image and steganalyte image of a fixed size of 512×512 pixels are randomly cropped, and the resulting rectangular cropped images of random size are input into the steganalysis model for training. For the validation set in the ALASKA_512 dataset, images of size 512×512 pixels are used directly for validation.

[0050] Cross-entropy loss is used to measure classification loss, and prototype contrast loss is used to assist cross-entropy loss in fitting the network. Simultaneously, Euclidean distance is used to measure the similarity between feature vectors extracted by two sub-networks, and the similarity loss is calculated. The formula for the similarity loss function is as follows:

[0051]

[0052] Among them, f i and f j These represent the feature vectors extracted by the two sub-networks, respectively. y∈{0,1} represents the label of the input image, with the carrier image labeled as 0 and the steganalysis image labeled as 1. Let f denote the L2 norm; m denotes the margin, which represents the threshold for measuring the similarity between the left and right sub-images of the input image. If the input image label is 1, when the feature vector f i and f j If the Euclidean distance is less than the threshold m, then the similarity loss is greater than 0, and the smaller the Euclidean distance, the greater the similarity loss; when the feature vector f i and f j If the Euclidean distance is greater than the threshold m, the similarity loss is 0; if the input image label is 0, the smaller the Euclidean distance, the smaller the similarity loss.

[0053] The fused feature vectors are passed through a multilayer perceptron (MLP) with a single hidden layer and an L2 regularization layer. The MLP transforms the image feature representation into a low-dimensional space, and the L2 regularization layer maps the feature vectors into a spherical embedding space. At this point, the dot product of the two vectors is the cosine similarity between them. Prototype contrastive loss is calculated in the spherical embedding space. Since different steganases have different contents, the location and quantity of steganases will also vary, which leads to certain differences in the residual features of the steganases. The prototype contrastive loss based on mini-batch is used to assist cross-entropy loss in training classification tasks with large inter-class differences, and guides the network to focus on inter-class differences, i.e., steganases in the high-frequency information of the image. The introduction of prototype vectors enhances the contrastive power of this contrastive learning and improves generalization.

[0054] To introduce the prototype contrast loss, we first introduce the supervised contrast loss as follows;

[0055]

[0056] For N sample / label pairs used in a batch for training, it can be represented as {x k ,y k} k＝1,···N Let i∈I≡{1,2…N} be the indices of N sample / label pairs, z i The anchor sample is denoted by τ. Note that this method does not use data augmentation to create positive and negative sample pairs as in self-supervised contrastive learning. Instead, it utilizes the label information from the supervised task, using samples with the same label in a batch as positive anchor samples and samples with different labels as negative anchor samples; · denotes the inner product, τ∈R. + Let be a scalar temperature parameter, A(i)≡I\{i} representing the index of all samples in a batch except the anchor sample, and P(i)≡{p∈A(i): y p =y i} represents the index of all positive samples with the same label as the anchor sample in a batch, and |P(i)| represents its cardinality; z p Represents anchor sample z i Corresponding positive samples;

[0057] The greater the similarity between the anchor sample and the positive sample, the smaller the loss; conversely, the greater the similarity between the anchor sample and the negative sample, the greater the loss. Contrastive loss allows the features corresponding to the carrier image and the stegana image to complete their respective spatial clustering in the embedding space, and feeds this back to the preprocessing and feature extraction modules, guiding the network to pay more attention to the stegana signals in high-frequency information.

[0058] The average value of the sample features for each category (also known as the prototype) is added to the loss calculation to transform the contrastive loss into a prototype-contrast loss. This allows the model to better learn the feature distribution of each category and pushes the feature representation to align with the prototype for each category. To ensure the accuracy of the prototype, the prototype value is updated in each batch to reflect the feature distribution of the current batch of data. Specifically, a fixed-size sample pool of n is introduced for both the carrier image and steganalyte image categories. i represents the label category; then, r samples are randomly selected from the sample pool of each category as the support set S, and the mean of the support set is taken as the prototype vector O. i The formula is expressed as follows:

[0059]

[0060] The prototype vectors O corresponding to the carrier image and the stegana are obtained through the above method.k and their class label k; now let F(z) i ,X)=exp(z i ·X / τ), Prototype contrast loss L spcl It can be described as:

[0061]

[0062] Where k represents the set of all class labels, O j This represents the prototype vector corresponding to the sample with label j, where j∈k. Representative and anchor sample z i Prototype features with the same labels;

[0063] In summary, the overall loss function comprises three parts: classification loss, similarity loss, and prototype comparison loss. The overall loss function is expressed as:

[0064] L = L CE +αL sml +βL spcl (6)

[0065] Among them, L CE The classification loss is represented by cross-entropy loss; α and β are both weighting factors. The model achieves optimal classification performance when α is 0.1 and β is 0.8.

[0066] Step 4: Save the optimized steganalysis model and test it on the test set. That is, use the optimized steganalysis model for steganalysis detection and obtain the detection results.

[0067] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. An image steganalysis method based on Siamese network contrastive learning, characterized in that, The method includes the following: A steganalysis model is constructed based on the SiaStegNet network. The steganalysis model includes two parallel sub-networks and a classifier. The image to be analyzed is divided into two sub-images, left and right, and input into the two sub-networks. The sub-networks extract feature vectors from the sub-images. The feature vectors extracted by the two sub-networks are classified in the classifier to determine whether the image to be analyzed is a steganalysis image. The sub-network comprises two stages: preprocessing and feature extraction. The feature extraction stage includes two sequential feature extraction modules and a global average pooling layer. Each feature extraction module includes a residual attention propagation module and a G-Block module. The residual attention propagation module is divided into two stages: channel weight learning and feature map downsampling. In the channel weight learning stage, the input feature map of the residual attention propagation module is multiplied by the input feature map of this module after global average pooling, convolution, and activation to obtain a weighted feature map. The feature map downsampling stage has two branches. The weighted feature map is passed through a point convolutional layer, a normalization layer, a depthwise convolutional layer, and a normalization layer, and then added to the feature map obtained after the weighted feature map has passed through a point convolutional layer and a normalization layer. The added feature map passes through an activation layer, then through a convolutional layer and a normalization layer to obtain the output feature map of the first branch. The weighted feature map passes through a point convolutional layer and a normalization layer to obtain the output feature map of the second branch. The output feature maps of the two branches are concatenated along the channel dimension and then passed through an activation layer to obtain the output feature map of the residual attention propagation module.

2. The image steganalysis method based on Siamese network contrastive learning according to claim 1, characterized in that, The classifier first performs element-wise operations on the two feature vectors to obtain four vectors: maximum value, minimum value, mean, and variance. These four vectors are then concatenated to obtain a fused feature vector. The fused feature vectors are passed through a fully connected layer and a classification layer to obtain the classification result.

3. The image steganalysis method based on Siamese network contrastive learning according to claim 1, characterized in that, The G-Block module includes two 3×3 convolutional layers. After the first convolutional layer, a normalization layer and an activation layer are added. After the second convolutional layer, a normalization layer is added. The output of the second convolutional layer is added to the input of the G-Block module, and then passed through the activation layer to obtain the output of the G-Block module.

4. The image steganalysis method based on Siamese network contrastive learning according to claim 1, characterized in that, The method also includes model training. During model training, the classification loss is measured by cross-entropy loss, the prototype contrast loss is used to assist the cross-entropy loss in fitting the network, and the Euclidean distance is used to measure the similarity between two feature vectors. The formula for calculating similarity loss is: Among them, L sml f represents the similarity loss. i and f j Let represent the feature vectors extracted by the two sub-networks, and y represent the label of the input image. Let m represent the L2 norm and m represent the margin. The fused feature vector is passed through a multi-layer perceptron and an L2 regularization layer, mapping the fused feature vector to a spherical embedding space. The prototype contrast loss is then calculated in the spherical embedding space using the following formula: Among them, L spcl Let z represent the prototype contrastive loss, I represent the index of the sample / label pair, P(i) represent the index of all positive samples with the same label as the anchor sample, |P(i)| represent the cardinality of P(i), A(i) represent the index of all other samples except the anchor sample, and z i Represents the anchor sample, z p Represents anchor sample z i The corresponding positive sample, z a This represents all samples other than the anchor sample, k represents the set of all class labels, and O j This represents the prototype vector corresponding to the sample with label j. Represents the relationship with anchor sample z i Prototype features with the same labels; In summary, the overall loss function is expressed as: L=L CE +αL sml +βL spcl (6) Among them, L CE This represents the cross-entropy loss; α and β are both weighting factors.

Images