Feature enhancement and sample expansion-based few-sample steganalysis method

Abstract

The invention provides a few-sample steganalysis method based on feature enhancement and sample expansion, and the method combines a few-sample learning theory, enhances the diversity of feature space, and uses the generated pseudo steganalysis sample to finely adjust the pre-training model, thereby improving the training effect and detection precision of the steganalysis model. Specifically, firstly, high-frequency noise features of an image are extracted through a spatial domain rich model and a 2D Gabor filter, and a steganographic feature map with the most representative is screened through normalization feature saliency measurement; thirdly, generating a steganography feature prototype by adopting a variational auto-encoder VAE, and generating a large number of pseudo-steganography samples through methods of sampling, threshold segmentation and the like, so as to enhance the diversity of training samples; and finally, in combination with the real steganography sample and the pseudo steganography sample, carrying out hierarchical training on the pre-trained steganography analysis model, and gradually optimizing the model performance. The method is suitable for various steganography analysis networks, and steganography images generated by different steganography algorithms can be effectively detected.

Description

Technical Field

[0001] The present invention belongs to the technical field of information security, and particularly relates to a few-shot steganalysis method based on feature enhancement and sample augmentation. Background Art

[0002] Steganography is a technique for hiding secret information by modifying digital carriers (such as images, audio, or video), aiming to achieve covert communication without arousing suspicion from third parties. With the development of big data and digital image processing technology, image steganography has gradually matured and diversified, becoming one of the important technologies in the field of information hiding. However, the abuse of steganography may lead to security risks. For example, criminals may use steganography for covert communication to evade supervision or carry out malicious acts. Therefore, steganalysis technology has emerged, aiming to detect and identify the secret information hidden in digital media to ensure the security of digital communication.

[0003] Steganalysis technology is mainly divided into methods based on handcrafted features and methods based on deep learning. Methods based on handcrafted features usually rely on manually designed features such as statistical features, filter kernels, or high-order moments. However, when facing complex or unknown steganographic algorithms, the detection performance of these methods is often low, and the generalization ability is limited. In recent years, the success of deep learning technology in the field of image processing has provided new solutions for steganalysis. By constructing deep neural networks, deep learning models can automatically extract the steganographic noise features of images, overcoming the limitations of traditional methods that require manual feature design. For example, generative adversarial networks (GANs) are used to generate stego images, while convolutional neural networks (CNNs) are widely applied to steganalysis tasks. In addition, deep learning models can combine spatial and frequency domain features to effectively improve the detection ability of steganalysis.

[0004] However, the performance of deep learning models highly depends on a large amount of training data. In practical applications, the number of stego images is often limited, which restricts the generalization ability of the models. Few-shot learning aims to improve the model's recognition ability for new categories using a small number of training samples, usually by methods such as data augmentation and meta-learning to expand the training set or construct support sets and query sets, so as to improve the model's generalization ability and metric ability. In the field of steganalysis, few-shot learning methods provide new ideas for solving the problem of insufficient training data. For example, by adding auxiliary information and data augmentation techniques, diverse training samples can be generated, thus improving the detection accuracy of the model. In addition, few-shot learning methods can also combine the prior knowledge of steganographic algorithms to enhance the steganalysis ability of the model.

[0005] In practical applications, steganalysis models often face challenges such as insufficient training data and unknown steganography algorithms. Existing deep learning methods usually require a large amount of labeled data, and the collection and annotation costs of steganographic data are relatively high, resulting in the model being prone to overfitting in the case of small samples.

[0006] In the prior art, Chinese Patent CN112785478A discloses a method and system for detecting hidden information based on generating an embedding probability map. The above method can generate a to-be-detected embedding probability map according to the to-be-detected image, convolve the to-be-detected embedding probability map and the to-be-detected image with a plurality of high-pass filter kernels respectively to obtain a first residual map corresponding to the to-be-detected embedding probability map and a second residual map corresponding to the to-be-detected image, fuse the first residual map and the second residual map to obtain a to-be-detected fused image, and use a pre-trained steganalysis model to learn the to-be-detected fused image and output the probability that the to-be-detected image contains secret information, so as to detect whether secret information is hidden in the to-be-detected image, which has relatively high detection accuracy and detection efficiency. However, this method relies on a large amount of training data for training the steganalysis model, and in practical applications, it is relatively difficult to obtain steganographic data, resulting in poor adaptability of this method in small sample scenarios.

[0007] To solve the above problems, there is an urgent need for a method that can effectively expand steganographic image samples under few-sample conditions and enhance the detection ability of steganalysis models, so as to improve the applicability of steganalysis technology in practical scenarios. Summary of the Invention

[0008] The purpose of the present invention is to provide a few-shot steganalysis method based on feature enhancement and sample expansion to overcome the defects of the above-mentioned prior art.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] On the one hand, the present invention provides a few-shot steganalysis method based on feature enhancement and sample expansion, including the following steps:

[0011] Obtain a set of steganographic images generated by an unknown steganography algorithm;

[0012] Extract features from each steganographic image in the set of steganographic images to obtain a plurality of steganographic noise feature maps corresponding to each steganographic image;

[0013] Use a normalized feature saliency function to screen the steganographic noise feature maps of each steganographic image, and screen out the significant feature maps of each steganographic image;

[0014] Input the significant feature maps into a pre-trained variational autoencoder model to generate an embedding probability map for each steganographic image in the set of steganographic images;

[0015] Steganography is performed on the non-steganographic image based on the embedding probability graph to generate pseudo-stego samples;

[0016] The pre-trained steganalysis model is fine-tuned by combining the pseudo-stego samples and the stego image set to obtain the target steganalysis network;

[0017] The test samples in the unknown set are input into the target steganalysis network, and it is judged whether the test samples hide secret information through the output labels.

[0018] Furthermore, the feature extraction of each stego image in the stego image set specifically includes:

[0019] Extract the high-frequency noise features in the stego image through the spatial rich model SRM. Specifically, perform convolution operations on the stego image through multiple high-pass filter kernels to obtain the residual feature map of the image. The formula is:

[0020]

[0021] Among them, S (i,j) represents the i-th stego image extracted by the spatial rich model SRM method and the j-th high-pass filter kernel s j The residual feature map obtained after convolution operation;

[0022] Perform convolution on the stego image through the Gabor filter in different directions and scales to extract the texture and edge features in the stego image. The formula is:

[0023]

[0024] Among them, represents the i-th stego image extracted by the Gabor filter and the j-th Gabor filter kernel g j The texture and edge feature map obtained after convolution operation;

[0025] The residual feature map S (i,j) extracted by SRM and Gabor filter and the texture and edge feature map are jointly used as the stego noise feature map of the stego image i.

[0026] Furthermore, the screening of the stego noise feature map of each stego image by using the normalized feature saliency function specifically includes:

[0027] Use the normalized feature saliency function A NES (x) to calculate the representativeness of the stego noise feature map of each stego image. The formula is:

[0028]

[0029] Among them, G M represents calculating the average gradient magnitude of each pixel in the steganographic noise feature map using the Sobel operator, where H and W are the height and width of the steganographic noise feature map, Gx and Gy are the gradient approximations along the horizontal and vertical directions respectively, and C mns represents the root mean square contrast of the steganographic noise feature map, and I ij is the gray value of the pixel (i, j) in the steganographic noise feature map, is the average gray value of the steganographic noise feature map, and H(p i ) represents the image entropy, which is used to measure the complexity of the image, and p i is the frequency of the i-th level gray value in the steganographic noise feature map, n is the number of gray levels, Norm represents the min-max normalization process, and A NES (x) is the normalized feature saliency value of the steganographic noise feature map;

[0030] According to the normalized feature saliency value A NES (x), select the steganographic noise feature map with a saliency value higher than the preset threshold ∈ as the saliency feature map of the steganographic image. The formula is:

[0031]

[0032] Among them, represents the j-th saliency feature map of the i-th steganographic image after being screened by the normalized feature saliency threshold, and x (i,j) is the j-th steganographic noise feature map of the steganographic image i, and A NES (x (i,j) ) is the normalized feature saliency value of the steganographic noise feature map x (i,j) .

[0033] Furthermore, the variational autoencoder model includes an encoder E(x, z) and a decoder D(z, x). Among them, the encoder is used to map the input saliency feature map to the latent variable z in the low-dimensional latent space, and the decoder is used to reconstruct the original image. The encoder part uses ResNet-152 as the feature extractor, removes its last fully connected layer, and connects two fully connected layers and a batch normalization layer respectively. The feature vectors output by ResNet are respectively transformed into two 256-dimensional vectors for output, one of which is the mean vector μ, and the other is the natural logarithm of the variance logvar = lnσ 2 .

[0034] Further, inputting the significant feature maps into a pre-trained variational autoencoder model to generate an embedding probability map for each stego-image in the stego-image set specifically includes:

[0035] Input all the significant feature maps of stego-image i into the pre-trained variational autoencoder model, and the encoder of the variational autoencoder model outputs the significant feature maps along with their corresponding mean vector μ and natural logarithm of variance logvar = lnσ 2 ;

[0036] Based on the mean vector μ and natural logarithm of variance logvar = lnσ 2 sample a latent space vector z from the standard normal distribution through the reparameterization trick (i,j) ;

[0037] Based on the latent space vector z of all the significant feature maps of stego-image i calculate the Gaussian mean (i,j) and covariance matrix Σ of stego-image i ; i ;

[0038] According to the calculated Gaussian mean and covariance matrix Σ of stego-image i i generate R latent variables z of stego-image i through uniform sampling (i,r) , input the latent variable z (i,r) into the decoder of the variational autoencoder model to generate the corresponding embedding probability map (i,r) for it, and the formula is: Formula:

[0039]

[0040] where Decoder represents the decoder operation of the variational autoencoder model.

[0041] Further, the generation formula for the latent space vector z (i,j) is:

[0042] z (i,j) = μ + eps × σ ~ N(μ, σ 2 )

[0043] where μ is the mean vector output by the encoder of the variational autoencoder model, σ is the exponential of the natural logarithm of variance, i.e., the standard deviation, eps is the noise sampled from the standard normal distribution N(μ, σ 2 ), and z (i,j) is the significant feature map Latent space vectors extracted by the encoder of the variational autoencoder model.

[0044] Furthermore, for all significant feature maps of the stego image i (i,j) , calculate the Gaussian mean and covariance matrix Σ i , and the calculation formula is:

[0045]

[0046] where z (i,j) is the latent space vector mapped by the encoder for the j-th significant feature map of the i-th stego image , and is the mean of the latent vectors of all significant feature maps of the i-th stego image , Σ i is the covariance matrix of the i-th stego image, and J is the number of significant feature maps of the i-th stego image.

[0047] Furthermore, based on the calculated Gaussian mean and covariance matrix Σ i of the stego image i, generate R latent variables z (i,r) of the stego image i through the uniform sampling method, specifically including:

[0048] Use the Gaussian mean and covariance matrix Σ i of the stego image i as the upper and lower limits of the uniform distribution to generate R latent space vectors z (i,r) , where r represents the r-th generated latent variable, and the generation formula is:

[0049]

[0050] where Φ -1 represents the inverse function of the standard normal distribution, α is the probability threshold used to control the sampling range of the generated latent variables, and U represents the uniform distribution.

[0051] Furthermore, perform steganography on the un-stegoed image based on the embedding probability map, specifically including:

[0052] Input each embedding probability map into the Otsu threshold segmentation method to obtain the segmentation threshold T (i,r) , and determine the segmentation result through the following formula

[0053]

[0054] where Denote the pixel value at position (x, y) of the r-th embedding probability map of the i-th stego-image as T (i,r) The segmentation threshold T is calculated by the Otsu thresholding method is the binary pixel value at position (x, y) of the segmented image, which is 0 if the original pixel value is less than the threshold, and 1 otherwise

[0055] Generate a random number matrix W = (w(x, y)) H×W and satisfy w(x, y) ~ U(0, 1). Based on the following rules and the segmentation result and the random number matrix W, perform embedding modification

[0056]

[0057] where respectively represent the probabilities that the pixel is modified to +1 or -1 at position (x, y), and N ±1 represents the number of pixels modified to +1 or -1 in the image is the segmented probability map the sum of all pixel values in, and H and W are the height and width of the image respectively

[0058] Adjust the random number matrix W through the following embedding mapping rules to obtain the embedding modification matrix M (i,r) (x, y):

[0059]

[0060] where M (i,r) (x, y) is the value at position (x, y) of the r-th embedding modification image of the i-th stego-image, and w x,y is the random value at position (x, y) in the random number matrix W

[0061] Add the embedding modification matrix M (i,r) to the non-stego image X us to obtain the pseudo-stego sample X' us The formula is

[0062] X' us = X us + M (i,r)

[0063] where X us is the non-stego image, and X' us is the pseudo-stego sample image after embedding modification

[0064] Furthermore, the loss function of the variational autoencoder model is

[0065]

[0066] Among them, is the loss function of the variational autoencoder model. The first term is the reconstruction loss, which measures the mean square error between the decoder output and the input saliency feature map. The second term is the Kullback-Leibler divergence loss, which measures the difference between the posterior distribution and the standard normal distribution p(z) ∼ N(0, I). The third term R cluster is the clustering regularization term, which encourages the latent variables to form a clearer clustering structure in the latent space. is the reconstruction loss, which measures the input saliency feature map and the decoder reconstruction output The mean square error between them. μ and σ are the mean and standard deviation of the encoder output respectively, and λ is a hyperparameter. Tr(Σ i ) is the trace of the covariance matrix, representing the total variance of the latent variables.

[0067] Compared with the prior art, the present invention has the following advantages:

[0068] (1) By generating an embedding probability map and fine-tuning with pseudo-stego samples and the original stego image set, the present invention enhances the learning ability of the model under few-shot conditions. This method effectively overcomes the challenge of limited stego data, enhances the generalization ability of the steganalysis model, and significantly improves the adaptability and detection accuracy in small-sample scenarios.

[0069] (2) By using the spatial rich model (SRM) and Gabor filters to extract high-frequency noise, texture, and edge features in the stego image, the present invention generates a stego noise feature map of the stego image. This technology can effectively capture the subtle changes in the stego image and improve the recognition ability of stego noise. Using the normalized feature saliency function to screen out the salient feature maps further improves the distinguishability of image features.

[0070] (3) The present invention adopts a pre-trained variational autoencoder model. By using the encoder to extract the saliency feature map, generate the latent space vector in the low-dimensional latent space, and generate an embedding probability map based on these latent space vectors. This technology not only improves the image reconstruction ability but also further enhances the recognition effect of the model on stego images, enabling the model to maintain a high accuracy even in the case of unknown stego algorithms.

[0071] (4) By fine-tuning the pre-trained steganalysis model, the present invention can effectively enhance the steganalysis ability with only a small number of stego-image samples. By combining pseudo-stego samples and saliency feature maps for training, the dependence on a large amount of labeled data is greatly reduced, and the applicability and flexibility in practical applications are improved. Description of the Drawings

[0072] Figure 1 is the input-output flowchart of the method of the present invention.

[0073] Figure 2 is the framework diagram of the present invention.

[0074] Figure 3 is the display of 30 SRM filter kernels used in the present invention.

[0075] Figure 4 is the display of 2D Gabor filter kernels with different directions and scales used in the present invention.

[0076] Figure 5 is the probability graph generation network based on the VAE architecture in the present invention.

[0077] Figure 6 is the experimental result of different normalized feature saliency thresholds.

[0078] Figure 7 is the detection accuracy corresponding to different numbers of support set sample pairs.

[0079] Figure 8 is the ROC curve of the present invention on two adversarial stego sample sets. Detailed Embodiments

[0080] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0081] Embodiment 1:

[0082] In this embodiment, by enhancing stego samples and hierarchical training of the model, the detection ability of the existing steganalysis model for unknown steganography algorithms is improved in the case of a small number of samples of unknown steganography algorithms. As Figure 1As shown in the figure, for a 2-way K-shot few-shot steganalysis task, it is assumed that there are K pairs of carrier images and stego-images of unknown steganographic algorithms. First, diverse high-pass filtering kernels are used to extract high-frequency stego-noise maps from these sample pairs. Based on these stego-noise maps, a generative model is constructed and trained to generate a large number of diverse pseudo-stego-noise maps through a series of operations. Subsequently, the pre-trained steganalysis model is fine-tuned by combining these pseudo-stego samples and real samples to obtain the target steganalysis network. Finally, the test samples of the unknown set are input into the target steganalysis network, and whether the test samples hide secret information is judged by the output labels.

[0083] The technical problem solved by the present embodiment is that in actual steganalysis tasks, due to unknown steganographic algorithms and scarce labeled samples, the detection ability of existing deep learning models for unknown steganographic algorithms decreases. For this reason, the purpose of the present invention is to provide a method that combines feature enhancement technology and few-shot learning strategy to optimize and improve the model detection ability by synthesizing pseudo-stego samples.

[0084] The present embodiment provides a few-shot steganalysis method based on feature enhancement and sample augmentation, as Figure 2 shown. First, the SRM and Gabor filtering kernels are used to capture the high-frequency noise features in the images, and the normalized feature saliency index is used to screen out the most representative feature maps to adjust the VAE model to generate pseudo-stego samples. Subsequently, the pre-trained model is adjusted in a coarse-grained and fine-grained manner by combining real stego samples to optimize the model performance. In addition, the present invention has the characteristics of plug-and-play and high universality, and does not need to change the structure of the steganalysis network itself. This method can greatly improve the detection accuracy of unknown steganographic algorithms with a small amount of labeled data, and also has good performance in the detection of adversarial stego samples. The specific steps are as follows:

[0085] Obtain a set of stego-images generated by an unknown steganographic algorithm;

[0086] Extract features from each stego-image in the set of stego-images to obtain multiple stego-noise feature maps corresponding to each stego-image;

[0087] Use the normalized feature saliency function to screen the stego-noise feature maps of each stego-image, and screen out the significant feature maps of each stego-image;

[0088] Input the significant feature maps into a pre-trained variational autoencoder model to generate an embedding probability map for each stego-image in the set of stego-images;

[0089] Perform steganography on the un-stegoed images based on the embedding probability maps to generate pseudo-stego samples;

[0090] Fine-tune the pre-trained steganalysis model by combining the pseudo-stego samples and the stego image set to obtain the target steganalysis network;

[0091] Input the test samples of the unknown set into the target steganalysis network, and judge whether the test samples hide secret information through the output labels.

[0092] Furthermore, the feature extraction of each stego image in the stego image set specifically includes:

[0093] Extract the high-frequency noise features in the stego image through the spatial rich model SRM. Specifically, perform convolution operations on the stego image through multiple high-pass filter kernels to obtain the residual feature map of the image. The formula is:

[0094]

[0095] where S(i, j) represents the i-th stego image extracted by the spatial rich model SRM method and the j-th high-pass filter kernel s j The residual feature map obtained after convolution operation;

[0096] Perform convolution on the stego image through the Gabor filter in different directions and scales to extract the texture and edge features in the stego image. The formula is:

[0097]

[0098] where represents the i-th stego image extracted by the Gabor filter and the j-th Gabor filter kernel g j The texture and edge feature map obtained after convolution operation;

[0099] The residual feature map S extracted by SRM and Gabor filter (i,j) And the texture and edge feature map Together as the stego noise feature map of the stego image i.

[0100] In steganography, the operation of embedding secret information can be regarded as adding very weak noise to the cover image, and this modification is only slightly adjusted at the pixel level. Different from the method of directly modeling the image content, the Spatial Rich Model (SRM) focuses on analyzing the noise components (i.e., noise residuals) in the image. Since the prediction error between local pixels can reflect the neighborhood correlation, SRM extracts various types of features through multiple sub-models to better describe the destruction of various local correlations between pixels caused by the steganography operation. The SRM method extracts the spatial feature information of the image and calculates the residuals by establishing different high-pass filtering kernels. Then, these residual information will be subjected to truncation and quantization processing, and the steganalysis features are calculated through the co-occurrence matrix. The obtained co-occurrence matrix is divided into 7 categories, as Figure 3 shown, showing 30 SRM filtering kernels containing the above 7 categories. These high-pass filtering kernels focus on extracting the embedding artifacts introduced by steganography, so as to obtain richer steganography features. Therefore, m filtering kernels are selected from the above filtering kernels, and all kernel sizes are unified to (5,5) by zero-padding, and then the stego image generated by the i-th unknown steganography algorithm is

[0101] convolved. Figure 4 In image processing, 2D Gabor filters are often used for texture analysis and are particularly suitable for detecting content with specific directions and frequencies in the image. By selecting specific Gabor functions, Gabor filters for multi-scale and multi-direction feature extraction can be designed. In , the 2D Gabor filtering kernels with different parameters and a size of (8,8) are visualized. We set four direction parameters (i.e., θ ∈ {0, π / 4, π / 2, 3π / 4}), and set the scale parameter σ to 0.5, 0.6, 0.7, 0.8, and the phase shift parameter In addition, we subtract the kernel mean from the 2D Gabor filter elements to make the filter mean zero. Similarly, in order to obtain diverse Gabor-rich features, we generate n Gabor filtering kernels and convolve the stego image

[0102] of the unknown steganography algorithm.

[0103] Furthermore, the use of the normalized feature saliency function to screen the steganographic noise feature maps of each stego image specifically includes: NES (x) calculates the representativeness degree of the steganographic noise feature map of each stego image, and the formula is:

[0104]

[0105] where G MDenotes calculating the average gradient magnitude of each pixel in the steganographic noise feature map using the Sobel operator, where H and W are the height and width of the steganographic noise feature map, Gx and Gy are the gradient approximations along the horizontal and vertical directions respectively, and C mns Denotes the root mean square contrast of the steganographic noise feature map, I ij Is the gray value of the pixel (i, j) in the steganographic noise feature map, Is the average gray value of the steganographic noise feature map, H(p i ) Represents the image entropy, which is used to measure the complexity of the image, p i Is the frequency of the i-th gray level value in the steganographic noise feature map, n is the number of gray levels, Norm represents the min-max normalization process, and A NES (x) is the normalized feature saliency value of the steganographic noise feature map;

[0106] According to the normalized feature saliency value A NES (x), select the steganographic noise feature map with a saliency value higher than the preset threshold ∈ as the salient feature map of the steganographic image. The formula is:

[0107]

[0108] Where, Denotes the j-th salient feature map of the i-th steganographic image after screening by the normalized feature saliency threshold, x (i,j) Is the j-th steganographic noise feature map of the steganographic image i, A NES (x (i,j) ) Is the normalized feature saliency value of the steganographic noise feature map x (i,j) .

[0109] This embodiment proposes a method for training and guiding a variational autoencoder (VAE) to generate more such samples using representative samples. These representative samples possess the relatively significant key features in a certain type of steganographic image, which helps to more accurately construct the potential space prototype of this type of steganographic features. Since the steganographic features among different steganographic noise maps in the stego feature set vary greatly, in order to select those most representative samples, we introduce the normalized feature saliency function A NFS (x) to calculate the representativeness degree of each sample. After obtaining the normalized feature saliency values of each type of image, a threshold can be set to filter out those samples with unobvious steganographic features, and then a set of sample sets with stronger representative features can be obtained. This method can not only effectively identify and select the most representative samples, but also more accurately simulate the feature distribution of this type of steganographic image in the potential space.

[0110] Furthermore, the variational autoencoder model includes an encoder E(x,z) and a decoder D(z,x), where the encoder is used to map the input saliency feature map to the latent variable z in the low-dimensional latent space, and the decoder is used to reconstruct the original image. The encoder part uses ResNet-152 as the feature extractor, removes its last fully connected layer, and connects two fully connected layers and a batch normalization layer respectively, and converts the feature vectors output by ResNet into two 256-dimensional vectors for output, one of which is the mean vector μ, and the other is the natural logarithm of the variance logvar = lnσ 2 .

[0111] Furthermore, inputting the saliency feature map into the pre-trained variational autoencoder model to generate the embedding probability map of each stego-image in the stego-image set specifically includes:

[0112] Input all the saliency feature maps of stego-image i into the pre-trained variational autoencoder model, and the encoder of the variational autoencoder model outputs the saliency feature map its corresponding mean vector μ and the natural logarithm of the variance logvar = lnσ 2 ;

[0113] Based on the mean vector μ and the natural logarithm of the variance logvar = lnσ 2 sample the latent space vector z from the standard normal distribution through the reparameterization trick (i,j) ;

[0114] Based on the latent space vector z of all the saliency feature maps of stego-image i calculate the Gaussian mean value (i,j) and covariance matrix Σ of stego-image i i ;

[0115] According to the calculated Gaussian mean value and covariance matrix Σ i of stego-image i, generate R latent variables z of stego-image i through the uniform sampling method (i,r) , input the latent variable z (i,r) into the decoder of the variational autoencoder model, and generate the embedding probability map (i,r) corresponding to it The formula is:

[0116]

[0117] where Decoder represents the decoder operation of the variational autoencoder model.

[0118] Through the aforementioned method, we obtained rich steganographic noise feature maps and selected representative samples using the steganographic feature saliency index. These samples enabled us to train a probability map generation model to generate a certain number of probability maps for each steganographic image in the unknown set. However, due to the limited sample size, it is difficult to directly train using GANs or Diffusion models. Therefore, we adopted a simple and effective architecture based on variational autoencoders (VAEs) to design the probability map generation network.

[0119] As Figure 5 shown, this VAE architecture includes an encoder E(x,z) and a decoder D(z,x). The encoder maps the input two-dimensional image to a latent variable z in the low-dimensional latent space, while the decoder reconstructs the original image. In the encoder part, ResNet-152 is used as the feature extractor. Its last fully connected layer is removed, and two fully connected layers and a batch normalization layer are connected respectively. The feature vector output by ResNet is transformed into two 256-dimensional vectors, one of which is the mean vector μ, and the other is the natural logarithm of the variance logvar = lnσ 2 . Since the latent variable is random, directly sampling from the probability distribution will lead to ineffective gradient calculation through backpropagation. Using the reparameterization method, a latent variable is sampled from a standard normal distribution noise eps, and the sampled latent variable can be expressed as z = μ + eps×σ ∼ N(μ,σ 2 ). In the decoder part, the latent variable first passes through two fully connected layers and a batch normalization layer, and the Leaky ReLU activation function is used. Subsequently, it is upsampled through three residual modules and a transposed convolution module to output a three-channel tensor. Finally, the image is adjusted to the input size through the sigmoid function and bilinear interpolation.

[0120] Furthermore, the generation formula of the latent space vector z (i,j) is:

[0121] z (i,j) = μ + eps×σ ∼ N(μ,σ 2 )

[0122] where μ is the mean vector output by the encoder of the variational autoencoder model, σ is the exponential of the natural logarithm of the variance, i.e., the standard deviation, eps is the noise sampled from the standard normal distribution N(μ,σ 2 ), and z (i,j) is the feature map of saliency the latent space vector extracted by the encoder of the variational autoencoder model.

[0123] Further, for all significant feature maps of the stego-image i in the latent space vector z (i,j) , calculate the Gaussian mean value and covariance matrix Σ i of the stego-image i. The calculation formula is as follows:

[0124]

[0125] where z (i,j) is the j-th significant feature map of the i-th stego-image mapped to the latent space vector by the encoder, is the mean value of the latent vectors of all significant feature maps of the i-th stego-image , Σ i is the covariance matrix of the i-th stego-image, and J is the number of significant feature maps of the i-th stego-image.

[0126] Further, based on the calculated Gaussian mean value and covariance matrix Σ i of the stego-image i, generate R latent variables z (i,r) of the stego-image i through the uniform sampling method. Specifically, it includes:

[0127] To ensure that the generated probability map is close to the representative latent space prototype, we set a probability threshold α and control the value range of random sampling by adjusting the threshold. To obtain r probability maps of a certain category, R latent vectors z (i,r) need to be generated from the latent space, where R is a constant.

[0128] Take the Gaussian mean value and covariance matrix Σ i of the stego-image i as the upper and lower limits of the uniform distribution, and generate R latent space vectors z (i,r) . Among them, r represents the r-th generated latent variable, and the generation formula is:

[0129]

[0130] where Φ -1 represents the inverse function of the standard normal distribution, α is the probability threshold for controlling the sampling range of the generated latent variable, and U represents the uniform distribution.

[0131] Further, perform steganography on the un-stegoed image based on the embedding probability map. Specifically, it includes:

[0132] Input each embedding probability map into the Otsu threshold segmentation method to obtain the segmentation threshold T (i,r) , and determine the segmentation result through the following formula

[0133]

[0134] Among them, represents the pixel value at the position (x, y) of the r-th embedding probability map of the i-th stego-image, T (i,r) is the segmentation threshold calculated by the Otsu threshold segmentation method, is the binary pixel value at the position (x, y) of the segmented image, which is 0 if the original pixel value is less than the threshold, otherwise 1;

[0135] Generate a random number matrix W = (w(x, y)) H×W , and satisfy w(x, y) ~ U(0, 1). According to the following rules, based on the segmentation result and the random number matrix W, perform embedding modification:

[0136]

[0137] Among them, respectively represent the probabilities that the pixel is modified to +1 or -1 at the position (x, y), N ±1 represents the number of pixels modified to +1 or -1 in the image, is the segmented probability map the sum of all pixel values in, H and W are the height and width of the image respectively;

[0138] Adjust the random number matrix W through the following embedding mapping rules to obtain the embedding modification matrix M (i,r) (x, y):

[0139]

[0140] Among them, M (i,r) (x, y) is the value at the position (x, y) of the r-th embedding modification image of the i-th stego-image, w x,y is the random value at the position (x, y) in the random number matrix W;

[0141] Add the embedding modification matrix M (i,r) to the non-stego image X us to obtain the pseudo-stego sample X' us , and the formula is:

[0142] X' us = X us + M (i,r)

[0143] Among them, X us is the non-stego image, X' usIs the pseudo-stego sample image after embedding modification.

[0144] After obtaining the embedding probability map through the above steps, in order to implement the embedding modification, we first apply the Otsu threshold segmentation method to divide the image into two regions according to the threshold. After determining the segmentation threshold T of each probability map, use the STC embedding simulator to generate the embedding modification image. This step first requires creating a random number matrix W=(w(x,y)) H×W , and satisfying w(x,y)~U(0,1). To ensure that the impact of the embedding modification on the statistical characteristics of the image is minimized, it is usually required that the number of modified pixels of +1 and -1 is roughly the same. Finally, adjust the random number matrix according to the embedding mapping rule of M (i,r) to complete the embedding modification process. Finally, the pseudo-stego sample can be obtained through X' us =X uc +M, and it is combined with the unknown set carrier image to form positive and negative sample pairs for fine-tuning the pre-trained steganalysis model.

[0145] Furthermore, we classify the stego feature maps corresponding to each unknown set stego image into one category (i.e., k-shot is divided into k categories). The VAE loss function for training the j-th feature map of the i-th category can be expressed as:

[0146]

[0147] Among them, is the loss function of the variational autoencoder model. The first term is the reconstruction loss, which measures the mean square error between the decoder output and the input saliency feature map. The second term is the Kullback-Leibler divergence loss, which measures the difference between the posterior distribution and the standard normal distribution p(z)~N(0,I). The third term R cluster is the clustering regularization term, which encourages the latent variables to form a clearer clustering structure in the latent space. is the reconstruction loss, which measures the input saliency feature map and the decoder reconstruction output The mean square error between them, μ and σ are the mean and standard deviation of the encoder output respectively, λ is a hyperparameter, and Tr(Σ i ) is the trace of the covariance matrix, which represents the total variance of the latent variables.

[0148] Example 2:

[0149] The dataset used in this embodiment is sourced from BOSSbase v1.01, BOWS2, and ALASKA#2. First, we determined the optimal value of the normalized feature significance threshold ε and demonstrated the effectiveness and broad applicability of this method through experiments on multiple baseline steganalysis models. In addition, we also studied the impact of changes in the number of support set samples on the performance of the model in detecting unknown samples. Considering that adversarial samples can mislead steganalysis detectors through specific embedding techniques, thereby enhancing the security of steganography, we also evaluated the effectiveness of the proposed method in detecting these adversarial steganography algorithms. The experimental results show that this method can provide stable performance improvement whether at different embedding rates or in the face of complex adversarial samples, demonstrating its consistency and reliability. To verify the rationality of the proposed feature-enhanced few-shot steganalysis method, we designed a series of comparative experiments, including different methods and parameter settings, and conducted a detailed experimental analysis. This series of experiments not only verified the effectiveness of the method but also provided a basis for further optimization.

[0150] According to Figure 3 the described architecture of the feature-enhanced steganalysis method based on a small number of stego sample pairs is constructed.

[0151] Use the pytorch framework to construct a probabilistic graph generation model with VAE as the core architecture. The ResNet-152 of the encoder is initialized with the pre-trained model weights for visual recognition on the ImageNet dataset, and the rest is initialized using the He initialization method. Use the Adam optimizer and set the learning rate to 0.001 and the batch size to 50. The hardware configuration used in this invention is as follows: the graphics card is an NVIDIA GeForce RTX 3090 with 24GB of video memory, the CUDA version is 12.2, the CPU is an Intel(R) Xeon(R) Silver 4314 CPU @ 2.40GHz 32-core processor, the memory size is 16GB, and the operating system is Ubuntu 18.04.6 LTS.

[0152] To explore the impact of different thresholds ε in the screening formula on the adjustment of steganalysis, and to determine the optimal threshold in the experiment, we conducted experiments on the BOWS2 dataset. We set 10 different thresholds and used the datasets generated by the S-UNIWARD steganographic algorithm at embedding rates of 0.2 bpp and 0.4 bpp to train CVTStego-Net as a pre-trained network. To evaluate the performance of the model in unknown domains, we selected the HUGO (Highly Undetectable Steganography) and MiPOD datasets as the unknown sets, and randomly selected 6 pairs of carrier and stego samples from them as the support set for few-shot learning. In addition, we also conducted zero-shot tests, that is, directly applying the pre-trained model to the test set of the unknown set to obtain its detection accuracy, and using it as a baseline for comparison. The experimental results are as Figure 6 shown, demonstrating the detection performance under different embedding payloads: (a) Embedding payload is 0.2 bpp; (b) Embedding payload is 0.4 bpp. The results show that the detection accuracy corresponding to almost all thresholds is higher than that of the zero-shot test, indicating that the feature enhancement method can improve the model's detection ability for steganographic algorithm images in the unknown set. However, there are significant differences in the performance of different thresholds. When the threshold ε is set to 0.5 or 0.6, the detection accuracy reaches the highest. As the threshold decreases from 0.5 or 0.6, the detection accuracy slowly decreases; while when it exceeds 0.5 or 0.6, the detection accuracy rapidly decreases. A low threshold results in too many non-representative samples being selected, affecting the quality of the latent space prototype, thus reducing the detection performance. As the threshold increases, non-representative samples are filtered out, making the steganographic features more concentrated and improving the detection accuracy. However, too high a threshold will lead to insufficient number of samples for training, affecting the model's ability to construct steganographic feature prototypes, and thus reducing the detection accuracy. Based on the above experimental results, we set the normalized feature significance threshold to 0.6 because this threshold can achieve the highest detection accuracy, further verifying the effectiveness of this threshold.

[0153] Based on the above experiments, the value of the threshold parameter ε was determined. We selected several previously excellent steganalysis models and pre-trained these models using the S-UNIWARD (Spatial Universal Wavelet Relative Distortion) dataset. Subsequently, we fine-tuned these pre-trained models twice and recorded the improvement in the detection accuracy of the fine-tuned models relative to the unadjusted models (i.e., zero-shot testing). The results are shown in Table 1. By analyzing the experimental data in the table, we can see that the pseudo-stego samples generated using the feature enhancement method significantly improved the cross-domain detection ability of different pre-trained models, with an average improvement of 1% to 2%. In particular, when testing on the stego sample pairs generated by 6 MiPOD (Minimizing the Performance of the Optimal Detector) steganography algorithms with a payload of 0.2 bpp using the CVTStego-Net model trained on the S-UNIWARD stego dataset, the present invention improved the steganalysis detection accuracy of the pre-trained model on the MiPOD test set by 2.42%. This result indicates that the present invention not only has wide applicability but also can achieve significant performance improvement under specific conditions. Therefore, this section fully demonstrates the effectiveness and universality of the few-shot image steganalysis method through experiments.

[0154] Table 1

[0155]

[0156] We explored the influence of the number of support set sample pairs on the model's performance in detecting the unknown set. Specifically, five experimental conditions of 2-shot, 4-shot, 6-shot, 8-shot, and 10-shot were set, that is, the support set contained 2, 4, 6, 8, and 10 pairs of labeled positive and negative samples respectively. Two steganalysis models, CVTStego-Net and GBRAS-Net, were selected and pre-trained on the S-UNIWARD dataset (payload 0.4 bpp). To evaluate the cross-domain detection ability of the models, datasets of two steganography algorithms, HUGO and MiPOD, were selected as the target domain datasets. In addition, we also tested the zero-shot testing accuracy of these two pre-trained models on the target domain as a benchmark for comparison. The experimental results are as Figure 7 shown, where Figure 7 a represents using the HUGO stego dataset as the unknown set, Figure 7b represents using the MiPOD stego dataset as the unknown set. The results show that when the number of support set sample pairs is 2 and 4, the detection accuracy of CVTStego-Net and GBRAS-Net on the HUGO and MiPOD test sets decreases compared to the baseline. However, when the number of support set sample pairs increases to 6 and above, the detection accuracy of the model adjusted by the method we proposed exceeds the baseline level, and as the number of sample pairs increases, the cross-domain detection performance of the model is further improved. When the number of support set sample pairs is small, the generated pseudo-stego samples are relatively single, resulting in the steganalysis model being unable to fully learn the data characteristics of the target domain during the fine-tuning process, thus reducing the detection performance for unknown set images. On the contrary, as the number of support set sample pairs increases, the generated pseudo-stego samples become more diverse, enabling the model to better capture the stego characteristics of the target domain, thereby enhancing its generalization ability and the detection ability for unknown set stego images. Based on the above experimental results, it can be concluded that appropriately increasing the number of support set sample pairs helps to improve the cross-domain detection performance of the model.

[0157] Since adversarial samples mislead steganalysis by using adversarial embedding techniques, thereby enhancing the security of steganography, we explored the performance improvement of the present invention in detecting adversarial steganographic algorithms. To this end, we generated samples of MAE and Steg-GMAN adversarial steganographic algorithms with embedding rates of 0.2 bpp and 0.4 bpp as the unknown set. The experimental results are as Figure 8 shown, where Figure 8 a represents the sample set of the MAE adversarial steganographic algorithm, Figure 8 b represents the sample set of the Steg-GMAN adversarial steganographic algorithm. In the figure, FAFSL represents the performance of the model adjusted by using the few-shot learning method based on feature enhancement. AUC (area under the receiver operating characteristic curve) is used to measure the detection performance, and the larger the AUC value, the better the classification performance. After adjusting the pre-trained model using the present invention, the ROC curve of the model is closer to the upper left corner, and the AUC value is significantly higher than the result of the zero-shot transfer test. This shows that the few-shot learning method based on feature enhancement can effectively improve the model's detection ability for adversarial samples. Specifically, the detection performance of the adjusted model on adversarial samples has been significantly improved, especially under the condition of high embedding rate (0.4 bpp), and the increase in the AUC value is particularly obvious. This result verifies the effectiveness of the present invention in detecting adversarial steganographic attacks. In addition, the experiment also shows that the present invention shows stable performance improvement at different embedding rates, demonstrating its consistency and reliability in dealing with complex adversarial samples.

[0158] To verify the rationality of the proposed feature-enhanced few-shot steganalysis method, we designed a variety of methods and parameters for experimental comparison. First, we explored the impact of data augmentation techniques on model performance. Data augmentation increases the diversity and quantity of the dataset by transforming existing data (such as adding zero-mean Gaussian noise with different intensities, salt-and-pepper noise with different ratios, adjusting image brightness, horizontal and vertical flipping, rotating at different angles, and grayscale inversion), thereby improving the generalization ability of the model. In this experiment, we performed the above data augmentation operations on 6 support set sample pairs, expanding the dataset to 100 times its original size to ensure consistency with the number of generated pseudo-stego sample pairs. In addition, we also studied the impact of the clustering regularization term on model performance. Specifically, a clustering regularization term was introduced into the loss function of the VAE model, and different regularization coefficients λ were set. When λ = 0, it means that the clustering regularization term is not included, that is, the standard VAE model is used. In this way, the impact of the clustering regularization term on the detection performance was evaluated. We measured the accuracy, false alarm rate, miss detection rate, and F1-score of the model under the conditions of payloads of 0.2 bpp and 0.4 bpp, and the target domain steganography algorithm was HUGO. Table 2 shows that when using the data augmentation method at payloads of 0.2 bpp and 0.4 bpp, the detection performance is significantly improved compared with the zero-shot test scheme. Specifically, the accuracy is increased by 0.81% and 0.52% respectively, the false alarm rate is reduced by 0.0079 and 0.0032 respectively, the miss detection rate is reduced by 0.0083 and 0.072 respectively, and the F1-score is increased by 0.008 and 0.0055 respectively. This indicates that data augmentation can effectively improve the diversity of a small number of samples and enhance the model's detection ability for unknown steganography algorithms. Further findings show that the detection performance of the present invention is better than that of the simple data augmentation method in both cases of λ = 0 and λ = 1. Especially for the CVTStego-Net+FAFSL (λ = 1) scheme, at a payload of 0.2 bpp, the accuracy is increased by 1.3% compared with the data augmentation scheme; at 0.4 bpp, it is increased by 0.82%. Compared with CVTStego-Net+FAFSL (λ = 0), CVTStego-Net+FAFSL (λ = 1) has further improvement at both payloads. This indicates that the clustering regularization term helps the model to more effectively aggregate the potential space distribution of each class of samples, generate more representative pseudo-stego samples, and thus improve the detection performance of the model.

[0159] Table 2

[0160]

[0161] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0162] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A few-shot steganalysis method based on feature enhancement and sample augmentation, characterized in that Including the following steps: Obtain a set of stego-images generated by an unknown steganographic algorithm; Extract features from each stego-image in the set of stego-images to obtain multiple stego-noise feature maps corresponding to each stego-image; Use a normalized feature saliency function to screen the stego-noise feature maps of each stego-image, and screen out the feature maps with saliency for each stego-image; Input the feature maps with saliency into a pre-trained variational autoencoder model to generate an embedding probability map for each stego-image in the set of stego-images; Perform steganography on the non-stego images based on the embedding probability map to generate pseudo-stego samples; Fine-tune the pre-trained steganalysis model by combining the pseudo-stego samples and the set of stego-images to obtain a target steganalysis network; Input the test samples of the unknown set into the target steganalysis network, and judge whether the test samples hide secret information through the output labels.

2. The few-shot steganalysis method based on feature enhancement and sample augmentation according to claim 1, wherein The extracting features from each stego-image in the set of stego-images specifically includes: Extract the high-frequency noise features in the stego-image through the Spatial Rich Model (SRM), specifically perform convolution operations on the stego-image through multiple high-pass filters to obtain the residual feature map of the image. The formula is: Among them, S (i,j) represents the i-th stego-image extracted by the spatial-domain rich model SRM method and the j-th high-pass filtering kernel s j the residual feature map obtained after convolution operation; Perform convolution on the stego-image through Gabor filters in different directions and scales to extract the texture and edge features in the stego-image. The formula is: Among them, represents the i-th stego-image extracted by the Gabor filter and the j-th Gabor filter kernel g j The texture and edge feature map obtained through the convolution operation; The residual feature map S extracted by SRM and Gabor filters (i,j) and the texture and edge feature map are jointly used as the steganographic noise feature map of the stego image i.

3. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 1, characterized in that, The using a normalized feature saliency function to screen the stego-noise feature maps of each stego-image specifically includes: Using the normalized feature saliency function A NES (x) to calculate the representativeness of the steganographic noise feature map of each stego-image, and the formula is: Among them, G M represents calculating the average gradient magnitude of each pixel in the steganographic noise feature map using the Sobel operator, where H and W are the height and width of the steganographic noise feature map, Gx and Gy are the gradient approximations along the horizontal and vertical directions respectively, and C mns represents the root mean square contrast of the steganographic noise feature map, I ij is the gray value of the pixel (i, j) in the steganographic noise feature map, is the average gray value of the steganographic noise feature map, H(p i ) represents the image entropy, which is used to measure the complexity of the image, p i is the frequency of the i-th level gray value in the steganographic noise feature map, n is the number of gray levels, Norm represents the min-max normalization process, and A NES (x) is the normalized feature saliency value of the steganographic noise feature map; According to the normalized feature significance value A NES (x), select the steganographic noise feature map with a significance value higher than the preset threshold ∈ as the feature map of the significance of the steganographic image. The formula is as follows: Among them, represents the j-th significant feature map of the i-th stego-image after being screened by the normalized feature significance threshold, and x (i,j) is the j-th stego-noise feature map of the stego-image i, and A NES (x (i,j) ) is the normalized feature significance value of the stego-noise feature map x (i,j) .

4. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 1, characterized in that The variational autoencoder model includes an encoder E(x, z) and a decoder D(z, x). The encoder is used to map the input saliency feature map to the latent variable z in the low-dimensional latent space, and the decoder is used to reconstruct the original image. The encoder part uses ResNet-152 as the feature extractor, removes its last fully connected layer, and connects two fully connected layers and a batch normalization layer respectively. The feature vectors output by ResNet are respectively transformed into two 256-dimensional vectors for output, one of which is the mean vector μ, and the other is the natural logarithm of the variance logvar = lnσ 2 .

5. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 1 or 4, characterized in that The inputting the feature maps with saliency into a pre-trained variational autoencoder model to generate an embedding probability map for each stego-image in the set of stego-images specifically includes: All significant feature maps of the stego image i are input into a pre-trained variational autoencoder model, and the encoder of the variational autoencoder model outputs significant feature maps along with their corresponding mean vector μ and the natural logarithm of variance logvar = lnσ 2 ; Based on the mean vector μ and the natural logarithm of the variance logvar = lnσ 2 Sample the latent space vector z from the standard normal distribution through the reparameterization trick (i,j) ; Feature maps based on all the saliencies of the stego image i The latent space vector z (i,j) , calculate the Gaussian mean and covariance matrix Σ i ; According to the calculated Gaussian mean value of the stego-image i and covariance matrix Σ i Generate R latent variables z of the stego-image i by the uniform sampling method (i,r) , and input the latent variable z (i,r) into the decoder of the variational autoencoder model to generate the latent variable z (i,r) Its corresponding embedding probability map The formula is: Where, Decoder represents the decoder operation of the variational autoencoder model.

6. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 5, characterized in that The latent space vector z (i,j) has the following generation formula: z (i,j) = μ + eps × σ ~ N(μ, σ 2 ) Among them, μ is the mean vector of the encoder output of the variational autoencoder model, σ is the exponential of the natural logarithm of the variance, i.e., the standard deviation, and eps is the noise sampled from the standard normal distribution N(μ, σ 2 ). (i,j) z is the feature map of significance and is the latent space vector extracted by the encoder of the variational autoencoder model.

7. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 5, characterized in that, The feature map of all saliencies of the stego image i (i,j) , calculate the Gaussian mean and covariance matrix Σ i of the stego image i, and the calculation formula is: Among them, z (i,j) is the j-th significant feature map of the i-th stego image is the latent space vector mapped by the encoder, is the mean of the latent vectors of all significant feature maps of the i-th stego image , Σ i is the covariance matrix of the i-th stego image, and J is the number of significant feature maps of the i-th stego image.

8. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 5, characterized in that The calculated Gaussian mean value of the stego-image i and the covariance matrix Σ i Generate R latent variables z of the stego-image i by the uniform sampling method (i,r) , specifically including: The Gaussian mean of the stego-image i and covariance matrix Σ i are used as the upper and lower limits of the uniform distribution to generate R latent space vectors z (i,r) , where r represents the r-th generated latent variable, and the generation formula is: where, Φ -1 represents the inverse function of the standard normal distribution, α is the probability threshold for controlling the sampling range of the generated latent variable, and U represents the uniform distribution.

9. A few-shot steganalysis method based on feature enhancement and sample augmentation according to claim 1, characterized in that The performing steganography on the non-stego images based on the embedding probability map specifically includes: Input each embedding probability map into the Otsu threshold segmentation method to obtain the segmentation threshold T (i,r) , and determine the segmentation result through the following formula Among them, represents the pixel value at the position (x, y) of the r-th embedding probability map of the i-th stego-image, and T (i ,r) is the segmentation threshold calculated by the Otsu threshold segmentation method, is the binarized pixel value at the position (x, y) of the segmented image, which is 0 if the original pixel value is less than the threshold, and 1 otherwise; Generate a random number matrix W = (w(x, y)) H×W , and satisfy w(x, y) ~ U(0, 1). According to the following rules, based on the segmentation result and the random number matrix W, perform embedding modification: Among them, respectively represent the probabilities that the pixel is modified to +1 or -1 at the position (x, y), and N ±1 represents the number of pixels modified to +1 or -1 in the image, is the probability map after segmentation is the sum of all pixel values in, and H and W are the height and width of the image respectively; Adjust the random number matrix W according to the following embedding mapping rules to obtain the embedded modification matrix M (i,r) (x,y): Among them, M (i,r) (x, y) is the value at the position (x, y) of the r-th embedded modified image of the i-th stego image, w x,y is the random value at the position (x, y) in the random number matrix W; Modify the embedding matrix M (i,r) and add it to the stego-free image X us to obtain the pseudo-stego sample X u ′ s , and the formula is as follows: X u ′ s = X us + M (i,r) Among them, X us is the un-stego image, and X u ′ s is the pseudo-stego sample image after embedding modification.

10. A few-shot steganography analysis method based on feature enhancement and sample augmentation according to claim 1, characterized in that The loss function of the variational autoencoder model is: Among them, is the loss function of the variational autoencoder model. The first term is the reconstruction loss, which measures the mean square error between the decoder output and the input saliency feature map. The second term is the Kullback-Leibler divergence loss, which measures the difference between the posterior distribution and the standard normal distribution p(z) ∼ N(0, I). The third term R cluster is the clustering regularization term, which encourages the latent variables to form a clearer clustering structure in the latent space. is the reconstruction loss, which measures the input saliency feature map and the decoder reconstruction output The mean square error between them, μ and σ are the mean and standard deviation of the encoder output respectively, λ is a hyperparameter, and Tr(Σ i ) is the trace of the covariance matrix, which represents the total variance of the latent variables.

Images