An image steganalysis method for carrier-selected steganography

Abstract

The application provides a steganalysis method for carrier selection steganography, comprising: acquiring a to-be-detected image and converting it into a grayscale image; constructing a corresponding gray level co-occurrence matrix in multiple directions to capture the texture features of the image; extracting corresponding texture features based on the gray level co-occurrence matrix and performing weighted summation on the texture features in different directions by using a multi-scale feature fusion method to obtain a texture complexity value of the to-be-detected image. A training set including a carrier image set and a stego image set is used to train a classifier; the to-be-detected image is input into the trained classifier to obtain a probability value. According to the texture complexity value, the probability value output by the classifier is dynamically adjusted, and a preset judgment rule is used to judge whether the to-be-detected image is a stego image. The method is suitable for steganalysis based on manually designed features and steganalysis based on deep learning, can effectively identify stego images in carrier selection steganography, and improves the accuracy and reliability of steganalysis.

Description

Technical Field

[0001] This invention belongs to the field of image steganalysis technology, specifically relating to an image steganalysis method for carrier-selective steganography. Background Technology

[0002] With the advent of the digital age, information security faces unprecedented challenges. Covert communication technologies, especially steganography, have rapidly developed in this context, becoming a crucial technology for information transmission. The core objective of steganography is to embed secret information into digital media (such as digital images and audio) and transmit it through public channels, making the transmitted secret information difficult for third parties to detect. Unlike traditional encryption technologies, steganography not only hides the content of the data but also the act of data transmission itself.

[0003] While steganography plays a crucial role in protecting the confidentiality of information transmission, it faces the challenge of steganalysis. Steganalysis, a countermeasure to steganography, aims to identify images with embedded secret information, reveal covert transmission behavior, and provide technical support for information security. The key task of steganalysis is to identify steganized images from a large dataset, thereby preventing illicit data transmission. With the continuous evolution of steganography techniques, methods have progressed from early least significant bit (LSB) steganography and matrix coding to modern steganography techniques focused on minimizing distortion. These techniques, through ingenious design, make the embedded secret information virtually undetectable, significantly increasing the undetectability of steganography.

[0004] Modern steganography techniques primarily rely on designing distortion functions (such as WOW, UNIWARD, HILL, and MiPOD) and combining them with efficient encoding methods (such as STC encoding) to minimize modifications to the carrier image when embedding data. For example, UNIWARD technology measures distortion cost by changing the wavelet transform directional filter, while HILL technology enhances the steganography's resistance to detection by optimizing the modification location strategy. These steganography techniques reduce image distortion while improving the concealment of steganographic information, thus becoming the mainstream steganography schemes.

[0005] With advancements in steganography, existing techniques have introduced cover selection methods. This involves steganalysts choosing images from available image sources that are more suitable for steganographic embedding, thereby enhancing the stealth of the steganography. Research indicates that steganalysts tend to choose images with high texture complexity as cover images because these images contain rich details and texture information. This makes minor image distortions less noticeable to the human eye, and the more complex statistical feature distribution of the image can effectively mask traces of steganographic operations, thus reducing the risk of detection by steganalysis. Specifically, images with high texture complexity, due to their rich details and significant texture variations, can effectively conceal minor alterations caused by steganographic operations, improving the stealth of the steganography. Related research shows that steganalysts often choose images with high texture complexity when selecting cover images, which creates a significant detection blind spot for traditional steganalysis methods when dealing with these images. For example, in the prior art, CN114972886A discloses an image steganalysis method, which includes: using the SMOTE algorithm to oversample minority class samples in grayscale image samples to obtain new image samples; adding the new image samples to the grayscale image samples to form an image sample to be detected; and inputting the image to be detected in the image sample to be detected into a pre-trained image steganalysis model for processing and analysis to determine whether the image to be detected is a steganalysis image.

[0006] However, existing steganalysis methods do not fully consider the anomalies in image statistical features caused by carrier selection, resulting in low accuracy in detecting carrier-selective steganography. Currently, steganalysis methods are mainly divided into two categories: one is based on hand-designed features, which extracts statistical features sensitive to steganography (such as SPAM, SRM, SRMQ1, TLBP, etc.) and then uses an ensemble classifier for detection; the other is based on deep learning, which automatically learns image steganalysis features through models such as convolutional neural networks (CNNs) (such as GNCNN, XuNet, TLU-CNN, Zhu-Net, etc.). Although deep learning methods can automatically learn and extract image features, existing steganalysis methods still do not fully consider the changes in image complexity caused by carrier selection, resulting in poor detection performance when dealing with highly complex images. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an image steganalysis method targeting carrier-selective steganography. This invention innovatively introduces image texture complexity into the steganalysis detection process. By constructing an accurate and comprehensive texture complexity measurement model, optimizing the classifier's decision logic, and strengthening the assessment of the suspiciousness of highly complex images, it improves the targeting and accuracy of carrier-selective steganography detection, providing a reliable technical solution for dealing with novel steganography attacks.

[0008] The objective of this invention can be achieved through the following technical solutions: This invention provides an image steganalysis method for carrier-selective steganography, comprising the following steps: Acquire the image to be detected and convert it to a grayscale image; For the grayscale image, construct corresponding grayscale co-occurrence matrices in multiple directions; Based on the constructed gray-level co-occurrence matrix, the corresponding texture features are extracted, and the texture features in different directions are weighted and summed through a multi-scale feature fusion method to obtain the texture complexity value of the image to be detected. Construct a steganalysis classifier and train it using a training set that includes a carrier image set and a dense image set. The image to be detected is input into the trained steganalysis classifier to obtain the probability values ​​of whether the image to be detected is a clean image or a dense image; Adjust the probability value output by the steganalysis classifier based on the calculated texture complexity value; Based on the adjusted probability value, and through preset judgment rules, it is determined whether the image to be detected is a cryptic image, thereby realizing steganalysis of the image to be detected.

[0009] Furthermore, the construction of corresponding gray-level co-occurrence matrices in multiple directions for the gray-level image specifically includes: The grayscale image is compressed from 256 grayscale levels to 16 grayscale levels to obtain the compressed grayscale image. A gray-level co-occurrence matrix is ​​constructed in multiple directions of the compressed grayscale image, including a 0° horizontal direction, a 45° upper right to lower left direction, a 90° vertical direction, and a 135° upper left to lower right direction.

[0010] Furthermore, the construction of the gray-level co-occurrence matrix specifically includes: For each direction Select the distance between pixel pairs Iterate through each pair of adjacent pixels in the compressed grayscale image and record the grayscale values ​​of the adjacent pixels. and ,in, Indicates grayscale level. The number of gray levels after compression; Statistics in each direction Below, the grayscale value is The pixel and its adjacent gray values The number of occurrences of pixel pairs is counted, and the statistical results are stored in the orientation. The corresponding gray-level co-occurrence matrix middle, Indicates in a given direction Below, grayscale value and The frequency of the combination.

[0011] Furthermore, the texture features include energy, contrast, entropy, correlation, homogeneity, and variance.

[0012] Furthermore, the energy, used to reflect the uniformity of image grayscale distribution and the coarseness of texture, is calculated using the following formula: in, Indicates in a given direction Below, grayscale value and The frequency of the combination; The number of gray levels after compression; Indicates direction The energy below; The contrast ratio measures the difference in grayscale values ​​between adjacent pixels in an image, reflecting the clarity of texture and the sharpness of edges. The calculation formula is as follows: in, Indicates direction Contrast ratio; The entropy, which characterizes the richness and randomness of image texture information, is calculated using the following formula: in, Indicates direction Entropy below; The correlation describes the linear similarity of gray levels in the row or column direction of the image, reflecting the directionality and regularity of the texture. The calculation formula is as follows: in, Indicates direction The correlation below; These are the mean gray values ​​of the rows and columns of the gray-level co-occurrence matrix, respectively. These are the standard deviations of the rows and columns of the gray-level co-occurrence matrix, respectively. Homogeneity measures the similarity of gray values ​​between adjacent pixels in an image, reflecting the smoothness of the texture. The calculation formula is as follows: in, Indicates direction Homogeneity; The variance, which measures the dispersion of image gray values ​​and reflects the centrality of texture gray value distribution, is calculated using the following formula: in, The mean gray level of the gray-level co-occurrence matrix; Indicates direction The variance below.

[0013] Furthermore, the step of using a multi-scale feature fusion method to weighted summation of texture features from different directions to obtain the texture complexity value of the image to be detected specifically includes: Calculate the average value of each texture feature parameter in four directions, where the average value is the arithmetic mean of the values ​​of the feature in the four directions; Calculate the 2-norm of the characteristic parameters in each direction and their corresponding average values; Calculate the sum of the 2-norms of the feature parameters in the four directions, that is, sum the 2-norms of each feature in the four directions; Calculate the weight of the feature parameter in each direction, where the weight is the ratio of the 2-norm of a single direction to the total norm; Based on the calculated weights, a weighted sum is performed to obtain the fusion result of each texture feature, where the feature values ​​in each direction are weighted according to their weights; Based on the weighted sum of the six texture features, the overall texture complexity value is calculated using the following formula: in, This represents the texture complexity value. , , , , , These are preset weighting coefficients; , , , , , This refers to the fused texture features.

[0014] Furthermore, the steganalysis classifier is either a classifier based on hand-designed features or a classifier based on deep learning. The classifier based on hand-designed features is an ensemble classifier based on Fisher's linear discriminant method, including: The subclassifier building module is used to generate several subclassifiers based on Fisher's linear discriminant criterion. Each subclassifier randomly extracts several dimensional features from the high-dimensional space with a set step size as its own training feature set. The sub-classifier selection module is used to train all the constructed sub-classifiers and select the sub-classifier with the smallest training error to form the final ensemble model. The feature input module is used to receive high-dimensional feature vectors of the data to be detected, providing detection input for each sub-classifier; The voting decision module is used to aggregate the output results of each sub-classifier during actual detection and to obtain the final classification decision through a voting mechanism. The deep learning-based classifier includes: The input layer is used to receive the feature vector of the image to be detected. Convolutional layers are used to extract local features from images; Pooling layers are used to pool the feature maps output by convolutional layers. Fully connected layers are used to map the features extracted by the convolutional layers to the final classification output; The output layer is used to output the probability values ​​of whether the image to be detected is a clean image or a dense image.

[0015] Furthermore, the training set includes a carrier image set and a dense image set, wherein: The carrier image set includes multiple natural images that have not undergone stegatization, serving as negative samples for the classifier; The cryptic image set includes images generated by embedding secret data into each natural image in the carrier image set. The secret data is embedded using the HILL and S-UNIWARD steganography methods with an embedding rate ranging from 0.1 to 0.5 bpp. The cryptic image set serves as positive samples for the classifier.

[0016] Furthermore, adjusting the probability value output by the steganalysis classifier based on the calculated texture complexity value specifically includes: Obtain the probability value of a clean image from the classifier output. ; The probability value of the clean image By texture complexity value Perform exponential calculations to obtain the adjusted probability value of a clean image. .

[0017] Furthermore, the step of determining whether the image to be detected is a dense image based on the adjusted probability value and a preset judgment rule specifically includes: Obtain the dense image probability value output by the classifier. Compared with the adjusted clean image probability value ; like < If so, the image to be detected is determined to be a dense image; like > If so, the image to be detected is determined to be a clean image; like = Then, a random number between 0 and 1 is generated using a random number generator, and this random number is compared with 0.5. If the random number is less than 0.5, the image to be detected is determined to be a dense image; If the random number is greater than or equal to 0.5, the image to be detected is determined to be a clean image.

[0018] Compared with the prior art, the present invention has the following advantages: (1) Existing steganalysis methods fail to adequately adapt to the unique characteristics of carrier-selective steganography, especially when dealing with images with high texture complexity, resulting in unsatisfactory detection performance. Traditional steganalysis methods often neglect the influence of image complexity during carrier-selective steganography, leading to insufficient assessment of the suspiciousness of these images in steganalysis. To address this issue, this invention proposes a texture complexity-based steganalysis method. This method calculates texture complexity using a multi-scale fusion approach, combining gray-level co-occurrence matrices in four directions to extract six independent texture features: energy, contrast, entropy, correlation, homogeneity, and variance. By weighted fusion of these features, the limitations of a single direction are avoided, and the accuracy of complexity assessment and sensitivity to carrier-selective steganographic images are significantly improved.

[0019] (2) Existing steganalysis methods often neglect the impact of image texture complexity on steganalysis detection, resulting in low accuracy for high-complexity images. This is especially true for steganography attacks using carrier selection methods, where the rich texture information of these images makes the steganalysis marks even more difficult to detect. To address this issue, this invention innovatively designs a probability adjustment mechanism based on texture complexity. By performing an exponential operation between the clean image probability value output by the classifier and the image's texture complexity value, the adjusted probability is obtained, effectively reducing the cleanliness determination probability of high-texture-complexity images. This improves the detection sensitivity of dense images, making the detection results more accurate and targeted.

[0020] (3) Existing steganalysis methods lack sufficient adaptability when dealing with special scenarios involving carrier-selective steganalysis, resulting in complex technical implementation and high difficulty in deployment. To improve the versatility and compatibility of this invention, the design fully considers the compatibility with existing mainstream steganalysis methods and proposes additional steps for texture complexity calculation and probability adjustment mechanisms. This method does not require reconstructing the existing classifier architecture; it can be seamlessly integrated into the existing classifier simply by adding texture complexity calculation and probability adjustment steps in the detection stage. This design greatly reduces the difficulty of technical implementation, making the deployment of this technical solution more convenient in practical applications, and ensuring compatibility with existing technologies, thus possessing stronger operability and universality.

[0021] (4) The detection performance of existing steganalysis methods is often limited by image quality and feature extraction capabilities, especially when the image complexity is high, resulting in severe feature redundancy and information loss, which affects the accuracy of the analysis. To solve this problem, this invention uses a weighted summation of texture features and a multi-scale fusion method to comprehensively evaluate texture features from different directions, avoiding feature redundancy and reducing information loss. This method effectively integrates image texture information while significantly improving the accuracy of image complexity assessment and more accurately determining whether an image is dense.

[0022] (5) In the prior art, steganalysis methods usually require manual feature design or rely on deep learning algorithms to automatically extract image features. However, these methods often fail to achieve the best balance between feature selection and adjustment. To overcome this problem, this invention solves the problem of insufficient processing of complex texture information in existing methods by using precise texture complexity measurement and multi-scale feature fusion. By employing six different texture features and utilizing a weighted fusion mechanism, this invention can fully capture the detailed changes in the image, improve the detection capability of complex images in steganalysis, and avoid detection blind spots caused by improper feature selection. Attached Figure Description

[0023] Figure 1 This is a flowchart of the image steganalysis method according to an embodiment of the present invention; Figure 2 Pixel pairs in embodiments of the present invention and A diagram illustrating the relationships between them; Figure 3 This is a schematic diagram of the gray-level co-occurrence matrix of a gray-level image in four directions according to an embodiment of the present invention; Figure 4 This is a comparison diagram of the texture complexity distribution of images selected by different carrier selection methods in embodiments of the present invention; Figure 5This is a visual representation of the images selected by different carrier selection methods in embodiments of the present invention; Figure 6 This is a comparison diagram of the steganalysis process based on hand-designed features in an embodiment of the present invention; Figure 7 This is a comparison diagram of the deep learning-based steganalysis process according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example 1: The core objective of this invention is to provide an image steganalysis method for carrier-selective steganography. Specific objectives include: 1. Constructing a multi-scale fusion image texture complexity measurement model to achieve accurate quantification of image texture richness; 2. Designing a probability adjustment mechanism based on texture complexity to optimize the classifier's decision logic and enhance detection sensitivity for highly complex, dense images; 3. Ensuring the method's compatibility and universality, adapting it to existing mainstream steganalysis classifiers; 4. Improving detection stability under different steganalysis methods and embedding rates, reducing detection errors. The proposed steganalysis method's workflow based on a manually designed feature steganalysis framework is as follows: Figure 6 As shown in (b); the process of the proposed steganalysis method on the deep learning-based steganalysis framework is as follows: Figure 7 As shown in (b).

[0026] The core technical idea of ​​this invention is to accurately calculate the texture complexity of an image using a gray-level co-occurrence matrix (GLCM) multi-scale fusion method, and then dynamically adjust the "clean" determination probability of the classifier using this complexity value. This makes the determination of dense text in high-complexity images more targeted, thereby improving the detection performance of carrier-selective steganography. Figure 1 As shown, the specific steps include: Step S1: Acquire the image to be detected and convert it into a grayscale image; Step S2: For the grayscale image, construct the corresponding grayscale co-occurrence matrix in multiple directions; A multi-scale fusion texture complexity evaluation method based on gray-level co-occurrence matrix (GLCM) is designed. Six key feature parameters are selected: energy (E), contrast (C), entropy (Ent), correlation (Cor), homogeneity (IDM), and variance (Var). These parameters comprehensively characterize the uniformity, detail richness, information entropy value, and spatial correlation of image texture, ensuring the comprehensiveness and accuracy of complexity evaluation.

[0027] The gray-level co-occurrence matrix (GLCM) is a core tool for describing the spatial distribution characteristics of image texture. Essentially, it statistically analyzes the probability of gray-level values ​​appearing in pairs of adjacent pixels at specific directions and distances. Figure 2 As shown. To comprehensively capture the texture information of an image in different directions, this invention selects four typical directions: 0° (horizontal direction), 45° (upper right to lower left direction), 90° (vertical direction), and 135° (upper left to lower right direction), while simultaneously setting the pixel pair distance. (Neighboring pixels), construct the gray-level co-occurrence matrix of the image, such as Figure 3 As shown.

[0028] Considering the balance between computational efficiency and evaluation accuracy, the image grayscale levels are compressed from the original 256 levels to 16 levels before constructing the matrix. This reduces the matrix dimension and computational complexity while preserving sufficient texture detail. For a grayscale image I of size M×N to be detected, with compressed grayscale levels L=16, the grayscale co-occurrence matrix constructed under direction θ is... ,in Indicates the grayscale value in the image i The pixels and their neighbors (direction) ,distance Grayscale value j The probability of occurrence of a pixel pair.

[0029] Step S3: Extract the corresponding texture features based on the constructed gray-level co-occurrence matrix, and perform weighted summation of texture features in different directions using a multi-scale feature fusion method to obtain the texture complexity value of the image to be detected; Six feature parameters—energy (E), contrast (C), entropy (Ent), correlation (Cor), homogeneity (IDM), and variance (Var)—are extracted from the gray-level co-occurrence matrices in four directions. The physical meaning and calculation formula of each parameter are as follows: Energy (E): Reflects the uniformity of grayscale distribution and the fineness of texture in an image. The more regular and uniform the texture, the higher the energy value; conversely, the more complex and disordered the texture, the lower the energy value. The calculation formula is: in, Indicates in a given direction Below, grayscale value and The frequency of the combination; The number of gray levels after compression; Indicates direction The energy below; Contrast Ratio (C): Measures the difference in grayscale values ​​between adjacent pixels in an image, reflecting the sharpness of textures and edges. The richer the texture details and the sharper the edges, the higher the contrast ratio; conversely, a smooth texture with no obvious edges results in a lower contrast ratio. The calculation formula is: in, Indicates direction Contrast ratio; Entropy (Ent): Characterizes the richness and randomness of image texture information. The more complex the texture, the more diverse and randomly distributed the gray-level combinations, the greater the entropy value; the simpler the texture and the more concentrated the gray-level distribution, the smaller the entropy value. When this condition is met, the value of this term is 0 (to avoid the meaninglessness of logarithmic operations), and the calculation formula is: in, Indicates direction Entropy below; Correlation (Cor): Describes the linear similarity of gray levels in rows or columns of an image, reflecting the directionality and regularity of the texture. A correlation value closer to 1 or -1 indicates a stronger linear relationship between pixels, and a more pronounced directionality and regularity in the texture; a value closer to 0 indicates a weaker linear relationship between pixels, and a more complex and disordered texture. The calculation formula is: in, Indicates direction The correlation below; These are the mean gray values ​​of the rows and columns of the gray-level co-occurrence matrix, respectively. These are the standard deviations of the rows and columns of the gray-level co-occurrence matrix, respectively. Homogeneity (Inverse Moment of Difference, IDM): Measures the similarity of gray values ​​between adjacent pixels in an image, reflecting the smoothness of the texture. The smaller the gray value difference between adjacent pixels, the smoother the texture, and the higher the homogeneity value; the larger the gray value difference between adjacent pixels, the coarser the texture, and the lower the homogeneity value. The calculation formula is: in, Indicates direction Homogeneity; Variance (Var): Measures the dispersion of image grayscale values, reflecting the centrality of grayscale distribution in texture. The more dispersed the grayscale distribution and the more complex the texture, the larger the variance value; the more concentrated the grayscale distribution and the simpler the texture, the smaller the variance value. When this condition is met, the value of this term is 0 (to avoid meaningless calculations), and the calculation formula is: in, The mean gray level of the gray-level co-occurrence matrix; Indicates direction The variance below.

[0030] To integrate texture information from four directions and avoid the limitations of features from a single direction, a multi-scale fusion (MSF) method is used to perform a weighted summation of the same feature parameters. The specific process is as follows: Calculate the average value of the same characteristic parameter in four directions: For energy parameters, the formula for calculating the average value is as follows: Consistent with the energy parameter, it is used to characterize the overall level of this characteristic parameter.

[0031] Calculate the 2-norm of each directional feature parameter and its corresponding average: the 2-norm reflects the degree of difference between a single directional feature and the overall feature level; the greater the difference, the larger the norm value. For the energy parameter, the 2-norm of direction θ is: in, For direction Elements of the upper energy parameter matrix, Average value The elements of the matrix; the 2-norm of the contrast, entropy, correlation, homogeneity, and variance parameters. , , , and The calculation method is consistent with the energy parameters.

[0032] Calculate the sum of the 2-norms of the four directional characteristic parameters: For the energy parameter, the total norm is: The total norm of contrast, entropy, correlation, homogeneity, and variance parameters. , , , and The calculation method is consistent with the energy parameters.

[0033] Calculate the fusion weights for feature parameters in each direction: the weights are the proportion of the norm of a single direction to the total norm. For energy parameters: Ensure that directional features with significant differences receive higher weights to improve the discriminative power of the fused features. The calculation methods for the fusion weights of contrast, entropy, correlation, homogeneity, and variance parameters are consistent with those for the energy parameter.

[0034] The weighted summation yields the fused feature parameters: For the energy parameter, the fusion result is: The calculation methods for the fusion results of contrast, entropy, correlation, homogeneity, and variance parameters are consistent.

[0035] To quantify the overall texture complexity of the image, the six fused feature parameters are weighted and summed to obtain the texture complexity value. T The weight allocation was determined through 10 independent random experiments: For each experiment, 1000 images were randomly selected from both the BOSSbasever.1.01 and BOWS2 datasets, and the average of the six feature parameters was calculated. The proportion of each feature's average to the sum of the total averages was used as the initial weight, and the average of the 20 experiments was taken as the final weight, ensuring the rationality and robustness of the weight allocation. The final texture complexity calculation formula is as follows: in, This represents the texture complexity value. , , , , , These are preset weighting coefficients; , , , , , This refers to the fused texture features. T The larger the value, the more complex the image texture, and the more likely it is to be chosen as a carrier by a stegan. For example... Figure 4 , Figure 5 As shown, in the images selected by the distortion minimization-based carrier selection method (Joint), most images have high texture complexity values ​​T, and their distribution is relatively concentrated in the high texture complexity range. This indicates that the distortion minimization method preferentially selects images with high texture complexity as steganographic carriers, and these images are more difficult to detect by traditional steganalysis methods. In contrast, the texture complexity of the images selected by the similarity-based carrier selection method (Similarity) is more dispersed, and many images have low texture complexity. This indicates that although the similarity selection method also considers image similarity, it does not particularly emphasize texture complexity, resulting in a large difference in complexity among the selected carrier images.

[0036] In addition, through Figure 5 It can be intuitively seen that in carrier-selective steganography, the images selected by the distortion minimization method exhibit more obvious high-complexity texture features. For these images, the method of this invention can more effectively identify dense images through accurate calculation of texture complexity. In contrast, traditional methods perform poorly in the detection of high-complexity images and are difficult to identify effectively.

[0037] Step S4: Construct a steganalysis classifier using a training set that includes a carrier image set and a dense image set; The steganalysis classifier in this embodiment includes a classifier based on hand-designed features or a classifier based on deep learning; Classifiers based on hand-designed features: SRMQ1, TLBP and SPAM were selected. SPAM captures pixel perturbations caused by steganography embedding by modeling the autocorrelation statistics of the spatial neighborhood of image pixels. SRMQ1 uses multiple sub-models to balance detection performance and computational efficiency. TLBP obtains image residuals through high-order differential filters and extracts second-order co-occurrence matrix features, resulting in excellent detection performance.

[0038] Deep learning-based classifiers: Ye-Net and Zhu-Net were selected. Ye-Net introduces truncated linear units (TLU) to improve the signal-to-noise ratio; Zhu-Net combines depthwise separable convolutions with cross-layer residual connections to enhance feature representation capabilities and performs well in low embedding rate scenarios.

[0039] Classifier Training: A training set is constructed, including a carrier image set and a hidden image set. The carrier images are from the BOSSbasever.1.01 and BOWS2 datasets; the hidden images are generated by embedding secret data into the carrier images, using HILL and S-UNIWARD steganography methods with embedding rates of 0.1~0.5 bpp. For classifiers based on hand-designed features, features are extracted from the training set images to construct a feature set, and an ensemble classifier (consisting of multiple sub-classifiers, with results output through a voting mechanism) is trained. For deep learning-based classifiers, the carrier-hidden image pair is used directly to train the network.

[0040] Performance evaluation index: Select detection error The core evaluation indicator is calculated using the following formula: in, (False alarm probability) is the probability of misclassifying a clean image as a dense image. (Missed detection probability) is the probability of misclassifying a dense image as a clean image. The value range is (0,1), and the closer it is to 0, the better the detection performance.

[0041] Step S5: Input the image to be detected into the trained steganalysis classifier to obtain the probability values ​​of whether the image to be detected is a clean image or a dense image; Image detection: The preprocessed image to be detected is input into a trained classifier to obtain the probability that the image is classified as "clean" (without secret data). The probability of "containing secrets" (having secret data) ,satisfy Among them, the ensemble classifier To determine the proportion of "clean" subclassifiers to the total number of subclassifiers, To determine the proportion of "dense" data; a deep learning classifier. and It is calculated by the softmax function of the network output layer.

[0042] Step S6: Adjust the probability value output by the steganalysis classifier based on the calculated texture complexity value; Probability adjustment: for The adjusted probability is obtained by performing an exponential operation with the texture complexity value T. .because And T>0, after exponentiation This reduces the effective probability of a "clean" determination and strengthens the suspicion of dense content in highly complex images. The core logic of this design is that the higher the texture complexity of an image, the greater the probability that it will be chosen as a carrier by a stegan, so it is necessary to reduce the confidence of its "clean" determination and increase the sensitivity of dense content detection.

[0043] Step S7: Based on the adjusted probability value, determine whether the image to be detected is a cryptic image according to the preset judgment rules, thereby realizing steganalysis of the image to be detected, specifically including: Obtain the dense image probability value output by the classifier. Compared with the adjusted clean image probability value ; like < If so, the image to be detected is determined to be a dense image; like > If so, the image to be detected is determined to be a clean image; like = Then, a random number between 0 and 1 is generated using a random number generator, and this random number is compared with 0.5. If the random number is less than 0.5, the image to be detected is determined to be a dense image; If the random number is greater than or equal to 0.5, the image to be detected is determined to be a clean image.

[0044] Example 2: This embodiment provides an image steganalysis system for carrier-selective steganography, including: Image preprocessing module: This module receives the image to be detected and converts it into a grayscale image, while compressing the grayscale levels from 256 to 16. This module can effectively reduce the complexity of image data and simplify the subsequent texture feature extraction calculation process.

[0045] Gray-level co-occurrence matrix construction module: Based on the image preprocessing module, this module constructs corresponding gray-level co-occurrence matrices in four directions: 0°, 45°, 90°, and 135°. This module is used to capture the texture information of the image and generate basic data for subsequent texture feature extraction by statistically analyzing the gray-level relationships between adjacent pixels.

[0046] Texture feature extraction module: Based on the constructed gray-level co-occurrence matrix, this module extracts six texture features from the image, including energy, contrast, entropy, correlation, homogeneity, and variance. It then uses a multi-scale fusion method to weightedly sum the texture features in each direction, ultimately generating the overall texture complexity value T of the image.

[0047] Classifier Training Module: A steganalysis classifier is trained using a training set containing both carrier images and hidden images. The carrier images in the training set are natural images without steganalysis, while the hidden images are images with hidden information embedded using steganography. This classifier learns the statistical features of the images to determine whether an input image is hidden.

[0048] Probability Adjustment Module: Based on the calculated image texture complexity value T, this module dynamically adjusts the clean image probability output by the classifier. It obtains the adjusted probability by performing an exponential operation on the clean image probability. This adjustment mechanism helps improve the assessment of the density suspicion of high-texture-complexity images, making steganalysis detection of high-complexity images more targeted.

[0049] Judgment Module: By comparing the probability of dense images with the output of the classifier, the module determines whether the image to be detected is a dense image based on the adjusted probability value and the magnitude of the probability of dense images.

[0050] Output module: Outputs the final judgment result, displaying whether the image to be detected is a dense image or a clean image. This module provides users with the judgment result to assist in steganalysis of the image to be detected.

[0051] If the aforementioned functions are implemented as 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 this invention, 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. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An image steganalysis method for carrier-selective steganography, characterized in that, Includes the following steps: Acquire the image to be detected and convert it to a grayscale image; For the grayscale image, construct corresponding grayscale co-occurrence matrices in multiple directions; Based on the constructed gray-level co-occurrence matrix, the corresponding texture features are extracted, and the texture features in different directions are weighted and summed through a multi-scale feature fusion method to obtain the texture complexity value of the image to be detected. Construct a steganalysis classifier and train it using a training set that includes a carrier image set and a dense image set. The image to be detected is input into the trained steganalysis classifier to obtain the probability values ​​of whether the image to be detected is a clean image or a dense image; Adjust the probability value output by the steganalysis classifier based on the calculated texture complexity value; Based on the adjusted probability value, and through the preset judgment rules, it is determined whether the image to be detected is a cryptic image, thereby realizing the steganalysis of the image to be detected.

2. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The construction of corresponding gray-level co-occurrence matrices in multiple directions for the gray-level image specifically includes: The grayscale image is compressed from 256 grayscale levels to 16 grayscale levels to obtain the compressed grayscale image. A gray-level co-occurrence matrix is ​​constructed in multiple directions of the compressed grayscale image, including a 0° horizontal direction, a 45° upper right to lower left direction, a 90° vertical direction, and a 135° upper left to lower right direction.

3. The image steganalysis method for carrier-selective steganography according to claim 2, characterized in that, The construction of the gray-level co-occurrence matrix specifically includes: For each direction Select the distance between pixel pairs Iterate through each pair of adjacent pixels in the compressed grayscale image and record the grayscale values ​​of the adjacent pixels. and ,in, Indicates grayscale level. The number of gray levels after compression; Statistics in each direction Below, the grayscale value is The pixel and its adjacent gray values The number of occurrences of pixel pairs is counted, and the statistical results are stored in the orientation. The corresponding gray-level co-occurrence matrix middle, Indicates in a given direction Below, grayscale value and The frequency of the combination.

4. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The texture features include energy, contrast, entropy, correlation, homogeneity, and variance.

5. The image steganalysis method for carrier-selective steganography according to claim 4, characterized in that, The energy, used to reflect the uniformity of image grayscale distribution and the coarseness of texture, is calculated using the following formula: in, Indicates in a given direction Below, grayscale value and The frequency of the combination; The number of gray levels after compression; Indicates direction The energy below; The contrast ratio measures the difference in grayscale values ​​between adjacent pixels in an image, reflecting the clarity of texture and the sharpness of edges. The calculation formula is as follows: in, Indicates direction Contrast ratio; The entropy, which characterizes the richness and randomness of image texture information, is calculated using the following formula: in, Indicates direction Entropy below; The correlation describes the linear similarity of gray levels in the row or column direction of the image, reflecting the directionality and regularity of the texture. The calculation formula is as follows: in, Indicates direction The correlation below; These are the mean gray values ​​of the rows and columns of the gray-level co-occurrence matrix, respectively. These are the standard deviations of the rows and columns of the gray-level co-occurrence matrix, respectively. Homogeneity measures the similarity of gray values ​​between adjacent pixels in an image, reflecting the smoothness of the texture. The calculation formula is as follows: in, Indicates direction Homogeneity; The variance, which measures the dispersion of image gray values ​​and reflects the centrality of texture gray value distribution, is calculated using the following formula: in, The mean gray level of the gray-level co-occurrence matrix; Indicates direction The variance below.

6. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The step of obtaining the texture complexity value of the image to be detected by weighted summation of texture features in different directions through a multi-scale feature fusion method specifically includes: Calculate the average value of each texture feature parameter in four directions, where the average value is the arithmetic mean of the values ​​of the feature in the four directions; Calculate the 2-norm of the characteristic parameters in each direction and their corresponding average values; Calculate the sum of the 2-norms of the feature parameters in the four directions, that is, sum the 2-norms of each feature in the four directions; Calculate the weight of the feature parameter in each direction, where the weight is the ratio of the 2-norm of a single direction to the total norm; Based on the calculated weights, a weighted sum is performed to obtain the fusion result of each texture feature, where the feature values ​​in each direction are weighted according to their weights; Based on the weighted sum of the six texture features, the overall texture complexity value is calculated using the following formula: in, This represents the texture complexity value. , , , , , These are preset weighting coefficients; , , , , , This refers to the fused texture features.

7. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The steganalysis classifier is either a classifier based on hand-designed features or a classifier based on deep learning. The classifier based on hand-designed features is an ensemble classifier based on Fisher's linear discriminant method, including: The subclassifier building module is used to generate several subclassifiers based on Fisher's linear discriminant criterion. Each subclassifier randomly extracts several dimensional features from the high-dimensional space with a set step size as its own training feature set. The sub-classifier selection module is used to train all the constructed sub-classifiers and select the sub-classifier with the smallest training error to form the final ensemble model. The feature input module is used to receive high-dimensional feature vectors of the data to be detected, providing detection input for each sub-classifier; The voting decision module is used to aggregate the output results of each sub-classifier during actual detection and to obtain the final classification decision through a voting mechanism. The deep learning-based classifier includes: The input layer is used to receive the feature vector of the image to be detected. Convolutional layers are used to extract local features from images; Pooling layers are used to pool the feature maps output by convolutional layers. Fully connected layers are used to map the features extracted by the convolutional layers to the final classification output; The output layer is used to output the probability values ​​of whether the image to be detected is a clean image or a dense image.

8. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The training set includes a carrier image set and a dense image set, wherein: The carrier image set includes multiple natural images that have not undergone stegatization, serving as negative samples for the classifier; The cryptic image set includes images generated by embedding secret data into each natural image in the carrier image set. The secret data is embedded using the HILL and S-UNIWARD steganography methods with an embedding rate ranging from 0.1 to 0.5 bpp. The cryptic image set serves as positive samples for the classifier.

9. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The step of adjusting the probability value output by the steganalysis classifier based on the calculated texture complexity value specifically includes: Obtain the probability value of a clean image from the classifier output. ; The probability value of the clean image By texture complexity value Perform exponential calculations to obtain the adjusted probability value of a clean image. .

10. The image steganalysis method for carrier-selective steganography according to claim 1, characterized in that, The determination of whether the image to be detected is a dense image based on the adjusted probability value and a preset judgment rule specifically includes: Obtain the dense image probability value output by the classifier. Compared with the adjusted clean image probability value ; like < If so, the image to be detected is determined to be a dense image; like > If so, the image to be detected is determined to be a clean image; like = Then, a random number between 0 and 1 is generated using a random number generator, and this random number is compared with 0.

5. If the random number is less than 0.5, the image to be detected is determined to be a dense image; If the random number is greater than or equal to 0.5, the image to be detected is determined to be a clean image.

Images