Color image cross-domain information steganography detection method and device, medium and program product

Abstract

The invention discloses a color image cross-domain information steganography detection method and device, a medium and a program product. The method comprises the following steps: taking a color image steganography analysis model UCNet after pre-training on a source domain as a feature extractor part; a self-adaptive weighting and confidence filtering mechanism is introduced to carry out weighting processing on the extracted features, and target domain input data is divided into HCS and common samples; constructing an adaptive alignment library to store the feature tag pair of each training round; using three loss functions of CE, DSA and CLMD to jointly restrain the model to update and optimize; retaining feature extraction capability trained on a source domain by adopting a CE loss function; dual alignment of samples in a self-adaptive alignment library and samples in a target domain is realized by adopting a DSA loss function; adopting a CLMD loss function to realize sub-domain-level fine-grained alignment of category perception; and transmitting a to-be-detected image to the optimized model, and outputting a corresponding prediction label. According to the method, the robustness and generalization ability of the model in a color image cross-domain steganography detection task can be improved.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, and in particular to methods, devices, media and program products for cross-domain information steganalysis detection of color images, achieving category-aware subdomain-level fine-grained alignment. Background Technology

[0002] With the rapid development of digital image processing technology and the widespread adoption of network communication, images have become the most commonly used information carrier on social media. Among these, image steganography, as a key technology for achieving covert communication and ensuring information security, hides secret information by embedding it into the redundant space of the image carrier, and its application scenarios are constantly expanding and deepening. Correspondingly, to combat potential threats to the transmission of covert information, image steganography detection technology, as a crucial defensive measure, is also continuously evolving and strengthening.

[0003] In recent years, thanks to the powerful feature extraction and learning capabilities demonstrated by deep learning models such as deep neural networks, image steganalysis detection methods based on deep learning have rapidly emerged and become the mainstream technical approach in this field.

[0004] However, when deploying and applying the above models in practice, researchers and engineers often face a core challenge: there is often a significant distributional difference between the dataset used to train the detection model (hereinafter referred to as the source domain) and the real-world image dataset that the model actually needs to detect (hereinafter referred to as the target domain). The source domain is usually a standard dataset constructed in the laboratory, containing both stegated images and normal images without hidden information, which serve as the basic training samples for the model to learn the rules of steganalysis feature recognition; the target domain specifically includes image samples from real-world scenarios such as images uploaded by social media users and images transmitted in actual communications.

[0005] The differences in distribution between the source and target domains manifest themselves in the diversity of image sources, the heterogeneity of acquisition devices, the wide variation in image compression parameters, and complex post-processing operations (such as scaling, cropping, and filtering). These factors collectively lead to inconsistencies between source and target domain data in statistical characteristics, noise patterns, and potential embedding distortion features. Consequently, steganalysis models that perform well trained in the source domain often experience a significant decrease in detection performance (such as detection accuracy and robustness) when directly applied to target domain image detection. This insufficient cross-domain generalization ability has become one of the key bottlenecks and major challenges currently restricting the practical effectiveness of image steganalysis detection technology. Please refer to [reference needed]. Figure 3 Currently, the main methods for solving the cross-domain steganography detection problem include: 1. J-Net Model: To alleviate the performance degradation problem caused by cross-domain steganalysis in grayscale images, related research has conducted a comprehensive survey of cross-domain steganalysis and confirmed that this performance degradation stems from differences in texture complexity between datasets, i.e., differences in data distribution. Based on this, this study proposes the J-Net model. This model uses the pre-trained grayscale image steganalysis model Yedroudj-Net as the feature extractor part, and introduces JMMD (Joint Maximum Mean Discrepancy) from the UDA (Unsupervised Domain Adaptation) domain into the model's classifier part to align the distribution of output features from the source and target domains of the feature extractor part. By simultaneously optimizing the differences in edge distribution and conditional distribution, this model effectively alleviates the performance degradation problem of cross-domain steganalysis in grayscale images.

[0006] 2. RCDD-Net Model: Based on the J-Net model and employing UDA technology to alleviate the performance degradation problem of cross-domain steganalysis detection in grayscale images, the RCDD-Net model was proposed. This model uses a reliable steganalysis annotation strategy to re-steganize the unlabeled target domain data, using the original unlabeled data as the carrier sample and the re-steganized sample as the steganized sample, thus generating reliable pseudo-labels for the target domain. Simultaneously, it introduces CDD (Contrastive Domain Discrepancy) to achieve category-aware sub-domain distribution alignment, that is, aligning the source and target domains according to categories, thus forming aligned carrier and steganized sub-domains. This model also uses Yedroudj-Net as the feature extractor, achieving optimal detection performance among existing grayscale image cross-domain steganalysis detection methods by aligning the distributions of the source and target domains into sub-domains.

[0007] Although existing methods for cross-domain steganalysis detection in grayscale images have largely mitigated the performance degradation problem, their detection performance drops significantly when directly applied to cross-domain steganalysis detection in color images. This limitation stems primarily from the following key drawbacks: 1. Existing grayscale image cross-domain steganalysis detection methods all use grayscale image steganalysis models as feature extractors. They lack targeted design for extracting the intrinsic features of color images, and ignore the rich correlations between different color channels (RGB) of color images, unique color distribution patterns, and steganalysis noise cues that may be embedded in specific color channels or color spaces. This results in a serious lack of depth and specificity in extracting key features for color image steganalysis detection. 2. Among existing grayscale image cross-domain steganalysis detection methods, J-Net adopts a global alignment strategy, which brings all categories of samples in the source and target domains closer together in a general way. It fails to identify and adapt to the inherent differences in feature alignment difficulty between different samples in the source and target domains. The inherent feature distribution of some complex samples is inherently more complex and more susceptible to steganographic perturbations than that of simple samples, resulting in a significantly higher feature alignment difficulty for complex samples than for simple samples. Although RCDD-Net adopts a sub-domain alignment strategy, it still does not further distinguish the target domain samples, and its re-steganography strategy destroys the features of the color image itself. As a result, the model performs well in grayscale image cross-domain detection tasks, but its performance drops significantly in color image cross-domain detection tasks.

[0008] In summary, existing grayscale image cross-domain steganalysis detection methods are unable to fully mine and utilize the unique information of color images to achieve fine-grained data distribution alignment adapted to the color space, and also lack the ability to differentiate and adaptively process samples of different complexities. Ultimately, this severely restricts the robustness and generalization ability of the model in diverse color image cross-domain steganalysis detection tasks. Summary of the Invention

[0009] This invention provides a method, device, medium, and program product for detecting cross-domain information stegatology in color images, which effectively alleviates the problem of performance degradation of existing grayscale image cross-domain stegatology detection methods in color image cross-domain stegatology detection tasks.

[0010] In a first aspect, the present invention provides a method for detecting cross-domain information stegatology in color images, comprising: Using the color carrier image and its corresponding steganalysis image in the source domain as training data, the color image steganalysis model UCNet is pre-trained in the source domain. The pre-trained model is then used as the feature extractor to extract features from the source domain input data and the target domain input data. An adaptive weighting mechanism is introduced to perform sample-by-sample and dimension-by-dimensional weighting of the extracted features, and a confidence filtering mechanism is introduced to divide the target domain input data into HCS and ordinary samples. Construct an adaptive alignment library to store source domain feature label pairs and HCS feature label pairs for each training round; The model is updated and optimized by jointly constraining the three loss functions: CE, DSA, and CLMD. Specifically, the CE loss function is used to preserve the feature extraction capability trained on the source domain; the DSA loss function is used to achieve dual alignment between source domain samples and target domain samples, and between HCS and ordinary samples in the target domain; and the CLMD loss function is used to minimize the distance between similar subdomain samples and dissimilar subdomain samples between source domain samples and HCS, and maximize the distance between dissimilar subdomain samples, thereby achieving category-aware subdomain-level fine-grained alignment. The image to be detected is transmitted to the optimized model, which outputs the corresponding predicted label.

[0011] Optional feature weighting processing steps of the adaptive weighting mechanism specifically include: After receiving the source domain features fs and target domain features ft from the feature extractor, the weight values ​​for each dimension of each feature are calculated as follows: weighted_fs=Sigmoid(fc2(Relu(fc1(fs)))), weighted_ft=Sigmoid(fc2(Relu(fc1(ft)))), As can be seen from the above operations, in the adaptive weighting mechanism, after dimensionality reduction through a fully connected layer fc1, the ReLU activation function is used for calculation, and then the activation value is passed to another fully connected layer fc2 for dimensionality increase. Finally, the Sigmoid activation function is used to obtain the weight value, where weighted_fs is the source domain weight and weighted_ft is the target domain weight. Multiply the weight values ​​by the corresponding original features to obtain the weighted source domain features and the weighted target domain features.

[0012] Optionally, after the weighting process, the following steps may also be included: Lightweight single-layer fully connected layers are used to process the weighted source domain features and the weighted target domain features respectively, and output the original probability values ​​of the source domain Logits_s and the original probability values ​​of the target domain Logits_t.

[0013] Optional, the sample partitioning steps of the confidence filtering mechanism specifically include: At different training stages, the maximum value of the original prediction result Logits_t for each sample in the target domain input data is taken as the confidence score of that sample. Based on the current model state and target domain data distribution during the training phase, a dynamic confidence threshold is calculated using the following formula: , In the above formula, This represents the confidence threshold for the current training round. This represents the minimum value of the confidence threshold. This represents the maximum value of the confidence threshold. This is the current training round. It refers to the total number of training rounds. Indicates coercion Not less than 0.5; The confidence level of each sample in the target domain input data is compared with the dynamic confidence threshold. Samples with a confidence level greater than or equal to the dynamic confidence threshold are classified as HCS, and the remaining samples are classified as ordinary samples.

[0014] Optionally, an adaptive alignment library stores the source domain feature label pairs and HCS feature label pairs for each training epoch as follows: Adaptive alignment library ∈ R (Vs·ns+Vt·nt)×(256+1) , where R represents the real number field, Vs and Vt represent the sample storage ratio of the source domain and the target domain respectively, ns and nt represent the total number of samples in the source domain and the target domain respectively, and 256+1 represents the joint dimension of the feature vector and the label. An adaptive alignment library is updated using a first-in, first-out (FIFO) strategy.

[0015] Optionally, the steps for achieving double alignment using the DSA loss function include: The source domain samples in the adaptive alignment library and HCS are integrated together to form the adaptive alignment library sample set; the target domain sample data input into the model in the current training round constitutes the target domain sample set in the current training round. We introduce a DSA loss function based on LMMD improvement in UDA to quantify the feature distribution difference between the adaptive alignment library sample set and the target domain sample set in the current training round. By incorporating this distribution difference into the model optimization objective, we use backpropagation to iteratively reduce the DSA loss value and simultaneously narrow the distribution distance between samples in the adaptive alignment library and samples in the target domain, thus achieving dual alignment: cross-domain alignment between source domain samples and target domain samples, and intra-domain alignment between HCS and ordinary samples within the target domain.

[0016] Optionally, the step of using the CLMD loss function to minimize the distance between samples in the same subdomain and samples in different subdomains of the source domain and maximize the distance between samples in different subdomains includes: The similarity calculation method in the InfoNCE loss function of contrastive learning is replaced with the kernel calculation method of UDA, thus forming the CLMD loss function; The CLMD loss function is used to calculate the distance between samples of the same subdomain and samples of different subdomains in the current training round; Minimize the distance between samples from the same subdomain and maximize the distance between samples from different subdomains.

[0017] In a second aspect, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned method for detecting cross-domain information stegatology in a color image.

[0018] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned method for detecting cross-domain information stegatology in color images.

[0019] Fourthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method for detecting cross-domain information stegatology in color images.

[0020] One or more technical solutions provided by this invention have at least the following technical effects or advantages: This invention can alleviate the problem of performance degradation of existing grayscale image cross-domain steganalysis detection methods in color image cross-domain steganalysis detection tasks, and effectively improve the robustness and generalization ability of the model in color image cross-domain steganalysis detection tasks. The specific analysis is as follows: To address the shortcomings of existing grayscale image cross-domain steganalysis methods, which fail to fully exploit and utilize the unique information of color images to achieve fine-grained, color space-adaptive data distribution alignment, this invention uses a pre-trained color image steganalysis model UCNet as the feature extractor. This model fully considers the rich correlations between different color channels of a color image, unique color distribution patterns, and steganalytic noise cues that may be embedded in specific color channels or color spaces. It obtains source domain features and target domain features containing color image-specific information, facilitating fine-grained, color space-adaptive data distribution alignment.

[0021] To address the shortcomings of existing grayscale image cross-domain steganalysis detection methods, which lack the ability to differentiate and adaptively process samples of varying complexity, this invention first employs an adaptive weighting mechanism. This mechanism assigns corresponding weights to different feature dimensions of different samples, enhancing the weights of source domain features with the same distribution as the target domain while suppressing the weights of target domain features with significantly different distributions, thus achieving adaptive enhancement and suppression of cross-domain features. Secondly, a confidence filtering mechanism is used for unlabeled target domain samples, dividing them into HCS (Hardware Classification) and ordinary samples. This ensures the reliability of the high-confidence pseudo-labels of the HCS used for alignment, enabling subsequent category-aware sub-domain fine-grained alignment. Thirdly, an adaptive alignment library is constructed to store source domain feature label pairs and HCS feature label pairs for each training epoch. This fully models the global feature distribution of the target domain, reducing dependence on specific training epochs. Finally, three loss functions were used to jointly constrain model updates and optimizations in the alignment strategy. Specifically, the CE loss function was used to preserve the feature extraction capability trained on the source domain; the DSA loss function was used to achieve dual alignment between source domain samples and target domain samples, and between HCS samples and ordinary samples in the target domain that are relatively close to the distribution of the source domain; and the CLMD loss function was used to maximize the distance between source domain samples and samples of the same subdomain in the HCS, and to maximize the distance between samples of different subdomains, thereby enhancing the clustering of samples within the subdomain corresponding to each category during the alignment process and giving the model a clear decision boundary. Attached Figure Description

[0022] Figure 1 This is a simplified flowchart of a color image cross-domain information steganalysis detection method according to the present invention; Figure 2 This is a processing logic diagram of a color image cross-domain information steganalysis detection method according to the present invention; Figure 3 This diagram illustrates a comparison of alignment strategies between the color image cross-domain information steganalysis detection method of this invention and existing grayscale image cross-domain steganalysis detection methods. Detailed Implementation

[0023] This invention provides a method, device, medium, and program product for detecting cross-domain information stegatology in color images, which effectively alleviates the problem of performance degradation of existing grayscale image cross-domain stegatology detection methods in color image cross-domain stegatology detection tasks.

[0024] To effectively improve the robustness and generalization ability of the model in cross-domain steganalysis detection of color images, the overall approach of this invention is as follows: First, the color image steganalysis model UCNet (Universal Color image steganalysis Network) is pre-trained on the source domain, and the pre-trained model is used as the feature extractor part of this method.

[0025] Secondly, an adaptive weighting and confidence filtering mechanism is introduced to perform sample-by-sample and dimension-by-dimensional weighting on the extracted features, and to divide the target domain input data into HCS (High Confidence Sample) and ordinary samples based on confidence, so as to provide a data foundation for subsequent subdomain-level fine-grained alignment of category perception.

[0026] Furthermore, an adaptive alignment library is constructed to store source domain feature label pairs and HCS feature label pairs for each training round, providing support for subsequent category-aware subdomain-level fine-grained alignment between the source and target domains.

[0027] Finally, three loss functions were used to jointly constrain model updates and optimizations in the alignment strategy, thereby achieving category-aware, fine-grained subdomain-level alignment. The CE (Cross Entropy) loss function was employed to preserve the feature extraction capabilities trained on the source domain; the DSA (Dual Subdomain Alignment) loss function was used to achieve dual alignment between source domain samples and target domain samples, and between HCS and ordinary samples in the target domain; furthermore, the CLMD (Contrastive Learning Maximum Discrepancy) loss function was used to measure the distance between source domain samples and HCS samples in the same subdomain (i.e., subdomains of the same class) and different subdomains (i.e., subdomains of different classes) within each training epoch, minimizing the distance between samples in the same subdomain and maximizing the distance between samples in different subdomains.

[0028] To better understand, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described in this invention are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0029] Please refer to the following as well. Figure 1 and Figure 2 The present invention provides a method for detecting cross-domain information stegatology in color images (hereinafter referred to as the method), which includes the following steps: (1) Pre-training: Referring to the dataset settings in existing grayscale image cross-domain steganalysis detection methods, this method first divides the source domain and target domain into training and test sets, resulting in source domain training set, source domain test set, target domain training set, and target domain test set. Then, the color image steganalysis model UCNet is pre-trained on the source domain training set following the original hyperparameter configuration, and the pre-trained model is used as the feature extractor part of this method. After each training round, the source domain test set and target domain test set are used for testing, thereby verifying the performance degradation of color image cross-domain information steganalysis detection and ensuring that the model has sufficient feature extraction capability in the source domain.

[0030] (2) Data initialization and feature extraction: The feature extractor part of this method is initialized using the parameters obtained by the feature extraction module (excluding the classifier) ​​during the nth training round of the model in the pre-training process. A portion of samples are selected from the source domain training set as source domain input data, and a portion of samples are selected from the target domain test set as target domain input data. The sample size of the source domain input data is equal to the sample size of the target domain input data. The source domain input data is a "carrier image-carrier image" (hereinafter referred to as carrier / carrier image pair) with real labels, and the target domain input data is two unlabeled images. Subsequently, the above source domain input data and target domain input data are simultaneously transmitted to the feature extractor part to extract source domain features and target domain features.

[0031] (3) Adaptive feature weighting and confidence filtering: The source domain features and target domain features extracted in step (2) are weighted on a sample-by-sample and feature-by-feature-dimension basis. This can enhance the weight of source domain features with the same distribution as the target domain, while suppressing the weight of target domain features with a large difference from the source domain distribution, thus achieving adaptive enhancement and suppression of cross-domain features.

[0032] Specifically, after receiving the source domain features fs and target domain features ft from the feature extractor, the weight value for each dimension of each feature is first calculated. The specific operations are as follows: weighted_fs=Sigmoid(fc2(Relu(fc1(fs)))), weighted_ft=Sigmoid(fc2(Relu(fc1(ft)))), As can be seen from the above operations, in the adaptive weighting mechanism, after dimensionality reduction through a fully connected layer fc1, the ReLU activation function is used for calculation, and then the activation value is passed to another fully connected layer fc2 for dimensionality increase. Finally, the Sigmoid activation function is used to obtain the weight value, where weighted_fs is the source domain weight and weighted_ft is the target domain weight.

[0033] The calculated weight values ​​are multiplied by the corresponding original features to obtain the weighted source domain features and the weighted target domain features. Specifically, the source domain features fs are multiplied by the source domain weights weighted_fs to obtain the weighted source domain features, and the target domain features ft are multiplied by the target domain weights weighted_fs to obtain the weighted target domain features.

[0034] Since using multiple fully connected layers in the adaptive weighting mechanism may weaken the weighting effect and lead to unsatisfactory detection results, a lightweight single fully connected layer is used to process the weighted source domain features and the weighted target domain features respectively, and output the original probability value Logits_s of the source domain and the original probability value Logits_t of the target domain.

[0035] Since the true labels of the source domain input data are known, there is no need to perform special processing on the original predicted values ​​Logits_s of the source domain. The final predicted values ​​of the source domain can be calculated directly using the Softmax function and then input into the subsequent CE loss function to calculate the classification loss of the source domain.

[0036] Since the true labels of the target domain input data are unknown, alignment without labels can only achieve domain-level global alignment using the existing J-Net model, which cannot achieve superior detection performance. If the reliable steganalysis annotation strategy of the existing RCDD-Net model is used to obtain pseudo-labels for the target domain input data, the subsequent steganography operation may destroy the statistical characteristics of the target domain input data in color image steganalysis detection tasks; that is, the subsequent steganography operation may weaken the correlation between different channels of the color image. If the prediction results of the UCNet model (i.e., the predicted labels corresponding to the original probability values ​​Logits_t of the target domain) are directly used as pseudo-labels for the target domain input data, a large number of incorrectly predicted labels will appear, thus affecting the sub-domain alignment effect. Therefore, this method designs a confidence filtering mechanism. Through this filtering mechanism, HCSs are selected based on the model's own prediction results, and high-confidence pseudo-labels corresponding to HCSs are obtained, facilitating subsequent category-aware sub-domain fine-grained alignment.

[0037] The sample partitioning steps of the confidence filtering mechanism specifically include: (1) At different training stages, the maximum value of the original prediction result Logits_t for each sample in the target domain input data is taken as the confidence of that sample; (2) Based on the model state and target domain data distribution during the current training phase, calculate a dynamic confidence threshold. The calculation formula is as follows: , In the above formula, This represents the confidence threshold for the current training round; This represents the minimum confidence threshold, typically set to 0.0. This represents the maximum confidence threshold, which is differentiated based on the sample type of the input data in the target domain, including the carrier image (cover sample). Set to 0.9, for densely packed images (stego samples). The differential setting is set to 0.8, which is based on experimental observations: the mean confidence score of carrier images being correctly classified at the beginning is higher than that of dense carrier images. It refers to the total number of training rounds; The current training round satisfies the constraints. ; Indicates coercion Not less than 0.5, for The lower bound constraint is imposed, and the rationale for this constraint stems from the probability distribution characteristics of the binary classification task—the minimum confidence level is 0.5.

[0038] As can be seen from the above formula, with the number of training rounds Gradually increase from 0 to , The value of decreases monotonically from 1 to 0, driving the confidence threshold. from Towards It decreases monotonically and never falls below the lower limit of 0.5.

[0039] (3) The confidence level of each sample in the target domain input data is compared with the dynamic confidence threshold. Samples with a confidence level greater than or equal to the dynamic confidence threshold are classified as HCS, and the remaining samples are classified as ordinary samples. HCS has a similar distribution to the source domain input data and has relatively reliable high-confidence pseudo-labels.

[0040] (4) Construction of Adaptive Alignment Library: If the features and labels obtained in each training round are directly used to calculate the loss and perform alignment, there is a challenge: the HCS selected in step (3) may only represent the local optimal hypothesis of the UCNet model in the current training round, and cannot fully model the global feature distribution of the target domain. This will lead to the model over-reliance on the high-confidence pseudo-labels of the current training round. To this end, this method constructs a dynamically updated adaptive alignment library, which not only stores the source domain features, HCS features, and real labels of the source domain input data extracted in each training round in step (2), but also stores the high-confidence pseudo-labels of HCS obtained in different training rounds in step (3). Among them, the source domain features and the real labels of the source domain input data form source domain feature label pairs; the HCS features and the high-confidence pseudo-labels of HCS form HCS feature label pairs.

[0041] The adaptive alignment library stores the source domain feature label pairs and HCS feature label pairs for each training round as follows: Adaptive alignment library ∈ R (Vs·ns+Vt·nt)×(256+1) , where Vs and Vt represent the sample storage ratio of the source domain and the target domain, respectively, ns and nt represent the total number of samples in the source domain and the target domain, respectively, and 256+1 represents the joint dimension of the feature vector and the label.

[0042] To ensure timely updates of samples in the adaptive alignment library based on the current model state and reduce computational resource consumption, source domain feature label pairs and HCS feature label pairs are stored in a fixed ratio. The adaptive alignment library is updated using a first-in, first-out (FIFO) strategy: all source domain feature labels and HCS feature label pairs for the current training epoch are stored first; when a source domain partition or target domain partition reaches its storage capacity, the oldest sample in that partition is discarded. This strategy promotes the modeling of the global distribution pattern of the target domain, reduces the model's dependence on specific training epochs, and ensures that the library always contains the most reliable high-confidence pseudo-labels from the current training process.

[0043] (5) Joint Constraint Alignment: In the classifier part of the model, in order to retain the feature extraction capability trained in the source domain while achieving category-aware subdomain-level fine-grained alignment, three loss functions are used to jointly constrain model updates and optimizations in the alignment strategy. Among them, the CE loss function plays a role in retaining the feature extraction capability (i.e., model classification capability) trained during the pre-training process. For example... Figure 3 As shown, category-aware subdomain-level fine-grained alignment is mainly achieved using the DSA loss function and the CLMD loss function. The specific descriptions of these three loss functions are as follows: First, the CE loss function calculated using the source domain input data ensures that the feature extractor always has sufficient feature extraction capability obtained from supervised training (with real labels) in the source domain.

[0044] Secondly, based on the DSA loss function calculated using feature label pairs stored in the adaptive alignment library, dual alignment is achieved between source domain samples and target domain samples, and between HCS and ordinary samples in the target domain. First, the source domain samples and HCS in the adaptive alignment library are formed into a set, and then dual alignment is performed with all target domain samples in the current training round, i.e., dual alignment between source domain samples and target domain samples, and between HCS and ordinary samples in the target domain. The DSA loss function is an improvement based on LMMD (Local Maximum Mean Discrepancy) in UDA. Its core improvement is that, when calculating distance, it does not use the distance between the respective subdomains of the source and target domains in traditional LMMD, but rather the distance between samples in the adaptive alignment library and all target domain samples. The advantage of this improvement is that by performing a union operation between HCS in the target domain with the source domain samples, it can perform subdomain alignment between source domain samples and target domain samples, and also allow HCS in the target domain with similar distributions to the source domain to perform subdomain alignment with ordinary samples in the target domain, thereby optimizing the alignment effect.

[0045] The steps for achieving dual alignment using the DSA loss function specifically include: integrating the source domain samples and HCS from the adaptive alignment library to form an adaptive alignment library sample set; inputting the target domain sample data into the model in the current training round to form the target domain sample set for the current training round; introducing the DSA loss function to quantify the feature distribution difference between the adaptive alignment library sample set and the target domain sample set in the current training round; incorporating this distribution difference into the model optimization objective; using backpropagation to iteratively reduce the DSA loss value; and simultaneously narrowing the distribution distance between the source domain samples and the target domain samples, as well as between the HCS and ordinary samples in the target domain, thereby achieving dual alignment.

[0046] Thirdly, the DSA loss function focuses on fusing similar subdomains (subdomains of the same category) and separating dissimilar subdomains (subdomains of different categories) within different domains, but neglects the local discriminativeness within subdomains. To overcome the limitation that the DSA loss function cannot achieve clustering between source domain samples and HCS, and can only separate dissimilar subdomain samples, resulting in the model lacking a clear decision boundary, this method replaces the similarity calculation method in the InfoNCE loss function of contrastive learning with the kernel calculation method commonly used in UDA, thus forming the CLMD loss function in this method. The CLMD loss function is used to calculate the distance between similar subdomain samples and dissimilar subdomain samples between the source domain samples and HCS within the current training epoch. Then, the distance between samples of the same type of subdomain is minimized, and the distance between samples of different types of subdomain is maximized. That is, the distance between the source domain samples and the samples of the same type of subdomain in the HCS is brought closer to the maximum extent, while the distance between samples of different types of subdomain is pushed apart to the maximum extent. This makes the samples within the same type of subdomain more clustered, thereby dividing the subdomains of different categories during the alignment process, realizing category-aware subdomain-level fine-grained alignment, and enhancing the decision boundary of the model.

[0047] By jointly constraining and optimizing the model using the three loss functions mentioned above, the performance degradation of existing grayscale image cross-domain steganalysis detection methods in solving cross-domain information steganalysis in color images is effectively alleviated, and the robustness and generalization ability of the model in color image cross-domain steganalysis detection tasks are improved.

[0048] (6) Target domain data detection: After each training round, the target domain test set is input into the optimized model for classification, generating a classification probability vector. The predicted label of the corresponding sample is obtained based on this vector. To achieve performance evaluation, the predicted label can be compared with the real label to obtain the detection accuracy.

[0049] This method achieves category-aware, fine-grained subdomain alignment by selecting HCSs from the target domain input data that are similar to the source domain distribution. Furthermore, by introducing an adaptive alignment library and employing iterative updates with joint constraints from three loss functions—CE, DSA, and CLMD—it effectively alleviates the performance degradation problem of existing grayscale image cross-domain steganalysis detection methods in color image cross-domain steganalysis detection tasks, and provides a solution for color image cross-domain information steganalysis detection.

[0050] Example 1: To verify the effectiveness of this method, this example designs experiments in embedding rate scenarios of 0.4 bpc and 0.2 bpc to verify two aspects: first, to verify the performance degradation of the existing UCNet model (not a model designed for cross-domain tasks) in cross-domain steganalysis of color images; and second, to verify the performance improvement of this method compared to existing methods in cross-domain steganalysis of color images.

[0051] To verify the performance degradation of the existing UCNet model in cross-domain steganography detection of color images, this embodiment uses three spatial domain color image datasets: BOSSBase, ALASKA II, and SZUBase. Each dataset contains 10,000 images, and all images are preprocessed to 256×256 pixel size ppm format. Using the GINA and CMD-C color steganography schemes, combined with the HILL and SUNIWARD steganography algorithms, secret information is embedded into the three datasets at embedding rates of 0.4 bpc and 0.2 bpc, respectively. Finally, 10,000 pairs of "carrier image-secret image" (abbreviated as carrier / secret image pairs) samples are generated for each dataset, for a total of 30,000 pairs of samples across the three datasets.

[0052] The existing UCNet model was used to complete the cross-domain steganography detection task for color images. The results are shown in Table 1-2. Underlined entries in the table indicate no performance degradation, while down arrows indicate performance degradation. As can be seen from Table 1-2, although the existing UCNet model performs well in intra-domain detection, it exhibits performance degradation in cross-domain detection.

[0053] Table 1 shows the performance degradation at an embedding rate of 0.4 bpc.

[0054] Table 2 shows the performance degradation at an embedding rate of 0.2 bpc.

[0055] To verify the performance improvement of our proposed method compared to existing methods in cross-domain steganalysis detection of color images, we used the same three datasets to compare the detection accuracy of our method with that of existing methods such as UCNet, J-Net, and RCDD-Net in various cross-domain scenarios. For existing methods like UCNet, J-Net, and RCDD-Net, we used their original learning rate, optimizer, and weight decay settings, and initialized their respective feature extractors using their original initialization methods. In our proposed method, we used the parameters from the 150th training epoch of the UCNet pre-training for initialization.

[0056] The process of implementing this method mainly includes the following steps.

[0057] (1) Pre-training: Referring to the existing methods for setting up the dataset, each dataset is randomly divided into a training set and a test set in a ratio of 8:2. The UCNet model is pre-trained using the source domain training set.

[0058] (2) Data Initialization and Feature Extraction: The feature extractor part of this method is initialized using the parameters from the 150th training epoch during pre-training to ensure that the neural network has sufficient feature extraction capabilities. 2000 carrier / carrier image pairs are randomly selected from each source domain training set as source domain input data to represent the source domain data, and 2000 carrier / carrier image pairs from the target domain test set are used as target domain input data to represent the target domain data. The source domain input data has real labels, while the target domain input data disables real labels during alignment training to better reflect real-world application scenarios, using real labels only when calculating detection accuracy. For each training epoch, the source domain input data contains one carrier / carrier image pair, while the target domain input data contains two unlabeled images. In each training epoch, after being input into the feature extractor part, source domain features fs and target domain features ft are obtained respectively.

[0059] (3) Adaptive feature weighting and confidence filtering: After receiving the source domain features fs and target domain features ft from the feature extractor, the HCS and its high-confidence pseudo-labels are obtained by using an adaptive feature weighting mechanism and a confidence filtering mechanism.

[0060] (4) Construction of the adaptive alignment library: Based on the processing in steps (2) and (3), an adaptive alignment library is constructed to store source domain feature label pairs with real labels and HCS feature label pairs with high-confidence pseudo labels.

[0061] (5) Joint constraint alignment: In the domain distance calculation and alignment stage, this method uses three loss functions, namely CE loss function, DSA loss function and CLMD loss function, to jointly constrain model update and optimization. While retaining the source domain feature extraction capability, it realizes subdomain distribution alignment of category division and refinement of decision boundary.

[0062] It should be noted that, in the specific implementation process, the similarities and differences between the alignment strategy of this method and the alignment strategies of existing methods are as follows: Figure 3 As shown, J-Net only implements simple domain-level alignment, while RCDD-Net implements category-aware subdomain-level alignment. In order to optimize the alignment effect, this method further divides all target domain input data into HCS and ordinary samples, and then uses DSA loss function and CLMD loss function to perform subdomain alignment, thus achieving more fine-grained category-aware subdomain alignment.

[0063] (6) Target Domain Data Detection: After updating and optimizing the model, at the end of each training epoch, i.e., when the parameters are updated, 2000 pairs of carrier / carrier images used for alignment are used as input. The detection accuracy is calculated based on the predicted and true labels of the output to evaluate the detection performance. In addition, the effect after each training epoch is recorded. It should be noted that the learning rate during training decreases as the number of training epochs increases, thereby avoiding model parameter oscillations.

[0064] The detection accuracy calculated by this method in various cross-domain scenarios, compared with existing methods such as UCNet, J-Net, and RCDD-Net, is shown in Table 3-4. B represents BOSSBase, A represents ALASKA II, and S represents SZUBase. B→A means the source domain is BOSSBase and the target domain is ALASKA II, and so on. An asterisk indicates the maximum detection accuracy in the current cross-domain scenario. The UCNet column represents the accuracy achieved by directly testing on the target domain test set during cross-domain detection.

[0065] As shown in Tables 3-4, the existing J-Net, which only implements global alignment, performs worse than the existing UCNet pre-trained and directly tested performance in most cases. In contrast, the existing RCDD-Net, which performs subdomain-level alignment, shows performance improvement in all cases. This method addresses the shortcomings of existing methods by adapting it for cross-domain steganography detection in color images, achieving optimal performance through subdomain alignment and decision boundary refinement.

[0066] Table 3. Performance of existing methods and our proposed method in cross-domain steganalysis of color images at an embedding rate of 0.4 bpc.

[0067] Table 4. Performance of existing methods and our proposed method in cross-domain steganalysis of color images at an embedding rate of 0.2 bpc.

[0068] In a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a method for detecting cross-domain information stegatology in color images.

[0069] The memory can be volatile or non-volatile, or a combination of both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memory of this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0070] The processor can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0071] The method steps of this invention can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0072] Software implementation can be achieved by executing functional modules (such as procedures, functions, etc.). Software code can be stored in memory and executed by the processor. Memory can be implemented in the processor or outside the processor.

[0073] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for detecting cross-domain information stegatology in a color image.

[0074] Computer storage media can include various media that can store program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] Fourthly, the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of a method for detecting cross-domain information stegatology in color images.

[0076] Specifically, computer program products include: data signals, data signals embodied in a carrier wave, or computer-readable storage media.

[0077] It should be noted that the technical solutions described in this invention can be combined arbitrarily without conflict.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention also includes such modifications and variations.

Claims

1. A method for detecting cross-domain information steganalysis in color images, characterized in that, include: Using the color carrier image and its corresponding steganalysis image in the source domain as training data, the color image steganalysis model UCNet is pre-trained in the source domain. The pre-trained model is then used as the feature extractor to extract features from the source domain input data and the target domain input data. An adaptive weighting mechanism is introduced to perform sample-by-sample and dimension-by-dimensional weighting of the extracted features, and a confidence filtering mechanism is introduced to divide the target domain input data into HCS and ordinary samples. Construct an adaptive alignment library to store source domain feature label pairs and HCS feature label pairs for each training round; The model is updated and optimized by jointly constraining the three loss functions: CE, DSA, and CLMD. Specifically, the CE loss function is used to preserve the feature extraction capability trained on the source domain; the DSA loss function is used to achieve dual alignment between source domain samples and target domain samples, and between HCS and ordinary samples in the target domain; and the CLMD loss function is used to minimize the distance between similar subdomain samples and dissimilar subdomain samples between source domain samples and HCS, and maximize the distance between dissimilar subdomain samples, thereby achieving category-aware subdomain-level fine-grained alignment. The image to be detected is transmitted to the optimized model, which outputs the corresponding predicted label.

2. The method as described in claim 1, characterized in that, The feature weighting processing steps of the adaptive weighting mechanism specifically include: After receiving the source domain features fs and target domain features ft from the feature extractor, the weight values ​​for each dimension of each feature are calculated as follows: weighted_fs=Sigmoid(fc2(Relu(fc1(fs)))), weighted_ft=Sigmoid(fc2(Relu(fc1(ft)))), As can be seen from the above operations, in the adaptive weighting mechanism, after dimensionality reduction through a fully connected layer fc1, the ReLU activation function is used for calculation, and then the activation value is passed to another fully connected layer fc2 for dimensionality increase. Finally, the Sigmoid activation function is used to obtain the weight value, where weighted_fs is the source domain weight and weighted_ft is the target domain weight. Multiply the weight values ​​by the corresponding original features to obtain the weighted source domain features and the weighted target domain features.

3. The method as described in claim 1, characterized in that, Following the weighted processing step, the following steps are also included: Lightweight single-layer fully connected layers are used to process the weighted source domain features and the weighted target domain features respectively, and output the original probability values ​​of the source domain Logits_s and the original probability values ​​of the target domain Logits_t.

4. The method as described in claim 1, characterized in that, The sample partitioning steps of the confidence filtering mechanism specifically include: At different training stages, the maximum value of the original prediction result Logits_t for each sample in the target domain input data is taken as the confidence score of that sample. Based on the current model state and target domain data distribution during the training phase, a dynamic confidence threshold is calculated using the following formula: ， In the above formula, This represents the confidence threshold for the current training round. This represents the minimum value of the confidence threshold. This represents the maximum value of the confidence threshold. This is the current training round. It refers to the total number of training rounds. Indicates coercion Not less than 0.5; The confidence level of each sample in the target domain input data is compared with the dynamic confidence threshold. Samples with a confidence level greater than or equal to the dynamic confidence threshold are classified as HCS, and the remaining samples are classified as ordinary samples.

5. The method as described in claim 1, characterized in that, The adaptive alignment library stores the source domain feature label pair and the HCS feature label pair of each training round in the following manner: adaptive alignment library ∈ R (Vs·ns+Vt·nt)×(256+1) Wherein R represents the real number field, Vs and Vt represent the sample storage ratio of the source domain and the target domain respectively, ns and nt represent the total number of samples of the source domain and the target domain respectively, and 256+1 represents the joint dimension of the feature vector and the label. An adaptive alignment library is updated using a first-in, first-out (FIFO) strategy.

6. The method as described in claim 1, characterized in that, The steps for achieving double alignment using the DSA loss function specifically include: The source domain samples in the adaptive alignment library and the HCS are integrated together to form the adaptive alignment library sample set; the target domain sample data input into the model in the current training round constitutes the target domain sample set in the current training round. We introduce a DSA loss function based on LMMD improvement in UDA to quantify the feature distribution difference between the adaptive alignment library sample set and the target domain sample set in the current training round. By incorporating this distribution difference into the model optimization objective, we use backpropagation to iteratively reduce the DSA loss value and simultaneously narrow the distribution distance between samples in the adaptive alignment library and samples in the target domain, thus achieving dual alignment: cross-domain alignment between source domain samples and target domain samples, and intra-domain alignment between HCS and ordinary samples within the target domain.

7. The method as described in claim 1, characterized in that, The steps of using the CLMD loss function to minimize the distance between samples in the same subdomain and maximize the distance between samples in different subdomains of the source domain and the HCS specifically include: The similarity calculation method in the InfoNCE loss function of contrastive learning is replaced with the kernel calculation method of UDA, thus forming the CLMD loss function; The CLMD loss function is used to calculate the distance between samples of the same subdomain and samples of different subdomains in the current training round; Minimize the distance between samples from the same subdomain and maximize the distance between samples from different subdomains.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images