Image steganography method and device based on residual guidance

Abstract

The application discloses a residual guide-based image steganography method and device, and belongs to the technical field of information hiding. A multi-feature fusion generation network including a residual guide module, a distance guide module and a local entropy guide module is constructed to form a non-adversarial deep generation framework. An original carrier image is preprocessed and input into the network to generate an embedding probability map. Then, the embedding probability map and a random coding feature are input into a deep feature embedding learning model to perform embedding simulation to generate a simulated stego image. The network is trained end to end by calculating the multi-guide joint loss between the simulated stego image and the preprocessed carrier image. Finally, the to-be-processed carrier image is input into the trained network to obtain the embedding probability map, and a final stego image with high concealment and security is generated accordingly. The application realizes intelligent optimization of an embedding region and multi-feature collaborative guidance, and improves the anti-detection capability and visual fidelity of a steganography algorithm.

Description

Technical Field

[0001] This invention relates to the field of information hiding technology, and in particular to an image steganography method, apparatus, and device based on residual guidance. Background Technology

[0002] With the deepening integration of artificial intelligence and information hiding technology, deep learning-based image steganography has become a cutting-edge research direction in the field of information security. In terms of technological development, the existing technology system has mainly undergone two stages of evolution.

[0003] The first stage is represented by traditional methods that manually design distortion functions or embedding costs, with typical algorithms including WOW (Wavelet Obtained Weights), HUGO (Highly Undetectable steGO), and S-UNIWARD (Spatial-UNIversal WAvelet Relative Distortion). The core idea of ​​these methods is to manually define the "cost" of pixel modifications based on the local texture complexity of the image, prioritizing the embedding of secret information in textured regions. While they have achieved some success in static image steganography, their fundamental limitation lies in the reliance on fixed, hand-crafted models for embedding strategies. These predefined, content-independent rules are ill-suited to the complex and varied statistical distributions and high-order semantic features of natural images, resulting in insufficient adaptability of the embedding process. More importantly, with the emergence of steganalyzers based on deep convolutional neural networks (such as SRNet and YeNet), the modification traces of steganalytes generated by these methods, which rely on shallow statistical regularities, are easily captured by deep analysis models, posing a serious challenge to security.

[0004] In the second phase, researchers introduced Generative Adversarial Networks (GANs) and Deep Convolutional Neural Networks (CNNs) architectures into the steganography task, marking the field's entry into a data-driven, end-to-end paradigm. Typical models, such as SteGAN, ASDL-GAN, and ReLOAD, typically consist of a generator (responsible for constructing the steganalyte) and a discriminator (usually acting as a steganalyzer) trained adversarially. The generator aims to successfully embed information while deceiving the discriminator, which strives to accurately distinguish between the carrier image and the steganalyte. Theoretically, this paradigm can automatically learn embedding strategies adapted to the image content. However, in practical applications, such adversarial training-based methods have revealed several inherent defects: First, the GAN framework itself suffers from high training instability and easy mode collapse, leading to difficulty in model convergence and poor reproducibility; second, the adversarial training process involves a dynamic game between the generator and the discriminator, consuming huge computational resources and resulting in low training efficiency; third, the embedding process of the generator often lacks clear and interpretable physical or statistical guidance, and the allocation of embedding resources has a "black box" characteristic, making it difficult to control precisely; finally, due to the inherent contradiction of the adversarial target, the model struggles to achieve a stable and excellent balance between the visual fidelity (quality) of the steganagraph and its anti-detection ability (security), often sacrificing one for the other.

[0005] In summary, existing technologies, whether relying on traditional methods with fixed rules or deep learning models based on adversarial game theory, all suffer from significant shortcomings in key dimensions such as adaptability, training stability, computational efficiency, and the security-quality balance. Therefore, there is an urgent need for a novel steganography generation framework that is stable in training, clearly guided, and capable of collaboratively optimizing multi-objective performance, in order to advance this technology towards greater practicality. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] To address this, this invention proposes a residual-guided image steganography method. It constructs a multi-feature fusion generative network comprising residual guidance, distance guidance, and local entropy guidance modules to form a non-adversarial deep generative framework. The original carrier image is preprocessed and input into this network to generate an embedding probability map. This embedding probability map, along with randomly encoded features, is then input into a deep feature embedding learning model to perform embedding simulation and generate a simulated steganographic image. The network is then end-to-end optimized by calculating the multi-guided joint loss between the simulated steganographic image and the original image. Finally, the fully trained network is used to process the image to be steganized, generating a final steganographic image with high concealment and security. This invention achieves adaptive information embedding under multi-feature collaborative guidance, effectively improving the anti-detection capability and visual fidelity of steganography operations.

[0008] Another objective of this invention is to provide an image steganography device based on residual guidance.

[0009] The third objective of this invention is to provide a computer device.

[0010] To achieve the above objectives, this invention proposes a residual-guided image steganography method, comprising: S1, preprocess the original carrier image to obtain the preprocessed carrier image; S2, a multi-feature fusion generation network containing a residual guidance module, a distance guidance module, and a local entropy guidance module is constructed using a non-adversarial deep generation framework, and the preprocessed carrier image is input into the multi-feature fusion generation network to generate an embedding probability map; S3, input the embedding probability map and the corresponding random coded features into the preset deep feature embedding learning model, perform embedding simulation, and generate simulated steganalysis image; S4, calculate the multi-guided joint loss between the simulated stegana and the preprocessed carrier image, and use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network. S5. Input the carrier image to be processed into the trained multi-feature fusion generative network to obtain a new embedding probability map, and obtain the final steganalysis image based on the new embedding probability map.

[0011] An image steganography method based on residual guidance according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, the preprocessing of the original carrier image to obtain a preprocessed carrier image includes: S11, the original carrier image is spatially standardized and adjusted to a fixed size that meets the input requirements of the multi-feature fusion generation network; S12, perform pixel value normalization processing on the carrier image after spatial size standardization, and convert the corresponding pixel value range to a predetermined numerical range.

[0012] In one embodiment of the present invention, the multi-feature fusion generation network, comprising a residual guidance module, a distance guidance module, and a local entropy guidance module, is constructed using a non-adversarial deep generation framework, and the preprocessed carrier image is input into the multi-feature fusion generation network to generate an embedding probability map, including: S21, the high-frequency residual features of the preprocessed carrier image are extracted through the residual guidance module to generate a residual guidance feature map; S22, through the distance guidance module, the depth features of the pre-processed carrier image are extracted using a pre-trained feature extraction network as the first depth feature, and a distance guidance feature map is generated based on the feature distance constraint between the first depth feature and the depth features corresponding to the stegana image to be generated. S23, the local information entropy distribution of the preprocessed carrier image is calculated through the local entropy guidance module to generate a local entropy feature map; S24. The residual guided feature map, distance guided feature map and local entropy feature map are fused and input into the backbone network of the multi-feature fusion generation network for feature mapping and probability normalization to generate an embedding probability map.

[0013] In one embodiment of the present invention, the step of inputting the embedding probability map and the corresponding random encoded features into a preset deep feature embedding learning model, performing embedding simulation, and generating a simulated steganalyte image includes: S31, Based on the embedding probability map, determine the embedding strength corresponding to each pixel position, and based on the corresponding random coding features, determine the modification direction to be simulated; S32, the embedding strength and modification direction are input to the dual Tanh simulation STC model. The dual Tanh simulation STC model performs nonlinear mapping on the input through two cascaded Tanh functions, and outputs a continuous and differentiable simulation modification. S33, the simulated modification amount is added to the pixel value of the preprocessed carrier image to generate the steganalytic image.

[0014] In one embodiment of the present invention, the step of calculating the multi-guided joint loss between the simulated stegana image and the preprocessed carrier image, and using the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain a trained multi-feature fusion generation network, includes: S41, calculate the residual guided loss, distance guided loss and local entropy guided loss between the simulated stegana and the preprocessed carrier image respectively; S42, the residual guided loss, distance guided loss and local entropy guided loss are weighted and summed to form the multi-guided joint loss, and the parameters of the multi-feature fusion generation network are iteratively optimized with the goal of minimizing the multi-guided joint loss until convergence.

[0015] In one embodiment of the present invention, the step of inputting the carrier image to be processed into a trained multi-feature fusion generation network to obtain a new embedding probability map, and obtaining the final steganalyte image based on the new embedding probability map, includes: S51, the carrier image to be processed is input into the trained generative network. The residual guidance module, distance guidance module and local entropy guidance module of the trained generative network work together to extract and fuse multi-scale depth features and guidance features, and output a new embedding probability map after decoding and mapping. S52, the new embedding probability map and the corresponding random coding features are input into the preset deep feature embedding learning model. The preset deep feature embedding learning model weights each bit of the corresponding random coding features according to the new embedding probability map, and simulates the coding process through a cascaded hyperbolic tangent function to calculate the pixel-level simulated modification amount. S53, the simulated modification amount is superimposed on the corresponding pixel value of the processed carrier image, and the final steganalytic image is directly synthesized and output.

[0016] In one embodiment of the present invention, it further includes: S6 simulates the STC embedding process in traditional steganography algorithms through a dual Tanh activation function structure; wherein, the first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on pixel-level embedding operations, so that the steganalyte maintains visual consistency with the carrier image in the pixel domain.

[0017] To achieve the above objectives, another aspect of the present invention provides an image steganography device based on residual guidance, comprising: The image preprocessing module is used to preprocess the original carrier image to obtain a preprocessed carrier image; The probability map generation module is used to construct a multi-feature fusion generation network that includes a residual guidance module, a distance guidance module, and a local entropy guidance module using a non-adversarial deep generation framework, and inputs the preprocessed carrier image into the multi-feature fusion generation network to generate an embedded probability map; The embedding simulation module is used to input the embedding probability map and the corresponding random coded features into a preset deep feature embedding learning model, perform embedding simulation, and generate simulated steganalysis images. The network training module is used to calculate the multi-guided joint loss between the simulated stegana image and the preprocessed carrier image, and to use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network. The steganalysis image generation module is used to input the carrier image to be processed into a trained multi-feature fusion generation network to obtain a new embedding probability map, and to obtain the final steganalysis image based on the new embedding probability map.

[0018] In one embodiment of the present invention, it further includes: The dual Tanh activation function structure module is used to simulate the STC embedding process in traditional steganography algorithms through the dual Tanh activation function structure. The first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on the pixel-level embedding operation, so that the steganalyte image maintains visual consistency with the carrier image in the pixel domain.

[0019] This invention discloses an image steganography method and apparatus based on residual guidance. By constructing a multi-feature fusion generation network integrating residual guidance, distance guidance, and local entropy guidance modules, and employing a non-adversarial deep generation framework and a dual-Tanh simulated embedding mechanism, it effectively solves the problems of poor concealment and weak anti-detection capability in existing technologies caused by inaccurate embedding regions, single feature representations, and unstable adversarial training. It achieves end-to-end optimization from multi-feature collaborative extraction and adaptive probabilistic graph generation to differentiable simulated embedding, significantly improving the concealment and security of steganography operations and enhancing the robustness and visual fidelity of the algorithm under various steganography attacks.

[0020] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a residual-guided image steganography method as described in the first aspect embodiment.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an image steganography method based on residual guidance according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the research process of another residual-guided image steganography method according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the model structure of another residual-guided image steganography method according to an embodiment of the present invention. Figure 4 This is a ResNet-50 architecture diagram after optimization of another residual-guided image steganography method according to an embodiment of the present invention; Figure 5 This is the ResNet-50 probability map generation process after optimization of another residual-guided image steganography method according to an embodiment of the present invention; Figure 6 This is a VGG16 model structure diagram after optimization of another residual-guided image steganography method according to an embodiment of the present invention; Figure 7 This is an experimental application diagram of residual-guided image steganography, based on an embodiment of the present invention. Figure 8 This is a comparison of PSNR and SSIM at different bit rates between RGIS and other methods, based on an experimental application of a residual-guided image steganography method according to an embodiment of the present invention. Figure 9 This is a comparison of PSNR and SSIM of RGIS and other methods under different noise attacks in an experimental application of an image steganography method based on residual guidance according to an embodiment of the present invention. Figure 10 This is an experimental application of RGIS, a residual-guided image steganography method according to an embodiment of the present invention, showing the detection error rate of RGIS and other methods in the face of different steganalysis detection methods. Figure 11 This is a schematic diagram of the structure of an image steganography device based on residual guidance according to an embodiment of the present invention; Figure 12 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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] The following description, with reference to the accompanying drawings, describes an image steganography method, apparatus, and device based on residual guidance according to embodiments of the present invention.

[0026] The core idea of ​​this invention is to construct a multi-feature fusion generation network that includes a residual guidance module, a distance guidance module, and a local entropy guidance module, and to adopt a non-adversarial deep generation framework to achieve collaborative perception of deep features and statistical properties of carrier images. Building upon this foundation, a residual guidance module extracts high-frequency residual features to locate complex texture regions, a distance guidance module constrains the semantic consistency between the steganalyte and the carrier image in the deep feature space, and a local entropy guidance module quantifies the local information complexity of the image to evaluate embedding adaptability. Furthermore, multi-source guidance features are fused and mapped to generate a refined embedding probability map. This probability map and randomly encoded features are then input into a deep feature embedding learning model, and a cascaded hyperbolic tangent function is used to continuously and differentiably simulate the embedding process, generating a simulated steganalyte image. Finally, a multi-task joint loss function fusing multiple guidance signals is constructed to perform end-to-end optimization training on the generator network. This transforms traditional steganography methods that rely on manual design or adversarial training into an intelligent steganography system capable of adaptively selecting embedding regions, precisely controlling modification intensity, and simultaneously optimizing visual concealment and statistical security, significantly improving the overall performance of steganography operations in terms of imperceptibility and resistance to steganalysis detection.

[0027] Example 1 To achieve the above invention, embodiments of the present invention provide an image steganography method based on residual guidance, such as... Figure 1 As shown, it includes: S1, preprocess the original carrier image to obtain the preprocessed carrier image.

[0028] Specifically, the principle behind carrier image preprocessing lies in constructing a standardized and uniformly distributed data front-end to eliminate the randomness of the original image data in terms of spatial scale and numerical range, providing a stable and standardized input for subsequent deep learning models. This processing, based on image signal processing and data normalization theory, projects multi-source heterogeneous original image data onto a unified feature space through geometric transformations and numerical mapping. This effectively reduces training oscillations and convergence difficulties caused by differences in input data distribution, providing a benchmark reference system for multi-feature fusion generation networks. This ensures that the extraction and calculation of subsequent guiding features such as residuals, distances, and local entropy are based on comparable and consistent data, which is the primary step in ensuring that the model can learn universal embedding rules.

[0029] Furthermore, the preprocessing mainly includes two sequential operations: spatial size standardization and pixel value normalization. First, a bicubic interpolation algorithm is used to uniformly adjust the size of the input carrier image to a preset fixed resolution, such as 256 pixels × 256 pixels, to meet the tensor dimension requirements of the generative network input layer. This operation eliminates image size differences while maintaining the main content and aspect ratio of the image. Second, the pixel intensity values ​​of the size-standardized image undergo a linear transformation, mapping them from the original integer representation range to a continuous, preset numerical interval. A typical approach is to normalize the pixel values ​​from the [0, 255] interval to a floating-point interval of [0.0, 1.0], or further standardize them to a distribution with a mean of 0 and a standard deviation of 1. This process, through numerical scaling, makes the model optimization process smoother and more stable.

[0030] Specifically, key parameters involved in preprocessing include the target input size and the normalization interval. The target input size is a hyperparameter predefined based on the network structure and computational resources. Its selection requires a trade-off between detail preservation and computational efficiency. For example, it can be set to a power of 2, such as 256×256 or 512×512, to accommodate the downsampling structure of convolutional networks. The normalization interval is another core parameter, typically mapping pixel values ​​to the range [0,1], or standardizing them using the mean and standard deviation of the dataset. The determination of these parameters directly affects the initial data distribution received by the network and serves as an important reference benchmark for learning rate setting, weight initialization, and other aspects of model training. Their consistency is fundamental to ensuring model reproducibility and performance comparability.

[0031] Specifically, this preprocessing workflow can convert digital images from different acquisition devices with varying resolutions and dynamic ranges, such as web images, surveillance screenshots, medical images, or digital copies of artworks, into a format that conforms to the model input specifications. This allows the method described in this invention to be flexibly applied to various information hiding scenarios involving image carriers, such as covert communication requiring high security, digital media copyright protection, and secure military intelligence transmission. Standardized preprocessing ensures the method's robustness to image data from different sources, which is a key prerequisite for achieving "train once, apply widely" engineering deployment.

[0032] Specifically, the system's preprocessing operations bring about several performance improvements. First, it significantly accelerates the convergence speed of model training by providing uniformly distributed input data, reducing gradient variance in the initial learning phase. Second, it enhances the model's generalization ability and robustness, enabling the model to focus on learning essential features relevant to image content and the steganography task, rather than irrelevant size or brightness differences. Third, it provides a fair and consistent basis for comparison in subsequent guidance modules (especially residual and local entropy calculations), ensuring that the extracted guidance features truly reflect the characteristics of the image content, rather than artifacts introduced by differences in data preprocessing. Therefore, this preprocessing step is an indispensable foundation for the overall method to achieve high visual quality, high security, and strong generalization ability.

[0033] Furthermore, S1 includes: S11 performs spatial size standardization on the original carrier image and adjusts the carrier image to a fixed size that meets the input requirements of the multi-feature fusion generation network.

[0034] Specifically, the input layers of fixed-structure models such as convolutional neural networks require tensors of a defined dimension, while natural images vary in size. The core principle of this process is to map the original image of arbitrary size to a unified, pre-defined two-dimensional grid space through geometric transformation, thereby eliminating the spatial dimensionality variability of the input data. This operation ensures that the data tensors are completely consistent across all spatial and channel dimensions outside the batch dimension during batch training. This provides a stable basis for the sliding of receptive fields in weight-sharing convolutional kernels within the network, a prerequisite for forward propagation and gradient backpropagation calculations, and the foundation for the correct operation of complex structures such as feature pyramids and skip connections.

[0035] Specifically, first, a fixed target size is set, such as a width of W pixels and a height of H pixels. Then, an interpolation algorithm is used to resample all pixels of the original image to this target grid. Commonly used interpolation methods include bilinear interpolation or bicubic interpolation. The latter considers the weighted contribution of pixels in a larger neighborhood when calculating new pixel values, better preserving edge information and texture details of the image, thus reducing blurring or jagged edges that may occur due to scaling. The processed image has a completely consistent width W and height H, and its spatial resolution strictly matches the expected dimension of the network input layer.

[0036] Furthermore, the choice of size in this step is a critical design parameter that requires a balance among several factors: too small a size leads to a significant loss of image detail, affecting the accuracy of high-frequency residual feature extraction; too large a size drastically increases the computational complexity and memory consumption of the model, reducing training and inference efficiency. Typically, this size is set to an integer power of 2 to accommodate downsampling operations such as pooling or stride convolution in the network. For example, 256×256 or 512×512 are common choices. In addition, the aspect ratio preservation strategy that may be involved in scaling, such as maintaining the original proportions through edge padding, is also an important implementation detail parameter.

[0037] Specifically, this step can seamlessly convert large images from high-resolution digital cameras, images of varying sizes crawled from the web, and low-resolution frames extracted from surveillance videos into a standard input format. This allows the steganography method to be seamlessly applied to diverse image carriers in open network environments, providing a reliable technical foundation for practical applications such as covert communication on social media, secure distribution of digital documents, and intelligence hiding in remote sensing images. This processing is a crucial bridge connecting heterogeneous real-world data with standardized deep learning models.

[0038] Specifically, implementing spatial size standardization ensures the stability and reproducibility of the training process, avoiding dynamic instability in training batches caused by random changes in input size. Secondly, by unifying the computation graph, it enables the graphics processor to perform efficient batch parallel computation, significantly improving the speed of model training and steganalysis. Finally, it provides deterministic operational benchmarks for each guiding module within the network (especially the local entropy guiding module that requires local window operations), guaranteeing the accuracy of feature extraction and computation, thereby ensuring the quality and security of the final steganalysis from the source.

[0039] S12, perform pixel value normalization processing on the carrier image after spatial size standardization, and convert the corresponding pixel value range to a predetermined numerical range.

[0040] Specifically, the pixel values ​​of the original image are typically represented as discrete integers (e.g., 0 to 255), and their absolute values ​​differ from their distribution range. The core principle of normalization is to transform the original pixel values ​​from their independent dynamic ranges to a unified, continuous, predetermined numerical interval using a linear or nonlinear mapping function. This process essentially standardizes the first-order statistics (mean) and second-order statistics (standard deviation) of the input data, ensuring that different image samples exhibit similar numerical centers and fluctuations when input into the network. This significantly reduces the model's sensitivity to differences in overall brightness and contrast of the input data, allowing the optimizer to focus more on learning structured features relevant to the steganography task itself.

[0041] Specifically, pixel value normalization typically employs a linear scaling method. For an image that has undergone spatial normalization, the intensity value of each pixel channel (e.g., R, G, B) is... (Values ​​range [0, 255]) The transformation is performed using the following formula: A common approach is to set the Offset to 0 and the Scale to 255, thus mapping pixel values ​​to a floating-point range of [0.0, 1.0]. A more refined implementation involves performing standard deviation normalization, which calculates the mean μ and standard deviation σ of all training image pixels in the dataset, and then uses the formula... The data is transformed to approximate a standard normal distribution with a mean of 0 and a standard deviation of 1. This process is computationally efficient and can be integrated into a data loading pipeline.

[0042] Furthermore, when using simple scaling to the [0,1] interval, the parameters are fixed. When using dataset-based normalization, the key parameters are the global pixel mean μ and standard deviation σ, statistically derived from the training set. These parameters are considered part of the model preprocessing pipeline, and their determinism and consistency are crucial. During deployment, the normalization applied to new input images must use the exact same parameters as in the training phase to ensure consistency in data distribution. The selection of these parameters directly affects the effectiveness of model weight initialization, the appropriate range of the learning rate, and the working saturation region of activation functions (such as Sigmoid and Tanh), forming the basis for model hyperparameter tuning.

[0043] Specifically, regardless of whether the input image comes from a professional camera, a smartphone, or a network image that has undergone different compression and rendering processes, its brightness, contrast, and color balance may vary significantly. This normalization step effectively eliminates these appearance variations unrelated to the steganographic content, bringing image data from different sources back to the same starting line. This allows the method to be stably applied to complex scenarios such as open internet environments, multi-source intelligence data aggregation, and cross-platform digital content protection, ensuring the consistency and reliability of the steganography algorithm's performance.

[0044] Specifically, pixel value normalization has several positive impacts on model training and final performance. First, it accelerates training convergence by providing a uniformly distributed input, avoiding drastic fluctuations in gradient update direction due to differences in data scale. Second, it improves the model's numerical stability and generalization ability, preventing overfitting to specific brightness or contrast features in the training set. Finally, it provides a standardized numerical foundation for all subsequent pixel-value-based calculation modules (such as residual calculation, local entropy calculation, and loss function), ensuring the accuracy and comparability of these calculation results. This, in turn, guarantees the effectiveness of the multi-guided joint optimization mechanism at the underlying data level, ultimately contributing to the high visual quality and high security of the generated steganalytic images.

[0045] S2, a multi-feature fusion generation network is constructed using a non-adversarial deep generation framework, which includes a residual guidance module, a distance guidance module, and a local entropy guidance module. The preprocessed carrier image is then input into the multi-feature fusion generation network to generate an embedding probability map.

[0046] Specifically, traditional generative adversarial networks (GANs) rely on a dynamic game between the generator and the discriminator, which can easily lead to training instability and pattern collapse. This invention abandons this approach. Its principle is to use a feedforward deep convolutional network to integrate three types of heterogeneous guiding features from the residual domain, deep semantic space, and local statistical domain in parallel. The network is based on an encoder-decoder structure, mapping the preprocessed image into high-dimensional features. The three guiding modules act as parallel feature extraction and modulation branches, respectively analyzing the embeddability of the image from different dimensions. Finally, these multi-source features are adaptively fused in the network's bottleneck layer or decoder. A normalized output layer (such as a sigmoid activation function) maps the fused high-dimensional representation into a probability map with the same resolution as the input image, where each pixel value is in the [0,1] interval. This probability map, in principle, represents the suitability of each pixel location to carry secret information. Its generation process is a deterministic forward inference, avoiding the uncertainty in adversarial training.

[0047] Furthermore, the multi-feature fusion generative network uses an improved U-Net or a deep ResNet as its backbone. The preprocessed carrier image is first input into the encoder part of the backbone network for downsampling and basic feature extraction. Simultaneously, three guiding modules work in parallel: the residual guiding module performs high-pass and low-pass filtering on the input image to generate residual feature maps emphasizing texture edges; the distance guiding module uses an intermediate layer of a pre-trained classification network (such as VGG16 or ResNet50) to extract a high-level semantic feature tensor from the input image, which serves as a reference for subsequent distance calculations; and the local entropy guiding module calculates the local information entropy of the image through a sliding window to generate an entropy distribution map. These guided feature maps are fused with the features extracted by the backbone network at specific intermediate layers of the backbone network (at the end of downsampling or the beginning of upsampling) through feature concatenation, element-wise addition, or attention gating mechanisms. The fused features are upsampled by the decoder to restore the original image size, and finally passed through a 1×1 convolutional layer and a sigmoid activation function to output the final embedding probability map.

[0048] Furthermore, the depth of the backbone network determines its receptive field size and feature abstraction capability; for example, an encoder with 4 to 5 downsampling stages can be used. The feature maps output by each bootstrapping module are typically designed as tensors with a specific number of channels (e.g., 64 or 128) to control their contribution weight to the backbone network. The choice of fusion point location (e.g., at which network layer feature injection is performed) is an important architectural design parameter. In addition, the number of trainable parameters for the entire network is a core engineering metric, directly affecting the model's computational complexity, memory usage, and training data requirements. Non-adversarial frameworks typically reduce the number of parameters by 15%-30% compared to equivalent GAN structures, demonstrating their lightweight advantage.

[0049] Specifically, this deterministic generative network architecture makes it particularly suitable for applications with high requirements for deployment efficiency and stability. Since it eliminates the need for complex adversarial inference loops, the network can quickly and stably generate embedding probability graphs. This allows the invention to be integrated into systems requiring real-time or near-real-time steganography, such as online secure communication software, streaming media information hiding plugins, or mobile privacy protection applications. Simultaneously, its deterministic output facilitates systematic testing and security assessment, meeting the stringent requirements of high-security fields (such as defense and finance) for predictable and auditable algorithm behavior. This design advances advanced deep learning steganography capabilities from laboratory research to the stage of engineering application.

[0050] Specifically, the adoption of a non-adversarial multi-feature fusion generative network brings about several performance breakthroughs, fundamentally eliminating training instability and the risk of pattern collapse, making the model training process converge faster and more reliably. Secondly, by explicitly introducing and fusing three types of guiding features—residual, semantic distance, and local entropy—the network generates an embedding probability map that combines high-frequency detail sensitivity, semantic consistency, and statistical security, achieving intelligent optimization of embedding regions. This directly leads to a better balance between visual quality, resistance to conventional steganalysis (such as SRM), and resistance to deep learning steganalysis (such as SRNet) in the generated steganographic images. Finally, this feedforward network is extremely efficient in the inference stage, significantly reducing the computational latency and energy consumption of steganography operations, providing technical feasibility for applications in large-scale or resource-constrained environments.

[0051] Furthermore, S2 includes: S21, the high-frequency residual features of the preprocessed carrier image are extracted through the residual guidance module to generate a residual guidance feature map.

[0052] Specifically, a digital image can be viewed as a superposition of low-frequency components representing gently changing structures and high-frequency components representing details, edges, and textures. Secret information embedding is essentially a weak signal superposition, which is more easily detected in smooth regions (dominated by low frequencies) and more easily masked in textured regions (rich in high frequencies). The core principle of this module is to decouple the preprocessed carrier image in the frequency domain by designing specific filter banks, actively separating and enhancing its high-frequency residual components. This process aims to explicitly capture the features of regions in the image that can naturally "hide" modifications, thereby providing the generative network with direct, image content-based physical guidance signals, shifting the network's learning objective from fuzzy overall fitting to targeted detail utilization.

[0053] Furthermore, the residual guidance module typically consists of a parallel dual-path filtering architecture. One path is a high-pass filtering path, which uses edge detection kernels such as the Kirsch operator for convolution operations, strongly responding to step edges and fine textures in the image to generate a residual map that highlights details. The other path is a low-pass filtering path, which uses a larger-sized mean filter to smooth the original image, obtaining a background image that mainly contains low-frequency information. Subsequently, the original image is subtracted from the low-pass filtering result by a residual merging unit, or the high-pass filtering result is nonlinearly enhanced, ultimately synthesizing a "residual guidance feature map". The intensity value of each pixel in this feature map reflects the "high-frequency activity" or "texture complexity" of the corresponding location in the original image, and is usually normalized for fusion with the network backbone features.

[0054] Specifically, the core adjustable parameters of this module include the type, size, and threshold of the filter kernel. The choice of the high-pass filter kernel (such as Prewitt, Sobel, or Kirsch) and its directionality determine the sensitivity and orientation of edge feature extraction. The size of the low-pass filter kernel (such as 7×7 or 11×11) directly affects the range of high-frequency components that are filtered out, thus determining the boundary between what constitutes "background" and what constitutes "residual." Furthermore, before generating the final feature map, thresholding or nonlinear activation functions may be introduced to suppress noise or enhance contrast; the threshold or function parameters are also important indicators. The combination of these parameters defines the module's quantitative perception of image "embeddability," and its settings must be adapted to the statistical detection characteristics of the steganalyzer.

[0055] Specifically, whether it's the rich textures of vegetation in natural landscape images, the details of hair and fabric in portrait photographs, or the edges of urban buildings in satellite imagery, the residual guidance module can effectively locate these complex areas with high information capacity. This enables the invention to maintain stable high performance when dealing with real-world images from a wide range of sources and with diverse content. It is suitable for various high-security scenarios, from covert communication in public network images to the embedding of confidential information in professional fields (such as remote sensing and medical) images, thus enhancing the breadth and depth of the technology's practical applications.

[0056] Specifically, the residual guidance module directly guides the generator network to concentrate modification resources on regions with complex textures. This mathematically and statistically makes the distribution of steganalytic modifications closer to natural image noise, significantly reducing the risk of detection by traditional rich model-based steganalysts. Simultaneously, by avoiding invalid or harmful modifications in flat regions, the visual quality of the steganalyte is better preserved, reducing potentially noticeable local distortions. The explicit physical feature signal provided by this module also serves as a strong constraint, simplifying the optimization objective of the generator network. Working in conjunction with other guidance modules, it contributes to a balance between model security, stealth, and visual fidelity.

[0057] S22, through the distance guidance module, the depth features of the pre-processed carrier image are extracted using a pre-trained feature extraction network as the first depth feature, and a distance guidance feature map is generated based on the feature distance constraint between the first depth feature and the depth features corresponding to the stegana image to be generated.

[0058] Specifically, the feature extraction network pre-trained for large-scale image classification tasks uses the activation responses of its deep convolutional layers to form a highly abstract and semantic representation of image content (such as object contours, texture combinations, and spatial structures). The core principle of this module is that an ideal steganalysis image should, while preserving the secret information, be as close as possible to the original carrier image in this high-level semantic feature space. By extracting and comparing the deep feature representations of the two, and quantifying this feature difference into an optimizable distance constraint, the generative network can be guided to prioritize the modification direction and intensity that has the least impact on the global semantics and main content of the image when making pixel-level modifications, thereby maintaining the visual naturalness and consistency of the steganalysis image at the perceptual level.

[0059] Furthermore, the distance-guided module typically includes a pre-trained deep network with frozen parameters (not involved in training) as a feature extractor, such as VGG16 or ResNet50. The preprocessed carrier image is first propagated forward through this network, extracting a high-dimensional feature tensor from its designated intermediate layer (such as the conv4_3 layer of VGG16 or the output of a residual block of ResNet50), which is the so-called "first deep feature". During the training phase, the simulated steganalyte output by the generator network is fed into the same location of the same feature extractor to extract the corresponding deep features. Subsequently, the difference between these two feature tensors is calculated, typically using the L2 norm (Euclidean distance) or cosine distance as a metric. This calculated distance value, or the gradient map generated from this distance, is transformed appropriately (such as inverting, scaling, or combining with an attention mechanism) to form a "distance-guided feature map", which is used to impose constraints in subsequent loss calculations or feature fusion.

[0060] Specifically, the key parameters of this module include the selection of the pre-trained network model, the specific network layer used for feature extraction, and the feature distance metric. Model selection (e.g., VGG, ResNet, Inception) determines the granularity and level of abstraction of semantic features; the selection of network layers requires a trade-off between the semantic abstraction of features and spatial resolution—shallower layers retain more spatial details but are sensitive to content changes, while deeper layers are semantically strong but less sensitive to spatial details. The distance metric directly affects the sensitivity and smoothness of the constraint. Furthermore, to control the strength of this constraint, an adjustable weighting coefficient is typically introduced when calculating the total loss. These parameters collectively define the concrete quantitative form of the abstract goal of "visual semantic consistency."

[0061] Specifically, the introduction of this module ensures that steganography does not compromise the visual integrity of key objects, faces, or scenes in an image. This is crucial for scenarios with strict requirements for the fidelity of the main content of an image, such as social media sharing, news image transmission, and digital archiving of artwork. Even with highly structured images (such as those containing text, faces, or specific logos), distance guidance effectively prevents perceptible semantic distortions or content aberrations caused by embedding operations. This allows steganography to be more deeply integrated into normal image use and circulation processes, greatly enhancing its concealment in open environments.

[0062] Specifically, on the one hand, in terms of visual quality, it effectively suppresses local artifacts or global tone shifts in steganographic images that are semantically inconsistent with the carrier image, making steganographic images more difficult to identify as modified under human observation. On the other hand, in terms of security, it increases the confusion between steganalytic modification and natural image content changes in the deep feature space, making it more difficult for advanced steganalysis tools based on deep learning and relying on high-level feature differences to effectively distinguish between the carrier image and the steganographic image. Therefore, this module is one of the core components for achieving high-quality, high-security steganography, complementing residual guidance and local entropy guidance to jointly construct a multi-layered, comprehensive feature protection system.

[0063] S23, the local information entropy distribution of the preprocessed carrier image is calculated through the local entropy guidance module to generate a local entropy feature map.

[0064] Specifically, in digital images, the entropy value of a local region quantifies the randomness and complexity of the gray-level distribution within that region. High-entropy regions represent rich textures, numerous details, or noise, and their pixel value distribution exhibits high uncertainty; low-entropy regions represent smooth, uniform, or gradually changing regions, and their pixel value distribution is highly predictable. Steganalysts often utilize modification behavior to detect perturbations to image statistical properties (especially those of smooth regions). The principle of this module lies in actively and quantitatively evaluating the inherent information complexity of each local region of the carrier image, and using this as a guiding signal: encouraging embedding in regions with high complexity and strong uncertainty, because additional modifications are easily masked by the original complex statistical properties; and suppressing embedding in regions with low complexity and strong determinism to avoid introducing easily detectable statistical anomalies. This is essentially an adaptive matching of the information hiding process with the information capacity distribution of the carrier itself.

[0065] Furthermore, the local entropy guidance module typically employs a sliding window method to traverse the preprocessed carrier image. For each pixel location in the image (or with a certain step size), a fixed-size neighborhood window is defined centered on that location. First, the gray-level histogram of all pixels within this window is calculated, and the probability of each gray level occurring is determined. Then, the entropy value of the window is calculated according to the information entropy formula. After traversal, a "local entropy feature map" corresponding to the spatial location of the original image is generated, where the value of each pixel represents the information entropy of the local window centered on that location. To improve computational efficiency, optimization techniques such as integral histograms can be used. The generated entropy feature map can be further normalized or adjusted through a nonlinear mapping function to make it more suitable for fusion with features of deep networks, or directly used to construct constraint terms in the loss function.

[0066] Furthermore, the core adjustable parameters of this module include the size and shape of the sliding window, the number of bits used for grayscale quantization, and possible subsequent transformation parameters. The window size (e.g., 5×5, 7×7, or 9×9) is a key parameter, determining the scale of "locality." Too small a size makes the system sensitive to noise, while too large a size blurs detail boundaries. Grayscale quantization (e.g., quantizing 256 levels to 64 levels) affects the accuracy and computational cost of probability estimation. Additionally, when generating the final guide signal, thresholding may be involved to delineate high / low entropy regions, or a sigmoid function may be used for soft partitioning; the threshold or function parameters also need to be set. The optimization goal for these parameters is to strike a balance between computational complexity and the accuracy of describing the statistical characteristics of the image, ensuring the effectiveness of the guide signal.

[0067] Specifically, this module can automatically identify and differentiate between smooth areas such as the sky and water surfaces and complex areas such as forests and grasslands in natural images, as well as regular and irregular areas in man-made images (such as architectural drawings and software interface screenshots). This allows the invention to maintain a high level of security when processing complex real-world images, making it particularly suitable for professional scenarios with combined security and capacity requirements, such as secure archiving of image libraries, copyright protection of commercial design drafts, and the need to achieve large-capacity covert embedding within a single image.

[0068] Specifically, by actively avoiding embedding in low-entropy smooth regions, it directly reduces the statistical anomaly of the steganalyte in the eyes of classic steganalysis methods (such as those based on SPAM and SRM features), as these analyzers are often extremely sensitive to statistical perturbations in smooth regions. Simultaneously, guiding modifications to high-entropy regions is equivalent to "hiding" secret information within the inherent uncertainty of the carrier image in an information-theoretic sense, improving the efficiency and security of the embedding operation. This module, along with the residual guidance module (focusing on spatial frequency) and the distance guidance module (focusing on semantic features), complements each other, forming a comprehensive, multi-layered guidance system from pixel statistics and spatial texture to high-level semantics. This is a key technical aspect ensuring that the steganalytes generated by this invention possess both high visual fidelity and strong anti-detection capabilities.

[0069] S24. The residual guided feature map, distance guided feature map and local entropy feature map are fused and input into the backbone network of the multi-feature fusion generation network for feature mapping and probability normalization to generate an embedding probability map.

[0070] Specifically, this step aims to integrate independent and complementary image embeddability evaluation information from different physical or statistical dimensions (high-frequency details in the residual domain, semantic distance in the feature space, and statistical entropy of local regions) into a unified, high-dimensional joint feature representation. The core principle is that the final embedding decision should not rely solely on a single feature, but rather be the result of a nonlinear comprehensive evaluation of multi-dimensional evidence. Through feature fusion, the network learns a complex mapping that weights, compares, and synthesizes the "local fit" information implicit in each guiding feature, ultimately converging at the network's output layer into a globally optimal, pixel-level probability distribution. This probability distribution not only reflects texture complexity but also accommodates semantic preservation requirements and statistical safety considerations, achieving an intelligent transformation from multi-perspective feature analysis to a single collaborative decision.

[0071] Furthermore, the three guiding feature maps—residual, distance, and local entropy—are initially fused at a specific level of the backbone network encoder (e.g., after shallow feature extraction) through channel concatenation or element-wise addition, forming a multi-channel fused feature tensor. This tensor is then fused a second time with features extracted from the original image by the backbone network at the same level (e.g., through attention gating or convolutional layers). The fused features undergo a series of upsampling, skip connections, and convolutional operations in the subsequent decoder section to gradually restore spatial resolution and deepen feature representation. Finally, at the end of the decoder, a 1x1 convolutional layer is used to reduce the number of channels to 1, and a Sigmoid activation function is applied to map the feature value of each pixel location to the [0, 1] interval, thereby generating an "embedding probability map" with the same size as the input image, where each value represents the final confidence that the pixel is selected for embedding secret information.

[0072] Furthermore, the fusion strategy involves choosing which network layer to fuse at, whether to use concatenation or summation, and whether to introduce adaptive weighting mechanisms (such as attention). The depth (number of downsampling iterations) and width (number of channels per layer) of the backbone network determine its ability to map and synthesize fused features, which is crucial for model capacity. Probability normalization typically uses the Sigmoid function, whose output range is fixed at [0,1]. In addition, during training, the model learns tens to hundreds of thousands of adjustable parameters to optimize the complex nonlinear function from multi-feature input to probability map output. The optimization goal of these parameters is to maximize the overall security, visual quality, and embedding efficiency of the generated probability map for subsequent stegographic images.

[0073] Specifically, this multi-feature fusion decision-making mechanism ensures the robustness and intelligence of the embedding strategy when faced with diverse and unknown real-world images. Regardless of the type of carrier image, the system can automatically and comprehensively evaluate its texture, semantic, and statistical properties and generate an adaptive embedding scheme that matches it. This enables the technology to be reliably deployed in automated or semi-automated steganography systems, such as secure transmission platforms for batch processing of image intelligence, covert watermark embedding services for social media content, or stealth identifier generation in digital rights management (DRM) systems. Its integrated intelligent decision-making process reduces the need for professional experience from operators and improves the feasibility and consistency of large-scale applications.

[0074] Specifically, implementing multi-feature fusion and collaborative mapping overcomes the limitations of single-feature guidance, resulting in a more reasonable and refined spatial distribution of the generated embedding probability map. This directly leads to three improvements: First, enhanced security, by integrating resistance to texture analysis (residual), semantic analysis (distance), and statistical analysis (local entropy), making the distribution patterns of steganalytic modifications more difficult for various steganalysts to model and capture; second, optimized visual quality, by incorporating semantic distance constraints, preventing damage to important visual content in advance during the probability allocation stage; and third, improved decision reliability, as multi-evidence fusion reduces the risk of misjudgment (such as high-probability embedding in seemingly complex but semantically sensitive regions). Ultimately, the high-precision probability map produced in this step is the most crucial basis for the entire steganalysis system to generate high-quality, high-security steganalytic images.

[0075] S3. Input the embedding probability map and the corresponding random encoded features into the preset deep feature embedding learning model, perform embedding simulation, and generate simulated steganalysis images.

[0076] Specifically, traditional steganography (such as STC) involves discrete, non-differentiable bit operations, making it impossible to directly integrate into the gradient backpropagation training framework of deep neural networks. The core principle of this step is to design a mathematically smooth surrogate function to approximate the mapping from "embedding probability" and "information to be embedded" to "pixel modification amount." The model receives two inputs: a probability map (representing the probability of modification at each position) and randomized encoded features (representing the bitstream of the secret information to be embedded). Through a carefully designed, parameter-learnable neural network structure, the model can compute a continuous and differentiable simulated modification amount corresponding to each combination of inputs. This allows the entire steganography system to be jointly optimized end-to-end; the generative network can adjust its parameters based on the results of the simulated embedding and its loss, thereby learning to generate a better embedding probability map.

[0077] Furthermore, the embedding probability map, after upsampling or maintaining its original size, is concatenated or multiplied pointwise with processed random encoded features (e.g., a multi-channel tensor adjusted to the same spatial size as the probability map, each channel potentially corresponding to a different embedding orientation or intensity). This combined feature is then fed into a core module consisting of two cascaded Tanh activation functions. The first Tanh function generates an initial, continuously modified orientation and intensity signal based on the secret information bits and probability intensity; the second Tanh function further smooths and amplitude-limits this signal, outputting a final signal whose range is constrained to a specific value. ( Let be a small constant, such as 1.5), which is a "simulated modification" tensor. Finally, this modification tensor is added point-by-point to the pixel values ​​of the preprocessed original carrier image to generate the "simulated steganalytic image". The entire process is fully differentiable.

[0078] Furthermore, the key parameters of this model include the mathematical form of the core simulation function (such as the double Tanh cascade structure) and the amplitude limiting parameter. This includes, and possibly, internal network weights used to adjust the simulation accuracy. The parameters of the dual Tanh structure itself (such as the linear transformation weights between cascades) are learnable. The parameter κ defines the maximum absolute value of the simulation modification, and its magnitude must be coordinated with the subsequent quantization step size used for the real embedding to ensure the effectiveness of the simulation. Furthermore, the dimension of the projection layer used to handle random encoded features is also a design parameter. The entire embedding learning model is typically small in scale and has minimal parameters. Its training objective is not to precisely reproduce every operation of the discrete STC, but rather to learn its statistically significant average effect, thereby providing a stable and meaningful gradient signal to the master generator network.

[0079] Furthermore, during the training phase, it allows the system to perform efficient end-to-end optimization in a continuous, differentiable simulation space, exploring the optimal embedding strategy. In the final practical application (inference) phase, the embedding probability map output by the trained generative network can be combined with a traditional optimal encoder (such as STC) to discretize and precisely embed real secret information. This design allows the invention to enjoy the advantages of automated embedding design in deep learning while ensuring full compatibility between the final steganalyte and standard steganalyte codecs. It is applicable to all practical covert communication scenarios requiring reliable information extraction, such as secure file transfer and digital rights management (DRM).

[0080] Specifically, introducing a deep feature embedding learning model for embedding simulation solves the gradient propagation bottleneck in end-to-end training, allowing gradients from the image quality and security loss function to propagate smoothly back to the front-end multi-feature fusion generative network, thus achieving direct and effective optimization of the generative network. Secondly, this continuous simulation, compared to simple binarization or sampling operations, provides the generative network with richer and lower-noise training signals, promoting stable model convergence. Finally, the generative network trained through simulation can generate embedding probability maps highly adapted to the optimal encoder, making the embedding position and modification strength of secret information more reasonable in real discrete embedding, thereby fundamentally improving the overall performance indicators of the final steganalyte image, including visual imperceptibility, resistance to statistical analysis, and embedding capacity effectiveness.

[0081] Furthermore, S3 includes: S31. Based on the embedding probability map, determine the embedding strength corresponding to each pixel position, and based on the corresponding random coding features, determine the modification direction to be simulated.

[0082] Specifically, this step works by converting the probabilistic assessment of "suitability for embedding" output by the generator network into two deterministic control variables required for the actual embedding operation: strength and direction. The embedding probability map provides the prior confidence level of each pixel location carrying secret information, but its value itself is not a direct operational instruction. The core principle of this step is to establish a continuous mapping function from probabilistic confidence level to physical modification magnitude, while simultaneously decoding the random encoded features, which serve as the carrier of secret information, into specific modification polarities. Essentially, this bridges the decision-making process for information hiding (completed by the generator network) with the execution process of signal modulation (completed by the embedding simulation module), ensuring that high-probability regions receive correspondingly stronger modification authorization, while the secret information bits unambiguously determine whether the modification is positive or negative, thus mathematically providing complete input conditions for subsequent differentiable simulated modifications.

[0083] Furthermore, the embedding strength is typically determined by performing linear or nonlinear transformations on the embedding probability map. For example, the probability value can be directly multiplied by a preset maximum strength coefficient α, i.e., Strength = Probability Map × α. The random coding feature is usually a binary sequence (e.g., ±1) corresponding to the secret information, or a modulated multi-valued sequence. To determine the modification direction, this sequence needs to be spatially aligned and dimensionally transformed, mapping it to a direction map of the same size as the embedding probability map, where each pixel value is +1 or -1, representing an increase or decrease in pixel value, respectively. In some implementations, the direction may be determined by more complex multi-level coding. Finally, the strength map and the direction map, as a pair of paired, same-sized control signals, are simultaneously fed into the subsequent analog modification calculation unit.

[0084] Specifically, the core parameters of this step include the intensity mapping function and its coefficients, the modulation method of the random coding features, and the alignment rules. The choice of the intensity mapping function (e.g., linear scaling, exponential scaling) and the maximum intensity coefficient α are crucial, as they jointly determine the dynamic range of the simulated modification amount, directly affecting the signal-to-noise ratio and visual quality of the steganalyte. The bit-to-direction symbol mapping rule of the random coding features (e.g., bit '0' maps to -1, bit '1' maps to +1) must be strictly consistent with the decoding rule during the final real embedding; this is the foundation for ensuring information extractability. Furthermore, the processing logic that ensures the spatial alignment of the coded features with the probability map (e.g., through space filling, interleaving, or convolutional projection) is also an important implementation parameter.

[0085] Specifically, whether in simple steganography tools for consumer applications or in specialized systems with high security requirements (such as military intelligence or trade secret protection), this step provides a standardized interface for connecting abstract strategies with specific signal modulation. This enables deep learning-based adaptive embedding strategies to seamlessly integrate with various mature steganography frameworks, greatly expanding the applicability of this invention. It allows for both experimental research and easy integration into existing secure communication and information protection product lines.

[0086] Specifically, the direct effect of this step is the generation of precise, differentiable embedding control signals. By transforming probability values ​​into continuous intensity, rather than simple binary decisions, more nuanced gradient feedback is provided to the generative network, which is beneficial to its optimization process. By explicitly separating intensity and direction, the simulation process can more realistically reflect the operation of actual encoders such as STC: that is, larger ±1 modifications may be made in high-probability regions, while smaller modifications or no modifications are made in low-probability regions. This fine-grained control enables the finally trained generative network to produce probability maps with a very high degree of matching with the optimal encoder, thereby achieving efficient and secure embedding of secret information in the application stage, while maintaining the visual fidelity and statistical naturalness of the carrier image to the greatest extent.

[0087] S32 inputs the embedding strength and modification direction into the dual-Tanh simulation STC model. The dual-Tanh simulation STC model performs nonlinear mapping on the input through two cascaded Tanh functions, and outputs a continuous and differentiable simulation modification.

[0088] Specifically, the core idea of ​​this step is to utilize the saturated nonlinearity of the hyperbolic tangent (Tanh) function to perform continuous reshaping and bounded processing on the original input signal, which consists of the embedding strength and modification direction, in two stages. The first-stage Tanh function simulates a preliminary, nonlinear "decision-modulation" process based on the secret information bits and local probability strength, mapping the input signal to an intermediate continuous space. The second-stage Tanh function further smooths and compresses the amplitude of this intermediate signal, acting like a soft limiter to ensure that the final output analog modification is strictly limited within a pre-defined bounded interval that conforms to the actual physical modification amplitude. This two-stage cascaded structure mathematically constructs a complex mapping function that is continuous everywhere from the control signal to the analog modification, thus enabling the entire analog embedding process to be seamlessly integrated into the gradient descent-based neural network training framework.

[0089] Furthermore, the model receives two inputs: the original driving signal obtained by element-wise multiplication of the embedding strength and the modification direction, and possible contextual features. This driving signal first passes through a linear transform layer and then is input to the first Tanh activation function to generate a preliminary modulation signal. This modulation signal then passes through a second linear transform layer and is input to the second Tanh activation function. The output of the second Tanh function is then multiplied by a preset, learnable, or fixed amplitude scaling factor κ to finally generate the simulated modification amount Δ. This modification amount Δ is added to the pixel values ​​of the carrier image to obtain the pixel values ​​of the simulated steganalyte image. The entire data stream can be formally represented as: Where W1, b1, W2, and b2 are learnable parameters, and ⊙ represents element-wise multiplication. This structure ensures that the output Δ is within the range of... Internal continuous change.

[0090] Furthermore, the core adjustable parameters of this model include the weights and biases of the two-stage linear transformation ( The model incorporates several parameters, including the amplitude limiting factor κ and possible intermediate layer dimensions. The amplitude limiting factor κ is a key hyperparameter that defines the maximum absolute value of the simulated modification, typically determined by the target embedding distortion level or pixel depth (e.g., for 8-bit images, κ is often set to 1.5 or 2.0). The dimensions of the two-level linear transformation determine the model's nonlinear fitting capability; their parameter count is usually kept small to prevent overfitting and maintain computational efficiency. These parameters are initialized early in training and jointly optimized via backpropagation throughout the end-to-end training of the generative network, aiming to approximate the distribution characteristics of the simulated modification to approximate the statistical characteristics of the real STC encoding under the guidance of the corresponding probabilistic map.

[0091] Specifically, the dual-Tanh simulated STC model enables highly approximate steganalytic performance evaluations without running time-consuming, discrete real STC encodings during the training phase, thus significantly accelerating model iteration and convergence. This design makes this method particularly suitable for steganalytic research and application scenarios requiring rapid prototyping, large-scale parameter search, or online adaptive learning. In the final deployment phase, although the simulator itself may be replaced with standard STC for real embeddings, it is its "teaching" during the training phase that enables the generative network to learn how to generate embedding probability maps that perfectly match the optimal encoding.

[0092] Specifically, employing a dual-Tanh simulation of the STC model provides a fully differentiable path from the loss function to the generator network, solving the core gradient propagation challenge of end-to-end steganalysis training and enabling global optimization based on image-level losses (such as perceptual loss and statistical loss). Secondly, its continuous and smooth characteristics provide the generator network with low-noise, high-information gradient signals, effectively avoiding excessive gradient estimation variance or gradient vanishing problems caused by discrete operations, thus promoting more stable and faster model convergence. Finally, through this precise simulation training, the generator network can learn to predict the probability map that minimizes detectable modifications to the real STC encoding, thereby fundamentally endowing the final stegana image with excellent anti-detection (security) and high visual fidelity.

[0093] S33, the simulated modification amount is added to the pixel value of the preprocessed carrier image to generate the steganalytic image.

[0094] Specifically, the finely controlled analog modification, generated through the joint decision-making of the aforementioned multi-feature guidance network and analog encoder, is considered as a weak perturbation signal with a specific spatial distribution and amplitude applied to the original carrier image. According to the additive noise steganography model, the generation process of the steganographic image can be mathematically modeled as the linear addition of the carrier image signal and the steganographic modulation signal. This linear addition operation, in principle, ensures that the secret information can be carried in a predictable and reversible manner within the minute changes in pixel values. Simultaneously, since the analog modification has been subjected to multiple nonlinear constraints during the generation process (such as amplitude saturation of double Tanh), its addition to the carrier image ensures at the signal level that the synthesized pixel values ​​do not exceed a reasonable range, and at the perceptual level, it utilizes visual masking effects to make the modification more easily integrated into the image background texture.

[0095] Specifically, this operation manifests as an efficient element-wise tensor addition. The preprocessed carrier image is typically represented as a tensor of size [size missing]. A floating-point tensor, where H is the height, W is the width, and C is the number of color channels (e.g., 3), whose pixel values ​​have been normalized to the target range (e.g., ...). The simulated modification amount is a variable with the same size. A tensor, where each element represents a suggested modification value for the corresponding pixel channel, typically ranging from... ( (A small constant). The operation of generating steganographic images is directly accomplished through broadcast addition in deep learning frameworks (such as PyTorch and TensorFlow): Steganographic image tensor = Carrier image tensor + Simulated modification tensor. To ensure that the result is still within the valid pixel value representation range (e.g., ... An overflow can be truncated to the boundary after addition, or overflow can be naturally avoided by relying on the saturation characteristics of the previous simulation module.

[0096] Furthermore, the parameter directly involved in this step is the dynamic range boundary of the simulated modification amount. This parameter is a key output constraint of the aforementioned dual-Tanh simulation STC model, and it determines the maximum possible magnitude of a single modification. The choice of values ​​needs to strike a balance between the reliability of information embedding (sufficient signal-to-noise ratio) and visual imperceptibility. Furthermore, while the addition operation itself is deterministic, the numerical precision (e.g., 32-bit or 16-bit floating-point numbers) of its inputs—the carrier image tensor and the analog modification tensor—constitutes the fundamental numerical metrics of this step, affecting the accuracy of the calculation results, the quality of the final image, and the hardware computational efficiency. During deployment, it may also be necessary to consider whether and how rounding or requantization to integer representation is needed after the addition operation (e.g., ...). This relates to compatibility with standard image formats.

[0097] Specifically, the addition step transforms all the complex feature extraction, probability evaluation, and analog encoding calculations into a single standard digital image file that can be stored, transmitted, or distributed. This step is an indispensable output interface, whether for generating large-scale steganalytical image datasets for security testing and evaluation, or for real-time generation of encrypted vectors in actual covert communication links. Its efficiency and determinism ensure that this invention can be integrated into various image processing pipelines or real-time communication systems, meeting the needs of diverse scenarios ranging from offline batch processing to online interactive applications.

[0098] Specifically, this pixel value addition operation faithfully executes the embedding strategy jointly optimized by the front-end network and the analog encoder, translating the strategy into actual image data changes. Secondly, because the modifications are continuous and constrained, the steganographic image generated by direct addition maintains the continuity of pixel statistical distribution with the original carrier to the greatest extent possible, avoiding additional statistical flaws that may be introduced by binarization or crude quantization. Finally, the efficiency of this operation ensures that the speed bottleneck of the entire steganographic image generation process is not here, thus highlighting that this invention maintains high performance while also possessing superior generation efficiency. This makes the computational cost of generating high-security steganographic images comparable to conventional image filtering processing, greatly enhancing the practical value of the technology.

[0099] S4. Calculate the multi-guided joint loss between the simulated steganalysis image and the preprocessed carrier image, and use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network.

[0100] Specifically, the core principle of this step is to jointly optimize a set of carefully designed sub-loss functions, each targeting different image characteristics (such as high-frequency residuals, deep semantics, local statistical distribution, and embedding capacity). This forces the multi-feature fusion generative network to simultaneously satisfy multiple interrelated but potentially mutually restrictive constraints during parameter updates. This joint loss serves as a comprehensive evaluation function for network performance, and its gradient is backpropagated to the generative network via a differentiable embedding simulation module through a backpropagation algorithm, driving the network parameters to adjust towards generating an embedding probability map that minimizes the joint loss. In principle, this process achieves a closed-loop learning process from "generation-simulation-evaluation" to "feedback-optimization," enabling the network to automatically explore and approximate Pareto optimal solutions under different constraints.

[0101] Furthermore, the multi-guided joint loss consists of a weighted sum of four core loss components. First, the residual-guided loss constrains the consistency of the simulated steganographic image and the carrier image in high-pass filtered residual domains by calculating the L1 norm difference between them. Second, the distance-guided loss utilizes a pre-trained deep network to extract high-level features from both images and calculates the L2 distance between their feature maps to maintain semantic similarity. Third, the local entropy-guided loss guides the embedding behavior towards regions with complex textures by penalizing high embedding probabilities generated in low-entropy regions. Finally, the embedding loss (such as cross-entropy loss) ensures that the simulated embedding process can effectively carry the capacity of secret information. The total loss function takes the form: ,in These are adjustable weight coefficients. During training, stochastic gradient descent or its variants (such as the Adam optimizer) are used to minimize... With the goal of iteratively updating all trainable parameters of the multi-feature fusion generative network, the model converges.

[0102] Furthermore, the core adjustable parameters of this training process include the weight coefficients of each loss term ( The optimization process involves several key parameters: the type of optimizer and its hyperparameters (such as initial learning rate, momentum, and weight decay coefficients), batch size, and the total number of training iterations. Setting the loss weights is crucial for balancing different optimization objectives and requires empirical or automated adjustments based on the task's focus (e.g., prioritizing security or visual quality). The learning rate is one of the most important optimization hyperparameters and often employs warm-up and decay strategies. The number of trainable parameters in the network, the size and diversity of the training dataset, and the final total loss value after convergence are key quantitative indicators for measuring model complexity and training effectiveness. The coordinated configuration of these parameters directly determines whether the model can successfully learn effective embedding strategies.

[0103] Specifically, this end-to-end training mechanism based on multi-guided joint loss enables the method to adaptively learn steganography strategies suitable for different image distributions and security levels through a data-driven approach. Without the need for manually designing complex distortion functions, the system can automatically summarize universal embedding patterns from a large number of image samples. This allows the technology to be easily applied to customized scenarios. For example, by adjusting the loss weights or fine-tuning on domain-specific image data (such as medical images or design drawings), a domain-optimized steganography model can be quickly obtained, greatly enhancing the method's flexibility and engineering practicality. It is suitable for various secure communication and information protection systems with dual requirements for security and adaptability.

[0104] Specifically, firstly, end-to-end training using multi-guided joint loss forces the generative network to learn a complex and robust embedding strategy that balances texture adaptability, semantic preservation, statistical security, and embedding effectiveness through multi-angle and comprehensive joint constraints, thereby fundamentally improving the overall performance of steganographic images. Secondly, the end-to-end training paradigm ensures high coordination throughout the entire process from feature extraction and probability prediction to simulated embedding, avoiding the inconsistencies in objectives and error accumulation that may occur with staged optimization. Finally, the trained generative network possesses strong generalization capabilities, capable of generating high-quality embedding probability maps even for images not present in the training set, thus ensuring that this invention can stably produce high-security, high-visual-fidelity steganographic images in various practical application scenarios.

[0105] Furthermore, S4 includes: S41, calculate the residual guided loss, distance guided loss and local entropy guided loss between the simulated steganalysis image and the preprocessed carrier image, respectively.

[0106] Specifically, the residual-guided loss is calculated based on high-frequency component analysis in signal processing, aiming to assess the degree to which steganography perturbs the detailed structure of image edges and textures. The distance-guided loss relies on semantic metrics in the deep feature space, evaluating the impact of steganography on the consistency of high-level visual content by comparing the differences in their high-dimensional semantic representations. The local entropy-guided loss is rooted in information theory, assessing the degree to which embedding behavior alters the inherent statistical properties of the image by quantifying the differences in the complexity of their local statistical distributions. These three types of losses, from pixel-level details and semantic content to statistical distribution, construct a hierarchical and comprehensive evaluation system.

[0107] Furthermore, the implementation of residual-guided loss is as follows: First, the simulated steganographic image and the carrier image are passed through the same set of high-pass filters (such as the Kirsch operator) to extract their high-frequency residual maps; then, the L1 norm of the pixel-wise difference between the two residual maps is calculated, and the mean is taken. The implementation of distance-guided loss is as follows: The two images are input into a pre-trained deep network (such as VGG16) with frozen parameters, and feature maps are extracted from specified intermediate convolutional layers; then, the square of the L2 norm of the position-wise difference between the two feature tensors in the channel dimension is calculated, and finally, the average is taken in the spatial dimension. The implementation of local entropy-guided loss is as follows: The local information entropy maps of the two images are calculated separately (usually using the sliding window method); then, the embedding probability map of the simulated steganographic image is weighted by an attention mask generated based on the entropy map, and the mean of the L1 norm of the weighted probability map is calculated to penalize embedding behavior in low-entropy regions.

[0108] Furthermore, for residual-guided loss, the core parameters are the type and size of the high-frequency filter kernel, which determines the scale of detail of interest. For distance-guided loss, key parameters include the selected pre-trained network model, the specific network layer used for feature extraction, and the distance metric norm used (L1 or L2). For local entropy-guided loss, the main parameters are the sliding window size, the gray-level quantization scheme for calculating entropy, and the threshold function or soft partitioning parameter used to generate the attention mask. Additionally, before the final weighted summation, the output value of each type of loss typically needs to be adjusted by an optional scaling factor to balance its numerical magnitude; this scaling factor is also an important adjustment parameter.

[0109] Specifically, this multi-dimensional loss calculation mechanism allows the model training process to be flexibly adapted to the core requirements of different application scenarios. For example, when combating traditional steganalysts that primarily analyze texture anomalies, the weight of the residual-guided loss can be appropriately increased; in situations requiring extreme preservation of the invariance of the main image content (such as covert communication using facial photos), the constraint of the distance-guided loss can be strengthened; and when combating statistical model-based analysis methods, the role of the local entropy-guided loss is crucial. This configurability enables the present invention to quickly customize the most targeted steganalysis models for different security levels and image types in military, commercial, and civilian applications, greatly expanding the applicability of the technology.

[0110] Specifically, the separate calculation and joint optimization of multi-dimensional guided losses are the fundamental driving force behind generative networks learning the optimal embedding strategy. This ensures that while reducing the total loss, the network must simultaneously consider and properly balance the protection of high-frequency details, the maintenance of high-level semantics, and adherence to local statistical regularities. This collaborative optimization mechanism forces the network to explore a solution located on a better Pareto front, directly resulting in superior performance of steganographic images in the following aspects: visually, the details are natural and the main content is distortion-free; statistically, its distribution characteristics highly match those of natural images; and in terms of security, it can simultaneously resist steganalysis attacks from multiple dimensions, including texture, semantics, and statistics. Therefore, this step is the core algorithmic step for comprehensively improving the overall performance of the steganography system.

[0111] S42, the residual guided loss, distance guided loss and local entropy guided loss are weighted and summed to form the multi-guided joint loss, and the parameters of the multi-feature fusion generation network are iteratively optimized with the goal of minimizing the multi-guided joint loss until convergence.

[0112] Specifically, the principle behind iteratively optimizing a multi-guided joint loss by weighted summation of various guided losses lies in applying the linear scalarization method from multi-objective optimization theory. This method merges previously interconnected and potentially conflicting multi-dimensional sub-objectives into a single optimizable scalar objective function. The core principle is that by assigning a non-negative weight coefficient to each sub-loss, the relative importance preference among different safety and quality dimensions in the overall optimization problem is defined. This joint loss function, as the final output of the entire differentiable computational graph, transmits its value, reflecting the comprehensive gradient information of constraints across dimensions, to all trainable parameters of the multi-feature fusion generation network through backpropagation. The iterative optimization process, based on gradient descent, continuously adjusts the network parameters to explore the parameter space, seeking the point that minimizes the weighted sum. In principle, this process aims to guide the network to converge to a Pareto optimal solution under given weight preferences, a solution that allows the network to generate an embedding probability graph that achieves optimal balance under multiple constraints.

[0113] Specifically, the multi-guided joint loss is calculated using the following formula: ,in These are preset or learnable positive real-valued weights. In each iteration of the training loop, a forward propagation is first performed to calculate the joint loss value with the current network parameters. Then, the automatic differentiation function of the deep learning framework (such as PyTorch or TensorFlow) is invoked to calculate the gradient of the joint loss with respect to the parameters of each network layer. Next, using a selected optimizer (such as Adam or SGD), all trainable parameters of the network are updated based on the calculated gradients and the optimizer's own state (such as momentum). This cycle of "forward calculation of loss - backpropagation of gradients - optimizer parameter update" is repeated continuously, using a large number of image samples for training, until the model's performance stabilizes on the validation set or reaches the preset number of iterations, which is considered convergence.

[0114] Furthermore, the core parameters of this optimization process include the weighting coefficients of each loss term ( The optimization process involves defining the optimizer type, its hyperparameter set, and convergence criteria. Setting the weighting coefficients is crucial for balancing the optimization direction and can be done through grid search, automatic tuning based on the Pareto front, or manual setting based on prior knowledge. Optimizer hyperparameters mainly include the base learning rate, momentum coefficient, and weight decay coefficient. The learning rate often employs scheduling strategies such as cosine annealing or step descent. Convergence criteria can include the validation set loss no longer significantly decreasing over multiple consecutive epochs, reaching the maximum number of training epochs, or a certain evaluation metric for the generated images reaching a threshold. The total number of iterations in the entire optimization process, the final stable loss value, and the training time are key indicators for measuring training efficiency and effectiveness.

[0115] Specifically, during the R&D phase, by adjusting the weight combinations, model variants with different focuses, such as "ultimate security," "optimal visual quality," or "large-capacity embedding," can be quickly explored. In actual deployment, pre-trained model weights with corresponding characteristic preferences can be selected and loaded based on the security level of the specific task, the type of carrier image, and channel conditions. This flexibility allows the same core technology to adapt to a wide range of application needs, from defense classified communications (high security weights) to civilian copyright protection (high visual quality weights), enhancing the transformation potential and market adaptability of technological achievements.

[0116] Specifically, firstly, end-to-end iterative optimization based on weighted joint loss explicitly and tunably defines the desired steganalysis performance equilibrium point through mathematical programming, guiding the network to systematically learn a complex mapping function to reach this equilibrium point. Secondly, the standard gradient descent optimization framework ensures the stability and reproducibility of the training process, making the acquisition of high-performance models no longer dependent on cumbersome manual debugging or unstable adversarial games. Finally, the network trained through this process converges, and its internal parameters encode the "knowledge" of solving the optimal embedding strategy under multiple constraints. This enables it to quickly and accurately output a matching embedding probability map for any input image, which is the fundamental reason why the final steganalysis image exhibits excellent performance in terms of security, imperceptibility, and robustness.

[0117] S5. Input the carrier image to be processed into the trained multi-feature fusion generative network to obtain a new embedding probability map, and obtain the final steganalysis image based on the new embedding probability map.

[0118] Specifically, the training process encodes the network's weight parameters with general knowledge learned from a large number of samples regarding the complex relationships between image features and optimal embedding locations. When the image to be processed is input into the network, the forward propagation process activates the feature extractors and decision functions learned in each layer of the network, calculates residuals, distances, and local entropy guidance signals in parallel, and fuses and maps them deep within the network. Finally, the network's output layer generates an optimal embedding probability map tailored to the specific input image based on these internal calculations. This probability map, along with the secret information to be hidden, is then processed by a standard encoder (such as STC) that is compatible with the training phase but no longer requires gradient backpropagation, to perform discrete and precise physical pixel modifications, generating a final steganalyte image that can be directly used for transmission or storage.

[0119] Specifically, first, the carrier image to be processed (usually preprocessed with the same standardization and normalization as in the training phase) is input into the trained multi-feature fusion generation network. The network operates in frozen parameter mode, and after a series of convolution, activation, pooling, and feature fusion operations, it outputs a floating-point "new embedding probability map" at the end of the network, with the same size as the input image and each pixel value in the range [0, 1]. Second, the final steganalyte image is obtained based on this probability map: the bitstream of the secret information to be hidden is combined with the generated probability... Figure 1 The input is a standard, optimal STC encoder or a similar embedding encoder. The encoder modulates the secret information bits into the carrier image according to the cost defined by the probabilistic graph, performs specific +1 or -1 modifications on selected pixels (or other modifications according to the embedding strategy), thereby generating an irreversible "final steganalyte image" that can be accurately extracted by the decoder.

[0120] Furthermore, the core "parameters" involved in this stage are all the fixed weights and structural hyperparameters of the trained generative network, which determine the model's capacity and behavior. The numerical range and resolution of the embedding probability map (consistent with the input image) are key intermediate output metrics. In addition, in the final embedding stage, the parameters of the specific STC encoder used (such as the check matrix and embedding efficiency) must be consistent with the target approximated by the simulated STC model during training to ensure performance consistency. The quality metrics of the final output image, such as peak signal-to-noise ratio, structural similarity index, and the detection error rate against various steganalyzers, are comprehensive quantitative indicators for measuring the actual effectiveness of this step.

[0121] Specifically, this step marks the formal transition of this technology from the model training phase to the practical deployment phase. Its efficient and stable forward inference characteristics allow this method to be packaged as a standalone software module or hardware acceleration IP core and integrated into various products requiring image steganography capabilities. For example, it can be integrated into secure instant messaging software to generate encrypted images in real time, embedded in digital rights management systems to add invisible watermarks to image assets, or used in intelligence analysis systems to annotate publicly available images with hidden information. Its processing speed (the number of images processed per second or the processing time per image) is a key engineering indicator for measuring whether it can meet the real-time requirements of practical applications.

[0122] Specifically, firstly, the intelligence is reflected in the automatic and adaptive generation of the optimal embedding strategy for each new image, without the need for manual intervention or complex parameter adjustments. Secondly, the efficiency is reflected in the fast inference speed and controllable resource consumption, enabling high-security steganography operations to be performed on a large scale or in real time. Finally, high performance is guaranteed because the probabilistic graph produced by the network undergoes end-to-end joint optimization with multi-dimensional security and quality indicators as the goal. Guided by this optimization, the final embedding is performed, and the generated steganographic image theoretically approaches the optimal security-distortion boundary under a given network structure and training data, thus possessing excellent concealment and robustness in the real world.

[0123] Furthermore, S5 includes: S51, the carrier image to be processed is input into the trained generative network. The residual guidance module, distance guidance module and local entropy guidance module of the trained generative network work together to extract and fuse multi-scale depth features and guidance features, and output a new embedding probability map after decoding and mapping.

[0124] Specifically, this step utilizes a pre-trained, fixed-parameter multi-feature fusion generative network as a deterministic multimodal feature perception and collaborative decision-making system to perform a complete forward inference on a new input image to generate a customized embedding strategy. The core principle is that the network's three guiding modules and the backbone network function as a whole, their internal parameters having been trained end-to-end, learning a complex, nonlinear mapping from image pixels to the optimal embedding probability. When a new carrier image is input, the residual guiding module, distance guiding module, and local entropy guiding module, based on their respective fixed (i.e., trained and no longer updated) weight parameters, extract complementary guiding features from the image in parallel: high-frequency details, deep semantics, and local statistical complexity. These features are collaboratively integrated at pre-defined fusion points in the network, and through a cascaded nonlinear transformation of the encoder-decoder architecture, the fused multi-scale deep features are ultimately decoded and mapped into a two-dimensional probability distribution map. Essentially, this process projects the feature space of the input image into an "optimal embedding probability" space defined by multi-dimensional security and quality constraints.

[0125] Further, the carrier image to be processed undergoes preprocessing consistent with the training phase to form a standard three-dimensional tensor. This tensor is first input into the backbone of the generative network (such as a ResNet encoder) for basic feature extraction. Simultaneously, image copies are fed into three parameter-frozen guidance modules: the residual guidance module calculates residual feature maps using a pre-defined high / low-pass filter bank; the distance guidance module extracts high-level semantic feature tensors using a frozen pre-trained network (such as VGG16); and the local entropy guidance module generates a local entropy distribution map using a fixed sliding window and entropy calculation formula. These guided feature maps are fused with the backbone features at a specified layer of the backbone network (such as at the end of the downsampling path), typically through channel concatenation or weighted summation. The fused feature tensor is then upsampled and convolved by the network decoder to gradually restore spatial resolution and refine feature mapping. Finally, through a 1×1 convolutional layer and a sigmoid activation function, a "new embedding probability map" of the same size as the input image is output, where each pixel value represents the recommendation strength of the embedding information at that location.

[0126] Furthermore, the core "parameters" of this step are the millions to tens of millions of trained and fixed weights of the entire generative network; these weights are the carriers of the network's intelligence. Key intermediate data metrics in the process include: the number of channels and spatial resolution of the feature maps output by each guiding module, the dimension of the feature tensors output by each fusion layer, and the data range of the final embedded probability map (typically floating-point numbers in the range of [0, 1]). The computational complexity (measured in FLOPs) and memory usage of the entire forward inference process are important engineering metrics for evaluating its deployment efficiency. Inference speed (e.g., frames per second, FPS) directly relates to the real-time performance of practical applications. These metrics are all defined by the network architecture before training and remain constant during the deployment phase.

[0127] Specifically, this step is the core operation for enabling the technology of this invention to move from the "training platform" to the "production environment." Its determinism and efficiency allow the technology to be encapsulated as a standard software function library or hardware acceleration module, seamlessly integrated into various practical systems. Whether used for batch adding invisible copyright watermarks to massive amounts of publicly available network images, embedding secret information into single images in real-time in secure instant messaging, or automatically and covertly marking specific target images in intelligence analysis systems, this step provides a stable, reliable, and fast strategy generation service. Its powerful adaptive capability ensures that high-quality embedding schemes are output regardless of changes in the input image content, greatly improving the system's practicality and reliability in open and complex environments.

[0128] Specifically, this step fully leverages the intelligent advantages inherent in a fully trained generative network. First, it achieves real-time and high-quality policy generation, generating near-theoretically optimal embedding probability maps for any new image within milliseconds. Second, it ensures the consistency of the output policy; due to fixed network parameters, the output probability map is completely deterministic for the same input image, meeting the requirements of high-security applications for predictable and reproducible algorithm behavior. Finally, it inherits the comprehensive performance optimizations from the training phase. The generated embedding probability map comprehensively considers texture complexity, semantic consistency, and statistical security. Using this as a blueprint for final information embedding ensures that the resulting steganographic image simultaneously possesses excellent visual imperceptibility, strong resistance to various steganalysis detection methods, and a high secret information recovery rate, thus comprehensively achieving the technical objectives of this invention.

[0129] S52, the new embedding probability map and the corresponding random coding features are input into the preset deep feature embedding learning model. The preset deep feature embedding learning model weights each bit of the corresponding random coding features according to the new embedding probability map, and calculates the pixel-level simulated modification amount by simulating the coding process through a cascaded hyperbolic tangent function.

[0130] Specifically, this step is implemented by using a differentiable, parameterized neural network module to perform a high-fidelity mathematical simulation of the probability-weighted information encoding process. The core principle is to treat the "new embedding probability map" as a prior confidence distribution of the importance of embedding each secret information bit. Based on this probability map, the deep feature embedding learning model performs spatially adaptive bit-by-bit weighting on the "random encoded features" (i.e., a vectorized representation of the secret information bitstream to be embedded). This weighting operation simulates the core idea of ​​dynamically allocating embedding resources according to distortion cost in practical optimal encoders (such as STC). Subsequently, the weighted signal is passed through a nonlinear mapping network composed of cascaded hyperbolic tangent functions. This network simulates the complex transformation process from the weighted information signal to continuous pixel modifications, including amplitude constraints and saturation effects, thereby approximating the statistical distribution of the expected modifications produced by discrete STC encoding under a given probability map within a continuously differentiable framework.

[0131] Specifically, the pre-defined deep feature embedding learning model (typically referring to the dual-Tanh simulated STC model used during training) runs with its parameters frozen. The input includes a new embedding probability map P (size: ...). ) and the corresponding random coding features M (which are usually processed as multichannel tensors of the same size as the image, such as (where C may be related to the number of embedding directions). First, the model performs a weighting operation: , where ⊙ represents broadcast multiplication or pointwise multiplication along the channel dimension, which gives higher weight to the encoded features of high-probability regions. Subsequently, The signal is fed into a network consisting of two cascaded Tanh functions and a linear transform layer in between. The first-stage Tanh function performs preliminary nonlinear modulation on the weighted signal, while the second-stage Tanh function further smooths and constrains the output range. Ultimately, the network outputs a signal with a size of [value missing]. (generally The "pixel-level analog modification" tensor (equal to the number of image channels) Each of its values ​​is strictly limited to Within a continuous interval.

[0132] Furthermore, the core parameters of the model involved in this step are fixed, including the weight matrix and bias of the two-stage linear transformation, as well as the final magnitude scaling factor. These parameters are optimized during the training phase to ensure that the statistical properties (such as mean, variance, and spatial correlation) of the simulated modification amount Δ approximate the statistical properties of the modification amount produced by the real STC encoding under the guidance of the same probability map P. Scaling factor This is a key constant that defines the maximum magnitude of the simulated modification, and its value is consistent with the target embedding strength set during the training phase. The model's processing speed (e.g., the time required to process a single image) and the numerical precision of the output modification Δ (e.g., 32-bit floating-point) are key metrics for evaluating the computational efficiency and accuracy of this step.

[0133] Specifically, in the final real-world steganography system, this simulation model is typically replaced by a standard, discrete STC encoder to generate the final image. However, it is indispensable in the following scenarios: 1) Algorithm research and optimization: quickly evaluating the performance of different embedding probabilistic graph schemes without running time-consuming full encoding; 2) Model security verification: in adversarial training or robustness testing, a differentiable embedding process is needed to generate adversarial examples or analyze model vulnerabilities; 3) Simulation of novel encoding strategies: when exploring new encoding frameworks beyond STC, this differentiable model can serve as a prototype verification tool.

[0134] Specifically, this step, acting as a "high-performance simulator" connecting intelligent policy generation and final physical implementation, primarily improves development efficiency and performance assurance. First, it provides extremely fast policy performance prediction capabilities, enabling researchers and engineers to evaluate the quality of newly generated probabilistic graphs within seconds, significantly accelerating the iteration cycle. Second, its differentiability ensures closed-loop optimization during the training phase, a prerequisite for the front-end generative network to obtain effective training signals. Finally, because its model parameters are carefully trained to match the statistical behavior of the optimal encoder, its simulation results have high fidelity. This allows evaluations and optimizations based on these simulation results to be reliably transferred to a real STC encoding environment, thereby ensuring the reliability and superiority of the final steganography system's overall performance.

[0135] S53, the simulated modification amount is superimposed on the corresponding pixel value of the processed carrier image, and the final steganalytic image is directly synthesized and output.

[0136] Specifically, this step is implemented by performing a deterministic signal synthesis operation. The simulated modification, optimized by a deep network, is used as a weak additive perturbation signal and linearly superimposed onto the pixel domain of the carrier image, thus physically carrying the secret information. Based on the additive noise model for information hiding, the steganographic image is considered to be formed by adding the carrier image and the encoded steganographic signal. Here, the simulated modification is constrained to a perturbation pattern that is small in amplitude and highly adaptive to the image content in spatial distribution through intelligent decision-making by the front-end network and continuous processing by the analog encoder. Mathematically, the superposition operation realizes the final transformation of the secret information from abstract probability distributions and encoded features to specific pixel value changes. Simultaneously, because the modification undergoes nonlinear saturation processing, it ensures that the synthesis result does not introduce invalid overflow noise and fully utilizes the characteristic of the human visual system being insensitive to small brightness changes in complex texture areas (i.e., the visual masking effect), thereby achieving the covert embedding of information.

[0137] Specifically, the carrier image has been preprocessed (size normalization and pixel value normalization) and represented as a floating-point tensor. Its range is usually The simulated modification amount Δ is a [missing value]. Tensors with the same height, width, and number of channels have a range of . ( (The constant is small). The synthesis operation is directly performed through broadcast addition using deep learning frameworks or numerical computation libraries: The resulting tensor This refers to simulated steganography. To ensure the output conforms to standard image formats (such as 8-bit depth), it is usually... Perform numerical clipping, clamping its range to The algorithm calculates the range, then performs an inverse normalization operation (such as multiplying by 255 and rounding), finally outputting a standard image file (such as PNG format) with integer pixel values. The entire process is computationally deterministic and efficient.

[0138] Furthermore, the core input parameter for this step is the dynamic range boundary of the simulation modification Δ. The theoretical maximum magnitude of pixel modification is determined by the preceding analog coding model and is a key factor controlling embedding distortion and concealment. Another important indicator is the numerical precision (e.g., 32-bit floating-point numbers) of the carrier image tensor and the modification tensor, which affects the final calculation accuracy and potential rounding errors. The quality of the output image can be quantitatively evaluated using objective indicators such as peak signal-to-noise ratio and structural similarity index. Furthermore, the processing latency of the synthesis operation itself (typically on the order of microseconds) is a key engineering indicator for measuring the system's real-time performance. These parameters and indicators collectively define the technical specifications of the final steganalyte output from this step.

[0139] Furthermore, this synthesis step is the "final assembly" stage of the entire steganography system, producing the final usable output. It directly serves any application scenario that requires the generation of specific steganalytic image files. For example, in secure communication, the sender uses this step to quickly generate encrypted images for transmission; in digital rights management, content providers use this step to batch embed invisible copyright markers into massive amounts of images or video frames; in intelligence processing, the system can automatically embed confidential annotation information into reconnaissance images. Its simplicity and efficiency ensure that even on resource-constrained mobile devices or servers that need to process high-throughput images, fast and reliable steganalytic image generation can be achieved, meeting a wide range of deployment needs from offline batch processing to online real-time interaction.

[0140] Specifically, performing this overlay synthesis operation is the final step in achieving the ultimate performance of the steganography scheme. First, it faithfully and losslessly executes the optimal embedding strategy defined by the front-end intelligent system, accurately converting the optimized simulated modifications into real changes to the image data. Second, because the modifications are fully optimized and constrained, the overlaid steganographic image is visually highly consistent with the original image, avoiding visible flaws or artifacts. Finally, the determinism of this step ensures that the generated steganographic image is reproducible and decodable, allowing the receiver to use the corresponding decoding algorithm to extract the image. This step is crucial for accurately extracting secret information. Therefore, it is a key final link in ensuring the entire steganography system moves from theoretical optimization to practical application, ultimately achieving the purpose of covert communication or information protection.

[0141] S6 simulates the STC embedding process in traditional steganography algorithms through a dual Tanh activation function structure; wherein, the first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on pixel-level embedding operations, so that the steganalyte maintains visual consistency with the carrier image in the pixel domain.

[0142] Specifically, the implementation principle of this dual-Tanh activation function structure lies in constructing a two-stage differentiable nonlinear function chain with saturation characteristics to mathematically approximate the discrete, nonlinear decision boundaries and amplitude constraints generated by the traditional STC encoder during decoding and modulation. The first-stage Tanh function receives the initial control signal (integrating probability intensity and information bits) from the front-end generator network and, through its "S"-shaped nonlinear response, simulates the soft decision process in STC encoding based on local distortion costs and parity check matrix constraints, thereby outputting a continuous, preliminary modulation intention signal. The second-stage Tanh function performs a second nonlinear transformation and amplitude compression on the output of the first stage. Its function is to precisely simulate the quantization saturation effect and perceptual threshold characteristics in actual pixel modification operations, ensuring that even under high-intensity driving signals, the generated simulated modification amount can be constrained within a low amplitude range insensitive to the human visual system. This mathematically transforms the discrete encoding process into a fully differentiable continuous function approximation problem.

[0143] Furthermore, the input signal (obtained by element-wise multiplication of the embedded intensity map and the orientation map) first undergoes feature transformation through a fully connected layer or convolutional layer, and its output is fed into the first Tanh activation function, which maps the input value to... The first interval generates an initial modulated signal with nonlinear tradeoffs. This signal is then passed through a second fully connected / convolutional layer and input into a second Tanh function. The second Tanh function further compresses the signal and finally adjusts the output to the target dynamic range (e.g., using a learnable or preset scaling factor). This serves as the final simulated pixel modification. The entire process can be formally represented as: Where I is the input, γ is the linear layer parameter, and γ is the amplitude scaling factor.

[0144] Furthermore, the core adjustable parameters of this structure include the dimension and bias of the weight matrix for the two-stage linear transformation, and the final amplitude scaling factor γ. The dimension of the linear layer determines the module's ability to fit complex nonlinear relationships and is typically set to a small hidden layer size to maintain lightweight design. The amplitude scaling factor γ is a crucial hyperparameter that directly determines the maximum absolute value of the simulated modification; its setting must be matched to the pixel depth of the carrier image (e.g., 8 bits) and the tolerance for peak signal-to-noise ratio (PSNR) degradation caused by target embedding. Additionally, whether to introduce a batch normalization layer between the two Tanh functions to stabilize training is also an optional design parameter. All these parameters are initialized early in model training and are co-optimized as the entire network is trained end-to-end.

[0145] Specifically, this dual-Tanh simulation structure, acting as a "trainable bridge" connecting high-level embedding strategies and low-level pixel operations, offers the greatest value in revolutionizing the steganography algorithm design process. It integrates the previously separate "cost design" and "encoding implementation" stages into an end-to-end learning process that automatically searches for optimal solutions. This makes the method not only suitable for academic research seeking extreme security performance but also directly applicable to industrial product development with high requirements for development efficiency and algorithm adaptability. Examples include rapid iteration of secure communication software, customized development of digital content protection systems, and adaptive defense systems that need to address new and unknown steganography threats.

[0146] Specifically, the simulated modifications generated by the dual-Tanh structure are mathematically continuous and smooth, providing a high-quality, abrupt gradient flow for the backpropagation algorithm. This is the technical foundation for ensuring the stable convergence of the deep generative network and the finding of the optimal solution. Secondly, through two-stage nonlinear saturation processing, it effectively simulates the physical constraints and perceptual characteristics of real embeddings, enabling the embedded probability map predicted by the trained generative network to produce steganalytic images with minimal visual perturbation and the least obvious statistical anomalies when used in conjunction with a real STC encoder. Finally, this ingenious design allows the entire system to achieve intelligent and optimized embedding strategies while maintaining full compatibility with traditional steganalytic frameworks. This is the key technical guarantee for the significant advantages of this invention in terms of security and visual quality.

[0147] This invention discloses a residual-guided image steganography method. By constructing a non-adversarial deep generative network guided by multiple features including residuals, distance, and local entropy, and employing a differentiable embedding mechanism that simulates STC using dual Tanh, it effectively solves the problems of poor concealment, weak anti-detection capability, and low generation efficiency caused by inaccurate embedding region selection, unstable training process, and single feature utilization in existing technologies. It achieves end-to-end integrated optimization from multi-feature collaborative perception and adaptive probabilistic map generation to continuous differentiable simulated embedding, significantly improving the overall performance of steganographic images in terms of visual quality, robustness against statistical analysis, and semantic content preservation. This enhances the practicality and reliability of the algorithm in complex and variable carriers and harsh detection environments.

[0148] Example 2 To achieve the above invention, embodiments of the present invention also provide another image steganography method based on residual guidance, including: Research on residual-guided image steganography methods includes: overall model structure design, ResNet-based generator design, design and optimization of the residual-guided module, extraction of distance features, construction of the local entropy-guided module, and design and optimization of the loss function. A flowchart of the specific research process is shown below. Figure 2 As shown.

[0149] S101, Overall structural design of the model.

[0150] Specifically, the image steganography generation model proposed in this invention adopts a non-adversarial deep generative framework, aiming to achieve high-quality and secure steganography generation without relying on generative adversarial networks (GANs). The model consists of three parts: a residual guidance module, a distance guidance module, and a local entropy guidance module, and uses an improved U-Net structure as the generator backbone network to achieve multi-feature fusion and end-to-end learning.

[0151] Furthermore, the input consists of a carrier image and secret information, and the output is the generated steganographic image. Three guiding features constrain the generation process from the residual domain, the deep feature space, and the statistical information domain, respectively, ensuring the consistency of the steganographic image with the original image in terms of structural features and visual perception. The overall framework is trained using a joint loss function, balancing image quality, embedding accuracy, and steganographic security.

[0152] Specifically, the overall structure and workflow of the model are as follows: Figure 3As shown, the steganalysis image is preprocessed first, then the generation is constrained by three guiding modules and the embedding loss module ensures the integrity of the embedding payload, generating an embedding probability map. Then, the STC process is simulated by double Tanh. A double Tanh activation function structure is introduced in the generator output stage to simulate the STC (Syndrome-Trellis Codes) embedding process in the traditional steganalysis algorithm. Finally, the corresponding steganalysis image is generated.

[0153] S102, a generator design based on ResNet.

[0154] Specifically, the generator employs an improved ResNet residual network structure to enhance the model's ability to preserve image details and reconstruct high-frequency information, enabling efficient feature extraction. Furthermore, it optimizes parameters for image steganography training, maintaining the quality and security of the generated images. The network consists of stacked residual modules, each composed of two convolutional layers, a batch normalization layer, and a ReLU activation function. A skip connection mechanism effectively mitigates the gradient vanishing problem in deep networks.

[0155] Specifically, through the residual learning mechanism, the model can more accurately detect subtle differences between the original image and the stegogram, improving image reconstruction accuracy and embedding precision. Meanwhile, the lightweight design of this structure reduces the size of the model parameters, lowers resource consumption, and makes the training process more stable and faster.

[0156] S103, Design and optimization of the residual guidance module.

[0157] Specifically, the core idea of ​​the residual guidance module is to use the high-frequency residual information of the image to guide the generator in steganographic embedding in regions with complex textures. This module first extracts the residual feature map of the original image through a multi-scale high-pass filter, capturing detailed changes in edges, textures, and noise. During training, the RGM inputs the residual map into the generator, fusing it with features from the intermediate layers of the backbone network. Channel concatenation or weighted attention mechanisms are used to enhance the model's responsiveness to high-frequency features. For optimization, a residual consistency loss function is defined to constrain the difference between the generated steganographic map and the original image in the residual domain, ensuring that the statistical characteristics of the steganographic map remain highly consistent with the original image, thereby improving its resistance to steganalysis.

[0158] S104, Distance feature extraction.

[0159] Specifically, the distance guidance module aims to introduce similarity constraints in the deep feature space to maintain visual consistency between the stegmap and the carrier image at the semantic level. This module utilizes the intermediate layer outputs of a pre-trained deep convolutional network (ResNet50) as feature maps to extract high-level features from both the carrier and stegmap images. Then, it calculates the Euclidean distance or cosine similarity between the two in the feature space, forming a deep distance loss function. This loss function is used during training to constrain the generator output, ensuring that the stegmap remains close to the original carrier image in terms of high-level semantic features. This reduces perceptual differences and avoids visual artifacts caused by the stegographic content. This design effectively improves the visual quality and stability of the stegmap image.

[0160] S105, Construction of the Local Entropy Guidance Module.

[0161] Specifically, the local entropy guidance module, based on information theory principles, guides the generator to embed information in high-complexity regions by calculating the information entropy distribution of local image regions. First, the input image is divided into fixed-size windows. The gray-level histogram distribution of each window is calculated, and the corresponding information entropy value is obtained according to the formula, generating an entropy feature map. This entropy map is then input into the generator's intermediate layer, dynamically adjusting the embedding probability of each region: high-entropy regions enhance embedding strength, while low-entropy regions suppress embedding behavior. This achieves adaptive adjustment of the embedding, which not only improves concealment but also prevents detectable noise in smooth regions. Finally, by defining an entropy-constrained loss, the embedding region distribution is optimized, making the overall steganography structure more reasonable.

[0162] S106, Design and optimization of loss function.

[0163] Specifically, to achieve a comprehensive balance between image quality, embedding accuracy, and steganalysis security during image steganography generation, this invention designs a multi-guided joint loss function optimization mechanism, which simultaneously constrains and coordinates multiple performance metrics during the generator training phase. This loss function optimizes the generation result from four aspects. First, residual consistency constraints ensure that the generated steganalysis image maintains a high degree of similarity to the original carrier image in the residual domain, preventing significant high-frequency distortion introduced during the steganography process. Then, depth distance constraints maintain the similarity between the steganalysis image and the original image in the deep semantic space, making the generated result visually similar to the original. Figure 1 Then, a local entropy constraint mechanism is introduced. By adjusting the embedding probability of different regions of the image, adaptive control of the embedding position is achieved, so that the secret information is preferentially embedded in regions with complex textures and difficult detection, thereby improving the security of steganography. Finally, embedding precision constraints are set to ensure that the secret information can be completely and accurately embedded and extracted during the steganography process, improving the reliability of information recovery.

[0164] Another residual-guided image steganography method in this invention constructs a non-adversarial generator structure that integrates residual, distance, and local entropy feature guidance, and employs a joint loss function with multiple constraint terms for end-to-end training. This effectively solves the problems of training instability, high computational cost, and difficulty in co-optimizing visual quality and security in traditional generative adversarial network-based steganography methods. It achieves system integration from multi-scale feature extraction and embedded region adaptive guidance to global optimization of the generated image, significantly improving the stability and convergence efficiency of the steganography model training. Simultaneously, it enhances the visual fidelity of the steganographic image, its resistance to advanced steganalyzers, and the reliability of secret information embedding, thereby comprehensively improving the practicality and robustness of this technology in complex covert communication scenarios.

[0165] Example 3 To achieve the above invention, embodiments of the present invention also provide an optimization process for another residual-guided image steganography method, including: Specifically, during training, the weights of various loss functions are dynamically adjusted to flexibly balance the relationship between image quality and steganalysis security according to the actual task requirements. In the optimization phase, an adaptive moment estimation algorithm is used as the optimizer, continuously updating the network parameters through backpropagation to ensure stable convergence of the generative model during training.

[0166] In embodiments of the present invention, the ResNet-based generator design includes: Specifically, the image steganography generator uses the ResNet-50 network structure as its core, achieving efficient extraction and information transfer of deep features through a residual learning mechanism. The ResNet structure effectively avoids the vanishing gradient problem in traditional deep networks, ensuring training stability and preserving rich high-frequency feature information in complex image textures. Its advantages are mainly reflected in solving the vanishing gradient problem in deep networks through skip connections, ensuring the training and propagation of image feature information within the deep network. Therefore, when processing complex texture regions of images, it can retain more feature information, enhancing the generator's ability to handle complex texture regions.

[0167] Furthermore, the ResNet-50 architecture design includes: ResNet-50 has advantages in high-level feature learning and detail preservation, therefore the generator uses ResNet-50 as the core network structure. ResNet-50 can effectively extract deep-level features of images through multiple residual modules, and the feature maps extracted in the network can be further transformed into embedding probability maps. Pre-training ResNet-50 ensures the efficiency of the training process, allowing for better capture of global and local texture features in the steganalyte during training. The ResNet-50 structure is as follows: Figure 4 As shown.

[0168] Furthermore, the embedding probability map generation includes: after feature extraction, the generator uses convolutional layers to fuse and sample multi-scale feature maps, mapping deep features into pixel-level embedding probability distributions. An activation function restricts the output values ​​to between 0 and 1, ensuring the embedding probabilities meet embedding requirements, thus obtaining the probability of embedding secret information at each location; the generated embedding probability map provides a reference for the likelihood of embedding secret information at each pixel location. This process can dynamically adjust the embedding strategy based on residual features, edge features, and other information of the image, achieving high-fidelity and high-security steganography image generation. Figure 5 The diagram shows the process of generating the ResNet-50 probability map.

[0169] In an embodiment of the present invention, the residual guidance module structure includes: Specifically, the residual guidance module is the core module in RGIS. It aims to guide the generator to fully utilize high-frequency residual information in the image, concentrating steganographic embeddings in textured regions of the image load and reducing perturbations to smooth areas, thereby improving steganographic security. During the model's design and optimization, various strategies are employed to enhance the module's feature representation capabilities and generalization performance.

[0170] Furthermore, the module structure includes a residual guidance module consisting of a high-pass filter layer, a low-pass filter group, and a residual fusion unit. The high-pass filter layer extracts high-frequency components of the image to capture detailed features; the low-pass filter group preserves low-frequency structural information, ensuring overall brightness and structure are not distorted; and the residual fusion unit weighted and merges high- and low-frequency features to generate a residual signal mapping map. Through this structure, the module can comprehensively consider local changes and global consistency of the image, achieving high-frequency detail enhancement and adaptive control of the embedding region, thereby improving the security of steganography embedding. During training, this module integrates with the intermediate layers of the main generator network through feature concatenation and attention mechanisms, allowing residual information to directly participate in the steganography probability prediction process, thereby reducing the model's incorrect embedding of smooth regions and lowering the risk of overfitting.

[0171] Furthermore, feature extraction includes: for the residual guidance module, high-pass filtering and low-pass filtering are used in the feature extraction process; the high-pass filtering group is used to extract high-frequency residual features of the carrier image and capture edge and texture information. The high-pass filtering formula is: ; in, For convolution, high-pass filtering extracts high-frequency residual features by using the Kirsch edge detection kernel; The Kirsch 3x3 edge detection kernel is used to detect edge information in an image. Its design aims to highlight edge features in the image while suppressing the influence of relatively flat areas.

[0172] Furthermore, low-pass filtering employs mean filtering to generate a smoother residual map. Mean filtering removes high-frequency noise and details from the residual map through smoothing operations, resulting in a smoother image structure. These operations can be implemented through convolution operations, thereby extracting residual features that better conform to the steganography strategy for the generator network to process. The formula is as follows: ; in, It is the total number of pixels in the filter, for Mean filtering, ,for Mean filtering, . This is the radius of the filter kernel, with values ​​respectively. and , which represents the size of the image neighborhood.

[0173] Furthermore, this process can effectively improve the network's sensitivity to texture details and edge variations, enhance its concealment, and make the generator perform more accurately and stably in complex image regions.

[0174] In an embodiment of the present invention, the distance guidance module includes: Specifically, the distance-guided module measures the similarity between the generated steganalyte and the original carrier image in a high-level semantic feature space. This module enhances the resistance to deep learning-based steganalysis by ensuring a high degree of consistency between the steganalyte and the original carrier image at the semantic feature level.

[0175] Furthermore, the module structure includes: a distance guidance module consisting of a feature extraction unit, a distance calculation unit, and a loss fusion unit. The input is the steganographic image output by the generator and the corresponding carrier image; the output is the depth feature distance between the two. This distance is used to constrain the generator output, ensuring it maintains deep semantic consistency with the original image.

[0176] Furthermore, the distance feature extraction includes: the module extracts semantic feature vectors from the stegana and the carrier image by loading the intermediate layer output features of the pre-trained model, and calculates the difference between the two in the feature space. The calculation results are fed back into the generator training process to adjust the network parameters to reduce the feature distance. This mechanism can effectively constrain the semantic shift of the model, making the stegana image almost indistinguishable from the original image in visual perception.

[0177] VGG16 was chosen as the distance feature extractor because it performs excellently in image classification tasks, and its intermediate layer features effectively capture object contours, texture structures, and spatial details in images. The VGG16 structure is as follows: Figure 6 As shown.

[0178] Furthermore, the distance feature extraction process mainly involves first performing input preprocessing, normalizing the image pixel values ​​to [0-1], and standardizing them to the VGG16 input format, and then processing the carrier image. and steganalysis Distance feature sampling is performed separately, and finally, L2 norm distance is used as the calculation carrier image. With Stegogram Distance metrics in the feature space. L2 distance measures the difference in semantic features between two images. Minimizing this distance allows the steganalyst to approximate the original carrier in a high-level semantic sense.

[0179] In an embodiment of the present invention, the local entropy guidance module includes: Specifically, the local entropy guidance module is used to analyze the complexity information of local regions of an image, more precisely adapt to the statistical characteristics of local regions of an image, avoid the embedding problem of traditional steganalysts for highly sensitive low-entropy (smooth) regions, and adaptively adjust the embedding strategy based on the information entropy of the region, thereby realizing the dynamic distribution of stegographic information and effectively improving the robustness and security of steganalysis.

[0180] Furthermore, the local entropy-guided module extracts local image features using a sliding window technique, analyzes the information complexity of local image regions, and dynamically adjusts the embedding probability to further optimize the distribution of embedded information. It prioritizes embedding in high-entropy regions, performing embedding in texture-rich, information-rich areas (high-entropy regions) to reduce the impact of embedding operations on statistical features. Embedding in low-entropy regions is suppressed, avoiding embedding in smooth or uniform regions (low-entropy regions) to reduce the risk of detection by steganalysis tools (such as SRM and SRNet).

[0181] Furthermore, the module structure comprises three parts: a local entropy calculation unit, a sliding window entropy extraction unit, and a dynamic thresholding unit. The input image is divided into several local windows, and the information entropy value of each window is calculated by the entropy calculation unit from its grayscale histogram. The module outputs an entropy distribution map and fuses it with the intermediate features of the generator to guide the generation of the embedding probability map.

[0182] Furthermore, local entropy feature extraction includes: calculating the information entropy of each local region to measure the uncertainty of grayscale value distribution in local image regions. The calculation formula is as follows: .

[0183] Furthermore, a high local entropy value indicates a complex local gray-level distribution, making it suitable for embedding, while a low entropy value indicates a uniform and smooth local gray-level distribution, requiring avoidance of embedding. When calculating local entropy, a sliding window process is first performed, traversing the image with a 5×5 window to calculate the local entropy value of each window. Then, gray-level histogram statistics are performed to calculate the probability distribution, and an integral histogram is used to reduce redundant calculations.

[0184] In embodiments of the present invention, the loss function includes: Specifically, the loss function system employs a multi-guided joint optimization mechanism, comprehensively optimizing four aspects: residual constraints, semantic consistency, regional complexity, and information embedding accuracy. This ensures a balance between visual quality, embedding accuracy, and security of the steganalyte. The loss function aims to guide the steganalyte embedding operation through image residual features, distance features, local entropy features, and the embedding loss function, thereby achieving high visual quality and high security for RGIS.

[0185] Furthermore, the residual guidance loss includes: using residual features obtained through high-pass and low-pass filtering, a residual guidance image is generated. This image is used to constrain the similarity between the stegmap and the carrier image in the residual domain, concentrating the stegographic information in complex texture regions of the image, reducing perturbation to smooth regions of the carrier image, and ensuring consistency between the two at the high-frequency feature level. This reduces detectable high-frequency differences and improves the visual quality of the stegmap. The loss function is defined as the L1 norm normalized result of the residual guidance image. .

[0186] Furthermore, by minimizing this loss, this model can focus the embedding operation on complex texture regions, thereby reducing the impact of the embedding operation on the statistical features of the image and enhancing the security of the steganographic image.

[0187] Furthermore, the distance-guided loss includes: minimizing the difference in spatial features between the stegana and the carrier image to constrain the distance between the generated image and the original image in the deep feature space, maintaining visual semantic consistency and avoiding perceptual differences in the generated image. This loss function uses a pre-trained VGG16 model to extract deep image features, calculates the L2 norm distance between the carrier image and the stegana in the feature space, and constrains the embedding region through probabilistic graphical weighting. Its formula is: .

[0188] Furthermore, this not only enhances the security of steganographic images but also strengthens their visual invisibility.

[0189] Furthermore, the local entropy-guided loss includes: It uses a sliding window technique to obtain the information complexity of local regions in the carrier image, dynamically adjusts the embedding probability, further optimizes the probability distribution of the embedding region, and constrains the local information entropy of the image. This guides the model to increase embedding strength in complex regions and decrease embedding in smooth regions, achieving adaptive distribution of embedding positions and improving anti-detection performance. This loss function calculates the entropy value of local image regions, generates binary masks to mark low-entropy regions, and penalizes the embedding probability of low-entropy regions, forcing the model to embed in high-entropy regions of the carrier image, i.e., information-complex regions. .

[0190] Furthermore, the design of the local entropy guidance module takes into account the complexity of local regions of the image, making the embedding operation more consistent with the statistical laws of natural images.

[0191] Furthermore, the embedding loss includes: the embedding loss function, which calculates the cross-entropy loss of the stegana image, constrains the embedding capacity and ensures the complete transmission of secret information. This ensures the integrity and accuracy of the stegana information during embedding and extraction, thereby improving the secret information recovery rate. Its formula is: .

[0192] Furthermore, the total loss function includes a weighted combination of the above four losses to form a multi-objective optimization system. By dynamically adjusting the weight parameters, image quality and steganography security can be flexibly balanced according to different task requirements. The total loss function is a weighted sum of the losses from each guidance module and the embedding loss: ; in, , , , These are weighting coefficients, which control the contribution of each loss term.

[0193] Furthermore, the design of the total loss function optimizes the embedding strategy through the design of four loss functions: residual guidance, distance guidance, local entropy guidance, and embedding loss. This improves the image security, visual invisibility, and embedding effectiveness of the reinforcement learning image steganography method based on multiple image features presented in this paper.

[0194] This invention provides an alternative residual-guided image steganography optimization process. By constructing a generator based on ResNet-50 and integrating multiple guidance modules for residual, distance, and local entropy, combined with a dynamically weighted multi-constraint joint loss function, it effectively addresses the inherent shortcomings of traditional adversarial training-based steganography models in terms of training stability, computational efficiency, and security-quality balance. It achieves collaborative modeling and guidance of high-frequency details, deep semantics, and local statistical features of images, enabling secret information to be adaptively embedded into textured regions while ensuring that the generated image is highly consistent with the carrier image in the pixel domain, residual domain, and feature space. This optimization process significantly improves the convergence speed and generation stability of the steganography model, simultaneously enhancing the visual imperceptibility of the steganographic image, its resistance to advanced analyzer detection, and the reliability of information embedding.

[0195] Example 4 To achieve the above invention, embodiments of the present invention also provide an experimental application of an image steganography method based on residual guidance, including: The experiments of this invention were implemented using PyTorch, with the Adam optimizer used during training and a learning rate of 1. The gradient is 0.25, and the batch size is 8. The specific implementation process is as follows: Dataset setup and preprocessing included: Dataset setup: This experiment used ImageNet and CoCo datasets as training and testing sets, respectively. For the ImageNet dataset, 50,000 images were selected as the training set, and 1,000 images were selected as the testing set from the CoCo dataset. Images were resized to a resolution of 256. A 256-bit grayscale image. A grayscale image is used as the carrier image, and secret information, also in grayscale, is embedded using image steganography techniques.

[0196] Specifically, dataset setup and preprocessing also include: Data preprocessing: To ensure the stability of model training and the effectiveness of subsequent feature extraction, and to provide consistent input to the generator, the following is a standardized preprocessing workflow designed for residual-guided image steganography models: (1) Image data selection. Select photographic images that cover natural scenes, textures and backgrounds.

[0197] (2) Standardize the data format for model input. First, perform grayscale processing on the image. Then, use the cv2.resize operation to uniformly adjust the size of the grayscale image to the specified value. Pixels, to avoid training instability caused by different input sizes.

[0198] (3) Pixel normalization processing normalizes pixel values ​​from the range [0, 255] to [-1, 1], using the following formula: ; This normalization can accelerate model convergence, reduce the risk of gradient explosion, and is compatible with the output range of the double Tanh activation function.

[0199] (4) Data Augmentation. During the training phase, data augmentation techniques are applied to improve the robustness and generalization ability of the model. Images are randomly flipped horizontally with a 2% probability and randomly rotated within ±5 degrees to simulate changes in natural images. During the data preprocessing phase, data augmentation is only applied to the training set, while the validation and test sets remain unchanged to ensure the fairness of the evaluation.

[0200] Furthermore, the ablation experiments and weight settings include: To further optimize the training effectiveness of the three guiding modules (residual guidance, distance guidance, and local entropy guidance) in the residual-guided image steganography method, this invention designed an ablation experiment to conduct independent or combined training on the three guiding modules in the steganography method. Table 1 shows the steganography performance comparison of the steganography results under each combination.

[0201] Table 1 Performance under various module combinations

[0202] Specifically, ablation experiments revealed that removing the residual guided loss from the generator significantly degrades the quality of the steganalyte, with the PSNR dropping from 40.19 to 23.48. This demonstrates that the residual guided loss is crucial for the proposed image steganography method. Therefore, in the experiments of this invention, the weights of the loss function are set to... , , , To balance the magnitude of each loss term and ensure the stability of the training process, the design of the total loss function fully considers factors such as security, semantic consistency, and embedding effectiveness, enabling the model to achieve efficient embedding in complex scenarios.

[0203] Furthermore, the quality of the steganalyte images is evaluated: This invention uses Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as objective evaluation metrics. RGIS is compared with several representative methods at different embedding rates (0.2 bpp, 0.4 bpp, and 0.5 bpp). The experiments are based on a residual-guided image steganography method. Steganalyte images are generated by a generator, and PSNR and SSIM are used to evaluate the steganography effect, thereby verifying the effectiveness of RGIS in maintaining image visual fidelity.

[0204] Specifically, PSNR is a commonly used metric for measuring image quality; a higher value indicates that the generated steganalysis image is less different from the original image. SSIM is used to measure the similarity of images in terms of structure, brightness, and contrast; the closer the value is to 1, the better the structure of the steganalysis image is preserved from the original image.

[0205] Table 2 shows the steganography performance indicators of RGIS under different embedding rates.

[0206] Table 2 Steganography Effects of RGIS

[0207] Specifically, the results show that the residual-guided image steganography method achieves a PSNR of 40.19 at an embedding rate of 0.2 bpp, 37.93 at 0.4 bpp, and 35.02 at 0.5 bpp; the SSIM is 0.945 at 0.2 bpp, 0.923 at 0.4 bpp, and 0.856 at 0.5 bpp. Although the steganography effect decreases significantly with increasing embedding rate, the newly generated steganographic image still exhibits good visual quality and a natural steganography effect under the general condition of satisfying the secret information embedding rate, without significantly affecting the content of the carrier image. Figure 7 This is a steganography effect for residual-guided images.

[0208] Furthermore, to evaluate the effectiveness of the method of the present invention, RGIS was compared with other existing steganography methods such as Steganite and SteganoGAN. Table 3 shows the final comparison results.

[0209] Table 3 Comparison of Steganography Effects of RGIS with Other Methods

[0210] Based on the PSNR and SSIM values ​​at different embedding rates (0.2 bpp, 0.4 bpp, 0.5 bpp) in the table, it can be seen that the residual guided steganography (RGIS) method proposed in this paper exhibits significant advantages under low embedding rate conditions. Specifically, at an embedding rate of 0.2 bpp, RGIS achieves a PSNR of 40.19 and an SSIM of 0.939, both of which are higher than all the comparison methods listed in the table, indicating that the steganographic images generated under this condition have the highest visual fidelity. In the same comparison, SteganoGAN, which performed second best, had a PSNR of 39.25 and an SSIM of 0.942 (SSIM is slightly higher than RGIS, but PSNR is significantly lower than RGIS), while other methods such as Steganite, DeepMIH, and DAH... The PSNRs of Net at 0.2 bpp were 37.23, 38.4 and 38.18, respectively, and the SSIMs were 0.931, 0.909 and 0.932, respectively, all of which were lower than those of RGIS.

[0211] While RGIS performance declined slightly with increasing embedding rates to 0.4 bpp and 0.5 bpp, it remained highly competitive. At 0.4 bpp, RGIS achieved a PSNR of 37.93 and an SSIM of 0.903; under the same conditions, SteganoGAN achieved a PSNR of 33.84 (significantly lower than RGIS) and an SSIM of 0.909 (slightly higher than RGIS), while DAH... Net's PSNR is 37.18 and SSIM is 0.907, very close to RGIS. At 0.5 bpp, RGIS's PSNR is 35.02 and SSIM is 0.856, although lower than DAH. Net's best performance at this embedding rate (PSNR 36.23, SSIM 0.887) and SteganoGAN's SSIM 0.881 are still significantly higher than traditional steganography methods TSR and LSB, and outperform most other deep learning-based comparison methods. Overall, RGIS performs best at 0.2 bpp. Although its performance degrades at higher embedding rates, the difference from the best method is small (e.g., at 0.4 bpp, the difference in SSIM between RGIS and SteganoGAN is only 0.006), proving that it can effectively complete image steganography tasks and generate reliable steganographic images at various embedding rates. Figure 8 This visually demonstrates the line graph comparison trend of PSNR and SSIM with other methods at different bit rates for RGIS, further confirming the above conclusions.

[0212] Furthermore, a comparative analysis of the current GPU usage was conducted: Specifically, Parameters refers to the number of parameters generated by the model during training, reflecting the total numerical values ​​of convolutional layers, image feature parameters, and weight matrices in the model. Parameter analysis using the values ​​of Parameters is used to measure the training cost or effort of RGIS in completing image steganography tasks.

[0213] Specifically, GPU memory is used to store the data being processed by the GPU, and its capacity determines the amount of data the GPU can access quickly. Memory usage reflects the amount of data the GPU processes per unit time during training. Table 4 compares the GPU memory size during training for different steganography methods under identical GPU conditions and experimental settings, reflecting to some extent the training cost of each method.

[0214] Table 4 Comparison of parameters and memory usage between RGIS and other methods

[0215] Specifically, RGIS generates 12.32M parameters during training, significantly fewer than the other four steganography methods. Specifically, compared to DeepMIH's 14.95M parameters, RGIS generates only about 80% of DeepMIH's parameters under the same conditions. This verifies that the method of this invention generates high-quality steganographic images while consuming fewer training resources. It demonstrates that the model has strong scalability under limited resources and is more suitable for practical applications.

[0216] Specifically, RGIS requires 1504 MiB of GPU memory during image training, compared to 2298 and 1893 MiB for SteganoGAN and DAH-Net, respectively. This invention's method has a lower GPU memory usage, indicating that it has lower computational resource requirements during training and can adapt to most image steganography tasks, demonstrating good performance in balancing steganography performance and training cost.

[0217] Furthermore, a security analysis of the steganographic image is performed: Specifically, the robustness analysis included: To evaluate the security of the steganographic image, we conducted adversarial attack tests. The adversarial attacks employed included median blur, Gaussian blur, salt-and-pepper noise, and Gaussian noise. Regarding the parameter settings for median filtering and Gaussian filtering: the filter size was... The standard deviation of the Gaussian filter is set to Regarding the parameter settings for salt and pepper noise: the noise density is set to... Regarding the parameter settings for Gaussian noise: the mean is... The standard deviation of Gaussian noise is set to .

[0218] Analysis of the tabular data revealed that the introduction of a discriminator improved the steganography performance of this method under all three types of attacks, and its performance was superior to most other steganography methods.

[0219] Table 5 compares the steganography performance of RGIS with other steganography methods such as SteganoGAN under different attacks.

[0220] Table 5 Steganography performance of RGIS and other methods under different attacks.

[0221] Experimental results show that RGIS outperforms most other deep learning-based image steganography methods in PSNR and SSIM performance under attacks from median filtering, Gaussian filtering, salt-and-pepper noise, and Gaussian noise. Figure 9 As shown, the steganographic images generated by the residual-guided image steganography method exhibit good robustness against adversarial attacks, effectively resisting adversarial attacks involving median filtering, Gaussian filtering, and Gaussian noise.

[0222] Specifically, the steganalysis resistance capability includes: the experiment uses common deep learning-based steganalysts such as StegExpose, YeNet, SRNet and SiaStegNet to test the steganalysis resistance capability of the generated steg images. Table 6 shows the error rate of RGIS and other methods in different steganalysis methods.

[0223] Table 6 Error rates of RGIS and other methods in different steganalysis methods.

[0224] The results show that at an embedding rate of 0.2 bpp, the error rates of images generated by the residual-guided image steganography method in the steganalysis model are 56.72%, 36.33%, 31.26%, and 32.93%, respectively. At an embedding rate of 0.4 bpp, the error rates are 47.48%, 20.95%, 19.52%, and 21.65%, respectively. Compared to other steganography methods, RGIS exhibits the best steganalysis error rate at an embedding rate of 0.2 bpp, demonstrating better steganalysis security. A detailed bar chart is shown below. Figure 10 As shown.

[0225] This invention presents an experimental application of a residual-guided image steganography method. By constructing a multi-module collaborative guidance mechanism integrating residuals, distance, and local entropy, and employing an adaptive weighted loss function and standardized preprocessing flow, it effectively solves the problems of degraded steganalyte quality and weak security caused by embedding distortion, feature coupling, and insufficient robustness against adversarial attacks in existing technologies. It achieves integrated modeling from multi-dimensional feature guidance and hidden information embedding to anti-steganography analysis, significantly improving the resistance of steganalytes to common filtering and noise attacks while maintaining high visual fidelity. Furthermore, it maintains excellent peak signal-to-noise ratio and structural similarity at various embedding rates, and has lower model parameter count and memory usage, enhancing the practicality and concealment of the steganography system in real-world complex scenarios.

[0226] Example 5 To achieve the above invention, such as Figure 11As shown, this embodiment also provides a residual-guided image steganography device 10, which includes: The image preprocessing module 100 is used to preprocess the original carrier image to obtain a preprocessed carrier image.

[0227] The probability map generation module 200 is used to construct a multi-feature fusion generation network including a residual guidance module, a distance guidance module, and a local entropy guidance module using a non-adversarial deep generation framework, and inputs the preprocessed carrier image into the multi-feature fusion generation network to generate an embedded probability map.

[0228] The embedding simulation module 300 is used to input the embedding probability map and the corresponding random coded features into a preset deep feature embedding learning model, perform embedding simulation, and generate a simulated steganalysis image.

[0229] The network training module 400 is used to calculate the multi-guided joint loss between the simulated stegana image and the preprocessed carrier image, and to use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network.

[0230] The steganalysis image generation module 500 is used to input the carrier image to be processed into a trained multi-feature fusion generation network to obtain a new embedding probability map, and to obtain the final steganalysis image based on the new embedding probability map.

[0231] In one embodiment of the present invention, it further includes: a dual Tanh activation function structure module, used to simulate the STC embedding process in the traditional steganography algorithm through the dual Tanh activation function structure; wherein, the first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on the pixel-level embedding operation, so that the steganographic image maintains visual consistency with the carrier image in the pixel domain.

[0232] This invention discloses a residual-guided image steganography device. By constructing a collaborative system comprising modules for image preprocessing, probabilistic graph generation, embedding simulation, network training, and steganography image generation, and employing a non-adversarial deep generative framework and multi-feature fusion mechanism, it effectively solves the problems of poor steganography concealment, weak anti-detection capability, and low generation efficiency caused by inaccurate embedding regions, unstable training processes, and single feature utilization in existing technologies. It achieves end-to-end joint training, integrating the entire process from carrier image standardization, adaptive fusion of multiple guided features, and differentiable embedding simulation, significantly improving the overall performance of steganography operations in terms of visual imperceptibility, statistical security, and semantic consistency, and enhancing the practicality and reliability of the device in diverse application scenarios.

[0233] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 12 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the residual-guided image steganography method described above.

[0234] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0235] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A residual-guided image steganography method, characterized in that, include: S1, preprocess the original carrier image to obtain the preprocessed carrier image; S2, a multi-feature fusion generation network containing a residual guidance module, a distance guidance module, and a local entropy guidance module is constructed using a non-adversarial deep generation framework, and the preprocessed carrier image is input into the multi-feature fusion generation network to generate an embedding probability map; S3, input the embedding probability map and the corresponding random coded features into the preset deep feature embedding learning model, perform embedding simulation, and generate simulated steganalysis image; S4, calculate the multi-guided joint loss between the simulated stegana and the preprocessed carrier image, and use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network. S5. Input the carrier image to be processed into the trained multi-feature fusion generative network to obtain a new embedding probability map, and obtain the final steganalysis image based on the new embedding probability map.

2. The method as described in claim 1, characterized in that, The preprocessing of the original carrier image to obtain the preprocessed carrier image includes: S11, the original carrier image is spatially standardized and adjusted to a fixed size that meets the input requirements of the multi-feature fusion generation network; S12, perform pixel value normalization processing on the carrier image after spatial size standardization, and convert the corresponding pixel value range to a predetermined numerical range.

3. The method as described in claim 1, characterized in that, The method employs a non-adversarial deep generative framework to construct a multi-feature fusion generative network comprising a residual guidance module, a distance guidance module, and a local entropy guidance module. The preprocessed carrier image is then input into this multi-feature fusion generative network to generate an embedding probability map, including: S21, the high-frequency residual features of the preprocessed carrier image are extracted through the residual guidance module to generate a residual guidance feature map; S22, through the distance guidance module, the depth features of the pre-processed carrier image are extracted using a pre-trained feature extraction network as the first depth feature, and a distance guidance feature map is generated based on the feature distance constraint between the first depth feature and the depth features corresponding to the stegana image to be generated. S23, the local information entropy distribution of the preprocessed carrier image is calculated through the local entropy guidance module to generate a local entropy feature map; S24. The residual guided feature map, distance guided feature map and local entropy feature map are fused and input into the backbone network of the multi-feature fusion generation network for feature mapping and probability normalization to generate an embedding probability map.

4. The method as described in claim 1, characterized in that, The step of inputting the embedding probability map and corresponding random encoded features into a preset deep feature embedding learning model, performing embedding simulation, and generating a simulated steganalyte image includes: S31, Based on the embedding probability map, determine the embedding strength corresponding to each pixel position, and based on the corresponding random coding features, determine the modification direction to be simulated; S32, the embedding strength and modification direction are input to the dual Tanh simulation STC model. The dual Tanh simulation STC model performs nonlinear mapping on the input through two cascaded Tanh functions, and outputs a continuous and differentiable simulation modification. S33, the simulated modification amount is added to the pixel value of the preprocessed carrier image to generate the steganalytic image.

5. The method as described in claim 1, characterized in that, The calculation of the multi-guided joint loss between the simulated stegana image and the preprocessed carrier image, and the end-to-end training of the multi-feature fusion generation network using the multi-guided joint loss to obtain the trained multi-feature fusion generation network, includes: S41, calculate the residual guided loss, distance guided loss and local entropy guided loss between the simulated stegana and the preprocessed carrier image respectively; S42, the residual guided loss, distance guided loss and local entropy guided loss are weighted and summed to form the multi-guided joint loss, and the parameters of the multi-feature fusion generation network are iteratively optimized with the goal of minimizing the multi-guided joint loss until convergence.

6. The method as described in claim 1, characterized in that, The step of inputting the carrier image to be processed into a trained multi-feature fusion generation network to obtain a new embedding probability map, and obtaining the final steganalyte image based on the new embedding probability map, includes: S51, the carrier image to be processed is input into the trained generative network. The residual guidance module, distance guidance module and local entropy guidance module of the trained generative network work together to extract and fuse multi-scale depth features and guidance features, and output a new embedding probability map after decoding and mapping. S52, the new embedding probability map and the corresponding random coding features are input into the preset deep feature embedding learning model. The preset deep feature embedding learning model weights each bit of the corresponding random coding features according to the new embedding probability map, and simulates the coding process through a cascaded hyperbolic tangent function to calculate the pixel-level simulated modification amount. S53, the simulated modification amount is superimposed on the corresponding pixel value of the processed carrier image, and the final steganalytic image is directly synthesized and output.

7. The method as described in claim 4, characterized in that, Also includes: S6 simulates the STC embedding process in traditional steganography algorithms through a dual Tanh activation function structure; wherein, the first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on pixel-level embedding operations, so that the steganalyte maintains visual consistency with the carrier image in the pixel domain.

8. An image steganography device based on residual guidance, characterized in that, include: The image preprocessing module is used to preprocess the original carrier image to obtain a preprocessed carrier image; The probability map generation module is used to construct a multi-feature fusion generation network that includes a residual guidance module, a distance guidance module, and a local entropy guidance module using a non-adversarial deep generation framework, and inputs the preprocessed carrier image into the multi-feature fusion generation network to generate an embedded probability map; The embedding simulation module is used to input the embedding probability map and the corresponding random coded features into a preset deep feature embedding learning model, perform embedding simulation, and generate simulated steganalysis images. The network training module is used to calculate the multi-guided joint loss between the simulated stegana image and the preprocessed carrier image, and to use the multi-guided joint loss to perform end-to-end training on the multi-feature fusion generation network to obtain the trained multi-feature fusion generation network. The steganalysis image generation module is used to input the carrier image to be processed into a trained multi-feature fusion generation network to obtain a new embedding probability map, and to obtain the final steganalysis image based on the new embedding probability map.

9. The apparatus as claimed in claim 8, characterized in that, Also includes: The dual Tanh activation function structure module is used to simulate the STC embedding process in traditional steganography algorithms through the dual Tanh activation function structure. The first Tanh function is used to generate an embedding probability map, and the second Tanh function is used to perform nonlinear mapping on the pixel-level embedding operation, so that the steganalyte image maintains visual consistency with the carrier image in the pixel domain.

10. An electronic device, comprising: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the residual-guided image steganography method as described in any one of claims 1-7.

Images