A blind extraction method for secret information from deep steganographic images based on self-supervised learning

Abstract

The present invention provides a method for blindly extracting secret information from deep steganographic images based on self-supervised learning, comprising an encoding phase, a self-supervised learning task to generate a supervisory signal, attention coupling via an attention coupling module, and a decoding phase. The encoding phase includes taking two secret images as input, performing Haar transform processing on each, splitting the input features generated by the Haar transform processing of the secret images into two paths through a bidirectional coupling mechanism of a coupling layer, and utilizing nonlinear transformations to achieve feature decoupling and cross-path interaction, so that different paths focus on different components of the secret image. A reversible design is also used to ensure information losslessness during the feature transformation process. This method innovatively utilizes self-supervised learning technology, enabling the model to deeply explore the characteristics and patterns of the secret image itself, thereby achieving blind secret information extraction that does not rely on a key.

Description

Technical Field

[0001] The present invention belongs to the field of computer science and technology, and in particular relates to a method for blindly extracting secret information from deep steganographic images based on self-supervised learning. Background Art

[0002] Steganography is a technique for hiding secret information within seemingly ordinary media (such as images, audio, and video). Its core goal is to ensure that the secret information remains undetected. Early steganography methods primarily constructed embedding and extraction models based on spatial or transform domains, such as least significant bit (LSB) substitution and discrete cosine transform (DCT) coefficient modification. While relatively easy to implement, these methods have limited hiding capacity and carry the risk of detection due to changes in statistical properties. With the development of deep learning technology, deep image steganography methods have emerged, significantly improving hiding capacity and security. For example, the HiNet method proposed by Jing et al. (Reference: J. Jing, X. Deng, M. Xu, J. Wang and Z. Guan, "HiNet: Deep Image Hiding by Invertible Network," 2021 IEEE International Conferenceon Computer Vision (ICCV), Montreal, QC, Canada, 2021, pp. 4713-4722.) uses an invertible neural network to model image hiding and recovery as a symmetric reversible process, achieving size-independent secret image hiding. At the same time, it uses discrete wavelet transform and low-frequency loss function to concentrate secret information in high-frequency subbands, reducing the risk of detection. The LiDiNet method proposed by Li et al. (Reference: F. Li, Y. Sheng, K. Wu, C. Qin and X. Zhang, "LiDiNet: A Lightweight Deep Invertible Network for Image-in-Image Steganography," in IEEE Transactions on Information Forensics and Security, vol. 19, pp. 8817-8831, 2024) further optimizes the network structure by employing lightweight reversible convolutional modules and a spatial attention mechanism. This reduces the number of parameters while maintaining high capacity, improving both the security and visual quality of the stegograph. However, this increase in stegograph capacity often comes with more noticeable embedding artifacts. High-capacity stegographing inevitably requires significant modifications to the carrier, which can easily lead to visual distortion and thus be detected by advanced steganalysis tools. Existing steganalysis methods are mainly divided into two categories: passive detection and active erasure.Wei et al. (Reference: K. Wei, W. Luo, S. Tan and J. Huang, "Universal Deep Network for Steganalysis of Color Image Based on Channel Representation," in IEEE Transactions on Information Forensics and Security, vol. 17, pp. 3022-3036, 2022.) point out that passive detection methods analyze the statistical properties of the carrier and train a classifier to determine whether a suspected image contains secret information. Active erasure methods, on the other hand, subtly modify the secret image (for example, through image compression or noise interference) to disrupt the extraction of the stegosaurus. However, existing active erasure methods focus solely on preventing the correct extraction of the secret information and are generally unable to recover the specific content of the secret information.

[0003] Self-supervised learning, an emerging paradigm in machine learning, offers new opportunities for addressing these challenges. It automatically generates supervisory signals based on the data's structure and characteristics, enabling the model to learn effective feature representations from the data without requiring extensive manual annotation. Introducing self-supervised learning in the context of blind image secret information extraction can fully exploit the rich, high-frequency hidden information within the secret image, enabling the model to learn the characteristic differences between the image in its normal and steganographic states, effectively avoiding reliance on large numbers of labeled samples and enabling autonomous learning and improved adaptability.

[0004] Current high-capacity steganography methods based on deep neural networks, through end-to-end training, can adaptively embed large amounts of secret information into carrier images, achieving hiding capacities several times greater than those of traditional LSB-based methods. However, this high-capacity steganography also presents new security challenges: on the one hand, the surge in information embedding leads to more obvious stegoscopy traces; on the other hand, the unique nonlinear embedding method of deep steganography makes the information distribution more complex. While existing deep detection networks have improved detection accuracy, these methods can only determine whether the carrier contains secrets or remove the stegoscopy information, but cannot fully recover the secret information, severely limiting the practical application value of steganalysis technology. Summary of the Invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and provide a method for blind extraction of secret information from deep steganographic images based on self-supervised learning, which includes an encoding stage, a self-supervised learning task to generate a supervision signal, attention coupling through an ACB (Attentive Coupling Block) attention coupling module, and a decoding stage.

[0006] The encoding stage includes the following steps:

[0007] Step a1: Take the two secret images as input and perform Haar Transform (HT) processing on them. Low-pass and high-pass filters are applied row by row and column by column to the two secret images. A first-level decomposition is performed to obtain low-frequency components (LL) containing global structure and high-frequency components (LH, HL, HH) containing edges, textures, and secret information. Repeating this operation can achieve multi-level decomposition to capture multi-scale features. Subsequently, convolution with a stride of 2 is used to replace traditional downsampling. Corresponding low-pass and high-pass convolution kernels are designed. The images are convolved and sub-bands are spliced ​​along the channels to achieve downsampling.

[0008] Step a2: Split the input features of the secret image generated by the HT Haar transform method into two paths through the bidirectional coupling mechanism of the coupling layer (CL). Utilize nonlinear transformation to achieve feature decoupling and cross-path interaction, so that different paths focus on different components of the secret image. A reversible design ensures information losslessness during the feature transformation process. The coupling layer (CL) process includes a forward process and a reverse process.

[0009] In step a2, the forward process achieves decoupling of the carrier image and secret information and multi-scale feature interaction by splitting the input features, cross-path nonlinear transformation and feature splicing. The formula is:

[0010] ,

[0011] ,

[0012] ,

[0013] ,

[0014] in The feature tensor representing the input secret image, , Represent the feature tensor The path features of the carrier image and the path features of the secret image obtained by the split feature splitting operation, 、 、 is a nonlinear transformation function; Path features representing the carrier image Path features with secret images The added eigenvectors adjusted by the nonlinear transformation function F; Represents multiplication operation; exp represents natural exponential function; is the path feature of the secret image With the eigenvector The adjusted feature vector; concat represents the concatenation operation, where In the forward process, the feature and The feature tensor of the concatenated output.

[0015] In step a2, the reverse process uses the forward output to reversely solve the original features, and forms a closed loop with the forward process through reversible algebraic operations. The formula is:

[0016] ,

[0017] ,

[0018] ,

[0019] ,

[0020] Among them, Y can be obtained by splitting the feature and Two eigenvectors, 、 、 It is a nonlinear transformation function, which works the same as the forward process and is used for feature adjustment; Indicates division operation; exp represents natural exponential function; concat is feature concatenation operation, X is the reverse process of feature and The feature tensor of the concatenated output.

[0021] The self-supervised learning task generates a supervisory signal comprising the following steps:

[0022] Step b1: In the preprocessing stage, a cascading operation of a DWT (Discrete Wavelet Transform) and a WRM (Wavelet Rich Model) wavelet domain filter is used to highlight the high-frequency information in the secret image, thereby preserving the characteristics of the secret information.

[0023] In step b2, the reconstructed image is input into the backbone network. After extracting the features of the reconstructed image, the features are mapped to the estimation branch. The estimation branch regresses the corresponding values ​​from the reconstructed image through learning to generate a supervision signal S with a value range of (0, 1), where 0 represents the carrier image and 1 represents the secret image.

[0024] Step b1 includes: first, performing a discrete wavelet transform (DWT) on an input secret image to decompose the secret image into low-frequency subbands (LL) and high-frequency subbands (LH, HL, HH). After decomposition, the dimensional parameters of the secret image will change from the original (B, C, H, W) to (B, 4×C, H / 2, W / 2), where B is the batch size, C is the number of channels, H is the height, and W is the width.

[0025] The 30-layer filtering of the WRM wavelet domain filter passes the weight coefficient Partially suppress the low-frequency sub-band to weaken the interference of the main content of the secret image, while retaining a small amount of low-frequency information to avoid structural collapse;

[0026] Apply directional differential filtering (such as horizontal, vertical, and diagonal kernels) to the high-frequency subbands to enhance the steganograph noise, and extract local statistical features through multi-channel residual maps to capture high-frequency anomalies introduced by the steganograph operation;

[0027] Finally, the high-frequency enhanced image is reconstructed using the processed low-frequency sub-band and high-frequency sub-band through inverse wavelet transform (IWT).

[0028] The formula for suppressing the low-frequency sub-band is:

[0029] ,

[0030] in, is the initial low-frequency subband, and the weight coefficient The value range is [0.1,0.3], represents the low-frequency subband retained after suppression processing;

[0031] The reconstruction loss function of the inverse wavelet transform IWT is:

[0032] ,

[0033] in, is the total reconstruction loss function, is the structural similarity loss, is the subband matching loss in the wavelet domain, is the edge-aware loss, is the dynamic weight coefficient, which controls the contribution of each loss term;

[0034] The formula of structural similarity loss function is:

[0035] ,

[0036] in, They represent the downsampled version of the original image at the kth scale and the downsampled version of the reconstructed image at the kth scale, respectively. K is the number of multi-scale pyramid layers (3). is the structural similarity index;

[0037] The formula of subband matching loss function in wavelet domain is:

[0038] ,

[0039] in, is the subband matching loss function in the wavelet domain, Represents the DWT coefficient of the image in subband s (low frequency LL or high frequency LH / HL / HH), is the original image, To reconstruct the image, for norm, is a gradient operator that acts on the high-frequency subband HH to enhance the diagonal edge. Represents the DWT coefficient of the original image in the high frequency subband HH, Represents the DWT coefficient of the reconstructed image in the high frequency subband HH, represents the high-frequency gradient penalty coefficient (0.5~1.0); the edge perception loss formula is:

[0040] ,

[0041] in, is the Laplace edge detection operator, M represents the dynamic edge mask, and the edge area is focused through the binary image [0,1]. represents the norm (the square root of the sum of the squares of the matrix elements);

[0042] The dynamic weight adjustment formula is:

[0043] ,

[0044] in, represents the dynamic weight of the k-th loss function, The temperature coefficient is used to adjust the weight distribution. is the value of the kth loss function.

[0045] In step b2, the backbone network Backbone is composed of a convolution operation Conv and a LeakyReLU activation function connected in series;

[0046] The estimation branch Estimation Branch includes two series structures, the first series structure consists of 1 Conv convolution and LeakyReLU activation function, and the second series structure consists of 3 Conv convolutions, InstanceNorm instance normalization and LeakyReLU activation function.

[0047] The attention coupling by the ACB attention coupling module includes the following steps:

[0048] In step c1, the output features processed by the encoding stage are processed by four consecutive Conv convolutional layers and ReLU activation functions, and the feature tensor containing the secret information area is extracted layer by layer. The feature tensor output by the third layer is denoted as F;

[0049] In step c2, the feature tensor F is processed through the Conv convolution layer and the Sigmoid activation function to generate a single-channel attention map A with a value range of [0,1], where the high-value area corresponds to the location of the secret information:

[0050] ,

[0051] Among them, the element value of A is used to measure the importance of the corresponding position feature. The original feature tensor F is weighted element by element through the single-channel attention map A to generate a new feature tensor ;

[0052] The attention loss function is calculated by the supervision signal during training:

[0053] ,

[0054] in, is the positive attention loss, which represents the degree of alignment between the single-channel attention map and the secret information mask; 、 、 They are the single-channel attention maps generated by the three coupling layer functions F, G, and H in the ACB attention coupling module, which are used to locate the secret information area.

[0055] The decoding stage includes the following steps:

[0056] In step d1, the path features of the carrier image and the secret image processed in the encoding phase are restored to their spatial dimensions through the invertible Haar transform (IHT). The path features of the carrier image and the secret image form dual path features. The coupling layer CL is used to make the dual path features interact across paths. After the path features of the carrier image are nonlinearly transformed, the path features of the secret image are scaled and offset to achieve feature fusion and refinement. The formula is:

[0057] ,

[0058] in, represents the path features of the initial secret image, is the path feature of the carrier image, is the path feature of the adjusted secret image;

[0059] In step d2, the path features of the carrier image after interaction are added element by element to the path features of the secret image to synthesize the reconstructed secret image. The consistency between the reconstructed image and the original input is supervised by the reverse loss to complete the reconstruction from the feature to the pixel space. The difference loss function between the reconstructed image and the original input image is:

[0060] ,

[0061] in, is the reconstruction loss, which is used to measure the difference between the reconstructed image and the real image. is the preliminary reconstructed secret image, For the real secret image, is the optimized reconstructed secret image, is the weight coefficient.

[0062] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0063] The present invention also provides a storage medium storing a computer program or instruction, which executes the steps of the method when the computer program or instruction is run on a computer.

[0064] Beneficial effects: The method for blind extraction of secret information from deep steganographic images based on self-supervised learning proposed in this invention innovatively utilizes self-supervised learning technology to enable the model to deeply explore the characteristics and patterns of the secret image itself, thereby realizing blind extraction of secret information without relying on a key. Through the designed self-supervised learning tasks, the model can autonomously learn on a large amount of secret image data without secret information annotations, and automatically capture the unique traces and characteristic patterns left by the secret information in the image. When faced with unknown secret images, the model can directly and accurately extract the secret information based on the learned feature knowledge, without relying on traditional key mechanisms or specific deep models.

[0065] Based on a self-supervised learning architecture and multimodal feature decoupling technology, this method not only significantly improves the accuracy of extracting strong embedded traces caused by high-capacity steganography, but also effectively achieves blind extraction of secret information from deep imagery without relying on a key mechanism. Therefore, this method can achieve a rough extraction of secret information without relying on the original carrier or the steganographic key, significantly enhancing the practical value of steganalysis technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of the coupling layer CL.

[0067] Figure 2 Schematic diagram of the encoding stage.

[0068] Figure 3 Schematic diagram of generating supervisory signals for self-supervised learning tasks.

[0069] Figure 4 Schematic diagram of the structure of the ACB attention coupling module.

[0070] Figure 5 Schematic diagram of the decoding stage.

[0071] Figure 6 This is the overall framework diagram for blind extraction of secret information from deep steganographic images based on self-supervised learning. DETAILED DESCRIPTION

[0072] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0073] The present invention provides a method for blindly extracting secret information from deep steganographic images based on self-supervised learning. This method aims to leverage self-supervised learning techniques, without relying on a key mechanism, to enable the model to autonomously learn from a large amount of unlabeled, secret-containing image data. This allows the model to automatically capture the unique traces of secret information within the image, thereby enabling the approximate extraction of secret information from images with obvious embedding traces. The method comprises four stages: encoding, generation of supervisory signals through a self-supervised learning task, an ACB attention coupling module, and decoding.

[0074] like Figure 1 As shown, the coupling layer CL is described in detail:

[0075] In the task of blind extraction of secret information from deep steganographic images, the coupling layer CL promotes the cross-path feature interaction between the path features of the carrier image and the path features of the secret image, helping to accurately separate the secret image from the carrier image. Its function is to perform nonlinear transformation on the features through functions F, G, and H to achieve feature transformation and fusion, enrich feature expression, and enhance the model's expressive ability.

[0076] like Figure 2 As shown, the encoding stage CIC-EB (Coupling-in-Coupling Encoding Block) is described in detail:

[0077] The first step is to process the secret image into two identical copies. In the early stages of the encoding phase, the HT Haar transform is first used to perform multi-scale feature decomposition on the two input secret images. The image is downsampled to 1 / 4 resolution through a convolution operation with a step size of 2, and the three RGB color channels are expanded into 12 feature channels (high and low frequency combinations in the three directions of LL, LH, HL, and HH). Subsequently, the network uses a channel segmentation strategy to divide the feature map into two parts: one part enters the secret image processing path, and the other part enters the carrier image processing path. Figure 2 The upper path in is the carrier image processing path, and the lower path is the secret image processing path.

[0078] In the second step, the dual-path network achieves progressive feature decoupling through multiple levels of coupling layers. The secret image processing path has two coupling layers, employing a 1:1 channel split ratio for rapid feature extraction. The carrier image processing path comprises 10 coupling layers, the first eight of which employ a 1:3 channel split ratio for refined feature extraction, while the last two share coupling layers with the secret image processing path for feature interaction. Each coupling layer contains a learnable affine transformation function (F, G, H) and employs a residual structure for nonlinear mapping. Specifically, within the shared coupling layer, the secret image path features are deeply fused with the carrier image path features via addition operations (Y1=X1+F(X2)) and affine transformations (Y2=X2⊙exp(G(Y1))+H(Y1)). This ensures sufficient feature decoupling and lossless information transfer through reversible transformations, laying the feature foundation for subsequent attention mechanism processing.

[0079] like Figure 3 、 Figure 4 、 Figure 6 As shown in the figure, the generation of supervision signals for self-supervised learning tasks and the attention coupling through the ACB attention coupling module are described in detail:

[0080] In the first step, the ACB attention coupling module receives the multi-scale features output from the encoding stage. These features contain both high-frequency and low-frequency information of the secret image. These features are then split along the channel dimension into a secret image processing path and a carrier image processing path. The secret image processing path focuses on high-frequency components to capture texture and edge details, while the carrier image processing path retains low-frequency components and represents background structure, clarifying the division of labor for subsequent processing.

[0081] In the second step, the secret image's path features enter a processing module consisting of a cascade of four layers (Conv and ReLU). The first convolution layer expands the number of channels and initially extracts edge information from the secret image. The second layer further enhances the features. The third layer integrates the features and outputs a feature tensor F that enhances the secret image response. The fourth layer introduces skip connections to prevent feature degradation and ensure the effective transmission of underlying secret image clues.

[0082] In the third step, based on the feature tensor F, the ACB attention coupling module generates a single attention map in parallel through three independent convolution branches. 、 、 Each branch is processed by 1×1 convolution and Sigmoid activation function to compress the value range to [0, 1]. These correspond to controlling the scaling of the secret image processing path, modulating the scaling factor of the secret image processing path, and supplementing the position and shape information of the secret image, representing the confidence that the pixel belongs to the secret image.

[0083] The fourth step utilizes a cascade of discrete wavelet transform (DWT) and wavelet-domain filter (WRM) to highlight high-frequency information in the secret image, thereby preserving the characteristics of the secret information. First, a discrete wavelet transform (DWT) is applied to the input secret image, decomposing it into low-frequency and high-frequency subbands. The WRM's 30-layer filter partially suppresses the low-frequency subbands using weight coefficients, reducing the interference from the main content of the secret image while retaining a small amount of low-frequency information to prevent structural collapse. Directional difference filtering (such as horizontal, vertical, and diagonal kernels) is applied to the high-frequency subbands to enhance the steganograph noise. Local statistical features are extracted using a multi-channel residual map, specifically capturing high-frequency anomalies introduced by the steganograph operation. Finally, an inverse wavelet transform (IWT) is used to reconstruct a high-frequency enhanced image from the processed low- and high-frequency subbands. The backbone network extracts features from the reconstructed image and maps them to the estimation branch. The estimation branch learns to regress the corresponding values ​​from the reconstructed image to generate a supervision signal S with a value range of (0,1), where 0 represents the carrier image and 1 represents the secret image.

[0084] The fifth step is to minimize the L2 loss by generating the supervision signal S through the self-supervised learning task. 、 、 The difference from the supervisory signal S drives the update of the convolutional layer parameters through backpropagation, forcing the attention map to accurately locate the secret information area, effectively suppressing the misjudgment of image texture as secret information, and improving the accuracy of secret information area positioning.

[0085] like Figure 5 As shown, the decoding stage CIC-DB (Coupling-in-Coupling Decoding Block) of the method of the present invention is described in detail:

[0086] In the first step, during the decoding phase, the path features of the carrier image processed by the ACB attention coupling module and the path features of the secret image are coupled to achieve cross-path feature interaction. The path features of the carrier image are used as the main path, and the path features of the secret image are nonlinearly transformed through functions F, G, and H, including scaling and offset operations. This allows the features of the two paths to be deeply fused, enhancing the ability to distinguish between secret information and carrier image content.

[0087] In the second step, the path features of the carrier image and the secret image after downsampling and cross-path interaction are synthesized by element-by-element addition. Subsequently, upsampling is performed using the inverse Haar transform (IHT) to restore the spatial size of the feature map to the original image size, and the number of channels is adjusted accordingly to reconstruct the carrier image and the secret image. At the same time, the reverse loss is used The reconstructed image is compared with the secret image, the difference between the two is calculated, and the network parameters are adjusted through backpropagation to ensure that the details and structural information of the carrier image are preserved as much as possible while extracting the secret information. Compared with existing passive detection methods that analyze the statistical characteristics of the carrier and train classifiers to determine whether the suspected image contains secret information, but are unable to correctly extract the secret information, this invention innovatively utilizes self-supervised learning technology. It not only significantly improves the accuracy of extracting strong embedded traces caused by high-capacity steganography, but also effectively achieves the approximate extraction of secret information in deep images without relying on key mechanisms, greatly enhancing the practical value of steganalysis technology.

[0088] The present invention provides a method for blindly extracting secret information from deep stego images based on self-supervised learning. There are many methods and approaches to implement this technical solution. The above is only a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A method for blind extraction of secret information from deep steganographic images based on self-supervised learning, characterized in that: It includes the encoding stage, the generation of supervision signals from the self-supervised learning task, attention coupling through the ACB attention coupling module, and the decoding stage; The encoding stage includes the following steps: Step a1: Take the two secret images as input and perform the HT Haar transform method on them. Apply low-pass and high-pass filters to the two secret images row by row and column by column, and decompose them into low-frequency components containing global structure and high-frequency components containing edges, textures, and secret information. Then, replace the traditional downsampling with convolution with a stride of 2. Design corresponding low-pass and high-pass convolution kernels, convolve the images, and splice subbands along the channels to achieve downsampling. Step a2: Split the input features of the secret image generated by the HT Haar transform method into two paths through the bidirectional coupling mechanism of the coupling layer CL. Utilize nonlinear transformation to achieve feature decoupling and cross-path interaction, so that different paths focus on different components of the secret image. A reversible design ensures information losslessness during the feature transformation process. The coupling layer CL process includes a forward process and a reverse process. The self-supervised learning task generates a supervisory signal comprising the following steps: Step b1: In the preprocessing stage, the cascade operation of DWT discrete wavelet transform and WRM wavelet domain filter is used to highlight the high-frequency information in the secret image and retain the characteristics of the secret information; In step b2, the reconstructed image is input into the backbone network. After extracting the features of the reconstructed image, the features are mapped to the estimation branch. The estimation branch regresses the corresponding values ​​from the reconstructed image through learning to generate a supervision signal S with a value range of (0, 1), where 0 represents the carrier image and 1 represents the secret image. The attention coupling by the ACB attention coupling module includes the following steps: In step c1, the output features processed by the encoding stage are processed by four consecutive Conv convolutional layers and ReLU activation functions, and the feature tensor containing the secret information area is extracted layer by layer. The feature tensor output by the third layer is denoted as F; In step c2, the feature tensor F is processed through the Conv convolution layer and the Sigmoid activation function to generate a single-channel attention map A with a value range of [0,1], where the high-value area corresponds to the location of the secret information: ， Among them, the element value of A is used to measure the importance of the corresponding position feature. The original feature tensor F is weighted element by element through the single-channel attention map A to generate a new feature tensor ; The attention loss function is calculated by the supervision signal during training as: ， in, It is positive attention loss; 、 、 They are the single-channel attention maps generated by the three coupling layer functions F, G, and H in the ACB attention coupling module.

2. The method according to claim 1, characterized in that In step a2, the forward process achieves decoupling of the carrier image and secret information and multi-scale feature interaction by splitting the input features, cross-path nonlinear transformation and feature splicing. The formula is: ， ， ， ， in The feature tensor representing the input secret image, , Represent the feature tensor The path features of the carrier image and the path features of the secret image obtained by the split feature splitting operation, 、 、 is a nonlinear transformation function; Path features representing the carrier image Path features with secret images The added eigenvectors adjusted by the nonlinear transformation function F; Represents multiplication operation; exp represents natural exponential function; is the path feature of the secret image With the eigenvector The adjusted feature vector; concat represents the concatenation operation, where In the forward process, the feature and The feature tensor of the concatenated output.

3. The method according to claim 2, characterized in that In step a2, the reverse process uses the forward output to reversely solve the original features, and forms a closed loop with the forward process through reversible algebraic operations. The formula is: ， ， ， ， in, Represents the division operation; exp represents the natural exponential function; concat is the feature concatenation operation.

4. The method according to claim 3, characterized in that Step b1 includes: first, performing a discrete wavelet transform (DWT) on an input secret image to decompose the secret image into low-frequency subbands and high-frequency subbands. After decomposition, the dimensional parameters of the secret image will change from the original (B, C, H, W) to (B, 4×C, H / 2, W / 2), where B is the batch size, C is the number of channels, H is the height, and W is the width. The 30-layer filtering of the WRM wavelet domain filter passes the weight coefficient Partially suppress the low-frequency sub-band; Directional differential filtering is applied to high-frequency subbands to enhance the steganograph noise, and local statistical features are extracted through multi-channel residual maps to capture high-frequency anomalies introduced by the steganograph operation. Finally, the high-frequency enhanced image is reconstructed using the processed low-frequency sub-band and high-frequency sub-band through inverse wavelet transform (IWT); The formula for suppressing the low-frequency sub-band is: ， in, is the initial low-frequency subband, and the weight coefficient The value range is [0.1,0.3], represents the low-frequency subband retained after suppression processing; The reconstruction loss function of the inverse wavelet transform IWT is: ， in, is the total reconstruction loss function, is the structural similarity loss, is the subband matching loss in the wavelet domain, is the edge-aware loss, is the dynamic weight coefficient, which controls the contribution of each loss term; The formula of structural similarity loss function is: ， in, They represent the downsampled version of the original image at the kth scale and the downsampled version of the reconstructed image at the kth scale, respectively. K is the number of multi-scale pyramid layers. is the structural similarity index; The formula of subband matching loss function in wavelet domain is: ， in, is the subband matching loss function in the wavelet domain, represents the DWT coefficient of the image in subband s, is the original image, To reconstruct the image, for norm, is the gradient operator, Represents the DWT coefficient of the original image in the high frequency sub-band, Represents the DWT coefficients of the reconstructed image in the high frequency subband, Represents the high-frequency gradient penalty coefficient; the edge perception loss formula is: ， in, is the Laplace edge detection operator, M represents the dynamic edge mask, and the edge area is focused through the binary image [0,1]. represents the norm; The dynamic weight adjustment formula is: ， in, represents the dynamic weight of the k-th loss function, The temperature coefficient is used to adjust the weight distribution. is the value of the kth loss function.

5. The method according to claim 4, characterized in that In step b2, the backbone network Backbone is composed of a convolution operation Conv and a LeakyReLU activation function connected in series; The estimation branch Estimation Branch includes two series structures, the first series structure consists of 1 Conv convolution and LeakyReLU activation function, and the second series structure consists of 3 Conv convolutions, InstanceNorm instance normalization and LeakyReLU activation function.

6. The method according to claim 5, characterized in that The decoding stage includes the following steps: In step d1, the path features of the carrier image and the secret image processed in the encoding phase are restored to their spatial dimensions through the inverse Haar transform (IHT). The path features of the carrier image and the secret image form dual path features. The coupling layer CL is used to make the dual path features interact across paths. After the path features of the carrier image are nonlinearly transformed, the path features of the secret image are scaled and offset to achieve feature fusion and refinement. The formula is: ， in, represents the path features of the initial secret image, is the path feature of the carrier image, is the path feature of the adjusted secret image; In step d2, the path features of the carrier image after interaction are added element by element to the path features of the secret image to synthesize the reconstructed secret image. The consistency between the reconstructed image and the original input is supervised by the reverse loss to complete the reconstruction from the feature to the pixel space. The difference loss function between the reconstructed image and the original input image is: ， in, is the reconstruction loss, which is used to measure the difference between the reconstructed image and the real image. is the preliminary reconstructed secret image, For the real secret image, is the optimized reconstructed secret image, is the weight coefficient.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 6.

8. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 6 are executed.

Images