Steganographic method and system for hiding region-oriented controllable image in image

Abstract

The application discloses a steganography method and system for controllable hidden areas. First, input the carrier image and multiple secret images, generate the corresponding hidden area center set according to the number of secret images, and generate the area convolution sampling offset parameter, so that different secret images correspond to different hidden areas. Then, input the carrier image and the secret image into the hidden network, complete the regional embedding of the secret image through discrete wavelet transform, deep feature fusion and reversible mapping, and generate a steganography image through inverse wavelet transform. Finally, input the steganography image into the recovery network, and perform reverse feature separation according to the corresponding area convolution sampling offset parameter to gradually recover each secret image. The application realizes the regional embedding of multiple secret images by controlling the convolution sampling position, reduces the mutual interference between different secret images, improves the recovery quality, and enhances the controllability, randomness and system security of the hidden process.

Description

Technical Field

[0001] This invention belongs to the field of information security technology and relates to a method and system for image-to-image steganography, specifically a large-capacity image-to-image steganography method and system with controllable hidden areas; it achieves image-to-image steganography through the cross-application of deep learning and image processing. Background Technology

[0002] Modern steganography is an information security technology that utilizes digital media carriers to achieve covert communication. Its core objective is to hide the existence of information itself in a public transmission environment, making the communication difficult for third parties to detect. With the development of deep learning technology, "image-in-image steganography" has become an important research direction in image steganography. This involves embedding a complete covert image into another normal carrier image to generate a steganographic image with a visual effect highly similar to the original carrier image, and then recovering the covert image under authorized conditions. Compared to traditional methods that only hide text or a small amount of bit information, "image-in-image steganography" has advantages such as high hiding capacity and wide application scenarios, and has significant application value in areas such as privacy protection, copyright authentication, secure communication, and secure storage of multimedia data.

[0003] Most existing methods for "image-to-image" mapping are based on deep convolutional neural networks. These networks fuse features between the carrier image and the secret image, writing the secret information into the feature space of the carrier image. A recovery network then extracts the secret image from the stegana. Because convolutional neural networks possess strong nonlinear fitting and local feature modeling capabilities, they can effectively improve the visual quality of the stegana and the accuracy of the secret image recovery.

[0004] To further enhance spatial modeling capabilities, researchers proposed a deformable convolution mechanism. This method introduces a learnable offset to the standard convolution's fixed sampling points, allowing the convolution kernel sampling position to be dynamically adjusted, thus more flexibly adapting to changes in the spatial structure of the image. In image steganography tasks, the offset not only enhances feature extraction capabilities but can also be used to control the writing location of secret information. Theoretically, by assigning different convolution sampling offset parameters to different secret images, different secret information can be mapped to different regions within the carrier image, achieving spatially separated hiding.

[0005] However, for "image-to-image" systems deployed in scenarios involving multiple secret image embedding, existing methods generally employ a shared feature space for information fusion. When multiple secret images are simultaneously embedded into a single carrier image, information from different secret images can easily compete and overlap within the same feature space, leading to increased mutual interference and further affecting the quality of the reconstructed image. This interference typically manifests as a decrease in the peak signal-to-noise ratio of some reconstructed secret images, reduced structural similarity, and uneven reconstructed quality among different secret images.

[0006] Existing protection and optimization methods can be broadly classified into three categories.

[0007] The first category is an extension method based on single-image hiding structures, which improves steganalysis performance by enhancing the expressive power of convolutional networks. This type of method can achieve high visual quality, but it is mainly designed for single-image hiding, has limited hiding capacity, and is difficult to adapt to the needs of high-capacity embedding of multiple images.

[0008] The second category is high-fidelity hiding methods based on reversible networks. These methods improve the quality of secret image recovery and enhance the overall hiding capacity by constructing a reversible mapping relationship shared by the hiding and recovery processes. However, these methods typically encode multiple secret images together in a unified feature space, lacking explicit control over the physical writing regions of different secret information. As the number of embeddings increases, feature competition easily arises between secret images, leading to a decrease in recovery quality.

[0009] The third category is capacity expansion methods for multi-image hiding, which enhance the information carrying capacity of a single carrier image through serial embedding or shared representation mechanisms. Although it can hide multiple secret images simultaneously, the information embedded later can easily overwrite the previously hidden content, resulting in significant differences in the recovery effects of different secret images and making it difficult to guarantee overall stability.

[0010] Therefore, existing technologies generally have two problems: First, different secret images share the same hidden space, lacking an effective region separation mechanism, which leads to serious mutual interference between multiple images; second, it is difficult to explicitly control the writing position of secret information in the carrier image, the hiding process lacks randomness, and it is difficult to further improve system security and recovery quality.

[0011] Based on this, a "image-to-image" steganography method is needed that can actively control the spatial distribution of secret information during the hiding stage. By adjusting the convolution sampling position, different secret images can be mapped to different hidden regions inside the carrier image, realizing large-capacity multi-image embedding with region controllability. This reduces mutual interference between secret images, improves the peak signal-to-noise ratio and structural similarity of the recovered image, ensures the visual quality of the carrier image, and enhances the randomness and security of the hiding process. Summary of the Invention

[0012] The purpose of this invention is to propose a high-capacity "image-to-image" steganography method and system with controllable hidden regions, which solves the problems of limited hidden capacity, severe feature interference between different secret images, uneven quality of restored images, and uncontrollable hidden regions in existing multi-image steganography techniques. Without relying on the similarity of image content, it can achieve high-capacity, high-fidelity, region-controllable embedding and high-quality extraction of multiple arbitrary natural images in a single carrier image.

[0013] The technical solution adopted by the method of the present invention is: a method for image-to-image steganography with controllable hidden areas, comprising the following steps: Step 1: Input the carrier image and several secret images. Generate a set of hidden region centers in the feature space of the carrier image according to the number of secret images, and impose constraints on each hidden region to ensure that the hidden regions of different secret images are separated from each other in the feature space of the carrier image and do not cross the boundary. Step 2: Generate corresponding region convolution sampling offset parameters based on the set of hidden region centers. The region convolution sampling offset parameters are used to control the sampling position of each sampling point of the convolution kernel of the region convolution layer in the hidden network, so that the information of each secret image is only written into its corresponding hidden region. Step 3: Input the carrier image and the secret image into the hidden network. First, they are mapped to the frequency domain feature space by discrete wavelet transform. Then, through deep feature fusion, the regional fusion of carrier features and secret image features is completed by reversible transformation layer by layer. Finally, the steganalysis image is reconstructed by inverse wavelet transform. Step 4: Input the steganalysis image into the recovery network. Based on the region convolution sampling offset parameters consistent with the deep feature fusion, perform regionalized inverse feature separation in reverse order of deep feature fusion to gradually recover each secret image and complete the "image-hiding-image" extraction process of multiple images.

[0014] Preferably, in step 1, the generation of the hidden region center set is achieved by using regular grid partitioning, farthest point sampling, random uniform sampling, or dynamic search, for any two region centers. and It should satisfy the minimum spacing constraint and the uniform dispersion criterion of maximizing the minimum distance between the centers of any two regions, where i≠j.

[0015] Preferably, in step 1, center constraints, radius constraints, boundary constraints, and kernel consistency constraints are applied to each hidden region. The center constraint ensures that the overall sampling center of the convolutional kernel of the region convolutional layer is consistent with the center of the hidden region preset in the secret image, that is, all sampling points are distributed around the corresponding region center. The radius constraint ensures that the offset distance of each sampling point relative to the region center does not exceed the preset radius. The boundary constraint ensures that the coordinates of all sampling points after offset do not exceed the effective range of the feature map, and the excess part is truncated to the boundary. The kernel consistency constraint ensures that all sampling points within the same convolutional kernel share the same hidden region center, and only allows local perturbations to produce small-range offsets near the center.

[0016] Preferably, in step 2, for the input feature map First, a set of corresponding region convolution sampling offset parameters is generated based on the set of hidden region centers. For the A secret image whose region convolution sampling offset parameter is determined by the center of the hidden region. With local disturbance terms jointly determined, that is The hidden region center is used to determine the overall sampling region of the convolution kernel, and the local perturbation term is used to distribute the sampling points within the convolution kernel around the same region center in a small range. Subsequently, during the region convolution calculation, the... The secret image calls the corresponding region convolution sampling offset parameters, and the output feature map is located at... Convolution results at the point ;in, Indicates the input feature map, Indicates the output feature map at position The convolution result at the point, The fixed sampling grid represents the convolution kernel. Represents the first in the sampling grid Each sampling location The convolution weight parameters represent the corresponding sampling positions; different secret images correspond to different region convolution sampling offset parameters, thereby realizing the regional embedding of secret information in the feature space of the carrier image.

[0017] Preferably, in step 3, the hiding network includes a first discrete wavelet transform (DWT) layer, a second discrete wavelet transform layer, a hidden channel concatenation (Concat) layer, a reversible hiding module, a hidden channel splitting (Split) layer, and a first inverse discrete wavelet transform (IWT) layer; each secret image is embedded sequentially in stages: when embedding the first secret image, the first secret image is input into the first discrete wavelet transform layer, and the carrier image is input into the second discrete wavelet transform layer; when embedding the i-th secret image, i≥2, the i-th... The secret image is input into the first discrete wavelet transform layer, and the steganographic image output from the previous stage is input into the second discrete wavelet transform layer. After both are subjected to discrete wavelet transform, their outputs are fused along the channel dimension by the hidden channel splicing layer and input into the reversible hiding module. The reversible hiding module consists of several sequentially connected reversible coupling blocks. The output of the reversible hiding module is divided into two paths along the channel dimension by the hidden channel splitting layer. The channel corresponding to the carrier branch is subjected to inverse discrete wavelet transform by the first inverse discrete wavelet transform layer to obtain the steganographic image of this stage, and the channel corresponding to the secret branch is used as redundant information. The above process is repeated until all secret images are embedded, and the steganographic image output from the last stage is used as the final steganographic image.

[0018] Preferably, the reversible coupling block includes an intra-block channel splitting layer, a first sub-network φ, a second sub-network ρ, a third sub-network η, and an intra-block channel splicing layer. The input feature is divided into a first branch feature x1 and a second branch feature x2 according to the channel dimension by the intra-block channel splitting layer. The second branch feature x2 is transformed by the first sub-network φ and then added element-wise with the first branch feature x1 to form an additive coupling, resulting in a first output feature y1. The first output feature y1 is transformed by the second sub-network ρ and the third sub-network η respectively. The second branch feature x2 is multiplied element-wise with the exponentialized result output by the second sub-network ρ, and then added element-wise with the output of the third sub-network η to form an affine coupling, resulting in a second output feature y2. The first output feature y1 and the second output feature y2 are spliced ​​by the intra-block channel splicing layer to serve as the output feature of the reversible coupling block. The first sub-network φ, the second sub-network ρ, and the third sub-network η are all composed of residual dense blocks.

[0019] Preferably, the residual dense block includes a Region Conv layer, a first LeakyReLU layer, a first Concat layer, a first Conv layer, a second LeakyReLU layer, a second Concat layer, a second Conv layer, a third LeakyReLU layer, a third Concat layer, a third Conv layer, a fourth LeakyReLU layer, a fourth Concat layer, and a fourth Conv layer. The input feature first passes through the Region Conv layer and the first LeakyReLU layer to obtain a first intermediate feature. The input feature and the first intermediate feature are concatenated by the first Concat layer according to the channel dimension, and then sequentially passed through the first Conv layer and the second LeakyReLU layer to obtain a second intermediate feature. This process continues, with the input of each Conv layer being the concatenation of the original input feature with the channels of all its preceding intermediate features, forming a dense connection. The input of the fourth Conv layer is the concatenation of the original input feature with the channels of all its intermediate features, and its output serves as the output feature of the residual dense block. The fourth Conv layer is not connected to any activation layer, and its weights are initialized to zero.

[0020] Preferably, in step 4, the recovery network includes a third discrete wavelet transform layer, a recovery channel stitching layer, a reversible reveal module, a recovery channel splitting layer, a second inverse discrete wavelet transform layer, and a third inverse discrete wavelet transform layer. During extraction, the final steganalyte image is used as the first round input: the input steganalyte image is mapped to the frequency domain by the third discrete wavelet transform layer, and its output, along with auxiliary noise, is fused by the recovery channel stitching layer according to the channel dimension and then input to the reversible reveal module. In actual extraction, the auxiliary noise is replaced by a random sampling amount that is distributed in the same way as the redundant information in the embedding stage. The reversible reveal module and the reversible hiding module share a common... Sharing the same set of network parameters, the reversible coupling blocks are executed in reverse order of their embedding execution order. Each reversible coupling block performs inverse operations with the same parameters, i.e., element-wise subtraction replaces element-wise addition, and element-wise division replaces element-wise multiplication. The output of the reversible revealing module is divided into two paths according to the channel dimension by the recovery channel splitting layer, and then subjected to inverse discrete wavelet transform by the second and third inverse discrete wavelet transform layers respectively to obtain the restored stegana and the restored secret image. The restored stegana is used as the input for the next round of extraction, and extraction is performed stage by stage in reverse order of embedding order until all secret images are restored.

[0021] Preferably, both the hidden network and the recovery network are pre-trained networks; during training, a joint loss function is used: ; Where H represents the carrier image and C represents the steganalytic image. Represents the i-th original secret image, Let represent the i-th recovered secret image output by the recovery network, n represent the total number of secret images, and LL(·) represent the low-frequency subband extraction operator of discrete wavelet transform. Indicates the center of the hidden region corresponding to the i-th and j-th secret images. This indicates the preset minimum safe distance between areas. Denotes the L2 norm; This represents the weighting coefficient used to balance the four losses mentioned above.

[0022] The technical solution adopted by the system of the present invention is: a steganography system for controllable hidden areas, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the image-to-image steganography method with controllable hidden region.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention proposes a region convolution mechanism. By controlling the overall distribution center and local offset range of the sampling points of the convolution kernel, the information of different secret images is written into separate spatial regions in the feature space of the carrier image, thereby realizing the physical decoupling of the regions in the multi-image hiding process. Compared with existing methods that use a shared feature space for multi-image fusion, this invention reduces feature competition and interference between different secret images from the root, and effectively improves the problem of uneven recovery quality of secret images in multi-image steganography scenarios.

[0024] (2) This invention introduces a hidden region allocation mechanism. By applying a minimum safety interval constraint to the distance between the center of the regions, it ensures that multiple hidden regions are separated from each other and do not overlap in the feature space. Compared with the existing multi-image steganography methods that lack explicit control over the information writing position, this invention realizes the controllable allocation of hidden regions, enabling the defender or the authorizing party to pre-specify the writing position of each secret image, thereby enhancing the manageability and flexibility of the steganography process.

[0025] (3) This invention combines discrete wavelet transform with a reversible coupled network to construct a hiding and restoration framework. Discrete wavelet transform maps the image to the frequency domain, enhancing the embedding capability of high-frequency detail information; the reversible coupled network allows the hiding and restoration processes to share the mapping relationship, eliminating the need to design a separate decoder for the extraction stage. Compared with existing methods based on ordinary convolutional networks, this invention improves the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the restored image while reducing the network parameter scale, which is beneficial to improving the stability of multi-image restoration.

[0026] (4) This invention designs a joint loss function that includes carrier fidelity loss, secret image recovery loss, low-frequency consistency loss, and region separation loss, which simultaneously constrains the visual fidelity of the stegana image, the accuracy of secret image recovery, and the spatial separation effect of multiple regions from multiple dimensions. Compared with existing methods that only optimize a single or a few loss targets, this joint loss function can take into account multiple quality indicators during training, enabling the model to achieve a more balanced overall performance in multi-image embedding scenarios.

[0027] (5) This invention supports any natural image as the image to be hidden, without relying on the similarity between image content or specific image category constraints. Compared with some existing methods that require images with similar content to achieve low-distortion embedding, this invention has stronger versatility and can be widely applied to various application scenarios such as privacy protection, copyright authentication, secure communication, and secure storage of multimedia data. Attached Figure Description

[0028] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0029] Figure 1 This is a flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the method principle provided by the embodiments of the present invention, which shows the complete process of generating a steganalysis image by regional fusion of a carrier image and multiple secret images through a hidden network, and then extracting each secret image in reverse order through a recovery network, as well as the distribution relationship of the hidden regions of different secret images in the feature space of the carrier image; Figure 3 This is a hidden-recovery network structure diagram provided in an embodiment of the present invention. Detailed Implementation

[0030] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0031] Please see Figure 1 and Figure 2 This embodiment provides a method for image-to-image steganography with controllable hidden areas, comprising the following steps: Step 1: Input the carrier image and several secret images. Based on the number of secret images, generate a set of corresponding hidden region centers in the feature space of the carrier image. Apply center constraints, radius constraints, boundary constraints and kernel consistency constraints to each hidden region to ensure that the hidden regions of different secret images are separated from each other in the feature space of the carrier image and do not cross the boundary. In one implementation, the set of hidden region centers is generated through one of the following methods: regular grid partitioning, farthest point sampling, random uniform sampling, or dynamic search; for any two region centers... and It should satisfy the minimum spacing constraint and the uniform dispersion criterion that maximizes the minimum distance between the centers of any two regions.

[0032] In one implementation, the four constraints apply to the sampling points of the Region Conv layer convolution kernel in step 2; the center constraint: the overall sampling center of the convolution kernel is consistent with the center of the hidden region preset in the secret image, that is, all sampling points are distributed around the corresponding region center; the radius constraint: the offset distance of each sampling point relative to the region center does not exceed the preset radius; the boundary constraint: the coordinates of all offset sampling points do not exceed the effective range of the feature map, and the excess part is truncated to the boundary; the kernel consistency constraint: all sampling points within the same convolution kernel share the same hidden region center, and only local perturbations cause small-range offsets near the center; in this embodiment, the local perturbation is implemented by truncated Gaussian perturbation, and the upper limit of the perturbation amplitude is 0.4; Step 2: Generate corresponding region convolution sampling offset parameters based on the set of hidden region centers, and use the region convolution mechanism to control the convolution sampling position of different secret images so that the information of each secret image is only written into its corresponding hidden region; In one implementation, for the i-th secret image, its region convolution sampling offset parameter Δp n (i) The hidden region center c generated in step 1 i With local disturbance term ε n (i) Superposition, i.e., Δp n (i) = c i + ε n (i) Among them, regional center c i Used to determine the overall sampling region of the convolution kernel, local perturbation term ε n (i)A truncated random perturbation with limited amplitude is used to distribute the sampling points within the convolution kernel in a small range near the region center, while maintaining the relative structure of the sampling points within the convolution kernel. During the region convolution calculation, the corresponding region convolution sampling offset parameter is called when embedding the i-th secret image. The convolution result of the feature map output by the Region Conv layer at position p0 is represented as follows: Where x(·) represents the input feature map, y(p0) represents the convolution result of the output feature map at position p0, R represents the fixed sampling grid of the convolution kernel, which is a 3×3 grid in this embodiment, p n w(p) represents the nth sampling position in the sampling grid. n ) indicates the sampling position p n The corresponding convolution weight parameters, Δp n (i) This indicates that the i-th secret image is at sampling position p. n The region convolution sampling offset parameter is used; for non-integer sampling positions, the sampling value is obtained through bilinear interpolation; different secret images correspond to different region centers, thus generating an independent set of region convolution sampling offset parameters ΔP = {Δp ( ¹ ) , Δp ( ² ) , …, Δp (n) This allows the information from each secret image to be mapped to different spatial regions of the carrier image feature space.

[0033] Step 3: Input the carrier image and the secret image into the hidden network. First, they are mapped to the frequency domain feature space by discrete wavelet transform. Then, through deep feature fusion, the regional fusion of carrier features and secret image features is completed by reversible transformation layer by layer. Finally, the steganalysis image is reconstructed by inverse wavelet transform. In one implementation, please see Figure 3The hiding network includes a first Discrete Wavelet Transform (DWT) layer, a second Discrete Wavelet Transform layer, a Concat layer, a reversible hiding module, a Split layer, and a first Inverse Discrete Wavelet Transform (IWT) layer. Each secret image is embedded sequentially in stages: when embedding the first secret image, the first secret image is input to the first DWT layer, and the carrier image is input to the second DWT layer; when embedding the i-th secret image (i≥2), the i-th secret image is input to the first DWT layer, and the steganographic image output from the previous stage is input to the second DWT layer; the DWT layer decomposes the image into a low-frequency subband LL. The number of channels is expanded to four times the original number with the high-frequency subbands LH, HL, and HH. The outputs of the two are fused according to the channel dimension by the hidden channel splicing layer and then input into the reversible hiding module. In this embodiment, the reversible hiding module consists of 16 sequentially connected reversible coupling blocks. The output of the reversible hiding module is divided into two paths according to the channel dimension by the hidden channel splitting layer. The channel corresponding to the carrier branch is subjected to inverse discrete wavelet transform by the first inverse discrete wavelet transform layer to obtain the steganalysis image of this stage. The channel corresponding to the secret branch is used as redundant information. The above process is repeated until all secret images are embedded. The steganalysis image output by the last stage is used as the final steganalysis image.

[0034] In one implementation, please see Figure 3 The reversible coupling block includes an intra-block channel splitting layer, a first sub-network φ, a second sub-network ρ, a third sub-network η, and an intra-block channel splicing layer. The input feature is divided into a first branch feature x1 and a second branch feature x2 according to the channel dimension by the intra-block channel splitting layer. The second branch feature x2 is transformed by the first sub-network φ and then added element-wise with the first branch feature x1 to form an additive coupling, resulting in a first output feature y1 = x1 + φ(x2). The first output feature y1 is transformed by the second sub-network ρ and the third sub-network η respectively. The second branch feature x2 is multiplied element-wise with the exponentialized result of the output of the second sub-network ρ, and then added element-wise with the output of the third sub-network η to form an affine coupling, resulting in a second output feature y2 = x2⊙ exp(ρ(y1)) + η(y1), where ⊙ represents element-wise multiplication. The first output feature y1 and the second output feature y2 are spliced ​​by the intra-block channel splicing layer to serve as the output feature of the reversible coupling block.

[0035] In one implementation, please see Figure 3The first sub-network φ, the second sub-network ρ, and the third sub-network η are all composed of residual dense blocks; the residual dense blocks include a region convolution (Region Conv) layer, a first LeakyReLU layer, a first Concat layer, a first Conv layer, a second LeakyReLU layer, a second Concat layer, a second Conv layer, a third LeakyReLU layer, a third Concat layer, a third Conv layer, a fourth LeakyReLU layer, a fourth Concat layer, and a fourth Conv layer; the input features first pass through the region convolution layer. The first intermediate feature is obtained by the Conv layer and the first LeakyReLU layer; the input feature and the first intermediate feature are concatenated by the first Concat layer according to the channel dimension, and then passed through the first Conv layer and the second LeakyReLU layer to obtain the second intermediate feature; and so on. The input of each Conv layer is the concatenation of the original input feature with the channels of all the intermediate features before it to form a dense connection; the input of the fourth Conv layer is the concatenation of the original input feature with the channels of all the intermediate features, and its output is used as the output feature of the residual dense block. The fourth Conv layer is not connected to the activation layer, and its weights are initialized to zero to ensure that the reversible coupled block approximates the identity mapping in the early stage of training; in this embodiment, the number of channels of each intermediate feature is 32, and the size of the Conv layer convolution kernel is 3×3. The Region Conv layer, which is the convolutional layer controlled by the region convolution sampling offset parameter described in step 2, is the improvement of this invention compared to existing reversible steganography networks: existing steganography networks use ordinary convolution with a fixed grid in their subnetworks, and the secret information is globally diffused in the carrier feature space, making the hidden region uncontrollable; this invention replaces the first convolution of the residual dense block with the Region Conv layer, so that the feature transformation of each reversible coupled block is constrained within the hidden region of the corresponding secret image, realizing controllable hidden regions and regional isolation of multiple secret images.

[0036] Step 4: Input the steganalysis image into the recovery network. Based on the region convolution sampling offset parameters consistent with the deep feature fusion, perform regionalized inverse feature separation in reverse order of deep feature fusion to gradually recover each secret image and complete the "image-hiding-image" extraction process of multiple images.

[0037] In one implementation, please see Figure 3The recovery network includes a third discrete wavelet transform layer, a recovery channel stitching layer, a reversible reveal module, a recovery channel splitting layer, a second inverse discrete wavelet transform layer, and a third inverse discrete wavelet transform layer. During extraction, the final steganalyte image is used as the first round input: the input steganalyte image is mapped to the frequency domain by the third discrete wavelet transform layer, and its output, along with auxiliary noise, is fused by the recovery channel stitching layer according to the channel dimension before being input to the reversible reveal module. The auxiliary noise is a random sample quantity with the same distribution as the redundant information in the embedding stage. The reversible reveal module shares the same set of network parameters as the reversible hiding module, and is executed by several reversible coupling blocks in reverse order of the embedding execution order. Each reversible coupling block performs inverse operations with the same parameters, i.e., element-wise subtraction replaces element-wise addition, and element-wise division replaces element-wise multiplication. Each reversible coupling block contains a Region... The Conv layers all use the same region convolution sampling offset parameters as the embedding stage for regional sampling; the output of the reversible revealing module is divided into two paths according to the channel dimension by the recovery channel splitting layer, and then subjected to inverse discrete wavelet transform by the second inverse discrete wavelet transform layer and the third inverse discrete wavelet transform layer respectively to obtain the restored stegana and the restored secret image; the restored stegana is used as the input for the next round of extraction, and extraction is performed in reverse order of the embedding order until all secret images are restored.

[0038] In one implementation, both the hidden network and the recovery network are pre-trained networks. During training, a joint loss function including carrier fidelity loss, secret recovery loss, low-frequency consistency loss, and region separation loss is used to jointly train the hidden network and the recovery network, while simultaneously constraining the visual quality of the stegographic image, the recovery accuracy of each secret image, and the spatial separation effect of multiple hidden regions. The joint loss function is: ; Where H represents the carrier image and C represents the steganalytic image. Represents the i-th original secret image, Let represent the i-th recovered secret image output by the recovery network, n represent the total number of secret images, and LL(·) represent the low-frequency subband extraction operator of discrete wavelet transform. Indicates the center of the hidden region corresponding to the i-th and j-th secret images. This indicates the preset minimum safe distance between areas. Let L2 be the norm; the first term is the carrier fidelity loss, used to constrain the visual fidelity between the stegana and the carrier image; the second term is the secret recovery loss, used to constrain the recovery accuracy of each secret image; the third term is the low-frequency consistency loss, used to constrain the structural consistency between the stegana and the carrier image in the wavelet low-frequency subband; the fourth term is the region separation loss, used to constrain the maintenance of a preset minimum safe distance between different hidden regions. This represents the weighting coefficient used to balance the four losses mentioned above.

[0039] In one implementation, the weighting coefficients for each loss term are set as follows: =1.0 (carrier fidelity loss weight). =1.5 (Secret Recovery Loss Weight) =1.0 (low-frequency consistency loss weight). =0.8 (Region Separation Loss Weight), Minimum Safe Distance for Region Separation =32 pixels. The above weight coefficients are for illustrative purposes only and can be adjusted during actual training based on the magnitude of each loss term and the convergence result.

[0040] In one implementation, training uses the Adam optimizer with an initial learning rate of 1×10⁻⁶. - 4. Set the batch size to 8 and train for a total of 100 epochs; at the 40th, 60th, 80th, and 90th epochs, reduce the learning rate to half its original value. The above training hyperparameters are for illustrative purposes only and can be adjusted according to the hardware environment and convergence status.

[0041] The invention will be further illustrated below through specific experiments.

[0042] This experiment collected image samples from publicly available natural image datasets (such as ImageNet, COCO, or DIV2K), and uniformly adjusted the resolution and normalized all images. Training sample pairs were randomly formed by combining each carrier image with n secret images, where n is the preset number of hidden images. In the experiment, n=3, meaning each steganography task included 3 secret images; the image resolution was uniformly cropped to 256×256 pixels; the training set contained 80,000 sample pairs, and the test set contained 2,000 sample pairs. After the sample construction was completed, the training and test sets were randomly divided. The training set was used to jointly train the hiding and recovery networks, and the test set was used for model performance evaluation.

[0043] In the experiment, based on the number of secret images n, n separate sets of hidden region centers are generated in the feature space of the carrier image. Region centers can be generated using regular grid division, farthest point sampling, or random uniform sampling. For any two region centers... and It needs to satisfy the minimum spacing constraint to ensure that the hidden regions of different secret images do not overlap; at the same time, it needs to satisfy the minimum distance criterion to make multiple regions as evenly distributed as possible in the feature space.

[0044] In the experiment, the radius r of each hidden region iThe sampling points are uniformly set to 16 pixels, the minimum safe distance d0 between any two region centers is set to 32 pixels, the convolution kernel uses a 3×3 fixed sampling grid, and there are a total of 9 sampling points; the actual position of each sampling point is determined by the region center c of the secret image. i The local perturbation term is determined by superimposing a truncated Gaussian perturbation, with the upper limit of the perturbation amplitude set to 0.4 pixels. This ensures that all sampling points of the same convolution kernel are clustered near the center of the region and do not exceed c in size. i Center r i The above parameters are for a circular region with radius denoted as ; these parameters are merely examples and can be adjusted based on network depth, feature map size, and the number of hidden images. After region partitioning, the set of region center values ​​and local perturbation terms are superimposed to generate the set of region convolution offset parameters corresponding to each secret image. This is used to control the convolution sampling position in subsequent steps.

[0045] In the experiment, the carrier image H was compared with the first secret image. The images are mapped to the frequency domain feature space via Discrete Wavelet Transform (DWT) to obtain their respective low-frequency and high-frequency sub-band representations. The frequency domain features of both are then fed into a feature fusion network composed of multiple invertible coupling blocks (InvBlocks) connected in series. The first layer of each invertible coupling block introduces a region convolution, based on the offset parameters corresponding to the first secret image. Control the sampling position so that Feature information is written to the area The corresponding feature subspace. After multiple reversible mappings, the fused features are reconstructed into an intermediate steganalyte image via inverse wavelet transform (IWT). Then with As a new carrier input, it is embedded sequentially. Each embedding operation uses its corresponding region offset parameters to ultimately obtain the steganalyte image. .

[0046] In the experiment, the hidden network consists of 16 reversible coupled blocks connected in series. The nonlinear transformation functions φ(·), ρ(·) and η(·) inside each coupled block are implemented using residual dense blocks. The residual dense blocks consist of 5 convolutional layers, with the first layer being a Region Conv layer. Each intermediate layer outputs 32 feature channels, and the convolutional kernel size is 3×3.

[0047] In the experiment, the final steganographic image was... Input the recovery network. The recovery network shares reversibly coupled block parameters with the hidden network, eliminating the need to retrain independent decoders. The recovery phase will... After being mapped to the frequency domain by DWT, it is concatenated with the auxiliary noise variable z and then fed into the reversible coupling block, in reverse order of the embedding sequence. The inverse feature separation is performed sequentially, with each step using the region convolution offset parameter corresponding to the embedding stage. Perform regional sampling. After each inverse operation, map the currently recovered secret image features to the image space using IWT to obtain the recovered secret image. Then, update the intermediate steganographic feature state and proceed to the next step. Finally, the complete set of recovered secret images is obtained.

[0048] Comparative experiments and performance evaluations were conducted. The experiments were performed on three public datasets: DIV2K, COCO, and ImageNet. The number of secret images embedded in each iteration was set to n=4, meaning four secret images were simultaneously embedded into a single carrier image. Evaluation metrics included the peak signal-to-noise ratio (PSNR) of the steganalyte. h Steganographic structural similarity (SSIM) h ), and the peak signal-to-noise ratio (PSNR) of each secret image recovery si (i=1,2,3,4) and Structural Similarity Restored (SSIM) si The experimental results are shown in Table 1.

[0049] Table 1

[0050] Experimental results show that the method of this invention can maintain high visual quality of steganalytes on different datasets. On the DIV2K dataset, the peak signal-to-noise ratio (PSNR) of the steganalytes is [not specified]. h The structural similarity (SSIM) reached 35.42 dB. h The PSNR reached 0.9549; on the COCO dataset, the PSNR was... h Reaching 34.83 dB, SSIM h The PSNR reached 0.9462 on the ImageNet dataset. h Reaching 34.18 dB, SSIM h The result of 0.9398 indicates that the method of the present invention can effectively maintain the visual fidelity of the carrier image under high-capacity embedding conditions.

[0051] Regarding the quality of secret image restoration, all secret images can be stably restored. On the COCO dataset, the peak signal-to-noise ratios (PSNRs) of the restored four secret images are 34.05 dB, 32.67 dB, 32.18 dB, and 31.28 dB, respectively, with restored structural similarities of 0.9458, 0.9377, 0.9275, and 0.9091. On the DIV2K and ImageNet datasets, the restoration quality trends of each secret image are basically consistent, maintaining a high overall PNR and structural similarity. Experimental results show that the method of this invention can achieve effective embedding and stable extraction of multiple secret images, and reduces mutual interference between different secret images through a regionalized separation hiding mechanism.

[0052] This invention enhances the controllability and randomness of the hiding process and improves the overall security of the system by controlling the convolution sampling offset parameters corresponding to different secret images, thus distributing the secret information in different spatial regions within the carrier image. This invention can be applied to scenarios such as privacy protection, copyright authentication, secure communication, and secure storage of multimedia data.

[0053] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0054] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for image-to-image steganography with controllable hidden areas, characterized in that, Includes the following steps: Step 1: Input the carrier image and several secret images. Generate a set of hidden region centers in the feature space of the carrier image according to the number of secret images, and impose constraints on each hidden region to ensure that the hidden regions of different secret images are separated from each other in the feature space of the carrier image and do not cross the boundary. Step 2: Generate corresponding region convolution sampling offset parameters based on the set of hidden region centers. The region convolution sampling offset parameters are used to control the sampling position of each sampling point of the convolution kernel of the region convolution layer in the hidden network, so that the information of each secret image is only written into its corresponding hidden region. Step 3: Input the carrier image and the secret image into the hidden network. First, they are mapped to the frequency domain feature space by discrete wavelet transform. Then, through deep feature fusion, the regional fusion of carrier features and secret image features is completed by reversible transformation layer by layer. Finally, the steganalysis image is reconstructed by inverse wavelet transform. Step 4: Input the steganalysis image into the recovery network. Based on the region convolution sampling offset parameters consistent with the deep feature fusion, perform regionalized inverse feature separation in reverse order of deep feature fusion to gradually recover each secret image and complete the "image-hiding-image" extraction process of multiple images.

2. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: In step 1, the generation of the hidden region center set is achieved by using regular grid partitioning, farthest point sampling, random uniform sampling, or dynamic search, for any two region centers. and It should satisfy the minimum spacing constraint and the uniform dispersion criterion of maximizing the minimum distance between the centers of any two regions, where i≠j.

3. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: In step 1, center constraints, radius constraints, boundary constraints, and kernel consistency constraints are applied to each hidden region. The center constraint ensures that the overall sampling center of the convolutional kernel of the region convolutional layer is consistent with the center of the hidden region preset in the secret image, that is, all sampling points are distributed around the center of the corresponding region. The radius constraint ensures that the offset distance of each sampling point relative to the center of the region does not exceed the preset radius. The boundary constraint ensures that the coordinates of all sampling points after offset do not exceed the effective range of the feature map, and the excess part is truncated to the boundary. The kernel consistency constraint ensures that all sampling points within the same convolutional kernel share the same hidden region center, and only allows local perturbations to produce small-range offsets near the center.

4. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: In step 2, for the input feature map First, a set of corresponding region convolution sampling offset parameters is generated based on the set of hidden region centers. For the A secret image whose region convolution sampling offset parameter is determined by the center of the hidden region. With local disturbance terms jointly determined, that is The hidden region center is used to determine the overall sampling region of the convolution kernel, and the local perturbation term is used to distribute the sampling points within the convolution kernel around the same region center in a small range. Subsequently, during the region convolution calculation, the... The secret image calls the corresponding region convolution sampling offset parameters, and the output feature map is located at... Convolution results at the location ;in, Indicates the input feature map, Indicates the output feature map at position The convolution result at the point, The fixed sampling grid represents the convolution kernel. Represents the first in the sampling grid Each sampling location The convolution weight parameters represent the corresponding sampling positions; different secret images correspond to different region convolution sampling offset parameters, thereby realizing the regional embedding of secret information in the feature space of the carrier image.

5. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: In step 3, the hiding network includes a first discrete wavelet transform layer (DWT), a second discrete wavelet transform layer, a hidden channel splicing layer (Concat), a reversible hiding module, a hidden channel splitting layer (Split), and a first inverse discrete wavelet transform layer (IWT). Each secret image is embedded sequentially in stages: when embedding the first secret image, the first secret image is input to the first discrete wavelet transform layer, and the carrier image is input to the second discrete wavelet transform layer; when embedding the i-th secret image, i≥2, the i-th... The secret image is input into the first discrete wavelet transform layer, and the steganographic image output from the previous stage is input into the second discrete wavelet transform layer. After both are subjected to discrete wavelet transform, their outputs are fused along the channel dimension by the hidden channel splicing layer and input into the reversible hiding module. The reversible hiding module consists of several sequentially connected reversible coupling blocks. The output of the reversible hiding module is divided into two paths along the channel dimension by the hidden channel splitting layer. The channel corresponding to the carrier branch is subjected to inverse discrete wavelet transform by the first inverse discrete wavelet transform layer to obtain the steganographic image of this stage, and the channel corresponding to the secret branch is used as redundant information. The above process is repeated until all secret images are embedded, and the steganographic image output from the last stage is used as the final steganographic image.

6. The image-to-image steganography method with controllable hidden regions according to claim 5, characterized in that: The reversible coupling block includes an intra-block channel splitting layer Split, a first sub-network φ, a second sub-network ρ, a third sub-network η, and an intra-block channel splicing layer Concat; The input features are divided into a first branch feature x1 and a second branch feature x2 according to the channel dimension by the intra-block channel splitting layer; the second branch feature x2 is transformed by the first sub-network φ and then added element-wise with the first branch feature x1 to form an additive coupling, resulting in a first output feature y1; the first output feature y1 is transformed by the second sub-network ρ and the third sub-network η respectively; the second branch feature x2 is multiplied element-wise with the exponentialized result of the output of the second sub-network ρ, and then added element-wise with the output of the third sub-network η to form an affine coupling, resulting in a second output feature y2; the first output feature y1 and the second output feature y2 are spliced ​​by the intra-block channel splicing layer to serve as the output features of the reversible coupled block; the first sub-network φ, the second sub-network ρ, and the third sub-network η are all composed of residual dense blocks.

7. The image-to-image steganography method with controllable hidden regions according to claim 6, characterized in that: The residual dense block includes a Region Conv layer, a first LeakyReLU layer, a first Concat layer, a first Conv layer, a second LeakyReLU layer, a second Concat layer, a second Conv layer, a third LeakyReLU layer, a third Concat layer, a third Conv layer, a fourth LeakyReLU layer, a fourth Concat layer, and a fourth Conv layer. The input features first pass through the Region Conv layer and the first LeakyReLU layer to obtain a first intermediate feature. The input features and the first intermediate feature are concatenated by the first Concat layer according to the channel dimension, and then sequentially passed through the first Conv layer and the second LeakyReLU layer to obtain a second intermediate feature. This process continues, with the input of each Conv layer being the concatenation of the original input feature with the channels of all its preceding intermediate features, forming a dense connection. The input of the fourth Conv layer is the concatenation of the original input feature with the channels of all its intermediate features, and its output serves as the output feature of the residual dense block. The fourth Conv layer is not connected to any activation layer, and its weights are initialized to zero.

8. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: In step 4, the recovery network includes a third discrete wavelet transform layer, a recovery channel stitching layer (Concat), a reversible reveal module, a recovery channel splitting layer (Split), a second inverse discrete wavelet transform layer, and a third inverse discrete wavelet transform layer. During extraction, the final steganalyte image is used as the first-round input: the input steganalyte image is mapped to the frequency domain by the third discrete wavelet transform layer, and its output, along with auxiliary noise, is fused with the recovery channel stitching layer according to the channel dimension before being input to the reversible reveal module. In actual extraction, the auxiliary noise is replaced by a random sampling amount with the same distribution as the redundant information in the embedding stage. The reversible reveal module and the reversible steganalyte... The hidden modules share the same set of network parameters and are executed by the reversible coupling blocks in reverse order of their embedding execution order. Each reversible coupling block performs inverse operations with the same parameters, i.e., element-wise subtraction replaces element-wise addition, and element-wise division replaces element-wise multiplication. The output of the reversible revealing module is divided into two paths according to the channel dimension by the recovery channel splitting layer, and then subjected to inverse discrete wavelet transform by the second inverse discrete wavelet transform layer and the third inverse discrete wavelet transform layer, respectively, to obtain the restored stegana and the restored secret image. The restored stegana is used as the input for the next round of extraction, and extraction is performed stage by stage in reverse order of embedding order until all secret images are restored.

9. The image-to-image steganography method with controllable hidden regions according to claim 1, characterized in that: Both the hidden network and the recovery network are pre-trained networks; during training, the joint loss function used is: ； Where H represents the carrier image and C represents the steganalytic image. Represents the i-th original secret image, Let represent the i-th recovered secret image output by the recovery network, n represent the total number of secret images, and LL(·) represent the low-frequency subband extraction operator of discrete wavelet transform. Indicates the center of the hidden region corresponding to the i-th and j-th secret images. This indicates the preset minimum safe distance between areas. Denotes the L2 norm; This represents the weighting coefficient used to balance the four losses mentioned above.

10. A steganography system for controllable hidden areas, characterized in that: include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the image-to-image steganography method for controllable hidden regions as described in any one of claims 1 to 9.

Images