Generative image steganography method based on robust position mapping

Abstract

The invention discloses a generative image steganography method based on robust position mapping. According to the technical scheme, a steganography system composed of a run length coding module, an initial Gaussian noise generation module, a candidate pool processing module, a pseudorandom sequence generation module, an index sorting module, an information embedding module and a stable diffusion model SD is constructed. The run length coding module carries out run length coding on the secret information to obtain a two-tuple Y; an initial Gaussian noise generation module generates a Gaussian noise tensor zinit; the candidate pool processing module performs positive and negative value division on the zinit; the pseudo-random sequence module generates a random sequence P, and an index matrix S'is obtained through the index sorting module. The information embedding module embeds the secret sequence after run length coding in the Y into an empty tensor zemed to obtain a tensor zemed conforming to Gaussian distribution; and the SD carries out denoising on the zzzled and decodes the zzzled to generate a steganographic image. According to the method, the secret information extraction accuracy and steganography capacity can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of information hiding, specifically relating to a generative image steganography method based on robust position mapping. Background Technology

[0002] In open network environments, information is vulnerable to eavesdropping, tampering, and even tracking during transmission, posing serious security risks. Steganography, as an information hiding technology, can embed secret information into digital carriers without being detected, thereby achieving covert communication. With the widespread application of digital media, images, due to their high data redundancy and strong information carrying capacity, have become the most commonly used steganographic carrier.

[0003] Traditional image steganography methods primarily modify existing images, directly embedding secret information at the pixel level, or transforming the image to the frequency domain before embedding information. With the development of deep learning, steganography has entered a new research paradigm, automatically learning the optimal embedding location and method through deep features. Overall, the core characteristic of these methods is making minor modifications to existing images to embed information. While this "passive modification" strategy is simple and controllable, it still has limitations in terms of security and robustness against complex attacks, thus providing impetus for the development of generative steganography.

[0004] The core idea of ​​generative image steganography is to directly generate steganographic images containing secret information, rather than modifying existing images. This type of method no longer relies on traditional manual feature design, but instead actively synthesizes images that are both natural and realistic while carrying steganographic information through a generative model, thus effectively overcoming the limitations of traditional methods. Current mainstream generative image steganography techniques generally embed the secret information into Gaussian noise, that is, transforming the initial Gaussian noise through a mapping mechanism, and then using a generative model to generate an image containing the secret information, achieving secure information hiding and transmission. This mapping-based method treats the embedding and extraction of secret information as an inverse process, relying on preset mapping rules to establish a bijective relationship between the secret information and the steganographic image.

[0005] In image steganography, improving steganography performance typically requires comprehensive consideration of key metrics such as extraction accuracy, image quality, and steganalysis capacity. Extraction accuracy refers to the ability to successfully recover secret information from a steganographic image: it measures whether the receiver can accurately reconstruct the secret information under lossless channel conditions, and reflects the robustness of the method in extracting secret information under non-ideal conditions such as disturbances, noise, compression, or attacks. Steganalysis capacity represents the amount of secret information that can be embedded in each image. Several studies have proposed corresponding optimization schemes for these performance metrics. The following is a brief evaluation of several existing methods.

[0006] First, the method proposed by Zhou et al. (Background 1: Z. Zhou, Y. Su, J. Li, K. Yu, QJ Wu, Z. Fu, and Y. Shi, “Secret-to-image reversible transformation for generative steganography,” IEEE Transactions on Dependable and Secure Computing, vol.20, no. 5, pp. 4118-4134, 2023.) utilizes the reversible property of a flow-based model to design a novel mapping mechanism. First, a set of Gaussian noise elements is randomly sampled from a standard normal distribution and grouped and filtered according to their value range. Then, guided by the secret information, the arrangement position of elements within each group in the latent vector is controlled, realizing a one-to-one mapping relationship between the secret bits and the positions of the Gaussian noise elements. This method has high extraction accuracy, but the generated images are prone to detail blurring or over-smoothing, and its ability to generate complex scenes is limited. Secondly, to improve the quality of steganalysts, diffusion models have been introduced into the field of steganography. Hu et al. (Background 2: X.Hu, S.Li, Q.Ying, W.Peng, X.Zhang, and Z.Qian, "Establishing robust generative image steganography via popular stable diffusion," IEEE Transactions on Information Forensics and Security, vol.19, pp.8094-8108, 2024. Translated as: Building a robust generative image steganography method based on popular stable diffusion) proposed a diffusion model-based steganography method that uses orthogonal matrices to implement mapping, which effectively improves the quality of steganalysts, but has a low steganalysis capacity.Additionally, Peng et al. (Background Artwork 3: Y. Peng, Y. Wang, D. Hu, K. Chen, X. Rong, and W. Zhang, "LDStega: Practical and robust generative image steganography based on latent diffusion models," in Proceedings of the 32nd ACM International Conference on Multimedia, pp.3001-3009, 2024. Translated as: LDStega: A practical and robust generative image steganography method based on latent diffusion models) proposed a steganography scheme based on latent diffusion models. By truncating the Gaussian distribution to map the secret information in the latent space, it expands the fault tolerance range of secret information extraction to a certain extent, thereby improving the extraction accuracy. However, since some embedding positions may fall in highly perturbation regions, it is prone to extraction errors, which still affects the extraction accuracy. Furthermore, this method sacrifices the available steganography capacity for reliability, resulting in a decrease in steganography capacity.

[0007] In summary, although existing generative image steganography methods based on latent space diffusion models have made some progress and effectively improved the quality of stegated images, problems such as low extraction accuracy and limited steg capacity still exist. Future research should ensure that the method maintains high extraction accuracy under lossless channel conditions and interference such as noise and compression, while exploring new reversible mapping mechanisms to further improve steg capacity, thereby comprehensively enhancing the overall performance of steganography methods. Summary of the Invention

[0008] The technical problems this invention aims to solve are twofold: First, improving the accuracy of extracting secret information from steganographic images. Under lossless channel conditions, existing diffusion-based methods struggle to guarantee accurate extraction of secret information. Furthermore, during image transmission, storage, or reprocessing, steganographic images are susceptible to compression, cropping, and noise interference, exhibiting insufficient robustness and leading to information loss or decoding errors, thus affecting the accuracy of secret information extraction. Second, increasing the steganographic capacity of image steganography methods. Existing generative image steganography methods based on mapping mechanisms typically have limited steganographic capacity while maintaining image quality, making it difficult to meet the application requirements of high-capacity covert communication. Therefore, there is an urgent need to propose a new steganography method that improves steganographic capacity while ensuring image visual quality. This necessitates designing a more stable steganographic mapping scheme to ensure accurate extraction of secret information under various interference conditions.

[0009] The technical solution of this invention is: The first step is to construct a generative image steganography system based on robust location mapping: Generative image steganography based on robust location mapping is a steganography system based on the Stable Diffusion (SD) model. SD is an image generation model that diffuses in the latent space, proposed by the CompVis team in 2022. The SD structure consists of three main modules: a text encoder, a U-Shaped Neural Network (U-Net), and a Variational Autoencoder (VAE). The text encoder is responsible for encoding the input text prompts (conditions for model generation, guiding the model to generate semantically consistent images or content) into high-dimensional semantic vectors to capture semantic information and guide subsequent image generation. U-Net uses time steps and text semantics as conditions to denoise the Gaussian noise tensor in the latent space, obtaining a structurally consistent and semantically matching latent space representation. The VAE consists of a VAE encoder and a VAE decoder, where the VAE encoder maps the image to a latent space tensor; the VAE decoder maps the latent space tensor back to the image. Based on this structure, SD implements two core processes: noise addition and denoising. The denoising process refers to converting the image into a Gaussian noise tensor. The denoising process refers to generating an image from the Gaussian noise tensor. The text encoder, U-Net, and VAE decoder jointly implement the denoising process of Stitching. First, the text encoder encodes the text prompt input by the sender into a high-dimensional semantic vector. Second, under the condition of the high-dimensional semantic vector, U-Net progressively denoises the Gaussian noise tensor to obtain a denoised latent space tensor. Finally, the VAE decoder maps the latent space tensor to the pixel space to obtain the final generated image. In this steganography method, the embedding and extraction of secret information mainly rely on the denoising process of Stitching. The model weights and related code of StableDiffusion can be obtained from the HuggingFace official repository (https: / / huggingface.co / stable-diffusion-v1-5).

[0010] Specifically, the generative image steganography system based on robust position mapping consists of the following components: a run-length encoding module, an initial Gaussian noise generation module, a candidate pooling module, a pseudo-random sequence generation module, an index sorting module, an information embedding module, a text encoder, a U-Net decoder, and a VAE decoder. In use, both the sending end (used by the sender) and the receiving end (used by the receiver) must deploy the aforementioned modules of the generative image steganography system based on robust position mapping to ensure the embedding and extraction of secret information. Throughout the steganography process, the sender and receiver jointly agree on the first shared parameter. Second shared parameter ( and (They are independent of each other and are all natural numbers). First shared parameter Used to generate the initial Gaussian noise tensor z init Second shared parameter Used to generate a random sequence P.

[0011] The run-length encoding module is connected to the information embedding module. The run-length encoding module encodes the binary secret information M input by the sender into a run-length encoding, obtaining a tuple Y, and then sends Y to the information embedding module.

[0012] The initial Gaussian noise generation module is connected to the candidate pool processing module and receives the first shared parameter input from the sender. In the first shared parameter Generate the initial Gaussian noise tensor z under constraints init , will z init Send to the candidate pool processing module.

[0013] The candidate pool processing module is connected to the initial Gaussian noise generation module and the information embedding module. The candidate pool processing module receives the initial Gaussian noise tensor z from the initial Gaussian noise generation module. init , for z init The positive candidate pool ρ is obtained by partitioning the pool. + and negative candidate pool ρ - , will ρ + and ρ - Send to the information embedding module.

[0014] The pseudo-random sequence generation module is connected to the index sorting module, and receives the second shared parameter input by the sender. Using the second shared parameter Generate a random sequence P and send the random sequence P to the index sorting module.

[0015] The index sorting module, pseudo-random sequence generation module, and information embedding module are interconnected. The index sorting module receives a random sequence P from the pseudo-random sequence generation module, sorts and transforms P to obtain an index matrix. , . Index matrix Send to the information embedding module.

[0016] The information embedding module is connected to the candidate pool processing module, the run-length encoding module, the index sorting module, and U-Net. The information embedding module receives the tuple Y from the run-length encoding module and the positive candidate pool ρ from the candidate pool processing module. + and negative candidate pool ρ - Receive the index matrix from the index sorting module The information embedding module employs a robust position mapping method to embed the run-length encoded secret sequence M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed If the positive candidate pool ρ + or negative candidate pool ρ - There are still remaining elements. Continue to process the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net.

[0017] The text encoder is connected to U-Net, receives the text prompt input by the sender, converts the prompt into a high-dimensional text embedding vector PT, and sends PT to U-Net.

[0018] U-Net connects to the information embedding module, the text encoder, and the VAE decoder. U-Net receives zzz from the information embedding module. embed The text embedding vector PT is received from the text encoder, and zzz is gradually removed under the guidance of PT. embed The noise component in zzz embed The process gradually transitions from a highly perturbed state to a latent space tensor z0 that is structurally clear and semantically stable and conforms to the text embedding vector PT. z0 is then sent to the VAE decoder.

[0019] The VAE decoder is connected to the U-Net, receives the latent space tensor z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego .

[0020] The second step involves the run-length encoding module receiving the binary secret information M input from the sender, performing run-length encoding on M to obtain the run-length encoded secret sequence M', and combining M' and the run-length sequence L to form a binary tuple set Y, which is then sent to the information embedding module; M is as shown in Formula 1: Formula 1 Where n = 1, 2, …, N, b n It is the value of the nth bit in M. N is the length of M, that is, the number of bits in M, and is a positive integer; Y is as shown in Formula 2: Formula 2 Where j = 1, 2, …, J, J is the number of pairs in Y, J ≤ N, and ; , This represents the value of the j-th bit in the secret sequence M'. ; , For natural numbers, Indicates b' j The corresponding run length.

[0021] The third step is the initial Gaussian noise generation module in the first shared parameter. The initial Gaussian noise tensor z is generated using a Gaussian random sampling method under constraints. init , z init The dimensions are represented as (C, H, W), where C is the number of channels, H is the latent space height (i.e., the number of rows), and W is the width (i.e., the number of columns). The elements of each channel follow a standard Gaussian distribution. Let z... init Send to the candidate pool processing module. init The element in the p-th channel, q-th row, and r-th column is represented by z. p,q,r It means that z p,q,r Theoretically, the range of values ​​is Most of the values ​​are concentrated in Within the range. 1≤p≤C, 1≤q≤H, 1≤r≤W, z init The number of elements is C×H×W, and C×H×W≥J.

[0022] Fourth, the candidate pool processing module receives the initial Gaussian noise tensor z from the initial Gaussian noise generation module. init According to z init The candidate pool is divided according to z. init The sign of the element in the middle will be z init The candidate pools are divided into two disjoint pools, namely the positive candidate pool ρ. + and negative candidate pool ρ - As shown in Formula 3: Formula 3 The positive candidate pool ρ + and negative candidate pool ρ - Send to the information embedding module.

[0023] Step 5: The pseudo-random sequence generation module uses a pseudo-random sequence generator (such as the MT19937 pseudo-random sequence generator (M. Matsumoto, T. Nishimura, “Mersenne Twister: A 623-DimensionallyEquidistributed Uniform Pseudo-Random Number Generator,” ACM Transactions on Modeling and Computer Simulation, Vol.8(1), pp 3-30, 1998. translated as: Mersenne Twister: A 623-DimensionallyEquidistributed Uniform Pseudo-Random Number Generator)) with a second shared parameter. Use it as the initial seed to generate a random sequence P. The length of P is equal to z init The number of elements is also C×H×W. The pseudo-random sequence generation module sends P to the index sorting module.

[0024] Step 6: The index sorting module receives P from the pseudo-random sequence generation module, sorts and transforms the elements in P, and obtains the index matrix. ,Will Sending to the information embedding module is done in the following way: (6.1) The index sorting module receives P from the pseudo-random sequence generation module; (6.2) The index sorting module sorts P in ascending order to obtain the index sequence S. s k Represents a sequence of unique integer indices from 1 to C×H×W; (6.3) Using tensor reconstruction (reshape), S is rearranged according to the dimensions C×H×W to obtain the result similar to z. init Three-dimensional index matrix with the same dimensions , , Let be the index value. =v, where v is an integer and 1 v C×H×W is used to guide the embedding location of secret information; (6.4) The index sorting module will Send to the information embedding module.

[0025] Step 7: The information embedding module uses a robust position mapping method to embed the run-length encoded secret sequence M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zzembed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net. The method is: (7.1) The information embedding module receives Y from the run-length encoding module and receives the positive candidate pool ρ from the candidate pool processing module. + and negative candidate pool ρ - Receive the index matrix from the index sorting module Initialize a... empty tensors of the same size ; (7.2) The information embedding module uses a robust position mapping method to embed the secret sequence M' encoded in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed The method is: for Y Perform embedding priority sorting to obtain the sorted secret sequence, which is also the sequence to be embedded, M''; based on the index value... Select in ascending order, one by one. The position (p, q, r) corresponding to =v is used as the embedding. The order of element positions. Following the embedding position order, select from the positive candidate pool ρ according to the sequence to be embedded M''. + and negative candidate pool ρ - Select element values ​​and embed them into We obtain the tensor zz embedded with M'. embed If the positive candidate pool ρ + or negative candidate pool ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed The specific method is as follows: (7.2.1) The information embedding module sorts Y by embedding priority to obtain... Specifically, b' j travel length The longer the run, the higher the priority, thus increasing the priority of b' corresponding to the longer run. j Prioritize mapping to more stable locations to improve the accuracy of secret information extraction, as shown in Formula 4: Formula 4 After prioritizing, the sequence to be embedded is obtained. , Let represent the value of the j-th bit in M''. , represents the corresponding run - length, which is a natural number and satisfies ; (7.2.2) According to the index matrix received from the index sorting module , select in ascending order of the index value and embed into the position order of the elements, that is each index value in corresponds to a specific position in the tensor z embed , and its index order is the embedding order of the sequence M'' to be embedded; (7.2.3) According to the element values in the sequence M'' to be embedded obtained in (7.2.1) , embed into z embed , and obtain the tensor z z embedding M' embed ; If there are remaining elements in the positive candidate pool ρ + or the negative candidate pool ρ - , then continue to fill the remaining elements into z z in ascending order of the element values, and obtain the tensor z z z embed that conforms to the Gaussian noise distribution, specifically: embed (7.2.3.1) Let v = 1; (7.2.3.2) Select the position (p, q, r) corresponding to = v from z embed for embedding, that is, embed into the position z (p, q, r); If embed = 1, take an element z with the largest absolute value from the positive candidate pool and embed it into the position z p,q,r (p, q, r); If embed = 0, take a value z with the largest absolute value from the negative candidate pool and embed it into p,q,r (p, q, r); (7.2.3.3) Let v = v + 1; If v < J, go to (7.2.3.2); If v = J, obtain the tensor z z embedding M' embed and go to (7.2.3.4); (7.2.3.4) Determine z z embedIs there any available position; if J < C×H×W, it means zz embed There is an available position, go to (7.2.3.5); if J = C×H×W, it means zz embed There is no available position in zz embed That is zzz embed , go to (7.2.3.6); (7.2.3.5) Continue to fill the remaining positions to obtain the tensor zzz embed , specifically: (7.2.3.5.1) Let v = J + 1; (7.2.3.5.2) According to the index matrix Embed the remaining elements in the positive candidate pool ρ + or the negative candidate pool ρ - into the remaining positions of zz embed . The specific method is: sort the absolute values of the elements in each candidate pool from large to small. First, select the elements in the positive candidate pool ρ + and fill them into zz embed , then select the elements in the negative candidate pool ρ - and fill them into zz embed ; (7.2.3.5.3) Let v = v + 1; if v < C×H×W, go to (7.2.3.5.2); if v = C×H×W, select the position when the index value is v for filling to obtain zzz embed , the embedding is completed, go to (7.2.3.6); (7.2.3.6) At this time, the element values in zzz embed are the same as those in z init , only the order is different. Therefore, zzz embed and z init each channel follows a Gaussian distribution. After the robust position mapping, the obtained zzz embed contains the information of the sequence M'' to be embedded, and at the positions representing the information of the sequence M'' to be embedded, the absolute values of the element values are larger, and the error tolerance ability is stronger when extracting the secret information, with stronger robustness; [[ID= (49)]] (7.3) The information embedding module sends zzz embed to the U-Net.

[0026] Step 8: The SD text encoder converts the text prompt into a text embedding vector PT, which represents a high-dimensional continuous vector of the input text semantics, containing information such as word meaning, contextual relationships, and semantic structure. The text embedding vector PT effectively captures the semantic information of the input text, providing guidance for subsequent image generation. For example, the input text "A fluffylong-eared rabbit on the grass." will be encoded as an embedding representation containing the semantics of a rabbit. The text encoder then sends the text embedding vector PT to U-Net.

[0027] Step 9: SD's U-Net receives zzz from the information embedding module. embed Receive text embedding vector PT from text encoder, and perform zzz processing under the guidance of text embedding vector PT. embed To denoise and obtain the latent space tensor z0, the method is as follows: (9.1) U-Net receives zzz from the information embedding module embed Receive the text embedding vector PT from the text encoder; (9.2) Initialize the denoising state. U-Net will use zzz embed z, the initial state for the SD denoising process T , let z T =zzz embed Where T represents the total number of steps in the denoising process, 10 T 50; (9.3) Guided by the text embedding vector PT, the noise components are progressively inferred and the latent space state is updated. The method is as follows: (9.3.1) Let the time step t = T; (9.3.2) z t Input U-Net, and have U-Net predict the noise component contained in the t-th latent space state. ; (9.3.3) Based on noise components The latent space state z of the previous time step was calculated using the DPM-Solver++ sampler (C. Lu, Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, "DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models," Machine Intelligence Research, vol. 22, pp.730-751, 2025). t-1 ; (9.3.4) Let t = T-1; if t > 0, go to (9.3.2); if t = 0, obtain the t-th latent space state z0, at which point z0 is basically free of Gaussian noise and contains a stable image semantic structure; go to (9.4); (9.4) U-Net sends z0 to the VAE decoder.

[0028] Step 10: The SD VAE decoder receives z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego： (10.1) The VAE decoder receives z0 from the U-Net; (10.2) The VAE decoder performs scaling or channel transformation on z0 so that z0 meets the input requirements of the decoding network in the VAE decoder, and obtains the normalized latent space tensor z'0. (10.3) By performing multi-layer convolution and upsampling operations on the latent space tensor z'0, the abstract features in the latent space are gradually mapped to higher resolution image features, thereby restoring the spatial structure and texture information of the image; (10.4) Map the higher resolution image features to the RGB image space to generate the final steganalyte I. stego At this point, the process of steganography for the secret information is complete.

[0029] From steganalysis image I stego The process of extracting secret information is the reverse of the steganography process; that is, the decryption process is the reverse execution of the encryption process described above. and For stegographic image I stego The reverse process of the above operations is performed in the reverse order of the steganography process to extract the secret information.

[0030] The following beneficial effects can be achieved by using the present invention: (1) In the seventh step of this invention, the secret sequence M' is embedded into the tensor z according to the designed robust position mapping method. embed By selecting larger values ​​for the embedded elements, it has stronger fault tolerance and effectively improves the accuracy of extracting secret information under lossless channel conditions and interference such as noise and compression. (2) In the seventh step of this invention, based on the designed robust position mapping method, through run-length encoding, z embed One element value can represent secret information of multiple bits, effectively increasing the steganography capacity; (3) Steps 8, 9 and 10 of this invention jointly realize the denoising process of Stable Diffusion, guide image generation through text prompts, generate diverse images, effectively improve the quality of steganographic images, and ensure the security of covert information transmission. (4) This invention proposes a generative image steganography method based on robust position mapping. This method designs an invertible mapping to embed secret information into a latent Gaussian vector of standard Gaussian distribution, and then uses StableDiffusion to generate a steganographic image, which effectively improves the overall performance of the steganographic image. Attached Figure Description

[0031] Figure 1 This is a logical structure diagram of a generative image steganography system based on robust position mapping constructed in the first step of this invention; Figure 2 This is the overall flowchart of the present invention; Figure 3 This is the set of steganalytic images obtained in step ten of Example 1 under different text prompts. Figure 3 (a) is the steganalysis image generated when the text prompt is “A fluffy long-eared rabbit on the grass.” Figure 3 (b) is the steganalysis image generated when the text prompt is "The sound of waves crashing against rocks." Figure 3 (c) is the steganographic image generated when the text prompt is “Portrait of a young woman with curly hair, studio lighting.” Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. To verify the effectiveness of the present invention, the following experiments were conducted in conjunction with the embodiments.

[0033] Experimental Environment: This embodiment was completed on a computing platform with basic image processing and data computation capabilities. This platform is equipped with a conventional CPU and sufficient memory to meet the operational requirements of the image steganography process. In the generative image steganography system based on robust position mapping, the binary secret information M=0110011111100… is input into the run-length encoding module, and the steganography process from step one to step ten is executed sequentially to generate a steganographic image, thus completing the image steganography. The generated steganographic image is basically consistent with the ordinary generated image in terms of texture, color, and overall semantic structure, without artifacts, noise, or structural distortion. It is difficult for the naked eye to distinguish whether secret information is embedded, exhibiting high visual naturalness and concealment, demonstrating the good applicability of the method in multi-scene content generation.

[0034] Using the above-described experimental environment, the present invention can be implemented for example... Figure 2 As shown, it includes the following steps: The first step is to construct a generative image steganography system based on robust location mapping: Generative image steganography systems based on robust location mapping, such as Figure 1 As shown, it consists of a run-length encoding module, an initial Gaussian noise generation module, a candidate pooling module, a pseudo-random sequence generation module, an index sorting module, an information embedding module, a text encoder, a U-Net decoder, and a VAE decoder. In use, both the sending end (used by the sender) and the receiving end (used by the receiver) must deploy the aforementioned modules of a generative image steganography system based on robust position mapping to ensure the embedding and extraction of secret information. Throughout the steganography process, the sender and receiver jointly agree on the first shared parameter. =10, second shared parameter =5. First shared parameter Used to generate the initial Gaussian noise tensor z init Second shared parameter Used to generate a random sequence P.

[0035] The run-length encoding module is connected to the information embedding module. The run-length encoding module encodes the binary secret information M input by the sender into a run-length encoding, obtaining a tuple Y, and then sends Y to the information embedding module.

[0036] The initial Gaussian noise generation module is connected to the candidate pool processing module and receives the first shared parameter input from the sender. In the first shared parameter Generate the initial Gaussian noise tensor z under constraints init , will z init Send to the candidate pool processing module.

[0037] The candidate pool processing module is connected to the initial Gaussian noise generation module and the information embedding module. The candidate pool processing module receives the initial Gaussian noise tensor z from the initial Gaussian noise generation module. init , for z init The positive candidate pool ρ is obtained by partitioning the pool. + and negative candidate pool ρ - , will ρ + and ρ - Send to the information embedding module.

[0038] The pseudo-random sequence generation module is connected to the index sorting module, and receives the second shared parameter input by the sender. Using the second shared parameter Generate a random sequence P and send the random sequence P to the index sorting module.

[0039] The index sorting module, pseudo-random sequence generation module, and information embedding module are interconnected. The index sorting module receives a random sequence P from the pseudo-random sequence generation module, sorts and transforms P to obtain an index matrix. . Index matrix Send to the information embedding module.

[0040] The information embedding module is connected to the candidate pool processing module, the run-length encoding module, the index sorting module, and U-Net. The information embedding module receives the tuple Y from the run-length encoding module and the positive candidate pool ρ from the candidate pool processing module. + and negative candidate pool ρ - Receive the index matrix from the index sorting module The information embedding module employs a robust position mapping method to embed the run-length encoded secret sequence M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed If the positive candidate pool ρ + or negative candidate pool ρ - There are still remaining elements. Continue to process the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net.

[0041] The text encoder is connected to U-Net, receives the text prompt input by the sender, converts the prompt into a high-dimensional text embedding vector PT, and sends PT to U-Net.

[0042] U-Net connects to the information embedding module, the text encoder, and the VAE decoder. U-Net receives zzz from the information embedding module. embedThe text embedding vector PT is received from the text encoder, and zzz is gradually removed under the guidance of PT. embed The noise component in zzz embed The process gradually transitions from a highly perturbed state to a latent space tensor z0 that is structurally clear and semantically stable and conforms to the text embedding vector PT. z0 is then sent to the VAE decoder.

[0043] The VAE decoder is connected to the U-Net, receives the latent space tensor z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego .

[0044] The second step involves the run-length encoding module receiving the binary secret information M input from the sender, performing run-length encoding on M to obtain the run-length encoded secret sequence M', and combining M' and the run-length sequence L to form a binary tuple set Y, which is then sent to the information embedding module; M is as shown in Formula 1: Formula 1 Where n = 1, 2, …, 15000, that is, the number of bits N in M ​​is 15000, b n It is the value of the nth bit in M. Y is shown in Formula 2: Formula 2 Where j = 1, 2, …, 3000, J = 3000, is the number of pairs in Y, and ; , This represents the value of the j-th bit in the secret sequence M'. ; , For natural numbers, Indicates b' j The corresponding run length.

[0045] The third step is the initial Gaussian noise generation module in the first shared parameter. The initial Gaussian noise tensor z is generated using a Gaussian random sampling method under constraints. init , z init The dimensions are represented as (4, 64, 64), where 4 represents the number of channels. The latent space has a height and width of 64, and the elements of each channel follow a standard Gaussian distribution. Let z... init Send to the candidate pool processing module. init The element in the p-th channel, q-th row, and r-th column is represented by z. p,q,r It means that z p,q,r Theoretically, the range of values ​​is Most of the values ​​are concentrated in Within the range: 1≤p≤4, 1≤q≤64, 1≤r≤64, zinit The number of elements is 4×64×64=16384, and 4×64×64≥J=3000.

[0046] Fourth, the candidate pool processing module receives the initial Gaussian noise tensor z from the initial Gaussian noise generation module. init According to z init The candidate pool is divided according to z. init The sign of the element in the middle will be z init The candidate pools are divided into two disjoint pools, namely the positive candidate pool ρ. + and negative candidate pool ρ - As shown in Formula 3: Formula 3 The positive candidate pool ρ + and negative candidate pool ρ - Send to the information embedding module.

[0047] Step 5: The pseudo-random sequence generation module uses the MT19937 pseudo-random sequence generator with the second shared parameter. Use it as the initial seed to generate a random sequence P. The length of P is equal to z init The number of elements is also 4×64×64. The pseudo-random sequence generation module sends P to the index sorting module.

[0048] Step 6: The index sorting module receives P from the pseudo-random sequence generation module, sorts and transforms the elements in P, and obtains the index matrix. ,Will Sending to the information embedding module is done in the following way: (6.1) The index sorting module receives P from the pseudo-random sequence generation module; (6.2) The index sorting module sorts P in ascending order to obtain the index sequence S. s k Represents a sequence of unique integer indices from 1 to 4×64×64; (6.3) Using tensor reconstruction (reshape), S is rearranged according to the dimension 4×64×64 to obtain the result similar to z. init Three-dimensional index matrix with the same dimensions , , Let be the index value. =v, where v is an integer and 1 v 4×64×64, used to guide the embedding location of secret information; (6.4) The index sorting module will Send to the information embedding module.

[0049] Step 7: The information embedding module uses a robust position mapping method to embed the run-length encoded secret sequence M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net. The method is: (7.1) The information embedding module receives Y from the run-length encoding module and receives the positive candidate pool ρ from the candidate pool processing module. + and negative candidate pool ρ - Receive the index matrix from the index sorting module Initialize a... empty tensors of the same size ; (7.2) The information embedding module uses a robust position mapping method to embed the secret sequence M' encoded in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed The method is: for Y Perform embedding priority sorting to obtain the sorted secret sequence, which is also the sequence to be embedded, M''; based on the index value... Select in ascending order, one by one. The position (p, q, r) corresponding to =v is used as the embedding. The order of element positions. Following the embedding position order, select from the positive candidate pool ρ according to the sequence to be embedded M''. + and negative candidate pool ρ - Select element values ​​and embed them into We obtain the tensor zz embedded with M'. embed If the positive candidate pool ρ + or negative candidate pool ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed The specific method is as follows: (7.2.1) The information embedding module sorts Y by embedding priority to obtain... Specifically, b' j travel length The longer the run, the higher the priority, thus increasing the priority of b' corresponding to the longer run. jPrioritize mapping to more stable locations to improve the accuracy of secret information extraction, as shown in Formula 4: Formula 4 After prioritizing, the sequence to be embedded is obtained. , Let represent the value of the j-th bit in M''. , express The corresponding run length is a natural number and satisfies ; (7.2.2) Based on the index matrix received from the index sorting module According to index value Select embedding from smallest to largest The order of elements, i.e. Each index value in Corresponding tensor z embed A specific position in the sequence M'' to be embedded, and its index order is the embedding order of the sequence M'' to be embedded; (7.2.3) Based on the element values ​​in the sequence M'' to be embedded obtained in (7.2.1) According to the embedding position order in (7.2.2), Embedded into z embed We obtain the tensor zz embedded with M'. embed If the positive candidate pool ρ + or negative candidate pool ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz with the ascending order of the element values. embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed Specifically: (7.2.3.1) Let v=1; (7.2.3.2) From z embed Select =v is embedded at the position (p, q, r), that is, embedded into z. embed At the position (p, q, r); if =1, from the positive candidate pool Take the element z with the largest absolute value from the list. p,q,r And embedded in z embed At the position (p, q, r); if =0, from the negative candidate pool Take the value z with the largest absolute value. p,q,r And embedded into At the position (p, q, r); (7.2.3.3) Let \(v = v + 1\); if \(v<3000\), go to (7.2.3.2); if \(v = 3000\), obtain the tensor \(zz\) embedded with \(M'\). embed , go to (7.2.3.4); (7.2.3.4) Determine \(zz\) according to the magnitude relationship between the value of \(J(3000)\) and \(C\times H\times W(4\times64\times64)\). embed Check if there is any empty position; if \(J<C\times H\times W\), it means there is an empty position, go to (7.2.3.5); if \(J = C\times H\times W\), it means there is no empty position in \(zz\). embed \(zz\) embed is \(zzz\). embed , go to (7.2.3.6); (7.2.3.5) Continue to fill the remaining positions to obtain the tensor \(zzz\). embed , specifically: (7.2.3.5.​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0050] Step 8: The SD text encoder converts the text prompt into a text embedding vector PT, which represents a high-dimensional continuous vector of the input text semantics, containing information such as word meaning, contextual relationships, and semantic structure. The text embedding vector PT effectively captures the semantic information of the input text, providing guidance for subsequent image generation. For example, the input text "A fluffylong-eared rabbit on the grass." will be encoded as an embedding representation containing the semantics of a rabbit. The text encoder then sends the text embedding vector PT to U-Net.

[0051] Step 9: SD's U-Net receives zzz from the information embedding module. embed Receive text embedding vector PT from text encoder, and perform zzz processing under the guidance of text embedding vector PT. embed To denoise and obtain the latent space tensor z0, the method is as follows: (9.1) U-Net receives zzz from the information embedding module embed Receive the text embedding vector PT from the text encoder; (9.2) Initialize the denoising state. U-Net will use zzz embed z, the initial state for the SD denoising process T , let z T =zzz embed Where T=20 represents the total number of steps in the denoising process; (9.3) Guided by the text embedding vector PT, the noise components are progressively inferred and the latent space state is updated. The method is as follows: (9.3.1) Let the time step t = T; (9.3.2) z t Input U-Net, and have U-Net predict the noise component contained in the t-th latent space state. ; (9.3.3) Based on noise components The latent space state z of the previous time step is calculated using the DPM-Solver++ sampler according to the inverse inference rule of SD. t-1 ; (9.3.4) Let t = T-1; if t > 0, go to (9.3.2); if t = 0, obtain the t-th latent space state z0, at which point z0 is basically free of Gaussian noise and contains a stable image semantic structure; go to (9.4); (9.4) U-Net sends z0 to the VAE decoder.

[0052] Step 10: The SD VAE decoder receives z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego： (10.1) The VAE decoder receives z0 from the U-Net; (10.2) The VAE decoder performs scaling or channel transformation on z0 so that z0 meets the input requirements of the decoding network in the VAE decoder, and obtains the normalized latent space tensor z'0. (10.3) By performing multi-layer convolution and upsampling operations on the latent space tensor z'0, the abstract features in the latent space are gradually mapped to higher resolution image features, thereby restoring the spatial structure and texture information of the image; (10.4) Map the higher resolution image features to the RGB image space to generate the final steganalyte I. stego At this point, the process of steganography for the secret information is complete.

[0053] Experimental results show that the performance improvements achieved by this invention are as follows: (1) The ACC (Accuracy of Extraction ACC = Number of Correctly Extracted Secret Information Bits / Total Number of Embedded Secret Information Bits) of the secret information can reach 99%, which can stably and effectively recover the embedded secret information, and is superior to background technology 2 (ACC = 98.05%) and background technology 3 (ACC = 98.50%). (2) The steganalysis capacity is increased by N / J times. The steganalysis capacity reflects the carrying capacity of the steganalysis system and can be measured by the number of secret information bits that can be hidden in a single image, i.e., steganalysis capacity = number of embedding positions × number of bits that can be hidden at each position. The steganalysis capacity of the present invention is C×H×W×N / J (the number of embedding positions is C×H×W, and the number of bits that can be hidden at each position is the average run length N / J corresponding to each embedding value), which is N / J times higher than other steganalysis methods based on Stable Diffusion (Background Art 2 and Background Art 3). (3) Relying on the powerful generation capabilities of Stable Diffusion, this method can generate diverse and high-quality steganographic images, which is superior to background technique 1, which can only generate images of one style. For example... Figure 3 As shown, when hiding the same secret information M, different text cues can generate different types of steganalogs. Figure 3 The set of steganographic images shown contains three steganographic images. Figure 3 (a) is the steganalysis image generated when the text prompt is “A fluffy long-eared rabbit on the grass.” Figure 3(b) is the steganalysis image generated when the text prompt is "The sound of waves crashing against rocks." Figure 3 (c) The steganographic image generated when the text prompt is “Portrait of a young woman with curly hair, studiolighting.” From Figure 3 As can be seen, the same secret information can generate different types of steganographic images under different text prompts, which has diversified generation capabilities to meet the needs of covert communication scenarios, and 99% of the secret information can be recovered through the reverse process. (4) Experimental results show that although background technique 1 may achieve 100% extraction accuracy without loss, the generated steganographic image type is limited. Although background techniques 2 and 3 can generate diverse steganographic images like the present invention, their extraction accuracy and steganographic capacity are inferior to the technical solution proposed in the present invention. Therefore, the technical solution proposed in the present invention has better overall performance than the solutions proposed by the other three background techniques, effectively improving the overall performance of steganographic images.

[0054] The foregoing has provided a detailed description of a generative image steganography system based on robust position mapping provided by this invention. This document elucidates the principles and implementation methods of this invention, and the above description is intended to aid in understanding the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A generative image steganography method based on robust position mapping, characterized in that... Includes the following steps: The first step involves constructing a generative image steganography system based on robust location mapping. This system is a steganography system based on the stable diffusion model (SD). It consists of a run-length encoding module, an initial Gaussian noise generation module, a candidate pooling module, a pseudo-random sequence generation module, an index sorting module, an information embedding module, a text encoder in SD, a U-Net in SD, and a VAE decoder in SD. This system is deployed on both the sending end (used by the sender) and the receiving end (used by the receiver). The sender and receiver jointly agree on the first shared parameter. Second shared parameter , and All are natural numbers; The second step is that the run-length encoding module receives the binary secret information M input by the sender, performs run-length encoding on M to obtain the run-length encoded secret sequence M', and combines M' and the run-length length sequence L to form a set of two tuples Y, which is then sent to the information embedding module. M is as shown in Formula 1: Formula 1 Where n = 1, 2, …, N, b n It is the value of the nth bit in M. N is the length of M, that is, the number of bits in M, and is a positive integer; Y is as shown in Formula 2: Formula 2 Where j = 1, 2, …, J, J is the number of pairs in Y, J ≤ N, and ; , This represents the value of the j-th bit in the secret sequence M'. ; , For natural numbers, Indicates b' j The corresponding run length; The third step is the initial Gaussian noise generation module. The initial Gaussian noise tensor z is generated using a Gaussian random sampling method under constraints. init , z init The dimension is represented as (C, H, W), where C is the number of channels, H is the latent space height (number of rows), and W is the width (number of columns). The elements of each channel follow a standard Gaussian distribution. init Send to the candidate pool processing module; z init The element in the p-th channel, q-th row, and r-th column is represented by z. p,q,r This means: 1≤p≤C, 1≤q≤H, 1≤r≤W, z init The number of elements is C×H×W, and C×H×W≥J; Fourth step, the candidate pool processing module receives z from the initial Gaussian noise generation module. init According to z init The sign of the element in the middle will be z init Divided into positive candidate pools ρ + and negative candidate pool ρ - , will ρ + and ρ - Send to the information embedding module; Step 5: The pseudo-random sequence generation module uses the MT19937 pseudo-random sequence generator. Use it as the initial seed to generate a random sequence P. The length of P is equal to C × H × W. Send P to the index sorting module; Step 6: The index sorting module receives P from the pseudo-random sequence generation module, sorts and transforms the elements in P, and obtains the index matrix. , , For index value, =v, where v is an integer and 1 v C×H×W, will Send to the information embedding module; Step 7: The information embedding module uses a robust position mapping method to embed M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net; the method is: (7.1) The information embedding module receives Y from the run-length encoding module and ρ from the candidate pool processing module. + and ρ - Receive from the index sorting module Initialize a with empty tensors of the same size ; (7.2) The information embedding module uses a robust position mapping method to embed the secret sequence M' encoded in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed The method is: for Y Perform embedding priority sorting to obtain the sorted secret sequence, which is also the sequence to be embedded, M''; based on the index value... Select in ascending order, one by one. The position (p, q, r) corresponding to =v is used as the embedding. The order of elements; according to the embedding position, from ρ according to M'' + and ρ - Select element values ​​and embed them into We obtain the tensor zz embedded with M'. embed If ρ + or ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed ; (7.3) The information embedding module will zzz embed send to the U-Net; Step 8: SD's text encoder converts the text prompt into a text embedding vector PT, which represents a high-dimensional continuous vector of the input text semantics, including word meanings, contextual relationships, and semantic structure; the text encoder sends the PT to U-Net; Step 9: SD's U-Net receives zzz from the information embedding module. embed Receive PT from text encoder, and under the guidance of PT, zzz embed Denoising is performed to obtain the latent space tensor z0, and z0 is sent to the VAE decoder; Step 10: The SD VAE decoder receives z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego The process of steganography of the secret information has ended.

2. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... In the generative image steganography system based on robust location mapping described in the first step: The run-length encoding module is connected to the information embedding module; the run-length encoding module encodes the binary secret information M input by the sender to obtain a tuple Y, and sends Y to the information embedding module; The initial Gaussian noise generation module is connected to the candidate pool processing module and receives the first shared parameter input from the sender. ,exist Generate the initial Gaussian noise tensor z under constraints init , will z init Send to the candidate pool processing module; The candidate pool processing module is connected to the initial Gaussian noise generation module and the information embedding module; the candidate pool processing module receives the initial Gaussian noise tensor z from the initial Gaussian noise generation module. init , for z init The positive candidate pool ρ is obtained by partitioning the pool. + and negative candidate pool ρ - , will ρ + and ρ - Send to the information embedding module; The pseudo-random sequence generation module is connected to the index sorting module, and receives the second shared parameter input by the sender. ,use Generate a random sequence P and send P to the index sorting module; The index sorting module is connected to the pseudo-random sequence generation module and the information embedding module; the index sorting module receives P from the pseudo-random sequence generation module, sorts and transforms P to obtain the index matrix. ,Will Send to the information embedding module; The information embedding module is connected to the candidate pool processing module, the run-length encoding module, the index sorting module, and U-Net; the information embedding module receives the tuple Y from the run-length encoding module and receives ρ from the candidate pool processing module. + and ρ - Receive from the index sorting module A robust position mapping method is used to embed the run-length encoded secret sequence M' in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed If there are still remaining elements in the candidate pool, continue to process the remaining elements according to... Fill the zz in ascending order embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed , zzz embed Send to U-Net; The text encoder is connected to U-Net, receives the text prompt input by the sender, converts the prompt into a high-dimensional text embedding vector PT, and sends the PT to U-Net; U-Net is connected to the information embedding module, the text encoder, and the VAE decoder; U-Net receives zzz from the information embedding module. embed PT is received from the text encoder, and zzz is gradually removed under the guidance of PT. embed The noise component in zzz embed Gradually transition from a highly perturbed state to a structurally clear, semantically stable latent space tensor z0 that conforms to PT, and send z0 to the VAE decoder; The VAE decoder is connected to the U-Net, receives the latent space tensor z0 from the U-Net, decodes z0, and obtains the steganalytic image I. stego .

3. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... The third step mentioned z p,q,r Theoretically, the range of values ​​is Most of the values ​​are concentrated in Within the range.

4. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... The candidate pool processing module in step four, based on z init The method for dividing the candidate pool is: based on z init The sign of the element in the middle will be z init The candidate pools are divided into two disjoint pools, namely the positive candidate pool ρ. + and negative candidate pool ρ - As shown in Formula 3: Formula 3.

5. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... The pseudo-random sequence generator mentioned in step 5 is the MT19937 pseudo-random sequence generator.

6. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... The sixth step, the index sorting module, sorts and transforms the elements in P to obtain the index matrix. The method is: (6.1) The index sorting module receives P from the pseudo-random sequence generation module; (6.2) The index sorting module sorts P in ascending order to obtain the index sequence. , s k Represents a sequence of unique integer indices from 1 to C×H×W; (6.3) Using tensor reconstruction, S is rearranged according to the dimensions C×H×W to obtain the result with z. init Three-dimensional index matrix with the same dimensions , ,make =v, where v is an integer and 1 v C×H×W is used to guide the embedding location of secret information; (6.4) The index sorting module will Send to the information embedding module.

7. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... (7.2) The information embedding module in step (7.2) uses a robust position mapping method to embed the secret sequence M' encoded in Y into the empty tensor z. embed We obtain the tensor zz embedded with M'. embed and zz embed Filling the space yields a tensor zzz that conforms to a Gaussian noise distribution. embed The specific method is: (7.2.1) The information embedding module sorts Y by embedding priority to obtain... As shown in Formula 4: Formula 4 make , Let M'' represent the value of the j-th bit; , express The corresponding run length is a natural number and satisfies ; (7.2.2) Based on the index matrix received from the index sorting module According to index value Select embedding from smallest to largest The order of elements, i.e. Each index value in Corresponding tensor z embed A specific position in the sequence M'' to be embedded, and its index order is the embedding order of the sequence M'' to be embedded; (7.2.3) Based on the element values ​​in the sequence M'' to be embedded obtained in (7.2.1) According to the embedding position order in (7.2.2), Embedded into z embed We obtain the tensor zz embedded with M'. embed If ρ + or ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz with the ascending order of the element values. embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed .

8. The generative image steganography method based on robust position mapping as described in claim 7, characterized in that... (7.2.3) The step described above uses the element values ​​in the sequence M'' to be embedded. According to the order of embedding position Embedded into z embed We obtain the tensor zz embedded with M'. embed If ρ + or ρ - If there are still remaining elements, then continue processing the remaining elements according to... Fill the zz with the ascending order of the element values. embed This yields a tensor zzz that conforms to a Gaussian noise distribution. embed The method is: (7.2.3.1) Let v=1; (7.2.3.2) From z embed Select =v is embedded at the position (p, q, r), that is, embedded into z. embed At position (p, q, r); if =1, from the positive candidate pool Take the element z with the largest absolute value from the list. p,q,r And embedded in z embed At the position (p, q, r); if =0, from the negative candidate pool Take the value z with the largest absolute value. p,q,r And embedded into At the position (p, q, r); (7.2.3.3) Let v = v + 1; if v < J, go to (7.2.3.2); if v = J, obtain the tensor zz embedded with M' embed , go to (7.2.3.4); (7.2.3.4) Based on the relationship between J and C×H×W values, determine zz. embed Are there any available spaces? If J < C × H × W, it indicates that there are vacant positions, and proceed to (7.2.3.5); if J = C × H × W, it indicates that embed there are no vacant positions in embed which is embed , and proceed to (7.2.3.6); (7.2.3.5) Continue filling the remaining positions to get zzz embed Embedding complete, proceed to (7.2.3.6); (7.2.3.6) zzz embed The element values ​​in z init They are the same, only the order is different.

9. The generative image steganography method based on robust position mapping as described in claim 8, characterized in that... (7.2.3.5) Continue filling the remaining positions as described in step (7.2.3.5) to obtain the tensor zzz. embed The method is: (7.2.3.5.1) Let v = J + 1; (7.2.3.5.2) According to the index matrix The positive candidate pool ρ + or negative candidate pool ρ - The remaining elements are embedded in zz. embed The remaining positions are determined by the following method: Sort the elements in each candidate pool from largest to smallest absolute value, and then select the positive candidate pool ρ according to the order of their absolute values ​​from largest to smallest. + Fill the middle element into zz embed Then, negative candidate pool ρ is selected in descending order of absolute value of the elements. - Fill the middle element into zz embed ; (7.2.3.5.3) Let v = v + 1; if v < C × H × W, go to (7.2.3.5.2); if v = C × H × W, select the position at index value v for filling to obtain zzz embed , the embedding is completed.

10. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... The ninth step describes how SD's U-Net embedding module, under PT guidance, performs operations on zzz. embed The method for denoising to obtain the latent space tensor z0 is as follows: (9.1) U-Net receives zzz from the information embedding module embed Receive the text embedding vector PT from the text encoder; (9.2) Initialize the denoising state; U-Net will denoise zzz embed z, the initial state for the SD denoising process T , let z T =zzz embed Where T represents the total number of steps in the denoising process, 10 T 50; (9.3) Guided by the text embedding vector PT, the noise components are progressively inferred and the latent space state is updated. The method is as follows: (9.3.1) Let the time step t = T; (9.3.2) z t Input U-Net, and have U-Net predict the noise component contained in the t-th latent space state. ; (9.3.3) Based on The latent space state z of the previous time step is calculated using the DPM-Solver++ sampler according to the inverse inference rule of SD. t-1 ; (9.3.4) Let t = T-1; if t > 0, go to (9.3.2); if t = 0, obtain the t-th latent space state z0, at which point z0 no longer contains Gaussian noise and contains a stable image semantic structure, go to (9.4); (9.4) U-Net sends z0 to the VAE decoder.

11. The generative image steganography method based on robust position mapping as described in claim 1, characterized in that... In step ten, the SD's VAE decoder decodes z0 to obtain the steganalytic image I. stego The method is: (10.1) The VAE decoder receives z0 from the U-Net; (10.2) The VAE decoder performs scaling or channel transformation on z0 so that z0 meets the input requirements of the decoding network in the VAE decoder, and obtains the normalized latent space tensor z'0. (10.3) By performing multi-layer convolution and upsampling operations on the latent space tensor z'0, the abstract features in the latent space are gradually mapped to higher resolution image features, thereby restoring the spatial structure and texture information of the image; (10.4) Map the higher resolution image features to the RGB image space to generate the final steganalyte I. stego .

Images