Robust secure image steganography method and system based on hybrid expert and latent space optimization

Abstract

The application discloses a robust secure image steganography method and system based on mixed experts and hidden space optimization, and belongs to the technical field of information security. The method comprises the following steps: converting a secret message into an anti-distortion message representation with a distributed structure; generating clean initial latent space features based on a pre-trained diffusion model, in combination with a given text prompt word and a deterministic sampling strategy; generating a latent space disturbance based on the initial latent space features and the anti-distortion message representation, and superimposing the latent space disturbance on the initial latent space features to obtain a steganography latent variable; and performing a decoding operation on the steganography latent variable to obtain a steganography image. The application can effectively support robust covert communication under complex real channel distortion.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and in particular relates to a robust secure image steganography method and system based on hybrid experts and latent space optimization. Background Technology

[0002] The main research direction of contemporary steganography is to explore how to hide data in digital media carriers such as images, audio and video, so as to ensure that the encrypted file is highly consistent with the original carrier in terms of audiovisual perception and underlying statistical properties and is difficult to distinguish.

[0003] In today's age of artificial intelligence, the creation of high-quality images is no longer the exclusive domain of professional designers or artists. The rapid development of generative models has greatly promoted the rise and popularity of AI-generated content applications, and generative image technology has thus provided a completely new way of producing digital images, becoming one of the most innovative and influential media for dissemination today. Furthermore, with the continuous advancement of technologies such as deep learning algorithms, high-performance computing, and large models, generative images can be rapidly created and widely disseminated on the internet while maintaining extremely high visual fidelity and artistic detail.

[0004] The core logic of covert communication lies in disguising it as routine, everyday data interaction. This requires the chosen communication medium to have extremely high penetration and frequency of daily occurrence in network transmission. Using generative images as a carrier of covert communication fully leverages the explosive growth of AI-generated content on the internet and its inherent camouflage advantages as an emerging medium. Steganographers can cleverly integrate confidential data into images by combining specific generative model characteristics. Subsequently, these coded generated images can be naturally distributed through AI art sharing communities, mainstream online social media, or cloud storage platforms, thereby achieving highly covert information transmission. Furthermore, the "generation is steganography" characteristic integrates the image with the secret information, further enhancing the security of the overall communication.

[0005] Therefore, generative imaging overcomes the limitations of traditional digital images in covert communication, providing a better path for achieving large-volume, highly covert data transmission. For this reason, this technology is considered a highly promising key communication method, playing an irreplaceable role in industries requiring strict confidentiality, such as the military and finance.

[0006] Image steganography techniques mainly include three basic properties: imperceptibility, embedding capacity, and robustness. The relevant definitions are as follows.

[0007] (1) Imperceptibility. This mainly includes sensory indistinguishability and statistical undetectability. The former refers to the visual difference between the stegana and the carrier image being indistinguishable to the human sensory system; the latter refers to the statistical properties of the stegana and the carrier image being approximately the same for a given feature.

[0008] (2) Embedding capacity. This refers to the maximum amount of data that can be embedded into the carrier image while ensuring a certain degree of imperceptibility.

[0009] (3) Robustness. Reflects the ability to resist lossy digital signal processing. It means that after the stegana has undergone lossy digital signal processing (such as blurring or lossy compression), the embedded information can still be reliably and completely recovered.

[0010] In conventional end-to-end steganography, the exposure of a single communication node often triggers a chain reaction, potentially leading to the entire communication link being compromised. In contrast, publishing steganographic multimedia content on social media platforms and utilizing shared public pools for covert data exchange naturally severs direct contact between senders and receivers, providing a higher level of covert security. However, existing social networks, to optimize storage and display, perform a series of lossy operations on user-uploaded media data by default, such as filtering enhancement, resizing, and secondary compression. Furthermore, the types and parameters of these underlying algorithms remain completely closed to the outside world. This situation inevitably leads to uncontrollable "black box" channel distortion during the distribution of steganographic carriers. Therefore, accurately reconstructing confidential information after undergoing such complex lossy digital signal processing is a prerequisite for establishing a highly secure, unconnected, covert communication system based on social platforms. Based on this, robust steganography techniques for unreliable lossy channels (especially in social network environments) are increasingly becoming a cutting-edge focus for both academia and industry in information hiding. Currently, generative image steganography technology is still in its early stages of development and can be divided into two categories according to the generative model: 1) Generative steganography based on generative adversarial networks: The main idea is to use GANs to map secret images into unrelated natural images for spoofing and transmission, and then accurately reconstruct the original secret image at the receiving end through an inverse network. 2) Generative steganography based on diffusion model: The main idea is to use the latent space characteristics and reverse denoising mechanism of diffusion model to embed secret information into the dynamic generation process of image by manipulating initial noise, fine-tuning latent features or intervening in denoising path, and rely on the corresponding inverse process or decoding steps to achieve accurate extraction of information.

[0011] Among them, Chinese patent application No. 202610035767.0 discloses a generative image steganography method based on robust position mapping. However, its embedding mechanism adopts run-length encoding and physical position mapping. This mapping method, which relies on fixed rules, is prone to exposing statistical features and has weak resistance to mixing distortion. Information concentrated in a specific position is vulnerable to attack.

[0012] Chinese patent application No. 202511808331.2 discloses a robust video steganography method and apparatus based on an audio side channel. However, it generates a reference image at the receiving end and performs complex gradient descent iterative optimization, which leads to huge computational overhead and high latency in the extraction process, making it difficult to meet the requirements of real-time applications. Furthermore, this optimization process depends on the quality of the initial estimate and is prone to getting trapped in local optima, resulting in poor robustness under channel distortion and easy message extraction failure. Summary of the Invention

[0013] This invention discloses a robust secure image steganography method and system based on hybrid experts and latent space optimization. It is a generative image steganography method that integrates learnable orthogonal transformation, residual hiding network covert embedding, hybrid expert (MoE) denoising and adaptive repair strategies, thereby effectively supporting robust covert communication under complex real channel distortion (such as compression, blurring, noise, etc.).

[0014] To achieve the above objectives, the technical solution of the present invention includes the following:

[0015] A robust and secure image steganography method based on hybrid expert and latent space optimization is applied at the sending end. The method includes: Transform secret messages into distributed, distortion-resistant message representations; Based on a pre-trained diffusion model, and combined with given text prompts and a deterministic sampling strategy, clean initial latent space features are generated. Based on the initial latent space features and the anti-distortion message representation, a latent space perturbation is generated and superimposed on the initial latent space features to obtain the stegatical latent variables; Decoding the steganalytic latent variables yields the steganalytic image.

[0016] Furthermore, the steganalytic latent variables are decoded to obtain the steganalytic image, including: The steganographic latent variables are decoded using a decoder of a variational autoencoder. The decoding result is mapped back to pixel space to obtain the steganalytic image.

[0017] Furthermore, the secret message is converted into a two-dimensional message sequence, including: Convert the secret message into binary; The binary secret information is converted into bipolar symbols and reshaped into a two-dimensional matrix; The two-dimensional matrix is ​​mapped to a distributed, distortion-resistant message representation using a learnable orthogonal weight matrix.

[0018] Furthermore, based on the initial latent space features and the anti-distortion message representation, latent space perturbations are generated, including: The initial latent space features are combined with the anti-distortion message representation through channel splicing; The splicing result is input into a hidden network based on the U-Net structure to obtain the latent space perturbation.

[0019] Furthermore, after decoding the steganalytic latent variables to obtain the steganalytic image, the method further includes: The steganographic image is sent to the receiving end so that the receiving end can extract the secret message in the steganographic image by combining the given text prompt words and a deterministic sampling strategy; The receiving end extracts the secret message from the steganalysis image by combining the given text prompts and sampling strategy, including: The received steganatical image is encoded using a VAE encoder to obtain the damaged steganatical latent variables; Based on a pre-trained diffusion model, and combined with given text prompts and a deterministic sampling strategy, clean initial latent space features are generated. The damaged stegana latent variables are denoised using a hybrid expert denoising module to obtain the repaired stegana latent variables; wherein, the hybrid expert denoising module includes a routing network, a denoising expert network and a sparse gating fusion network; The initial latent space features, the repaired steganalytic latent variables, and the difference between the initial latent space features and the repaired steganalytic latent variables are all input into the extraction network to obtain the anti-distortion message representation; The message representation is converted into binary secret information by applying an inverse orthogonal transformation matrix to resist distortion. Restore the binary secret information to the secret message.

[0020] Furthermore, the damaged steganalytic latent variables are denoised to obtain the repaired steganalytic latent variables, including: Spatial degradation features and local distortion clues of damaged stegana latent variables are extracted by routing network to obtain a set of probability vector sequences representing the composition of distortion; wherein, the k-th component in the probability vector sequence represents the confidence that the damaged stegana latent variable contains the k-th type of distortion; A set of denoising expert networks is obtained based on the components that are greater than the set threshold. Forward inference of damaged stegana latent variables based on a denoising expert network ensemble; The output features of each denoising expert network are weighted and summed based on a sparse gated fusion network to obtain the repaired steganalytic latent variables.

[0021] Furthermore, the training process for the orthogonal weight matrix, the hidden network, the hybrid expert denoising module, and the extraction network includes: The orthogonal weight matrix and expert denoising module are frozen, and the hidden network and extraction network are jointly trained using the L2 message reconstruction loss function to optimize message recovery accuracy. The orthogonal weight matrix and expert denoising module are frozen, and perceptual loss and diffusion latent space distance loss are introduced to jointly train the hidden network and the extraction network to constrain the magnitude of the latent space perturbation predicted by the hidden network. The hidden network is frozen, and the MoE denoising module, orthogonal weight matrix, and extraction network are jointly trained based on end-to-end message recovery loss, so that the MoE denoising module, orthogonal weight matrix, and extraction network have the ability to adaptively repair and extract various unknown mixed distortions with high precision.

[0022] A robust and secure image steganography system based on hybrid expert and latent space optimization, the system comprising: The message conversion module is used to convert secret messages into a distributed, distortion-resistant message representation. Initial latent variable generation is used to generate clean initial latent space features based on a pre-trained diffusion model and combined with given text prompts and a deterministic sampling strategy. The hidden embedding module is used to generate latent space perturbations based on the initial latent space features and the anti-distortion message representation, and to superimpose the latent space perturbations onto the initial latent space features to obtain the stegatical latent variables; The decoding module is used to decode the steganalytic latent variables to obtain the steganalytic image.

[0023] A computer device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the robust secure image steganography method based on hybrid expert and latent space optimization as described above.

[0024] A computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the robust secure image steganography method based on hybrid expert and latent space optimization as described above.

[0025] Compared with the prior art, the present invention has at least the following beneficial effects.

[0026] 1) Enhanced Steganography Robustness and Security Based on Learnable Orthogonal Transformation. Most current generative image steganography methods based on latent space directly encode the secret information and modify the corresponding latent variables or initial noise. This centralized feature embedding method is extremely sensitive to local distortions in the channel, and is prone to bit errors once the image is damaged. In contrast, this invention designs a distortion-resistant message representation strategy based on learnable orthogonal transformation. By converting centralized discrete bit information into a globally distributed continuous feature representation, the message energy is effectively dispersed into a high-dimensional latent space. This not only ensures that local channel distortion only causes partial signal degradation, thus significantly improving fault tolerance and anti-interference capability, but also makes the transformed feature representation statistically approximate isotropic Gaussian noise, further enhancing the security of steganography.

[0027] 2) Enhanced Steganalysis Resistance Based on Residual Hidden Embedding. Existing generative image steganography methods often lack strict constraints on the diffusion path when modifying the latent space features of the generative model. This can easily lead to the generated steganalytic latent variables deviating from the natural prior distribution of the original diffusion model, leaving statistical anomalies that can be detected by advanced steganalysis algorithms (especially latent space domain detectors). In contrast, this invention designs a hidden network hidden embedding strategy based on residual learning. Instead of directly generating and replacing complete steganalytic latent variables, it uses the network to predict only the minimal latent space perturbations used to carry secret information (…). This design strictly limits steganalytic modifications to a very small range, preserving the original statistical distribution of the clean latent space features of the diffusion model to the greatest extent, effectively reducing detectable statistical shifts, and significantly enhancing security performance against advanced steganalysis algorithms.

[0028] 3) Enhanced Robustness of Complex Channels Based on Hybrid Expert Adaptive Denoising. In lossy transmission channels such as real-world social networks, steganographic images often undergo complex and unpredictable heterogeneous hybrid distortions such as compression, scaling, noise addition, and blurring. Current methods typically employ a single decoding network at the receiver for coarse-grained, uniform feature extraction, lacking an adaptive mechanism to address unknown channel degradation, leading to a significant drop in extraction accuracy under real-world hybrid distortion scenarios. This invention innovatively designs and introduces a latent variable denoising module based on a hybrid expert (MoE) architecture at the receiver, combining anti-distortion message representation and residual covert embedding. This module dynamically evaluates the distortion composition of damaged latent variables through a routing network and uses weighted sparse activation of specific denoising experts for forward targeted fine-tuning, effectively filtering out complex interference from heterogeneous channels before message decoding. This adaptive repair mechanism effectively supports highly reliable steganographic communication under unpredictable real-world lossy channels and possesses strong plug-and-play generalization capabilities as an independent component. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall message embedding and extraction of the present invention.

[0030] Figure 2 This is a schematic diagram of the adaptive denoising strategy based on a multi-expert network of the present invention.

[0031] Figure 3 This is a schematic diagram of the embedding and extraction network of the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, and to make the objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] According to research, most current generative image steganography methods based on latent space rely heavily on fragile direct latent variable modification or diffusion path manipulation to resist channel distortion. At the receiver, a single decoding network is typically used for unified extraction, making them highly susceptible to bit errors when faced with mixed distortions such as compression, blurring, and noise. Furthermore, these methods lack adaptability to unknown channel degradation. Moreover, existing methods often struggle to effectively suppress statistical offsets when directly modifying latent variables, making the generated steganographic images easily detectable by steganalysis algorithms. To address these shortcomings, this invention offers the following improvements: First, a distortion-resistant message representation strategy based on learnable orthogonal transformation is designed, distributing secret information into a high-dimensional latent space, effectively improving fault tolerance and anti-interference capabilities under local distortion. Second, a latent variable covert fusion strategy based on residual learning is designed, predicting only minimal latent space perturbations. This effectively reduces detectable statistical anomalies. Furthermore, based on the joint optimization approach of the transmitting and receiving ends, an innovative adaptive latent space repair scheme based on a hybrid expert (MoE) denoising module is designed at the receiving end, thus providing reliable technical support for robust image steganography communication under complex and unpredictable real channels (such as lossy channels in social networks).

[0034] like Figure 1 As shown, the overall process of this invention includes the embedding process at the transmitting end, lossy channel transmission, and the extraction process at the receiving end.

[0035] (a) Embedding process at the sending end.

[0036] Step S1. Convert the secret message into a distributed, distortion-resistant message representation.

[0037] First, before embedding message data using this invention, the secret data can be preprocessed to enhance security and robustness, such as by compression, encryption, and error correction coding, to obtain a length of [length missing]. Binary secret message (assuming it is a perfect square) In one embodiment, the secret data may be text data.

[0038] Then, the binary secret message Convert to bipolar sign And reshape it into a two-dimensional matrix .

[0039] Finally, a learnable orthogonal weight matrix is ​​initialized. (in To align the target dimension of the latent space features, a continuous, distortion-resistant message representation is computed via matrix multiplication. This design utilizes the concept of "spread spectrum" to globally distribute each bit of information in a high-dimensional latent space, thereby improving robustness against local distortions.

[0040] In this invention, the matrix is ​​strictly constrained during the model optimization training process. Maintaining the orthogonality property, i.e., satisfying ( (As an identity matrix). This globally disperses message energy into a high-dimensional latent space, ensuring that local channel distortion only causes partial signal degradation, thereby significantly improving robustness.

[0041] Step S2. Based on the pre-trained diffusion model, and combined with the given text prompts and deterministic sampling strategy, generate clean initial latent space features.

[0042] This embodiment is based on a pre-trained Latent Diffusion Model (LDM). It obtains the clean latent space features in the final step before generation through the deterministic sampling process of DDIM based on the text prompt words. Step S3. Based on the initial latent space features and the anti-distortion message representation, generate a latent space perturbation and superimpose the latent space perturbation onto the initial latent space features to obtain the steganalytic latent variables.

[0043] Existing methods that directly modify latent variables are prone to causing statistical feature shifts. This invention uses a hidden network based on residual learning to implement hidden embedding by predicting only minimal latent space perturbations, thus preserving the original prior distribution of the diffusion model to the greatest extent.

[0044] Specifically, given the clean latent space features of the step before generating the final image. The aforementioned anti-distortion message indicates The latent variable hidden embedding strategy based on residual learning includes the following steps: 1) Feature stitching: combining clean latent space features With anti-distortion message Concatenate along the channel dimension. 2) Minimal Perturbation Prediction: The concatenated features are input into a pre-trained hidden network (based on the U-Net architecture with residual blocks and spatial attention mechanisms). The network employs a residual learning mechanism, not directly generating steganalytic latent variables, but instead predicting only the minimal latent spatial perturbations used to carry secret information. ,Right now ; 3) Stamped latent variable synthesis: The predicted minimal latent space perturbation is synthesized. Directly superimposed on the initial latent space features Above, we obtain the steganographic latent variables. .

[0045] This strategy effectively restricts steganalytic modifications to a very small range, ensuring their distribution closely follows the original diffusion latent space prior, thereby significantly improving the ability to resist steganalysis detection.

[0046] Step S4. Decode the steganalytic latent variables to obtain the steganalytic image.

[0047] The above-mentioned latent variables after steganography The image is decoded by a variational autoencoder (VAE) decoder, mapped back to pixel space, and finally obtained as a steganalyte.

[0048] (ii) Lossy channel transmission.

[0049] In practical communication systems, lossy channel transmission is more common, mainly due to physical limitations, the balance between cost and efficiency, and the requirements of the application scenario.

[0050] First, the inherent characteristics of physical channels dictate that signals will inevitably experience attenuation, noise, and interference. For example, multipath effects and electromagnetic interference in wireless channels, or resistive losses and crosstalk in wired channels, can all lead to signal distortion or partial information loss.

[0051] Secondly, completely lossless transmission requires extremely high resource investment. To eliminate all noise and attenuation, high-power transmission, complex coding, and redundant error correction are required, which significantly increases equipment cost, energy consumption, and bandwidth usage, while absolute reliability is not required in most scenarios.

[0052] Furthermore, human senses are highly tolerant of information errors. By compressing media such as images and audio, discarding some data that is not sensitive to the human eye / ear, the amount of data transmitted can be significantly reduced while maintaining subjective quality, thus improving efficiency.

[0053] In another preferred embodiment, the invention is also applicable to lossless channel transmission.

[0054] (III) Extraction process at the receiving end.

[0055] Step P1. Encode the received steganatical image using a VAE encoder to obtain the damaged steganatical latent variables.

[0056] This embodiment uses a VAE encoder to extract the lossy steganographic latent variables from the received steganographic image. .

[0057] Step P2: Based on the pre-trained diffusion model, and combined with the given text prompts and deterministic sampling strategy, generate clean initial latent space features.

[0058] This embodiment utilizes the same text prompts and deterministic sampling strategy at the sending end to reconstruct the features of the unbuffered clean reference latent space.

[0059] Step P3. Denoise the damaged steganalytic latent variables using the hybrid expert denoising module to obtain the repaired steganalytic latent variables.

[0060] During the extraction process, this invention innovatively introduces a Mixture-of-Experts (MoE) denoising architecture from the field of computer vision to perform forward adaptive repair of damaged latent variables. On this basis, an extraction mechanism resistant to complex distortion is established. By dynamically responding to various heterogeneous channel degradations such as compression, ambiguity, and noise, it supports accurate reasoning of continuous message representations and reliable reconstruction of secret information during the extraction process.

[0061] The MoE denoising module is a neural network architecture based on a dynamic routing mechanism in deep learning, used to adaptively process data with diverse degradation distributions. Its design goal is to improve the robustness and generalization ability of the model when faced with unknown mixed distortions. For example... Figure 2 As shown, the MoE module consists of a routing network (ResNet-34 noise classifier), multiple dedicated denoising expert networks (NAFNet), and a sparse gated fusion network.

[0062] Routing Network Evaluation: The MoE module uses a routing network based on a multi-label ResNet-34 classifier. Since actual channel distortions (such as JPEG compression and Gaussian blur) leave specific spatial features and local degradation cues in the latent space, the routing network captures these features to predict a set of probability vectors representing the distortion components for the input impaired latent variables. Denoising Expert Activation: To achieve targeted repair of different types of distortion, MoE includes multiple independent denoising expert networks (such as those based on the NAFNet architecture). Each expert focuses on learning to repair a single specific channel distortion during the pre-training phase.

[0063] Sparse Gated Fusion: After obtaining the probability vector of the routing network, the MoE module performs sparse activation through a set threshold, selects only expert networks that are highly correlated with the current distortion for forward inference, and performs weighted summation of the output features of each activated expert according to the normalized probability weights to generate the latent variables for final fine repair.

[0064] This module is completely independent of the specific steganography embedding rules of the sending end and only operates in the feature recovery stage of damaged latent variables. Therefore, it can be seamlessly integrated as a general component into other latent space-based generative image steganography frameworks.

[0065] Overall, the MoE denoising module provides a flexible yet robust approach to repairing damaged latent spatial features transmitted through heterogeneous channels by using dynamic route evaluation and expert-weighted fusion, resulting in finely repaired latent variables. .

[0066] Step P4. Input the initial latent space features, the repaired steganalytic latent variables, and the difference between the initial latent space features and the repaired steganalytic latent variables into the extraction network to obtain the anti-distortion message representation.

[0067] like Figure 3 As shown on the right, using synchronously generated reference latent variables ,Will , and its explicit difference Common Input Extraction Network Extracting the network Based on the difference information, irrelevant image semantics and redundant distortion are filtered out, and continuous anti-distortion message representations are inferred to predict the values. This multi-channel input strategy utilizes residual inductive bias, enabling the network to more accurately isolate secret signals that have been distorted or corrupted.

[0068] Step P5. Apply an inverse orthogonal transformation matrix to anti-distort the message representation and convert it into binary secret information, then restore the binary secret information to the secret message.

[0069] This step is for Multiplying the transpose of an orthogonal matrix by left and then performing an inverse orthogonal transformation is calculated, i.e. Then, by using a threshold decision, it is reverse-engineered into the original binary bit stream. After the accurate extraction of the original binary secret information is completed, the binary secret information is restored into a secret message.

[0070] In a preferred embodiment of the present invention, in order to ensure that the hidden network, MoE denoising module and extraction network can work effectively together, this embodiment adopts a three-stage progressive joint training scheme: Phase 1 (Basic Reconstruction Optimization): Under lossless ideal channel conditions, the hidden network and the revealed network are jointly trained, and the L2 loss function is used to optimize message recovery accuracy and establish basic encoding and decoding feature mapping.

[0071] The second stage (fine-tuning of concealment and security): Introducing perceptual loss (LPIPS) and diffusion latent spatial distance loss to strictly constrain the perturbations in the hidden network predictions. The range of the steganalytic latent variables is adjusted to ensure that they do not deviate from the original diffusion prior, thereby improving the ability to resist steganalysis detection.

[0072] The third stage (hybrid expert robustness enhancement): freeze the hidden network parameters, and introduce a simulated layer of real channel distortion (such as JPEG compression QF=50~100, Gaussian blur, etc.) into the training pipeline. The focus is on end-to-end joint training of the MoE denoising module, orthogonal weight matrix, and extraction network, enabling it to have adaptive repair and high-precision extraction capabilities for various unknown hybrid distortions.

[0073] As can be seen from the above specific implementation methods, in the method flow of the present invention, the amplitude constraint parameters when the hidden network predicts the minimal latent space disturbance can be adjusted according to the complexity of the target lossy channel and the actual communication requirements, so as to achieve a dynamic balance between steganography security and distortion resistance robustness; the length of the secret information to be embedded can also be flexibly adjusted. Furthermore, the activation threshold of the routing network and the number of expert networks participating in the fusion in the hybrid expert (MoE) denoising module at the receiver are adjusted to achieve a trade-off between steganalysis capacity, extraction accuracy, and algorithm inference time. Therefore, this invention has strong adaptability and flexibility.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. A robust and secure image steganography method based on hybrid expert and latent space optimization, characterized in that, Applied to the sending end, the method includes: Transform secret messages into distributed, distortion-resistant message representations; Based on a pre-trained diffusion model, and combined with given text prompts and a deterministic sampling strategy, clean initial latent space features are generated. Based on the initial latent space features and the anti-distortion message representation, a latent space perturbation is generated and superimposed on the initial latent space features to obtain the stegatical latent variables; Decoding the steganalytic latent variables yields the steganalytic image.

2. The method according to claim 1, characterized in that, Decoding the steganalytic latent variables yields the steganalytic image, including: The steganographic latent variables are decoded using a decoder of a variational autoencoder. The decoding result is mapped back to pixel space to obtain the steganalytic image.

3. The method according to claim 1, characterized in that, Convert the secret message into a two-dimensional message sequence, including: Convert the secret message into binary; The binary secret information is converted into bipolar symbols and reshaped into a two-dimensional matrix; The two-dimensional matrix is ​​mapped to a distributed, distortion-resistant message representation using a learnable orthogonal weight matrix.

4. The method according to claim 3, characterized in that, Based on the initial latent space features and the anti-distortion message representation, latent space perturbations are generated, including: The initial latent space features are combined with the anti-distortion message representation through channel splicing; The splicing result is input into a hidden network based on the U-Net structure to obtain the latent space perturbation.

5. The method according to claim 4, characterized in that, After decoding the steganalytic latent variables to obtain the steganalytic image, the method further includes: The steganographic image is sent to the receiving end so that the receiving end can extract the secret message in the steganographic image by combining the given text prompt words and a deterministic sampling strategy; The receiving end extracts the secret message from the steganalysis image by combining the given text prompts and sampling strategy, including: The received steganatical image is encoded using a VAE encoder to obtain the damaged steganatical latent variables; Based on a pre-trained diffusion model, and combined with given text prompts and a deterministic sampling strategy, clean initial latent space features are generated. The damaged stegana latent variables are denoised using a hybrid expert denoising module to obtain the repaired stegana latent variables; wherein, the hybrid expert denoising module includes a routing network, a denoising expert network and a sparse gating fusion network; The initial latent space features, the repaired steganalytic latent variables, and the difference between the initial latent space features and the repaired steganalytic latent variables are all input into the extraction network to obtain the anti-distortion message representation; The message representation is converted into binary secret information by applying an inverse orthogonal transformation matrix to resist distortion. Restore the binary secret information to the secret message.

6. The method according to claim 5, characterized in that, The damaged steganalytic latent variables are denoised to obtain the repaired steganalytic latent variables, including: Spatial degradation features and local distortion clues of damaged stegana latent variables are extracted by routing network to obtain a set of probability vector sequences representing the composition of distortion; wherein, the k-th component in the probability vector sequence represents the confidence that the damaged stegana latent variable contains the k-th type of distortion; A set of denoising expert networks is obtained based on the components that are greater than the set threshold. Forward inference of damaged stegana latent variables based on a denoising expert network ensemble; The output features of each denoising expert network are weighted and summed based on a sparse gated fusion network to obtain the repaired steganalytic latent variables.

7. The method according to claim 5, characterized in that, The training process for the orthogonal weight matrix, the hidden network, the hybrid expert denoising module, and the extraction network includes: The orthogonal weight matrix and expert denoising module are frozen, and the hidden network and extraction network are jointly trained using the L2 message reconstruction loss function to optimize message recovery accuracy. The orthogonal weight matrix and expert denoising module are frozen, and perceptual loss and diffusion latent space distance loss are introduced to jointly train the hidden network and the extraction network to constrain the magnitude of the latent space perturbation predicted by the hidden network. The hidden network is frozen, and the MoE denoising module, orthogonal weight matrix, and extraction network are jointly trained based on end-to-end message recovery loss, so that the MoE denoising module, orthogonal weight matrix, and extraction network have the ability to adaptively repair and extract various unknown mixed distortions with high precision.

8. A robust and secure image steganography system based on hybrid expert and latent space optimization, characterized in that, The system includes: The message conversion module is used to convert secret messages into a distributed, distortion-resistant message representation. Initial latent variable generation is used to generate clean initial latent space features based on a pre-trained diffusion model and combined with given text prompts and a deterministic sampling strategy. The hidden embedding module is used to generate latent space perturbations based on the initial latent space features and the anti-distortion message representation, and to superimpose the latent space perturbations onto the initial latent space features to obtain the stegatical latent variables; The decoding module is used to decode the steganalytic latent variables to obtain the steganalytic image.

9. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the robust secure image steganography method based on hybrid expert and latent space optimization as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the robust secure image steganography method based on hybrid expert and latent space optimization as described in any one of claims 1-7.

Images