A generative steganography method and system based on diffusion model and semantic prompt

Abstract

The application provides a generative steganography method and system based on a diffusion model and semantic prompt words, relates to the technical field of image steganography, and includes embedding and extraction of secret information. The embedding of secret information is: obtaining secret information to be embedded, converting the secret information to binary and segmenting the binary to obtain a segmented binary sequence; using a pre-constructed semantic prompt word dataset, mapping the binary segments in the sequence into semantic prompt words to form a semantic prompt word sequence, and enhancing the semantic prompt word sequence; inputting the enhanced semantic prompt word sequence into an image generation diffusion model as a condition to guide the model to generate secret images related to the semantics of the prompt words; extracting the latent space representation vector of the secret images, inputting the vector into the image steganography diffusion model after shuffling, and generating a stego image unrelated to the semantics of the prompt words. The application can effectively improve the secret information capacity of a single generated image and improve the overall quality and naturalness of the generated image by the diffusion steganography model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image steganography technology, in particular to a generative steganography method and system based on diffusion model and semantic prompt words. BACKGROUND

[0002] Diffusion models have strong generation capabilities and are widely used in the field of image steganography technology, such as model-driven robust information embedding and high-quality carrier synthesis. Their performance is superior in complex scenarios. In the field of anti-steganalysis, when traditional methods cannot guarantee the reliable extraction of information in high compression or strong interference environments, steganography technology based on diffusion models can still maintain high concealment and recovery accuracy. However, the steganography method based on diffusion models uses semantics and labels as public and private keys to guide image generation and encryption, aiming to use the asymmetric encryption of diffusion model-generated content and semantics and labels to achieve information hiding. The label requirement of diffusion model-generated images results in similar semantics of public and private keys, and potential attackers can deduce the private key from the public key semantics, leading to insufficient security. Therefore, the study of generative steganography methods based on diffusion models can not only fully utilize the high-quality image generation capability of diffusion models, but also achieve traceless embedding of information, significantly reducing the risk of detection by steganalysis tools.

[0003] The generative image steganography method based on diffusion models uses diffusion models to synthesize stego carriers, avoiding modification traces from the source to improve concealment. However, existing generative steganography methods based on diffusion models cannot embed binary hash sequences or text to realize secret information transmission like traditional models (such as generative adversarial networks), resulting in the use of a mechanism that maps a single image and a binary hash sequence to deliver secret information in the generative steganography method based on diffusion models. The encryption capacity is extremely small, and a large amount of resources is required to transmit and store single images that have encoded secret information during communication. Therefore, the existing generative steganography method is not practical, and it is unrealistic to achieve large-scale information transmission.

[0004] At the same time, the existing steganography framework based on diffusion models directly uses the latent space representation vector of the secret image as input when processing images, resulting in the generated stego image containing the potential space features of the secret image, which leads to poor image quality and low naturalness. The generated stego image is easily noticed by attackers during transmission, and has low security and concealment.

[0005] Therefore, the existing generative steganography method based on diffusion models has deficiencies in encryption capacity, resource consumption, practicality, generation quality, security, and concealment. SUMMARY

[0006] The application proposes a generative steganography method and system based on a diffusion model and semantic prompt words to solve the above problems, which can effectively improve the secret information capacity of a single generated image and improve the overall quality and naturalness of the generated image.

[0007] According to some embodiments, the application adopts the following technical solutions:

[0008] A generative steganography method based on a diffusion model and semantic prompt words includes embedding and extracting secret information, and the embedding of the secret information is:

[0009] Obtaining secret information to be embedded, converting it into binary and segmenting it to obtain a segmented binary sequence;

[0010] Using a pre-constructed semantic prompt word dataset, mapping the binary segments in the sequence into semantic prompt words to form a semantic prompt word sequence, and enhancing the semantic prompt word sequence;

[0011] Inputting the enhanced semantic prompt word sequence into an image generation diffusion model as a condition to guide the model to generate a secret image related to the semantic prompt word;

[0012] Extracting the latent space representation vector of the secret image, inputting it into an image steganography diffusion model after shuffling to generate a stego image unrelated to the semantic prompt word.

[0013] According to some embodiments, the application adopts the following technical solutions:

[0014] A generative steganography system based on a diffusion model and semantic prompt words includes:

[0015] The acquisition module is configured to obtain secret information to be embedded, convert it into binary and segment it to obtain a segmented binary sequence;

[0016] The mapping module is configured to use a pre-constructed semantic prompt word dataset to map the binary segments in the sequence into semantic prompt words to form a semantic prompt word sequence, and enhance the semantic prompt word sequence;

[0017] The generation module is configured to input the enhanced semantic prompt word sequence into an image generation diffusion model as a condition to guide the model to generate a secret image related to the semantic prompt word;

[0018] The steganography module is configured to extract the latent space representation vector of the secret image, input it into an image steganography diffusion model after shuffling to generate a stego image unrelated to the semantic prompt word.

[0019] According to some embodiments, the present application adopts the technical scheme as follows:

[0020] A computer program product comprising a computer program which, when executed by a processor, implements the diffusion model and semantic prompt word-based generative steganography method.

[0021] According to some embodiments, the present application adopts the technical scheme as follows:

[0022] A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the diffusion model and semantic prompt word-based generative steganography method.

[0023] According to some embodiments, the present application adopts the technical scheme as follows:

[0024] An electronic device comprising a processor, a memory, and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the diffusion model and semantic prompt word-based generative steganography method.

[0025] Compared with the prior art, the present application has the beneficial effects that:

[0026] The present application provides a diffusion model and semantic prompt word-based generative steganography method and system, which is different from the method of using a single image and a binary hash sequence to form a mapping to transmit secret information when mapping information to a produced secret-containing carrier using a diffusion model in the traditional method, but mapping secret information into semantic prompts, guiding the high-quality images generated by the semantic prompts, and through the change of this mapping method, the secret information hidden in the generated single image is greatly improved, reducing the transmission and storage cost of secret information, and through the secondary processing of the generated image based on the diffusion model of the diffusion steganography framework, the semantic association degree of the image used for covert communication and the semantic prompt is further reduced. The concealment of communication and the security of the steganography system are improved, and at the same time, by combining the advantages of the diffusion model and the prompt, the generated image not only has high authenticity and controllability, but also can resist common image processing attacks such as compression, cropping, rotation, etc. The image generated by the diffusion model has high diversity and complexity in details, making it difficult to detect and extract the hidden secret information.

[0027] This invention provides a generative steganography method and system based on a diffusion model and semantic cue words, which improves upon the traditional diffusion model-based diffusion steganography framework. Unlike the traditional diffusion model steganography framework, which directly reconstructs the secret image from the latent space representation vector of the secret image, resulting in the generated secret image containing some features of the latent space of the secret image, this invention scrambles the latent space representation vector of the secret image before the forward operation of reconstructing the secret image. The new latent space representation vector obtained by this method can generate new images with high controllability and diversity when used for diffusion steganography, and can also reconstruct the secret image more efficiently and simply. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a schematic diagram of the method flow in Example 1;

[0030] Figure 2 This is a flowchart illustrating the embedding of secret information in Example 1;

[0031] Figure 3 This is an example diagram of the mapping group in Example 1;

[0032] Figure 4 This is a flowchart of the secret information extraction process in Example 1;

[0033] Figure 5 This is a flowchart of the diffusion steganography framework in Example 1. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] Example 1

[0038] In an embodiment of the present application, a generative steganography method based on diffusion model and semantic prompt word is provided, including embedding and extracting of secret information, the embedding of secret information is:

[0039] Step one: obtain the secret information to be embedded, convert it into binary and segment it to obtain the segmented binary sequence;

[0040] Step two: using the pre-constructed semantic prompt word dataset, map the binary segments in the sequence into semantic prompt words to form a semantic prompt word sequence, and enhance the semantic prompt word sequence;

[0041] Step three: input the enhanced semantic prompt word sequence as a condition into the image generation diffusion model to guide the model to generate secret images related to the semantics of the prompt words;

[0042] Step four: extract the latent space representation vector of the secret image, and input it into the image steganography diffusion model after shuffling to generate a stego image unrelated to the semantics of the prompt words.

[0043] As an embodiment, the secret information (assuming the text information "The attack starts at ten o'clock" is the secret information to be steganographed) is hidden and covert communication is carried out using the generative steganography method based on diffusion model and semantic prompt word of the present application, as shown in Figure 1 The specific implementation process is as follows:

[0044] I. Secret information embedding, as shown in Figure 2

[0045] Step S1, construct a lightweight semantic prompt word dataset to obtain the mapping of semantic prompt words to binary segments, which is used to map secret information to a semantic prompt word sequence;

[0046] A lightweight semantic prompt word dataset is constructed, which contains five mapping groups of prompt words-binary segments. The first mapping group is a background class prompt word, which is used to control the background content of the generated image; the second mapping group is a subject class prompt word, which is used to control the subject of the generated image; the third, fourth and fifth mapping groups are state prompt word 1, state prompt word 2 and state prompt word 3, which are used to control the state of the subject in the generated image. In the mapping group, each semantic prompt word corresponds to an 8-bit binary number (i.e. a binary segment), wherein the first mapping group corresponds to the first binary segment, i.e. the semantic prompt word corresponding to the first binary segment is obtained from the first mapping group, the second mapping group corresponds to the second binary segment, the third mapping group corresponds to the third binary segment, the fourth mapping group corresponds to the fourth binary segment, and the fifth mapping group corresponds to the fifth binary segment.

[0047] ​Step S2, after preprocessing the secret information to get the segmented binary sequence, map it to the semantic cue sequence, and appropriately enhance it. The enhancement here includes random shuffling and confusing cues. Specifically:

[0048] Step S2.1, the text information "The attack starts at ten o'clock" is converted to binary numbers through ASCII encoding, and the binary numbers are segmented. Each data segment is 8 bits. If the length of the binary number sequence is not a multiple of 8, 0 needs to be added until it can be divided by 8.

[0049] Step S2.2, the segmented binary numbers are randomly shuffled to disrupt the order of the binary number segments, and the shuffle key corresponding to the correct binary number segment order is recorded, which is assumed to be "52341" to restore the secret information later.

[0050] Step S2.3, the shuffled binary number sequence is mapped to the semantic cue data set, as shown in Figure 3 The first binary segment "00110101" corresponds to the cue "towering tree" in the first mapping group, the second binary segment "00110100" corresponds to the cue "Early morning mist" in the second mapping group, the third binary segment "00001111" corresponds to the cue "snow mountain" in the third mapping group, the fourth binary segment "00111100" corresponds to the cue "Calm lake surface" in the fourth mapping group, and the fifth binary segment "00000100" corresponds to the cue "blue sky" in the fifth mapping group. The mapping of the secret binary number sequence to the semantic cue sequence is obtained.

[0051] Step S2.4, based on the constructed semantic cue data set, introduce confusing cues to improve the complexity of the generated image, including two ways:

[0052] 1) Introduce randomly selected semantic cues in the above data set that are not mapped to binary numbers as confusing cues. After embedding the semantic cue sequence, combine a new semantic cue sequence.

[0053] 2) Introduce randomly selected cues outside the above data set as confusing cues, which can be embedded in the semantic cue sequence to combine a new semantic cue sequence.

[0054] Assuming "sunlight exposure" and "the Tyndall effect" are embedded as confusing cues in the cue sequence, the final enhanced semantic cue sequence is obtained.

[0055] Step S3, input the enhanced semantic prompt word sequence (for example, "Early morning mist, Early morning mist, snow mountain, Calm lake surface, blue sky, sunlight exposure, the Tyndall effect") as a condition into an image generation diffusion model (DDPM Mode), and place the subject class prompt words in the front part of the above prompt word sequence to maximize the image generation weight. The weight of each group of prompt words can be adjusted by modifying the parameters to ensure that the generated image has high quality and high naturalness, and at the same time guide the model to generate a secret image with high semantic relevance to the prompt words.

[0056] Step S4, based on the image output by the diffusion model, use the diffusion steganography framework to process the image to generate a stego image with low semantic relevance to the semantic prompt words for transmission.

[0057] Specifically, the secret image generated by the diffusion model is input into the diffusion steganography framework for processing, as shown in FIG. 2, in this example, the secret image is input into the image steganography diffusion model, and a stego image with high quality and high naturalness with low semantic relevance to the semantic prompt words is generated through the forward process. Figure 5

[0058] The proposed steganography diffusion model is improved compared to the traditional one: before the forward process, the secret image is preprocessed, the latent space representation vector of the secret image is extracted using the variational autoencoder (VAE), and the latent space representation vector is shuffled using a predefined shuffling sequence to ensure consistent deployment at the sender and receiver, eliminating the need for additional communication. The shuffling operation disrupts the feature space of the latent space representation vector, and the shuffled latent space representation vector is input into the image steganography diffusion model DDIM (Denoising Diffusion Implicit Models). After iteration processing through the forward process of the image steganography diffusion model, a stego image is generated. This improved method can generate a stego image with high controllability and diversity, while having low semantic similarity to the original image.

[0059] The forward process of DDIM aims to generate x T ∈ R D generate x 0∈ R D In the above diffusion steganography framework, the shuffled latent space representation vector is used as the initial noise x T stego image x ​0 represents the image to be generated. x Sampling to 0 can be achieved by solving the diffusion ordinary differential equation (ODE), which is a probability flow ODE derived from the diffusion process. Solving this ODE can be done from the noise. x T Deterministic recovery x 0, the equation is expressed as follows:

[0060] (1)

[0061] in, For continuous time variables, =0 represents the data distribution, i.e., the density-carrying image to be generated. =T represents the noise distribution, i.e., the initial noise. , The drift coefficient defines the evolution of the ODE dynamical system without a diffusion term. The diffusion coefficient controls the change in noise intensity over time and determines the contribution of the score function to the dynamic system. , These are noise scheduling parameters, which are predefined in the diffusion model. It is usually related to the original signal that is preserved. The lower the value, the less signal there is. It is related to the amount of noise added, and Follow Strictly decreasing; This is a forward noise prediction model, and its input is noise. and time Its goal is to predict x0. It is the core of the entire system, guiding the generation process.

[0062] Given time s Initial value at >0 DDIM used an exponential integrator to derive the solution of the diffusion ODE (Equation (1)) at time t. :

[0063]

[0064] in, It is the logarithm of the signal-to-noise ratio. The variable form represents the state x when the logarithm of the signal-to-noise ratio is λ. This is a variable substitution that allows the integration to be performed in the signal-to-noise ratio space. yes The inverse function of .

[0065] Utilize at point The DDIM diffusion model approximates the exact solution at time under the condition that the value of the DDIM diffusion model at time is given, to obtain the approximate solution at time :

[0066] (3)

[0067] represents the span of adjacent time steps in space, since the third term of formula (3) can be calculated analytically and omitted (represented by ), the only requirement is to find at time

[0068] The simplest approximation occurs when = 1, in which case:

[0069]

[0070] where This is an extrapolation term that scales the state at the last time step by the proportion of noise in the plan and directly projects it to the current time step; This is a noise correction term that corrects the state obtained by extrapolation above according to the direction and amplitude of the clean data predicted by the network at the last time step, so that it is closer to the real data distribution; is a factor, since = <0 (the signal-to-noise ratio is decreasing), so the factor is negative, and is positive, which constitutes a positive correction as a whole.

[0071] Similar to LDM (latent diffusion model), once the scrambled hidden space representation vector is obtained, it is input into DDIM as noise, and in addition, the stego prompt word selected by the user is input into DDIM as a condition to control the content of the generated stego image, and the prompt word is public in the stego image transmission process. The algorithm shown in Table 1 describes the generation process of the stego image in the above diffusion steganography framework:

[0072] Table 1 Stego image generation algorithm

[0073]

[0074] II. Secret information extraction, as shown in Figure 4 :

[0075] Step S5: The user impersonates an ordinary user of the AIGC community and transmits the secret image and shuffling key. The receiver uses the same input model and parameter settings to recover the secret image through the reverse process.

[0076] Specifically, such as Figure 5 As shown, the above inverse process includes the DDIM model continuously inverting the secret image through its inverse process to obtain the scrambled latent space representation vector, and then using the scrambled sequence of the latent space representation vector to recover the original latent space representation vector, thereby recovering the secret image.

[0077] The first and most crucial step in extracting the secret image is the inversion of DDIM. Each step in the DDIM inversion starts from the next time step. status To make a preliminary estimate of the previous time step status It can be expressed in the direct inverse form of formula (4) as follows:

[0078] (5)

[0079] in, In time and The noise standard deviation is a predefined noise scheduling parameter. In time The signal coefficients are also predefined scheduling parameters; For the inversion noise prediction model, the input is the current state. and time step Its goal is to predict in Noise at any given moment.

[0080] In the above inversion process, in order to obtain accurate inversion results... The post-Euler method, an iterative correction algorithm, is employed. Its idea is: if the estimated... If it is accurate, then substitute it into the forward generation formula, i.e., formula (4), with the goal of perfectly reproducing it. If it cannot be perfectly reproduced Then correct Specifically:

[0081] Inversion calculation :

[0082] (6)

[0083] Generate forward using formula (4) :

[0084]

[0085] Update using forward stepping:

[0086] (8)

[0087] Repeat formulas (7) and (8) until convergence, and the accurate .

[0088] where, is the noise predicted by the inverse noise prediction model at time and state , is the estimate that needs to be corrected, in this embodiment, the , is the updated, more accurate estimate calculated from the error, is the error term, the difference between the forward generated result and the known true value ; is the learning rate, a hyperparameter, controls the magnitude of each correction, needs to be carefully selected to ensure the algorithm converges stably.

[0089] The algorithm shown in Table 2 describes the reverse inversion process of the above diffusion model steganography framework:

[0090] Table 2 DDIM reverse inversion algorithm

[0091]

[0092] After the above process, the secret image with the shuffled latent space representation vector is obtained ; using the predefined shuffling sequence on the shuffled latent space representation vector , the latent space representation vector z is recovered, and finally, the latent space representation vector z is sent into the VAE decoder to reconstruct the original secret image , and the algorithm in Table 3 describes the process of extracting the original secret image in the above diffusion steganography framework:

[0093] Table 3 Secret image extraction algorithm

[0094]

[0095] Step S6, based on the recovered secret image, extract its semantic features, recover the secret binary sequence with the help of the semantic prompt word dataset, and further recover the secret information.

[0096] Specifically, in the original recovery, the semantic features in the image are extracted by means of the open source large language model, and are used as the enhanced semantic prompt word sequence (assuming Early morning mist, a towering tree grow on the snowmountain besides the Calm lake surface, With the blue sky as the backdrop), the semantic prompt word sequence “Early morning mist, towering tree, snowmountain, Calm lake surface, blue sky” is obtained after removing the confusion prompt word, the binary sequence is recovered by means of the semantic prompt word data set, the secret binary sequence “00000100 00110100 00001111 0011110000110101” is recovered by means of the shuffle key “52341”, and then the secret text information “The attack starts at ten o’clock” is recovered.

[0097] As another implementation, the method proposed in the embodiment is not only suitable for text transmission, but also suitable for symbol transmission and digital transmission, or transmission of secret images, and only the improved diffusion steganography framework is used; since only secret images are transmitted, it is simpler and more accurate, in order to improve the performance and security of the model, the embodiment is actually still mainly focused on the transmission of the secondary encrypted secret image.

[0098] Embodiment 2

[0099] In an embodiment of the present application, a generative steganography system based on a diffusion model and semantic prompt words is provided, comprising:

[0100] The acquisition module is configured to acquire secret information to be embedded, convert it into binary and segment it to obtain a segmented binary sequence;

[0101] The mapping module is configured to map the binary segments in the sequence into semantic prompt words by using a pre-constructed semantic prompt word data set, to form a semantic prompt word sequence, and to enhance the semantic prompt word sequence;

[0102] The generation module is configured to input the enhanced semantic prompt word sequence into an image generation diffusion model as a condition, to guide the model to generate a secret image related to the semantic prompt words;

[0103] The steganography module is configured to extract the latent space representation vector of the secret image, and input it into an image steganography diffusion model after shuffling, to generate a stego image unrelated to the semantic prompt words.

[0104] Embodiment 3

[0105] In an embodiment of the present application, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the diffusion model and semantic prompt word based generative steganography method.

[0106] Embodiment 4

[0107] In an embodiment of the present application, a non-transitory computer readable storage medium is provided, for storing computer instructions which, when executed by a processor, implement the diffusion model and semantic prompt word based generative steganography method.

[0108] Embodiment 5

[0109] In an embodiment of the present application, an electronic device is provided, comprising a processor, a memory, and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device implements the diffusion model and semantic prompt word based generative steganography method.

[0110] The present application is described in reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flow(s) or block(s).

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed by the computer or other programmable device provide a process for implementing the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 a step that performs the functions specified in the flow(s) or block(s).

[0112] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A generative steganography method based on a diffusion model and semantic cue words, characterized in that, This includes the embedding and extraction of secret information, wherein the embedding of the secret information is as follows: Obtain the secret information to be embedded, convert it into binary and segment it to obtain the segmented binary sequence; Using a pre-built semantic cue word dataset, binary segments in a sequence are mapped to semantic cue words to form a semantic cue word sequence, and the semantic cue word sequence is enhanced. The semantic cue word dataset includes a one-to-one mapping relationship between binary segments and semantic cue words. The enhanced semantic cue word sequence is used as a condition in the image generation diffusion model to guide the model to generate secret images that are semantically related to the cue words. Extract the latent space representation vector of the secret image, scramble it, and input it into the image steganalysis diffusion model to generate a secret image that is semantically unrelated to the clue words; The enhancements include random shuffling and obfuscating prompts; The random shuffling involves scrambling the order of the binary bits in the binary sequence and generating a shuffling key based on the correct order. The obfuscation prompts are selected from prompts that do not appear in the semantic prompt sequence and added to the semantic prompt sequence; The input to the image steganalysis diffusion model is a scrambled latent space representation vector and a preset public key hint. The calculation steps for the scrambled latent space representation vector are as follows: Extract the latent space representation vector of the secret image using a variational autoencoder (VAE); The latent space representation vector is scrambled using a predefined scrambling sequence; The extraction of the secret information is the reverse process of the embedding of secret information, and the specific steps are as follows: Based on the secret image and the shuffled key, the secret image is reconstructed through the inverse inversion method of the image steganalysis diffusion model, scrambling recovery and VAE decoder; Semantic cue word sequences are extracted from the secret image, and the binary data is recovered using the semantic cue word dataset, ultimately revealing the secret information.

2. The generative steganography method based on a diffusion model and semantic cue words as described in claim 1, characterized in that, The semantic prompt word dataset includes several mapping groups, each corresponding to a different position in a binary sequence.

3. The generative steganography method based on a diffusion model and semantic cue words as described in claim 1, characterized in that, It also includes the sender sending the shuffling key and the final generated encrypted image to the receiver, who then recovers the secret information by extracting the secret information.

4. A generative steganography system based on a diffusion model and semantic cue words, characterized in that, The generative steganography method based on a diffusion model and semantic cue words as described in any one of claims 1-3 includes: The acquisition module is configured to: acquire the secret information to be embedded, convert it into binary and segment it to obtain the segmented binary sequence; The mapping module is configured to: use a pre-built semantic cue word dataset to map binary segments in the sequence into semantic cue words, form a semantic cue word sequence, and enhance the semantic cue word sequence; The generation module is configured to take the enhanced semantic cue word sequence as a condition and input it into the image generation diffusion model to guide the model to generate secret images that are semantically related to the cue words. The steganography module is configured to extract the latent space representation vector of the secret image, scramble it, and input it into the image steganography diffusion model to generate a secret image that is semantically unrelated to the hint words.

5. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the generative steganography method based on a diffusion model and semantic prompt words as described in any one of claims 1-3.

6. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a generative steganography method based on a diffusion model and semantic prompt words as described in any one of claims 1-3.

7. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a generative steganography method based on a diffusion model and semantic prompt words as described in any one of claims 1-3.

Images