Generative implicit representation steganography method based on message mapping

Abstract

The invention relates to a message mapping-based generative implicit representation steganography method, which comprises the following steps of: representing multimedia data by using an implicit neural representation technology, and converting the multimedia data into uniform point cloud data for message embedding; a secret message needing to be embedded is expressed in a binary mode and then converted into noise data through mapping, the noise data are input into a function generator to be subjected to message embedding, a message extractor is used for extracting the noise data, and the secret-containing message is recovered through inverse mapping; the function generator and the message extractor are obtained through training of a training set; in a message embedding process, a noise space of the function generator applies a single-bit and multi-bit message mapping scheme to embed noise data; in a noise extraction process, a message extractor extracts noise data by setting a plurality of point convolution layers and a full connection layer. According to the method, multimedia data of different sizes can be received, the robustness is good, the message extraction accuracy reaches the baseline level, and the method also has the advantage in the quality of the generated secret-containing carrier.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a generative implicit representation steganography method based on message mapping. Background Art

[0002] Steganography is an information hiding technique used in covert communication, aiming to transmit secret messages over public channels without detection. Current research focuses on deep neural network-based steganography, which often employs a codec architecture. The concept of generative steganography has been proposed, using generative models as the codec. Unlike traditional embedded steganography, generative steganography consists of a generator and a message extractor. Instead of requiring the original carrier, the secret message is directly driven to generate the secret carrier. This approach significantly improves security against steganalysis.

[0003] At present, generative steganography can be divided into three categories according to different structures. The first category is generative steganography based on message mapping, such as Figure 1 (a) shows that the secret message is mapped to the noise vector space through a mapping scheme, and the mapped noise is used to generate the secret carrier through the generator. This method is also the most commonly used method in generative steganography. The second type is generative steganography based on synthesis, such as Figure 1 (b) As shown. The secret message and random noise are used as input and encoded in the generator to generate a secret carrier. Among them, the generator usually uses a generative adversarial network as a generation model. Diffusion model, flow model, VAE and other generation models are also used as generators. The third type is generative steganography based on a fixed message extractor, such as Figure 1 (c) This type of method no longer requires transferring the message extractor. Instead, the extractor is fixed, shifting the training process to other aspects. Specifically, there are two types: one, shown by the black line, embeds the secret message by modifying the noise; the other, shown by the blue line, directly modifies the secret carrier by adding small perturbations to the secret carrier.

[0004] However, all of these methods share a common problem: they all represent data in a grid-based manner, which limits the types and sizes of data that can be used within the scheme. This grid-based representation also strongly couples the model size of the generator and message extractor to the data size, causing model parameters to increase exponentially with increasing resolution. Furthermore, both the generator and message extractor can only process data of a single type and size, significantly limiting the flexibility and versatility of the steganographic scheme.

[0005] A new implicit neural representation technology has emerged to address the issue of monotonous data representation, becoming a significant trend in data representation development. Implicit neural representation is a representation of a continuous function that takes grid coordinates as input and returns features. Inspired by INR, existing technical solutions attempt to incorporate implicit neural representation technology into steganography schemes. Using implicit neural representation technology expands the data types used for steganography, addressing the monotonous nature of steganography schemes. However, these techniques merely modify the data representation and still only modify the parameter space, thus failing to apply them to generative steganography. Summary of the Invention

[0006] To address the problems of the prior art described above, the present invention proposes a generative implicit representation steganography method based on message mapping. Based on message mapping, a generative steganography scheme is constructed that simultaneously addresses the versatility of both the generator and the extractor. Considering the use of implicit neural representations, a universal and carrier-independent generator is constructed in generative steganography. Message mapping is then performed in the function domain, allowing the mapped noise to generate a variety of data types and formats. Finally, a universal message extractor is designed to match the generator and process various multimedia data. Furthermore, the present invention takes into account practical deployment and real-world applications, requiring efficiency and robustness.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A generative implicit representation steganography method based on message mapping, comprising:

[0009] Use implicit neural representation technology to represent multimedia data and convert it into unified point cloud data for message embedding;

[0010] The secret message to be embedded is represented in binary and converted into noise data through mapping. The noise data is input into a function generator for message embedding. The noise data is extracted by a message extractor, and the secret message is recovered through inverse mapping. The function generator and the message extractor are both trained using a training set; the training set includes: a generated point cloud and a real point cloud;

[0011] During the message embedding process, the noise space of the function generator applies single-bit and multi-bit message mapping schemes to embed the noise data;

[0012] During the noise extraction process, the message extractor extracts the noise data by setting multiple point convolution layers and fully connected layers.

[0013] Optionally, obtaining the encrypted point cloud data includes:

[0014] Collecting the multimedia data, performing pixel-wise discrete representation on the multimedia data, and obtaining grid data;

[0015] The grid data is implicitly represented and converted into coordinates and eigenvalues ​​represented in the form of set pairs, which is the dense point cloud data.

[0016] Optionally, extracting the noise data using a message extractor includes:

[0017] Inputting the noise data into a function generator for message embedding to obtain a continuous function;

[0018] According to the target task, corresponding point cloud data is sampled in the continuous function, the corresponding point cloud data is input into the message extractor, and the noise data is extracted.

[0019] Optionally, applying the single-bit and multi-bit message mapping scheme to embed the encrypted point cloud data includes:

[0020] During the message embedding process, when the maximum information hiding capacity of the noise data is less than a target threshold, a single-bit message mapping scheme is applied to map the noise data to a preset value; when the maximum information hiding capacity of the noise data is greater than the target threshold, the noise data is grouped and mapped to noise values ​​in different ranges.

[0021] Optionally, the message extractor includes: a plurality of point convolutional layers and fully connected layers;

[0022] The point convolution layer is used to extract high-dimensional features by downsampling and increasing feature dimensions layer by layer, while retaining local information and obtaining output features;

[0023] The fully connected layer is used to flatten the output features and reduce the dimensions layer by layer to obtain a feature vector of the target dimension.

[0024] Optionally, training the function generator using the training set includes:

[0025] In the process of training the generator using the training set, a discriminator is introduced, the generated point cloud and the real point cloud are input into the discriminator, the function generator is trained through an adversarial game process, and a regularized equivalent penalty loss function for point cloud data is introduced to penalize the gradient of the discriminator.

[0026] Optionally, the penalty loss function includes:

[0027]

[0028] Among them, D(s) is the loss of the discriminator, v ij is the eigenvalue, Eigenvalues of different coordinates.

[0029] The beneficial effects of the present application are:

[0030] The present application can realize secret message mapping at the noise level of the function domain, introduce it into generative steganography, and construct implicit representation generative steganography (GIS) based on secret message mapping:

[0031] Based on the characteristics of message mapping-based generative steganography, the strong coupling relationship between model size and data size is eliminated by introducing a function generator, which can process various types of multimedia data. At the same time, by designing a message extractor specially for point cloud data, the model size of the message extractor is not limited by the data resolution, and can receive multimedia data of different sizes, has good robustness, the message extraction accuracy reaches the baseline level, and the quality of the generated secret-containing carrier also has an advantage. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0033] Figure 1 is a generative steganography flowchart of the background art of the present application; Figure 1 (a) is a generative steganography flowchart based on message mapping, Figure 1 (b) is a generative steganography flowchart based on synthesis, Figure 1 (c) is a generative steganography flowchart based on a fixed message extractor;

[0034] Figure 2 is a generative implicit representation steganography method flowchart based on message mapping of an embodiment of the present application;

[0035] Figure 3 is a data representation form schematic diagram of an embodiment of the present application: Figure 3 (a) is grid data, Figure 3 (b) is point cloud data, Figure 3 (c) is function data;

[0036] Figure 4 is a multi-bit mapping scheme schematic diagram of an embodiment of the present application;

[0037] Figure 5 is a function generator neural network structure schematic diagram of an embodiment of the present application;

[0038] Figure 6 Traditional convolution and point cloud convolution schematic diagram for embodiments of the application;

[0039] Figure 7 Point cloud message extractor structure schematic diagram for embodiments of the application;

[0040] Figure 8 Visual security contrast diagram for embodiments of the application

[0041] Figure 9 Unstructured pruned visual image for embodiments of the application;

[0042] Figure 10 Unstructured pruned visual image for embodiments of the application;

[0043] Figure 11 Generation and extraction efficiency schematic diagram for embodiments of the application;

[0044] Figure 12 Super-resolution sampling schematic diagram for embodiments of the application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0046] In order to make the above-mentioned purposes, features and advantages of the application more apparent and easy to understand, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0047] The embodiment proposes a message mapping-based generative implicit representation steganography method, including: on the basis of the message mapping-based generative steganography scheme, replacing the traditional generator with a function generator, and designing a message mapping scheme in the noise space of the function generator. Meanwhile, a message extractor specially designed for point cloud data is designed. In the whole process, point cloud data is used as the actual processing object, and the purpose of directly generating a secret-containing carrier driven by a secret message is achieved:

[0048] 1. The function generator is used for the first time in the message mapping-based generative steganography, solving the problem that the size of the generator model increases with the increase of the data size, and being able to represent different types of multimedia data, breaking through the resolution limit in the traditional grid data.

[0049] 2. The single-bit and multi-bit message mapping schemes are respectively designed in the noise space of the function generator, and the embedding of the message is completed at the noise level of the function generator.

[0050] 3. Design a message extractor specifically for point cloud data, avoiding the strong coupling relationship between the extractor size and the data size, and improving the versatility of the extractor.

[0051] The goal of the present invention is to construct a universal generative steganography scheme based on message mapping technology. A function generator is used as the generative model of the steganography scheme in the scheme to eliminate the strong coupling between the generative model and the data size and expand the carrier types used for steganography. At the same time, single-bit and multi-bit mapping schemes are designed in the noise space of the function domain, and the message mapping process is embedded in the noise level of the function generator. In order to match the generator, a message extractor specifically for point cloud data is designed to extract noise from the secret point cloud data and restore the secret message through inverse mapping. The advantage of this method is that the message extractor can receive data of different types and sizes. By using a function generator and a message extractor specifically for point cloud data, the steganography scheme unifies the data format, can efficiently generate secret carriers, and construct a simple but universal generative steganography scheme.

[0052] like Figure 2 As shown, this embodiment discloses a generative implicit representation steganography method based on message mapping, including: using implicit neural representation technology to represent multimedia data, converting it into unified point cloud data for message embedding; representing the secret message to be embedded in binary and then converting it into noise data through mapping, inputting the noise data into a function generator for message embedding, using a message extractor to extract the noise data, and recovering the secret message through inverse mapping; both the function generator and the message extractor are trained using a training set; the training set includes: generated point cloud and real point cloud; in the process of message embedding, the noise space of the function generator applies single-bit and multi-bit message mapping schemes to embed the noise data; in the process of noise extraction, the message extractor extracts the noise data by setting multiple point convolution layers and fully connected layers.

[0053] Specifically, this embodiment discloses a generative implicit representation steganography method based on message mapping. The method includes: first, mapping a secret message into noise, then inputting it into a function generator to generate a continuous function, from which point cloud data is sampled for transmission as needed. Upon receiving the point cloud data sampled by the function, the receiver uses a message extractor specifically designed for point cloud data to extract the noise and recover the secret message through inverse mapping. In this scheme, the present invention uses point cloud data as the actual stegoscopy carrier. Unlike mesh data and function data, point cloud data, as a common form of implicit neural representation, can be arbitrarily converted to either type of data, making the scheme more versatile for different types of multimedia data. Furthermore, the present invention uses a function generator instead of a traditional mesh generator as the generative model, advantageously eliminating the coupling between the generator model size and resolution. The point cloud data sampled by the generated function can be visualized as various multimedia objects, such as 2D images, flow data, and 3D data, improving the model's versatility. Secondly, different message mapping schemes are designed within the noise space of the function generative model. The projected noise is used as input to the generator, which is then driven by the noise to directly generate the secret carrier. Finally, in order to enable the message extractor to process different types of data like the generator, a message extractor specifically for point cloud data is designed, and the accuracy of message extraction is improved through training.

[0054] Furthermore, obtaining dense point cloud data includes: collecting multimedia data, pixelating and discretely representing the multimedia data to obtain grid data; implicitly representing the grid data, and converting the grid data into coordinates and eigenvalues ​​represented in the form of set pairs, which is the dense point cloud data.

[0055] Specifically, most of the current multimedia data are represented explicitly, for example, images are represented by grid data. Grid data refers to the pixelated discrete representation of multimedia data. Its disadvantage is that the data size is closely related to the resolution. Another form of data representation is implicit representation, which mainly uses implicit neural representation technology to parameterize multimedia data into continuous function data. This method gets rid of the strong coupling relationship between data size and resolution. Usually when implicitly representing multimedia data, the grid data must first be converted into set representation data. This data represented in the form of set pairs is called point cloud data. Point cloud data is a set pair of coordinates and eigenvalues, which is a data representation between explicit data and implicit data. The image represented by grid data is composed of discrete pixel points. Assuming that each pixel point is (x i ,y j ), then the image can be represented as like Figure 3 (a) is shown. This is an explicit representation method. In order to perform parameterized representation, the present invention first converts the explicit data into a set representation of point cloud data. Assume that ui,j =(x i ,y j ) is the coordinate value, v ij =(r,g,b) is the eigenvalue, a set consisting of coordinates and features It is the point cloud formal representation of image I, such as Figure 3 (b) shows that the point cloud data is then used to construct the function data according to the implicit input-output relationship. The function data f(θ) is determined by the parameter θ, I = f (θ) :R 2 →R 3 , given the pixel position, get the RGB coordinate value f (θ) (x,y)=(r,g,b), such as Figure 3 (c) shows that r is the red color among the three primary colors, g is the green color among the three primary colors, b is the blue color among the three primary colors, n is the number of point clouds, f(θ) is the implicit representation of the data, that is, the function, θ is the function parameter, R 2 is a two-dimensional space, R 3 For three-dimensional space.

[0056] The above three are three different representations of an image. Point cloud data serves as a bridge between implicit and explicit data. It can be obtained by sampling from function data, or it can be visualized as grid data through conversion. Combining the characteristics of the above three types of data, the present invention represents the data implicitly and uses a generative model based on implicit neural representation as a generator to solve the problem that the model size is closely related to the resolution of the grid data. In order to make the message extractor universal, the present invention designs a message extractor based on point cloud data. In this way, the steganographic data type is expanded, and the versatility of the generator and message extractor is improved.

[0057] Furthermore, extracting noise data using a message extractor includes: inputting the noise data into a function generator for message embedding to obtain a continuous function; sampling corresponding point cloud data in the continuous function according to a target task, inputting the corresponding point cloud data into a message extractor, and extracting noise data.

[0058] Furthermore, the application of single-bit and multi-bit message mapping schemes to embed encrypted point cloud data includes: during the message embedding process, when the maximum information hiding capacity of the noise data is less than the target threshold, the single-bit message mapping scheme is applied to map the noise data to a preset value; when the maximum information hiding capacity of the noise data is greater than the target threshold, the noise data is grouped and mapped to noise values ​​in different ranges.

[0059] Specifically, the present invention designs a message mapping scheme in the function domain, and designs a multi-bit mapping scheme based on a single-bit mapping scheme to achieve a higher message embedding rate.

[0060] Single bit mapping: the present application refers to the single bit mapping scheme as M S , the core of which is to map the same message to consistent values. When the secret message value is 0, the value of Z is also 0; when the secret message value is 1, the value of Z is Or The noise value generated in this way is in the same space as the distribution of Gaussian noise. Only by passing the obtained Z value through the inverse mapping , the 0, 1 message value can be recovered.

[0061] Multi-bit mapping: the scheme of the present application also realizes large capacity embedding by designing a multi-bit mapping scheme M L . The channel number of the generated model input noise is 1, and the maximum information hiding capacity is 1 bpp when the single bit mapping scheme is applied. When the hiding capacity is higher than 1 bpp, more than 1 bit of secret message needs to be hidden in a single signal. For this purpose, the present application uses the message grouping method to realize multi-bit mapping. First, the secret message is grouped, and every two bits are a group, a total of four cases: 00, 01, 10, 11, which are mapped to four different ranges of noise values, as shown in Figure 4 . By hiding 2 bits in each bit, a 2 bpp message is hidden, and the receiver can recover the secret message by extracting the noise and passing it through the inverse mapping .

[0062] Further, the training set is used to train the function generator, comprising: in the process of training the generator by using the training set, a discriminator is introduced, the generated point cloud and the real point cloud are input into the discriminator, the function generator is trained through the process of adversarial game, and a regularization equivalent penalty loss function for point cloud data is introduced to punish the gradient of the discriminator.

[0063] Specifically, the function generator: the purpose of the present application is to apply implicit neural representation to generative steganography, so the present application needs a function generator that can continuously generate function instances. The neural network architecture of the function generator used by the present application is as shown in Figure 5 . The target of the present application is to obtain the generator g φ by training, so as to continuously generate the function f (θ) . And the parameter p(θ) of f (θ) determines f φ , so the target of the present application becomes how to obtain p(θ) from p(z). By sampling z from p(z) and inputting it into the generator , a series of weights can be obtained, wherein g φ : Z→Θ. For the generated function, the input coordinate u and the output feature value v; wherein p(θ) is the function parameter, p(z) is the random Gaussian noise distribution, z is the noise, and Θ is the parameter space.

[0064] To train the function generator, this paper proposes a modeling approach based on learning function distribution and using Generative Adversarial Strategy (GASP). During the training process, a discriminator is introduced, and the generated point cloud and the real point cloud are input into it, and the function generator is trained through an adversarial game process. Specifically, given the latent vector z and the coordinate u i,j , generate feature v ij :

[0065]

[0066] Among them, γ represents random Fourier encoding, which is used to learn the high-frequency information of the function. i,j} and the generated series of eigenvalues ​​{v ij}Evaluate the generator to train the model.

[0067] In the process of applying the discriminator training, the proposed R1 regularization is equivalent to the penalty for point cloud data. After representing the image as point cloud data, for a set The penalty is defined as follows:

[0068]

[0069] Where D is the loss of the discriminator. In this process, the present invention penalizes the gradient of the discriminator through the eigenvalue. The training process gets rid of the discretized grid data and directly acts on the point cloud data type.

[0070] Furthermore, the message extractor includes: multiple point convolution layers and fully connected layers; the point convolution layer is used to extract high-dimensional features by downsampling and increasing feature dimensions layer by layer, while retaining local information to obtain output features; the fully connected layer is used to flatten the output features and reduce the dimension layer by layer to obtain the feature vector of the target dimension.

[0071] Specifically, the message extractor for point cloud data includes: Unlike the neat arrangement of gridded data, the representation of point cloud data has the concept of potential distance, which makes its arrangement in space more disordered, such as Figure 6 Therefore, the traditional gridded convolution is no longer applicable. The present invention uses a neural network based on point convolution as the framework of the message extractor to construct a point cloud message extractor. The network structure is as follows: Figure 7As shown. Four point convolutional layers and three fully connected layers are defined therein, and the numbers on the point convolutional layers represent the output channel number, the output point number, the number of adjacent points for point convolution, and the intermediate channel number, respectively, wherein the design of reducing the sampling and increasing the feature dimension layer by layer preserves the local information while extracting high-dimensional features; the numbers under the fully connected layers represent the input and output dimensions, and the method of flattening the output of the convolutional layer and reducing the dimension layer by layer finally outputs a feature vector of a specified dimension. The model input is point cloud data of different quantities, and the data size can be modified according to the image resolution, and the output is 1*64 noise data.

[0072] The training message extractor includes: the message extractor is a model constructed based on point cloud convolution, in the training process, the batch size is 32, the learning rate is 0.001, 1000 groups of data are randomly sampled from the generative model as training samples, and the epoch is set to 10000 rounds, in each round of training, the average loss is calculated and compared with the previous training results to retain the best model.

[0073] The loss function includes: in the process of training the message extractor, the mean square error (MSE) is used to constrain the loss:

[0074]

[0075] Wherein, y i is the feature value of the point cloud data converted from the real image, is the feature value of the point cloud data sampled from the function, and n is the sample quantity.

[0076] Experiment and analysis: in order to directly see the experimental results, the generated data is visualized as a two-dimensional image by the data converter for analysis.

[0077] Evaluation index: the steganographic ability of the present application is evaluated from the aspects of visual security, message extraction accuracy, undetectability, robustness, efficiency and ablation experiment, and the super-resolution sampling ability of the implicit neural representation is evaluated. In order to quantitatively evaluate the effect of the scheme, the following standards are defined:

[0078] Message extraction accuracy (Acc):

[0079]

[0080] Wherein, BER is the bit error ratio (bit error ratio); XOR is the exclusive or operation, which is used to calculate the number of error bits; M(l) is the original secret message, and M'(l) is the extracted secret message; L m is the length of the secret message.

[0081] Unstructured pruning rate (P R):

[0082]

[0083] where P N is the number of pruned elements, P T is the total number of elements.

[0084] Experimental setup: The experiments of the present application were all carried out on an NVIDIA GeForce RTX 4060Ti graphics card. The experimental environment pytorch version is 2.5.1, torchvision version is 0.20.1, and cuda version is 12.1. The dataset is Celebahq, and the image size is uniformly adjusted to 64x64 before the experiment.

[0085] The present application carries out experiments on the Celebahq dataset, and randomly selects 6 times from 10000 experimental results for display, as shown in Figure 8 , the first row is the steganographic carrier generated by message mapping noise, the second row is the ordinary carrier generated by Gaussian noise, and the third row is the real image in the Celebahq dataset. It can be seen that the steganographic carrier generated by the noise through the message mapping scheme is almost indistinguishable from the other two types of images by the naked eye.

[0086] Message extraction accuracy: The accuracy of noise extraction is the core of recovering secret data. In order to improve the accuracy of secret message recovery, the present application carries out experiments by recovering secret data from the output of the extractor E. The present application uses the function generator G trained by the generative adversarial random process to generate the steganographic carrier, and uses E to recover the secret data according to the inverse mapping rule.

[0087] In order to illustrate the influence of the number of training rounds on the accuracy of the extractor, the present application sets the learning rate to 0.001, and sets the message embedding capacity to 64 and 128 bits, and uses the trained point cloud message extractor to carry out extraction experiments. The definition of message extraction accuracy is the ratio of the number of binary bits correctly recovered from the steganographic carrier to the length of the original secret message. Figure 9 The message extraction accuracy curve under different payloads is shown, and as the number of training rounds increases, the extraction accuracy also increases. When the message embedding capacity is 64 bits, the message extraction accuracy is 96.88%, and the smaller the payload, the higher the extraction accuracy. Table 1 shows the comparison of the present application with other schemes.

[0088] Table 1 Comparison of message extraction accuracy of different schemes

[0089]

[0090] Undetectability: To assess undetectability, we used the open-source steganalysis tool StegExpose for analysis. The results are shown in Table 2. A dataset of 10,000 steganalysis images with a resolution of 64*64 was constructed for analysis. StegExpose detected 624 of these images as secrets, yielding a detection rate of 6.24%. SamplePairs is a steganalysis method based on sample pair analysis; values ​​closer to 0 indicate a lower likelihood of steganalysis. RSanalysis is a steganalysis method based on Regular-Singular analysis; results closer to 0 indicate a lower likelihood of steganalysis. Fusion (mean) combines the average of multiple analysis methods. While the combined result is slightly higher than the previous two methods, it remains relatively low, indicating a low overall likelihood of steganalysis. Overall, all indicators are close to 0, indicating no significant steganalysis was detected in the analyzed data.

[0091] Table 2

[0092]

[0093] Robustness: The present invention uses continuous point cloud data in its steganographic scheme, so the traditional grid-based data-based noise-adding method to evaluate robustness is no longer applicable. For the point cloud data used in the present invention, the present invention adopts unstructured pruning technology to evaluate its robustness. Figure 10 A visualization of point cloud data at different pruning rates is provided. The first column shows the unpruned image, followed from left to right by images at pruning rates of 0.01, 0.05, 0.1, 0.3, 0.5, and 0.8. Message extraction was performed on these images at different pruning rates, and the message extraction accuracy is shown in Table 3. The results show that even after the secret carrier has been pruned, the secret message can still be extracted from it, demonstrating the robustness of the proposed solution.

[0094] Table 3 Message extraction accuracy under different pruning rates

[0095]

[0096]

[0097] Efficiency: When the message embedding size is 1bpp, the training time of the generative model for 10,000 epochs is 24 hours, and when the sample size is 1,000 and the batch size is 32, the training time of the message extractor is 120 hours. Figure 11 As shown in Figure 3, after training, the time required for the generator to generate data of different resolutions is almost the same, and the efficiency of the extractor in extracting noise from images of different resolutions is also basically the same.

[0098] Super-resolution: Figure 12 The image shows the encrypted vectors obtained by super-resolution sampling from the model when the training dataset is at a resolution of 64*64. From left to right, the image resolutions are 64, 128, 256, 512, and 1024, respectively. For ease of viewing, all images are resized to 256*256 resolution. The present invention uses implicit neural representations to replace traditional gridded data. This frees images from resolution constraints and instead implements vector encoding and decoding in a high-dimensional latent space. This allows the present invention to sample vectors of different resolutions from the function generator, which provides high flexibility and generalization for processing image data of different resolutions.

[0099] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A generative implicit representation steganography method based on message mapping, characterized in that: include: Use implicit neural representation technology to represent multimedia data and convert it into unified point cloud data for message embedding; The secret message to be embedded is represented in binary and converted into noise data through mapping, the noise data is input into the function generator for message embedding, the noise data is extracted by the message extractor, and the secret message is restored through inverse mapping; The function generator and the message extractor are both trained using a training set; the training set includes: a generated point cloud and a real point cloud; During the message embedding process, the noise space of the function generator applies single-bit and multi-bit message mapping schemes to embed the noise data; During the noise extraction process, the message extractor extracts the noise data by setting multiple point convolution layers and fully connected layers.

2. The generative implicit representation steganography method based on message mapping according to claim 1 is characterized in that: Acquiring the encrypted point cloud data includes: Collecting the multimedia data, performing pixel-wise discrete representation on the multimedia data, and obtaining grid data; The grid data is implicitly represented and converted into coordinates and eigenvalues ​​represented in the form of set pairs, which is the dense point cloud data.

3. The generative implicit representation steganography method based on message mapping according to claim 1 is characterized in that: Extracting the noise data using a message extractor includes: Inputting the noise data into a function generator for message embedding to obtain a continuous function; According to the target task, corresponding point cloud data is sampled in the continuous function, the corresponding point cloud data is input into the message extractor, and the noise data is extracted.

4. The generative implicit representation steganography method based on message mapping according to claim 1 is characterized in that: Applying the single-bit and multi-bit message mapping scheme to embed the encrypted point cloud data includes: During the message embedding process, when the maximum information hiding capacity of the noise data is less than a target threshold, a single-bit message mapping scheme is applied to map the noise data to a preset value; when the maximum information hiding capacity of the noise data is greater than the target threshold, the noise data is grouped and mapped to noise values ​​in different ranges.

5. The generative implicit representation steganography method based on message mapping according to claim 1 is characterized in that: The message extractor includes: multiple point convolution layers and fully connected layers; The point convolution layer is used to extract high-dimensional features by downsampling and increasing feature dimensions layer by layer, while retaining local information and obtaining output features; The fully connected layer is used to flatten the output features and reduce the dimensions layer by layer to obtain a feature vector of the target dimension.

6. The generative implicit representation steganography method based on message mapping according to claim 1 is characterized in that: Training the function generator using the training set includes: In the process of training the generator using the training set, a discriminator is introduced, the generated point cloud and the real point cloud are input into the discriminator, the function generator is trained through an adversarial game process, and a regularized equivalent penalty loss function for point cloud data is introduced to penalize the gradient of the discriminator.

7. The generative implicit representation steganography method based on message mapping according to claim 6 is characterized in that: The penalty loss function includes: Among them, D(s) is the loss of the discriminator, v ij is the eigenvalue, are the eigenvalues ​​at different coordinates.

Images