Steganography and extraction method of lossless neural network model

By using partial derivatives to distinguish parameters in neural network models and applying wet paper encoding to embed secret information, the problem of task performance degradation caused by information embedding is solved, lossless steganography is realized, and steganography security and scope of application are enhanced.

CN119011187AActive Publication Date: 2024-11-22SHANGHAI UNIV
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410926541.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-22
Estimated Expiration
2044-07-11

AI Technical Summary

Technical Problem

The existing neural network model steganography method will cause the performance of the original task to decline during the information embedding process, affecting the security and efficiency of hidden communication.

Method used

By obtaining the partial derivative of neural network model parameters, setting a modification threshold, distinguishing parameters into modifiable and non-modified, using wet paper encoding to embed secret information into modifiable parameters, and restoring the original parameters through dimensionality reduction and binary conversion to ensure model performance Not affected.

Benefits of technology

The lossless neural network model steganography is realized, which significantly enhances the steganography security and transmission efficiency. It is suitable for different neural network architectures and data sets, ensuring the hidden transmission of secret information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119011187A_ABST
    Figure CN119011187A_ABST
Patent Text Reader

Abstract

The invention relates to the field of secret data transmission, in particular to a steganography method of a lossless neural network model. The method comprises the following steps: setting a modification threshold value, and dividing neural network model parameters into modifiable parameters and unmodifiable parameters; performing dimensionality reduction on the multi-dimensional neural network model parameters, and converting the multi-dimensional neural network model parameters into neural network model parameters in a one-dimensional vector form; setting a judgment threshold, and converting the neural network model parameters in the one-dimensional vector form into a binary sequence; the sender generates a secret-containing binary sequence from the binary sequence, the secret information and the pseudo-random parity check matrix through wet paper coding; through the inverse process of binary conversion and dimension reduction processing, the secret-containing binary sequence is recovered to a normal model parameter form, and original model parameters are replaced to complete the original task of the model. Compared with the prior art, the method has the advantages that the problem of original task performance reduction caused by information embedding in neural network model steganography is solved, transmission of secret information is more hidden, and steganography safety is remarkably enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of secret data transmission, and in particular to a steganographic method of a lossless neural network model. Background Art

[0002] Steganography aims to conceal communication by hiding confidential information in a carrier. In order to share core data and improve transmission efficiency, enterprises need to transmit information in public channels. However, direct transmission or encrypted transmission will allow third parties to know that they are "communicating secretly", which is not secure enough. Steganography can conceal the behavior of "communicating secretly" and better prevent data leakage. At the same time, covert communication helps to understand the development trends in politics, economy, military and cyberspace, and thus seize the dominance of cyberspace.

[0003] According to the different carriers, steganography can be divided into image steganography, text steganography, video steganography, etc. The forms of carriers are varied, but all of them hide secret information by modifying the carrier. Modification of the carrier will undoubtedly affect the security of covert communication, so the design of steganography methods usually embeds secret information while minimizing damage to the original carrier. In image steganography, the most advanced steganography methods are implemented under the distortion minimization framework.

[0004] With the rapid development of artificial intelligence, neural networks have become an emerging carrier, and neural network steganography methods have emerged. Neural networks were initially used as auxiliary tools in steganography, such as helping to design the embedding cost of pixels in image steganography to reduce the damage and distortion caused by steganography to the original image. Later, it developed into a carrier that directly carries confidential information. Like other types of carrier steganography methods, the embedding of information will lead to a decrease in the performance of the original model task (such as image classification), which is not conducive to covert communication. Summary of the invention

[0005] The purpose of the present invention is to provide a lossless neural network model steganography method in order to overcome the defects of the above-mentioned prior art.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] The present invention provides a steganographic method of a lossless neural network model, comprising the following steps:

[0008] S1. Obtain the partial derivative of the neural network model parameter relative to the loss function, set the modification threshold, and divide the neural network model parameters into modifiable parameters and non-modifiable parameters according to the relationship between the partial derivative and the modification threshold;

[0009] S2, preprocessing process: reduce the dimension of the multi-dimensional neural network model parameters and convert them into the neural network model parameters in the form of one-dimensional vectors; set the decision threshold and convert the neural network model parameters in the form of one-dimensional vectors into binary sequences through the decision threshold;

[0010] S3, embedding process: generating a secret binary sequence by wet paper coding the binary sequence, secret information, and pseudo-random parity check matrix;

[0011] S4. Through the inverse process of binary conversion and dimensionality reduction, the secret binary sequence is restored to the normal neural network model parameter form, replacing the original neural network model parameters to complete the original task of the model.

[0012] The partial derivatives of the neural network model parameters relative to the loss function are obtained through gradient information.

[0013] The partial derivatives are normalized and their value range is [0,1].

[0014] A COST vector is generated according to the partial derivative. If the partial derivative is greater than the modification threshold, the neural network model parameters corresponding to the partial derivative are set as unmodifiable parameters through the COST vector; if the partial derivative is less than or equal to the partial derivative of the modification threshold, the neural network model parameters corresponding to the partial derivative are set as modifiable parameters through the COST vector.

[0015] Mark modifiable parameters as 1 and non-modifiable parameters as infinity.

[0016] The values ​​of the modifiable parameters that are greater than the decision threshold are set to 1, and the values ​​of the modifiable parameters that are less than or equal to the decision threshold are set to 0, thereby obtaining a binary sequence consisting of 1 and 0.

[0017] The value of a modifiable parameter may change from 1 to 0 or from 0 to 1, or it may not be modified. The value of an unmodifiable parameter will not be modified during the embedding process.

[0018] The inverse process of binary conversion includes: for the unmodified parameters in the normal model parameter form, they are directly used; for the modified parameters, the modified amplitude method is set to make the restored confidential parameter value close to the original parameter value; the modification amplitude method includes: if the confidential parameter is 0, it is modified to the decision threshold; if the confidential parameter is 1, it is modified to a value greater than the decision threshold and within the set range.

[0019] In the second aspect, the present invention provides an extraction method based on lossless neural network model steganography. The extraction method is applied to the steganography method of any of the above-mentioned lossless neural network models. First, after the receiver receives the secret model transmitted by the sender, the secret model parameters are extracted, and then the secret binary sequence is obtained through dimensionality reduction and binary conversion. After the secret binary sequence is calculated with the pseudo-random parity check matrix, the secret information is extracted.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. The present invention solves the problem of original task performance degradation caused by information embedding in neural network model steganography, makes the transmission of secret information more covert, and significantly enhances the security of steganography.

[0022] 2. The method provided by the present invention can use different neural network models as carriers by normalizing the partial derivatives of different neural network model parameters. It has a wide range of applications. Models trained for different neural network architectures and data sets are suitable for the steganography method, and the performance in all aspects is improved. The two parties to the secret information transmission can select different neural network carriers according to different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a structural diagram of the method of the present invention;

[0024] Figure 2 is a flow chart of the information embedding and extraction process of the present invention;

[0025] Figure 3 It is a comparison chart of the effect of changing the threshold value on the accuracy under different models or different capacities of the present invention;

[0026] Figure 4 This is a graph showing how the accuracy of the AlexNet model on different datasets changes as the threshold is modified;

[0027] Figure 5 This is a graph showing how the accuracy of the VGGt model on different datasets changes as the threshold is modified;

[0028] Figure 6 This is a graph showing how the accuracy of the GoogLeNet model on different datasets changes as the threshold is modified;

[0029] Figure 7 This is a graph showing how the accuracy of the ResNet model on different data sets changes as the threshold is modified;

[0030] Figure 8 This is a graph showing how the accuracy changes with the decision threshold in different models and data sets;

[0031] Fig. 9 This is a graph showing how the accuracy changes with the embedding capacity for different models and datasets;

[0032] Fig.10 It is the probability density estimation diagram of model parameters before and after embedding in different models and data sets. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0034] Example

[0035] This embodiment provides a steganographic method for a lossless neural network model. The implementation of this method is mainly divided into two steps: one is to select appropriate model parameters for loading confidential information, and the other is to use wet paper coding to complete information embedding and extraction. Figure 1 and Figure 2 As shown in the figure, the parameter screening process is achieved by comparing the size of the partial derivatives of each model parameter. Since most neural network models are trained using the gradient descent method, the size of the partial derivatives can reflect the degree of influence of the change of model parameters on the performance of the original model. The parameters that have a greater impact on the performance of the model are screened out and set to be unmodifiable, and only the parameters with smaller partial derivatives are modified. At the same time, the appropriate modification range is set, and the secret information is embedded using wet paper coding, which can achieve lossless model steganography and ensure that the original task performance of the neural network model is not affected.

[0036] The sender first performs parameter screening by obtaining the partial derivatives of each model parameter relative to the loss function, and selects the parameters with small partial derivatives to load confidential information. A small partial derivative indicates that the parameter is less correlated with the model performance, and modifying it will have little impact on the original task performance. During the embedding process, parameters with large partial derivatives are set to be unmodifiable. This process is consistent with wet paper coding, so the embedding of secret information is achieved by wet paper coding, and the receiver can extract the secret information without knowing the modification location. Since parameters with high correlation with the original task performance are not modified, the modification range can be set to limit the degree of modification when modifying parameters with less correlation, so the original task performance is not affected.

[0037] In this embodiment, the sender uses classic convolutional neural networks (AlexNet, VGG16, GoogLeNet, ResNet) as carriers, trained on the MNIST and CIFAR10 datasets, and obtained eight neural network models for carrying secret information. Information embedding is achieved by wet paper coding, and the receiver can completely extract secret information from the secret model without knowing the specific modification position of the model parameters.

[0038] like Figure 2 As shown, the specific steps of this embodiment are as follows:

[0039] 1. When the training is completed, the model gradient information is saved to obtain the partial derivatives of each model parameter with respect to the loss function, so as to select the model parameters with the least impact on the original model task performance and load secret information, such as Figure 1 As shown in the "Parameter Screening" step in

[0040] Since the gradient descent method is used to update the model parameters during training, taking the stochastic gradient descent method SGD as an example, the model parameters θ (θ includes weights and biases) in each iteration are calculated by formula (1), where η is the learning rate, x (i) is the training sample, y (i) It's a label. is the gradient information of the loss function J(θ), which is calculated from the partial derivatives of all parameters, as shown in formula (2). Therefore, the partial derivative information of each model parameter can be obtained through the gradient information.

[0041] A modification threshold is set, and the neural network model parameters are divided into modifiable parameters and non-modifiable parameters according to the size relationship between the partial derivative and the modification threshold. Specifically, a COST vector is generated based on the partial derivative information obtained by parameter screening, which is defined as formula (9) and is used to mark modifiable parameters and non-modifiable parameters (modifiable and non-modifiable refers to whether they can be modified during the secret embedding process, and can be modified during the preprocessing process). The purpose is to allow the computer in the wet paper code to identify which parameters can be modified, where the elements are related to X B One-to-one correspondence. First, normalize the partial derivatives of each model parameter to a value range of [0,1], which is greater than the modification threshold t v The COST vector sets the modifiable part of the binary sequence to 1 and the unmodifiable part to infinity, as shown in formula (9). The 1 and infinity here are only used for the identification of computing devices.

[0042] Before embedding the information, the sender sets the model parameters with large partial derivatives to be unmodifiable to ensure that the task performance of the original model is not affected.

[0043]

[0044] 2. After parameter screening, the sender can use wet paper coding to embed information, such as Figure 1 "Wet paper coding" is shown. The principle of wet paper coding is to convert binary carrier units It is divided into a modifiable part and an unmodifiable part. The sender and the receiver agree in advance on a pseudo-random parity check matrix A k×p , defined as formula (3). The sender sends k bits of secret information M = (m1, m2, …, mk ) T Embed X B After that, the receiver can extract the secret information through the pseudo-random parity check matrix and the secret carrier, as shown in formula (4). In this embodiment, the modification threshold t1 is set to distinguish between modifiable parameters and non-modifiable parameters. Since the model parameters are not binary carrier units that can be directly used for wet paper coding, they also need to be processed by dimensionality reduction and binary conversion. Dimensionality reduction and binary conversion belong to the preprocessing process.

[0045]

[0046] M=A k×p ·X B (4)

[0047] 3. Dimensionality reduction: The original model parameter X is defined as formula (5), where X n is a high-dimensional matrix with the same dimension as the tensor in the neural network, as shown in formula (6), where p is the matrix from X1 to X n The sum of all parameter values. In order to adapt to the encoding, X needs to be converted into a vector form, that is, the high-dimensional X n Convert it to a p×1 vector and get X′, which is defined as equation (7).

[0048] X=(X1,X2,…,X n ) T (5)

[0049]

[0050] X′=(x1,x2,…,x p ) T (7)

[0051] 4. Binary conversion: After obtaining X′, it needs to be converted into a binary sequence. p The value of is any real number, and we need to set rules to map its value space to {0,1}, that is, X′:X B →{0,1}. The specific measure is to set the decision threshold t2, and in X′, the value greater than t2 is set to 1, and the value less than or equal to t2 is set to 0, as shown in formula (8). In this way, the processing of the model parameters is completed and converted into a binary sequence that can be used for wet paper coding.

[0052]

[0053] 5. Specific measures for the embedding process such as Figure 2As shown in Figure 1, the model parameters are converted into binary sequences through dimensionality reduction and binary conversion. The parameter values ​​of the modifiable part set to 1 include 0 and 1. Parameters marked as non-modifiable will not be modified during the embedding process. Modifiable parameters may change from 1 to 0 or from 0 to 1 or may not be modified. Wet paper encoding is based on the binary sequence X B , secret information m, pseudo-random parity check matrix A k×p Generate a secret binary sequence As shown in formula (10). After the information is embedded, the encrypted binary sequence needs to be restored to the normal model parameter form to replace the original model parameters to complete the original task of the model.

[0054] The recovery process is divided into numerical recovery and dimensional recovery, which are the inverse processes of binary conversion and dimensionality reduction respectively. When performing numerical recovery, for the unmodified part of the model parameters, just make it equal to the original model parameters. For the changed parameters, if the value is 0 in the secret binary sequence, let it be equal to the decision threshold t2; if the value is 1, let it take a value greater than t2 and infinitely close to t2, such as t2+(1e-10). This value-taking method is to set the modification amplitude to make the restored secret parameter value as close to the original parameter value as possible to ensure the original task performance. Dimension recovery is to convert the secret sequence into the same dimension as the original model parameter X, and reshape can be completed. The embedding process finally obtains the secret model parameter X stego , defined as formula (11), X stegon Yes and X n Matrices of the same dimension but different values.

[0055]

[0056] X stego =(X stego1 ,X stego2 ,…,X stegon ) T (11)

[0057] 6. Extraction process Figure 2 As shown in the figure, after the receiver receives the secret model transmitted by the sender, it extracts the secret model parameter X stego , obtained by dimensionality reduction and binary transformation With the pseudo-random parity check matrix A k×p After calculation, the secret information can be extracted, as shown in formula (12).

[0058] The extraction error is defined as (13), which is the difference between the real information M and the extracted information M′. An extraction error of 0 indicates that the information is extracted completely and without error.

[0059]

[0060] In order to verify the effect of the method provided by this embodiment, the secret information embedding capacity of different network models is demonstrated through experiments. Then the undetectability of the steganographic image is tested.

[0061] (1) The neural networks AlexNet, VGG16, GoogLeNet, and ResNet are trained on the MNIST and CIFAR10 datasets, respectively, to obtain eight neural network models. During the training process, the last round of gradient information is saved for parameter screening.

[0062] (2) The experimental environment was built with PyTorch, and the training was performed on a single NVIDIA GeForce RTX 3090 GPU running on a Linux Centos 7.9 system. The wet paper coding was implemented on Matlab 2016a in a Windows 10 system.

[0063] (3) Determine the modification threshold t1 and the decision threshold t2. These two parameters are crucial in the embedding process. Generally speaking, the more secret information is embedded, the greater the degradation of the original task performance. Adjusting these parameters can adjust the relationship between the embedding capacity and the degradation of the original task performance. Figure 3 It can be seen that for different models with the same capacity, the impact of modifying the threshold on the original task performance is significantly different; for the same model with different capacities, the impact of modifying the threshold on the original task performance is consistent. Therefore, different models should use different parameter values. Figure 4-Figure 7 The performance degradation of the original tasks of the eight models under different modification thresholds is shown respectively. The modification threshold with the smallest degradation is selected as the optimal modification threshold of the model. The eight models are in the order of AlexNet-MNIST, AlexNet-CIFAR10, VGG-MNIST, VGG-CIFAR10, GoogLeNet-MNIST, GoogLeNet-CIFAR10, ResNet-MNIST, and ResNet-CIFAR10. The modification threshold values ​​are 0.1, 0, 0, 0, 0, 0.05, 0, and 0 respectively. Similarly, Figure 8 When the modification threshold has been determined, the one with the smallest impact on the original task performance at different decision thresholds is selected as the optimal decision threshold of the model. That is, the decision threshold values ​​when the accuracy is the highest in the figure are 0, 0, 0.001, 0, 0.05, 0.001, 0, and 0.01 respectively.

[0064] (4) Based on the parameters selected in the above experiment, this example will demonstrate the application results of this method under different models, that is, how many bits of secret information are successfully embedded without affecting the performance of the original task. Fig. 9As shown, without reducing the accuracy of the original task of the model, the maximum capacity of secret information that can be embedded in the eight models is 0.6M, 0.08M, 3.25M, 0.075M, 0.9M, 0.025M, 0.09M, and 0.015M bits respectively.

[0065] (5) Verify the undetectability of the method of this embodiment from two aspects: KL divergence and probability density function (PDF). KL divergence is defined as formula (14), which reflects the degree of difference between two probability distributions. The larger the KL divergence, the greater the degree of difference, and vice versa. The probability density function describes the probability density distribution of variables. If the probability density functions of two variables are the same or similar, it means that the statistical characteristics of the two are the same or similar. The purpose of neural network steganography is to communicate covertly, so the original model parameters and the secret model parameters should be as similar as possible to ensure the security of steganography. In the method of this embodiment, the KL divergence of the original model parameters and the secret model parameters of different models at the maximum embedding capacity are 0.1067, 0.0101, 0.0231, 0.0019, 0.0193, 0.0018, 0.0095, and 0.0001, respectively. The PDF curve is as follows: Fig.10 As shown, the distribution of the original parameters and the encrypted parameters of each model is almost consistent, indicating that the method of this embodiment has strong security performance.

[0066]

[0067] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A steganographic method for lossless neural network models, It is characterized in that The following steps are involved: S1. Obtain the partial derivative of the neural network model parameter relative to the loss function, set the modification threshold, and divide the neural network model parameters into modifiable parameters and non-modifiable parameters according to the relationship between the partial derivative and the modification threshold; S2, preprocessing process: reduce the dimension of the multi-dimensional neural network model parameters and convert them into the neural network model parameters in the form of one-dimensional vectors; set the decision threshold and convert the neural network model parameters in the form of one-dimensional vectors into binary sequences through the decision threshold; S3, embedding process: generating a secret binary sequence by wet paper coding the binary sequence, secret information, and pseudo-random parity check matrix; S4. Through the inverse process of binary conversion and dimensionality reduction, the secret binary sequence is restored to the normal neural network model parameter form, replacing the original neural network model parameters to complete the original task of the model.

2. The steganographic method of a lossless neural network model according to claim 1, It is characterized in that The partial derivatives of the neural network model parameters relative to the loss function are obtained through gradient information.

3. The steganographic method of a lossless neural network model according to claim 1, It is characterized in that The partial derivatives are normalized and their value range is [0,1].

4. The steganographic method of a lossless neural network model according to claim 1, It is characterized in that A COST vector is generated according to the partial derivative. If the partial derivative is greater than the modification threshold, the neural network model parameters corresponding to the partial derivative are set as unmodifiable parameters through the COST vector; if the partial derivative is less than or equal to the partial derivative of the modification threshold, the neural network model parameters corresponding to the partial derivative are set as modifiable parameters through the COST vector.

5. The steganographic method of a lossless neural network model according to claim 4, It is characterized in that Mark modifiable parameters as 1 and non-modifiable parameters as infinity.

6. The steganographic method of a lossless neural network model according to claim 1, It is characterized in that The values ​​of the neural network model parameters in the form of one-dimensional vectors greater than the decision threshold are set to 1, and the values ​​of the neural network model parameters in the form of one-dimensional vectors less than or equal to the decision threshold are set to 0, to obtain a binary sequence consisting of 1 and 0.

7. The steganographic method of a lossless neural network model according to claim 6, It is characterized in that The value of a modifiable parameter may change from 1 to 0 or from 0 to 1, or it may not be modified. The value of an unmodifiable parameter will not be modified during the embedding process.

8. The steganographic method of a lossless neural network model according to claim 1, It is characterized in that The inverse process of binary conversion includes: for the unmodified parameters in the normal model parameter form, they are directly used; for the modified parameters, the modification amplitude method is set to make the restored encrypted parameter value close to the original parameter value.

9. The steganographic method of a lossless neural network model according to claim 8, It is characterized in that The modification amplitude method includes: if the confidentiality parameter is 0, modifying it to a decision threshold; if the confidentiality parameter is 1, modifying it to a value greater than the decision threshold and within a set range.

10. A steganographic extraction method based on a lossless neural network model, It is characterized in that The extraction method is applied to the steganography method of the lossless neural network model described in any one of claims 1-8. First, after the receiver receives the secret model transmitted by the sender, the secret model parameters are extracted, and then the secret binary sequence is obtained by dimensionality reduction and binary conversion. After the secret binary sequence and the pseudo-random parity check matrix are calculated, the secret information is extracted.

Citation Information

Patent Citations

  • Image steganalysis method and system based on convolutional neural network

    CN108961137A

  • Information steganography method and device based on video carrier

    CN114339258A

  • Image steganography method based on LS-GAN

    CN114676446A

  • Adaptive double-image reversible information hiding method based on Sudoku matrix

    CN116800902A

  • Image super-resolution steganography method based on reversible neural network

    CN117078517A