Generative image information steganography method

Abstract

According to the generative image information steganography method provided by the invention, a memristor is implanted into an existing chaotic system, a new memristor chaotic system NCSM is constructed, the system can generate a chaotic sequence with higher randomness, and when the system is applied to image encryption, the security of an encrypted image can be improved; in the aspect of image steganography, a ciphertext image is converted into a generative steganography image through a generative reversible neural network model Glow, so that the information security of a plaintext image is ensured, and the visual security of the ciphertext image is also ensured; different from a current model based on irreversible adversarial generation, the generative reversible neural network does not need to train a decoder, so that the complexity of the model and the training difficulty of the model are reduced; besides, by optimizing the input of the reversible generation model, the content of the generative steganographic image can be controlled to a certain extent, and the method can adapt to wider application scenes.

Description

Technical Field

[0001] The present invention relates to the field of image encryption technology, and more specifically, to a generative image information steganography method. Background Art

[0002] In today's digital society of the information age, with the rapid development of computer technology and the widespread use of mobile devices, a large amount of information is transmitted and stored through the network. As one of the important carriers of information transmission, digital images play a vital role in people's daily communication process. However, in the network environment, there are a large number of private or confidential images, such as medical images of patients, military images, and defense confidential images. If such images are illegally stolen or disseminated without authorization, it will lead to the leakage of important personal information, and even in serious cases may pose a threat to the security of specific units or departments. Therefore, it is crucial to ensure the security and encrypted transmission of image information.

[0003] Encrypting a secret plaintext image into a noise-like ciphertext image can directly protect its secret information, but this method cannot ensure the visual security of the ciphertext image. With the development of deep learning, it has been widely used in computer vision, natural language processing, image classification, image generation and other fields, solving complex tasks in many application fields. In particular, the use of generative adversarial networks can convert noise-like ciphertext images into visually meaningful carrier images to achieve the goal of protecting the visual security of ciphertext images.

[0004] For example, an existing patent document discloses an image encryption method based on GAN and chaotic system, in which a quantum cellular neural network generates a quantum chaos control table and a quantum exchange table to scramble the plaintext image within and between blocks, and uses multiple chaotic diffusion steps such as forward and reverse diffusion and dynamic diffusion to remove the correlation between image pixels. Although the method of using a generative adversarial network to hide images does not require modifying the carrier image, this method requires training multiple models such as a generator, a discriminator, and a decoder, which makes model training extremely difficult. In addition, the performance of the chaotic system in the prior art is relatively low, and the randomness of the generated chaotic sequence is not high. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned prior art in that model training is difficult and the chaotic system has poor performance when performing image encryption, the present invention provides a generative image information steganography method, which can improve the randomness of the chaotic system generated sequence, reduce the difficulty and complexity of model training, and thus improve the efficiency and security of image encryption.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A generative image information steganography method comprises the following steps:

[0008] S1: Construction of a four-dimensional memristor chaotic system NCSM;

[0009] S2: Obtain the color plaintext image P and convert it into three channel components R, G and B of red, green and blue;

[0010] S3: randomly generate a key and input it into the memristor chaotic system NCSM for iteration to generate a first random sequence A and a second random sequence C;

[0011] S4: According to the first random sequence A and the second random sequence C, the three channels R, G and B are encrypted respectively to obtain encryption matrices R′, G′ and B′;

[0012] Combine the encryption matrices R′, G′ and B′ of the three channels R, G and B to obtain the color ciphertext image G;

[0013] S5: Obtain the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4;

[0014] S6: sampling the elements of the ciphertext image G in a standard normal distribution to obtain a sampled probability density matrix L, and sequentially replacing the elements in ε1, ε2 and ε3 with the elements in the probability density matrix L to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ of the sampling;

[0015] S7: ε1′, ε2′, ε3′ and ε4 are combined as the forward input of the pre-trained reversible generation model Glow, and finally the generated stego-image H is obtained.

[0016] Preferably, in step S1, the memristive chaotic system NCSM is specifically:

[0017]

[0018]

[0019] Wherein, a, b, c, k and ω are respectively the first to fifth preset hyperparameters; x, y, z and u are respectively the first to fourth input variables of the memristor chaotic system NCSM; and They are the first to fourth output variables of the memristor chaotic system NCSM respectively.

[0020] Preferably, the step S3 comprises:

[0021] The size of the color plaintext image P is m×n×3; wherein m and n are first and second positive integers;

[0022] Randomly generate keys x0, y0, z0 and u0, and use them as the initial values ​​of x, y, z and u and input them into the memristor chaotic system NCSM. Use the fourth-order-fifth-order Runge-Kutta algorithm to calculate the numerical solution of the memristor chaotic system NCSM, and obtain 4 random chaotic sequences. The first two random chaotic sequences are saved as the first random sequence A and the second random sequence C. The lengths of the first random sequence A and the second random sequence C are both m×n.

[0023] Preferably, step S4 comprises:

[0024] S4.1: Preprocess the first random sequence A and the second random sequence C respectively to obtain a third random sequence D and a fourth random sequence J;

[0025] S4.2: Sort the third random sequence D to obtain an index sequence I, and use the index sequence I to scramble and encrypt the three channels R, G and B respectively to obtain a scramble matrix S for the corresponding channel;

[0026] S4.3: Expand the scrambled matrices S of the three channels R, G and B into a one-dimensional sequence F, and perform XOR operation on the elements of the one-dimensional sequence F and the elements in the fourth random sequence J to obtain a fifth random sequence E of the corresponding channel;

[0027] S4.4: Restore the fifth random sequence E of the three channels R, G and B into m×n matrices respectively, and obtain encryption matrices R′, G′ and B′ of the corresponding channels.

[0028] Preferably, in step S4.1, the following operation is performed on the first random sequence A to obtain a third random sequence D:

[0029]

[0030] Among them, A i is the i-th element in the first random sequence A; D i is the i-th element in the third random sequence D;

[0031] The following operations are performed on the second random sequence C to obtain the fourth random sequence J:

[0032]

[0033] Among them, C i is the i-th element in the second random sequence C; J i is the i-th element in the fourth random sequence J.

[0034] Preferably, in step S4.2, the third random sequence D is sorted in descending order to obtain an index sequence I.

[0035] Preferably, in step S4.3, the scrambled matrices S of the three channels R, G and B are expanded into a one-dimensional sequence F, and the elements of the one-dimensional sequence F are XORed with the elements in the fourth random sequence J to obtain the fifth random sequence E of the corresponding channel, specifically:

[0036]

[0037] Among them, E i is the i-th element in the fifth random sequence E; F i is the i-th element in the one-dimensional sequence F; J i is the i-th element in the fourth random sequence J.

[0038] Preferably, in step S5, the type of the carrier image Q is the same as the image type of the training reversible generative model Glow, and the size is s×q×3, wherein p and q are the third and fourth positive integers, and p and q are divisible by 32;

[0039] The sum of the sizes of the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4 is equal to s×q×3;

[0040] The sum of the sizes of the first, second and third latent variables ε1, ε2 and ε3 is greater than the sum of all elements of the ciphertext image G to ensure the complete hiding of the ciphertext information.

[0041] Preferably, the step S6 includes:

[0042] The elements of the ciphertext image G are sampled from a standard normal distribution, the domain of which is [-4, 4], and the domain is evenly divided into 256 intervals, each of which has a length of 0.03125; the intervals are sorted from left to right, with serial numbers ranging from 0 to 255;

[0043] According to the value of the element of the ciphertext image G, the corresponding interval is found, and the probability density of the center point of the corresponding interval is taken as the probability density of the sampling. Finally, the probability density matrix L of the sampling is obtained, as shown below:

[0044]

[0045] L i =λ(-4+(1+G i )×0.03125-0.015625),i=1,2,…,m×n×3

[0046] Among them, G i is the i-th element in the ciphertext image G; L i is the i-th element in the sampled probability density matrix L;

[0047] The elements in ε1, ε2 and ε3 are replaced with the elements in the probability density matrix L in sequence to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ sampled.

[0048] Preferably, the method further comprises:

[0049] S8: Use an asymmetric encryption algorithm to send the key and the generated stego-image H to the communication recipient.

[0050] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0051] The present invention provides a generative image information steganography method. First, the present invention implants a memristor into a three-dimensional chaotic system to construct a new four-dimensional memristor chaotic system NCSM. The new system NCSM adds a dimension on the basis of the original three-dimensional chaotic system, making the dynamic behavior of the system more complex. The introduced memristor not only increases the nonlinear characteristics of the system, but also enhances its chaotic characteristics, thereby improving the unpredictability and security of the system. The present invention generates a random sequence based on the NCSM system and applies it to the image encryption process.

[0052] Secondly, the present invention adopts the Glow generative model and designs an image steganography method. Different from the traditional embedded hiding method, the generative image hiding method can achieve deeper concealment and higher quality image hiding. The generative method no longer relies on embedding secret information in the original image, but directly generates images containing secret information through the generator network, making the hiding process more flexible and difficult to detect. In contrast, the traditional embedding method is vulnerable to statistical analysis and attacks, while the generative method of the present invention utilizes the adversarial training mechanism of the generative adversarial network to enhance the security and robustness of image steganography.

[0053] At the same time, the present invention designs a generative reversible image hiding method. Unlike the existing hiding method based on generative adversarial network, the generative reversible image hiding method of the present invention does not need to train models such as discriminators and decoders, thereby reducing the difficulty of model training; in particular, the generated steganographic image of the present invention is almost the same as the real image in visual quality, and after the hidden information is extracted, the plaintext image can be restored almost losslessly;

[0054] In addition, the present invention optimizes the input of the generative reversible model so that the method has a certain control ability over the output content of the model; in certain scenarios where there are requirements for generated content, the method of the present invention has a wider applicability, thereby meeting the specific requirements for generated image content in different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a generative image information steganography method provided in Example 1.

[0056] Figure 2 This is a framework diagram of a generative image information steganography method provided in Example 2.

[0057] Figure 3 This is a Glow structure diagram of the reversible generation model provided in Example 2.

[0058] Figure 4 This is a structural diagram of the Flow block in the reversible generation model Glow provided in Example 2.

[0059] Figure 5 This is a schematic diagram of the carrier image provided in Example 2.

[0060] Figure 6 This is a schematic diagram of the plaintext image provided in Example 2.

[0061] Figure 7 This is a schematic diagram of the generative steganographic image provided in Example 2.

[0062] Figure 8 This is a diagram of the generative steganographic image recovery framework provided in Example 3.

[0063] Fig. 9 This is a schematic diagram of the restored plaintext image provided in Example 3. DETAILED DESCRIPTION

[0064] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0065] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0066] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0067] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0068] Example 1

[0069] like Figure 1 As shown, this embodiment provides a generative image information steganography method, comprising the following steps:

[0070] S1: Construction of a four-dimensional memristor chaotic system NCSM;

[0071] S2: Obtain the color plaintext image P and convert it into three channel components R, G and B of red, green and blue;

[0072] S3: randomly generate a key and input it into the memristive chaotic system NCSM for iteration to generate a first random sequence A and a second random sequence C;

[0073] S4: According to the first random sequence A and the second random sequence C, the three channels R, G and B are encrypted respectively to obtain encryption matrices R′, G′ and B′;

[0074] Combine the encryption matrices R′, G′ and B′ of the three channels R, G and B to obtain the color ciphertext image G;

[0075] S5: Obtain the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4;

[0076] S6: sampling the elements of the ciphertext image G in a standard normal distribution to obtain a sampled probability density matrix L, and sequentially replacing the elements in ε1, ε2 and ε3 with the elements in the probability density matrix L to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ of the sampling;

[0077] S7: ε1′, ε2′, ε3′ and ε4 are combined as the forward input of the pre-trained reversible generation model Glow, and finally the generated stego-image H is obtained.

[0078] In the specific implementation process, firstly, a four-dimensional memristor chaotic system NCSM is constructed;

[0079] In recent years, the performance of nonlinear chaotic systems has been continuously improved from low-dimensional chaos to high-dimensional chaos, and people have proposed many image encryption algorithms based on chaotic systems. To further improve the performance of chaotic systems, an effective method is to introduce memristors. Memristors are a new type of electronic component with a memory effect that can record the amount of charge passing through. Their unique properties enable them to introduce nonlinear and time-varying characteristics into circuits, thereby increasing the complexity of chaotic systems.

[0080] If the memristor is combined with the traditional chaotic system, the chaotic behavior of the system can be significantly improved, which is more conducive to the application in the field of image encryption. Therefore, this method implants the memristor into an existing three-dimensional chaotic system to construct a new chaotic system (named NCSM), and makes the generated chaotic sequence have better randomness.

[0081] Then the color plaintext image P is obtained and converted into three channel components R, G and B of red, green and blue;

[0082] A key is randomly generated and input into the memristor chaotic system NCSM for iteration to generate a first random sequence A and a second random sequence C;

[0083] According to the first random sequence A and the second random sequence C, the three channels R, G and B are encrypted respectively to obtain encryption matrices R′, G′ and B′;

[0084] Combine the encryption matrices R′, G′ and B′ of the three channels R, G and B to obtain the color ciphertext image G;

[0085] Get the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4;

[0086] This method uses the Glow generative reversible neural network as the generator of the steganographic model. Due to its reversibility, the ciphertext image can be restored without training the decoder, and the complexity and training difficulty of the model are effectively reduced.

[0087] The elements of the ciphertext image G are sampled from the standard normal distribution to obtain the sampled probability density matrix L, and the elements in ε1, ε2 and ε3 are replaced with the elements in the probability density matrix L in order to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ of the sampling;

[0088] Finally, ε1′, ε2′, ε3′ and ε4 are combined as the forward input of the pre-trained reversible generation model Glow, and the generated stego-image H is finally obtained to complete the image encryption.

[0089] Example 2

[0090] like Figure 2 As shown, this embodiment provides a generative image information steganography method, comprising the following steps:

[0091] S1: Construction of a four-dimensional memristor chaotic system NCSM;

[0092] S2: Obtain the color plaintext image P and convert it into three channel components R, G and B of red, green and blue;

[0093] S3: randomly generate a key and input it into the memristive chaotic system NCSM for iteration to generate a first random sequence A and a second random sequence C;

[0094] S4: According to the first random sequence A and the second random sequence C, the three channels R, G and B are encrypted respectively to obtain encryption matrices R′, G′ and B′;

[0095] Combine the encryption matrices R′, G′ and B′ of the three channels R, G and B to obtain the color ciphertext image G;

[0096] S5: Obtain the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4;

[0097] S6: Sample the elements of the ciphertext image G in a standard normal distribution to obtain a sampled probability density matrix L, and replace the elements in ε1, ε2 and ε3 with the elements in the probability density matrix L in order to obtain the first, second and third latent variables ε1′, ε2′, ε3′ of the sampling;

[0098] S7: Combining ε1′, ε2′, ε3′ and ε4 as the forward input of the pre-trained reversible generation model Glow, and finally obtaining a generated stego image H;

[0099] S8: using an asymmetric encryption algorithm to send the key and the generated stego-image H together to the communication recipient;

[0100] In step S1, the memristor chaotic system NCSM is specifically:

[0101]

[0102]

[0103] Wherein, a, b, c, k and ω are respectively the first to fifth preset hyperparameters; x, y, z and u are respectively the first to fourth input variables of the memristor chaotic system NCSM; and They are the first to fourth output variables of the memristor chaotic system NCSM;

[0104] The step S3 comprises:

[0105] The size of the color plaintext image P is m×n×3; wherein m and n are first and second positive integers;

[0106] Randomly generate keys x0, y0, z0 and u0, and use them as initial values ​​of x, y, z and u and input them into the memristor chaotic system NCSM. Use the fourth-order-fifth-order Runge-Kutta algorithm to calculate the numerical solution of the memristor chaotic system NCSM, and obtain 4 random chaotic sequences. Save the first two random chaotic sequences as the first random sequence A and the second random sequence C. The lengths of the first random sequence A and the second random sequence C are both m×n.

[0107] The step S4 comprises:

[0108] S4.1: Preprocess the first random sequence A and the second random sequence C respectively to obtain a third random sequence D and a fourth random sequence J;

[0109] S4.2: Sort the third random sequence D to obtain an index sequence I, and use the index sequence I to scramble and encrypt the three channels R, G and B respectively to obtain a scramble matrix S for the corresponding channel;

[0110] S4.3: Expand the scrambled matrices S of the three channels R, G and B into a one-dimensional sequence F, and perform XOR operation on the elements of the one-dimensional sequence F and the elements in the fourth random sequence J to obtain a fifth random sequence E of the corresponding channel;

[0111] S4.4: Restore the fifth random sequence E of the three channels R, G and B into m×n matrices respectively, and obtain encryption matrices R′, G′ and B′ of the corresponding channels;

[0112] In step S4.1, the following operations are performed on the first random sequence A to obtain a third random sequence D:

[0113]

[0114] Among them, A i is the i-th element in the first random sequence A; D i is the i-th element in the third random sequence D;

[0115] The following operations are performed on the second random sequence C to obtain the fourth random sequence J:

[0116]

[0117] Among them, C i is the i-th element in the second random sequence C; J i is the i-th element in the fourth random sequence J;

[0118] In step S4.3, the scrambled matrices S of the three channels R, G and B are expanded into a one-dimensional sequence F, and the elements of the one-dimensional sequence F are XORed with the elements in the fourth random sequence J to obtain the fifth random sequence E of the corresponding channel, specifically:

[0119]

[0120] Among them, E i is the i-th element in the fifth random sequence E; F i is the i-th element in the one-dimensional sequence F; J i is the i-th element in the fourth random sequence J;

[0121] In step S5, the type of the carrier image Q is the same as the image type of the training reversible generative model Glow, and the size is s×q×3, wherein p and q are the third and fourth positive integers, and p and q are divisible by 32;

[0122] The sum of the sizes of the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4 is equal to s×q×3;

[0123] The sum of the sizes of the first, second and third hidden variables ε1, ε2 and ε3 is greater than the sum of all elements of the ciphertext image G to ensure complete hiding of the ciphertext information;

[0124] The step S6 includes:

[0125] The elements of the ciphertext image G are sampled from a standard normal distribution, the domain of which is [-4, 4], and the domain is evenly divided into 256 intervals, each of which has a length of 0.03125; the intervals are sorted from left to right, with serial numbers ranging from 0 to 255;

[0126] According to the value of the ciphertext image G element, the corresponding interval is found, and the probability density of the center point of the corresponding interval is taken as the probability density of the sampling. Finally, the probability density matrix L of the sampling is obtained, as shown below:

[0127]

[0128] L i =λ(-4+(1+G i )×0.03125-0.015625),i=1,2,…,m×n×3

[0129] Among them, G i is the i-th element in the ciphertext image G; L i is the i-th element in the sampled probability density matrix L;

[0130] The elements in ε1, ε2 and ε3 are replaced with the elements in the probability density matrix L in sequence to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ sampled.

[0131] In the specific implementation process, firstly, a four-dimensional memristor chaotic system NCSM is constructed;

[0132] In this embodiment, the memristor chaotic system NCSM is specifically:

[0133]

[0134]

[0135] Among them, a, b, c, k and ω are respectively the first to fifth preset hyperparameters, specifically a=0.31, b=4.5, c=5, k=-1.2, ω=0.199; x, y, z and u are respectively the first to fourth input variables of the memristor chaotic system NCSM; and They are the first to fourth output variables of the memristor chaotic system NCSM;

[0136] Then, the color plaintext image P is obtained and converted into three channel components R, G and B of red, green and blue. In this embodiment, the size of the color plaintext image P is m×n×3.

[0137] Randomly generate keys x0, y0, z0 and u0, and use them as initial values ​​of x, y, z and u and input them into the memristor chaotic system NCSM. Use the fourth-order-fifth-order Runge-Kutta algorithm to calculate the numerical solution of the memristor chaotic system NCSM, and obtain 4 random chaotic sequences. Save the first two random chaotic sequences as the first random sequence A and the second random sequence C. The lengths of the first random sequence A and the second random sequence C are both m×n.

[0138] The first random sequence A and the second random sequence C are preprocessed respectively to obtain a third random sequence D and a fourth random sequence J, which are specifically:

[0139] Perform the following operations on the first random sequence A to obtain the third random sequence D:

[0140]

[0141] Among them, A i is the i-th element in the first random sequence A; D i is the i-th element in the third random sequence D;

[0142] The following operations are performed on the second random sequence C to obtain the fourth random sequence J:

[0143]

[0144] Among them, C i is the i-th element in the second random sequence C; J i is the i-th element in the fourth random sequence J;

[0145] The third random sequence D is sorted in descending order to obtain an index sequence I, and the index sequence I is used to scramble and encrypt the three channels R, G, and B to obtain the scrambling matrix S of the corresponding channel. For example, the scrambling matrix S of the red channel is expressed as: The same goes for the green and blue channels;

[0146] The scrambled matrices S of the three channels R, G and B are expanded into a one-dimensional sequence F, and the elements of the one-dimensional sequence F are XORed with the elements in the fourth random sequence J one by one to obtain the fifth random sequence E of the corresponding channel, which is specifically:

[0147]

[0148] Among them, E i is the i-th element in the fifth random sequence E; F i is the i-th element in the one-dimensional sequence F; J i is the i-th element in the fourth random sequence J;

[0149] Restore the fifth random sequence E of the three channels R, G and B into m×n matrices respectively, and obtain encryption matrices R′, G′ and B′ of the corresponding channels;

[0150] Then, the encryption matrices R′, G′ and B′ of the three channels R, G and B are combined to obtain the color ciphertext image G with a size of m×n×3;

[0151] Get the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4;

[0152] like Figure 3 The structure diagram of the Glow model is shown below. Figure 4 The figure shows the structure of the Flow block in the Glow model. The expression of the Actnorm layer in the Flow block of the Glow model is Among them, α and β are input and output, and γ are trainable parameter matrices, ⊙ represents element-by-element multiplication; the expression of the Invertible 1×1Convolution layer is η=Ψσ, where η and σ are input and output, Ψ is a trainable parameter matrix and is reversible; the expression of the Affine coupling layer is:

[0153]

[0154] Among them, v = (v1, v2) and φ = (φ1, φ2) are input and output, ⊙ represents element-by-element multiplication, exp represents the exponential function, r and t are convolution functions;

[0155] In the Glow model, the Squeeze operation transforms the input data. Assuming the input data size is w×h×δ, the output data size after the Squeeze operation is w2×h2×4δ. The Split operation divides the input into two parts: one part is used as the output, and the other part continues to enter the next layer of calculation. Assuming the input data size is w2×h2×4δ, the output sizes of both parts are w2×h2×2δ.

[0156] Assume that the latent variable μ follows the distribution μ~ξ θ (μ), and ρ=g θ (μ), g θ is a reversible function, g θ The inverse of θ ,Right now In the Glow model, g θ That is the generator, which is composed of a series of functions, denoted as For the convenience of representation, the parameter θ is omitted here. When generating images, the relationship between μ and ρ is expressed as follows:

[0157]

[0158] The inverse process is expressed as follows:

[0159]

[0160] Where μ is a randomly generated distribution, such as a normal distribution, and ρ is the target distribution, then the objective function of the Glow model is:

[0161]

[0162] in The training loss of the Glow model is obtained by log maximum likelihood;

[0163] In this embodiment, the type of the carrier image Q is the same as the image type of the training reversible generative model Glow, and the size is s×q×3, where p and q are the third and fourth positive integers, and p and q are divisible by 32;

[0164] The sum of the sizes of the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4 is equal to s×q×3;

[0165] The sum of the sizes of the first, second and third hidden variables ε1, ε2 and ε3 is greater than the sum of all elements of the ciphertext image G to ensure complete hiding of the ciphertext information;

[0166] Then, the elements of the ciphertext image G are sampled from a standard normal distribution, the domain of which is [-4, 4], and the domain is evenly divided into 256 intervals, each of which has a length of 0.03125; the intervals are sorted from left to right, with serial numbers of 0 to 255;

[0167] According to the value of the ciphertext image G element, the corresponding interval is found, and the probability density of the center point of the corresponding interval is taken as the probability density of the sampling. Finally, the probability density matrix L of the sampling is obtained, as shown below:

[0168]

[0169] L i =λ(-4+(1+G i )×0.03125-0.015625),i=1,2,…,m×n×3

[0170] Among them, G i is the i-th element in the ciphertext image G; L i is the i-th element in the sampled probability density matrix L;

[0171] Replace the elements in ε1, ε2 and ε3 with the elements in the probability density matrix L in order to obtain the first, second and third latent variables ε1′, ε2′ and ε3′;

[0172] Finally, ε1′, ε2′, ε3′ and ε4 are combined as the forward input of the pre-trained reversible generation model Glow, and the generated stego-image H is finally obtained to complete the image encryption;

[0173] After the encryption is completed, the initial values ​​x0, y0, z0 and u0 of the memristive chaotic system NCSM are used as keys, and the keys and the generated stego-image H are sent to the communication receiver using an asymmetric encryption algorithm. The communication receiver can decrypt the generated stego-image H using the keys;

[0174] like Figure 5 As shown as the carrier image Q, Figure 6 As the plaintext image P, after image steganography by this method, we get Figure 7 The secret carrier image, i.e., the generated stego-image H, shows that the method proposed in this embodiment is effective and has a good image information hiding effect;

[0175] This method implants a memristor into an existing chaotic system to construct a new memristor chaotic system NCSM. The chaotic sequence generated by this system has good randomness and is applied to image encryption. In terms of image steganography, this method converts the ciphertext image into a generative stego image through a generative reversible neural network model, which ensures the information security of the plaintext image while also ensuring the visual security of the ciphertext image. Unlike the current non-reversible adversarial generation model, the generative reversible neural network does not need to train a decoder, thereby reducing the complexity of the model and the difficulty of model training. Furthermore, this method optimizes the input of the reversible generation model, so that it has a certain control ability over the content of the generative stego image.

[0176] Example 3

[0177] like Figure 8 As shown, this embodiment provides a generative stego image recovery method for recovering the generative stego image H generated by Embodiment 2.

[0178] In the specific implementation process, the generated steganalysis image H is firstly input into the Glow model in reverse to obtain its corresponding hidden variables ε1′, ε2′, ε3′ and ε4, and then the hidden variables ε1′, ε2′ and ε3′ are inversely sampled to obtain hidden variables ε1, ε2 and ε3, and ε1, ε2, ε3 and ε4 are combined to obtain the ciphertext image G;

[0179] After decrypting the received key, the initial value of the memristor chaotic system NCSM is obtained. After the initial value is input into the memristor chaotic system NCSM for iteration to generate a chaotic sequence, the ciphertext image G is inversely diffused and inversely scrambled to obtain the plaintext image P, as shown in Fig. 9 The figure shows the restored plaintext image P. It can be seen that the method in this embodiment can correctly restore the plaintext image.

[0180] In this embodiment, the generative reversible neural network Glow not only does not need to train a decoder, reducing the complexity of the model and the difficulty of training the model, but also reduces the restoration error, so that the final restored image presents a better effect.

[0181] The same or similar reference numerals correspond to the same or similar components;

[0182] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;

[0183] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A generative image information steganography method, characterized in that: The following steps are involved: S1: Construction of a four-dimensional memristor chaotic system NCSM; S2: Obtain the color plaintext image P and convert it into three channel components R, G and B of red, green and blue; S3: randomly generate a key and input it into the memristor chaotic system NCSM for iteration to generate a first random sequence A and a second random sequence C; S4: According to the first random sequence A and the second random sequence C, the three channels R, G and B are encrypted respectively to obtain encryption matrices R′, G′ and B′; Combine the encryption matrices R′, G′ and B′ of the three channels R, G and B to obtain the color ciphertext image G; S5: Obtain the carrier image Q and reversely input it into the pre-trained reversible generative model Glow to obtain the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4; S6: sampling the elements of the ciphertext image G in a standard normal distribution to obtain a sampled probability density matrix L, and sequentially replacing the elements in ε1, ε2 and ε3 with the elements in the probability density matrix L to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ of the sampling; S7: ε1′, ε2′, ε3′ and ε4 are combined as the forward input of the pre-trained reversible generation model Glow, and finally the generated stego-image H is obtained.

2. A generative image information steganography method according to claim 1, characterized in that: In step S1, the memristor chaotic system NCSM is specifically: Wherein, a, b, c, k and ω are respectively the first to fifth preset hyperparameters; x, y, z and u are respectively the first to fourth input variables of the memristor chaotic system NCSM; and They are the first to fourth output variables of the memristor chaotic system NCSM respectively.

3. A generative image information steganography method according to claim 2, characterized in that: The step S3 comprises: The size of the color plaintext image P is m×n×3; wherein m and n are first and second positive integers; Randomly generate keys x0, y0, z0 and u0 and input them as initial values ​​of x, y, z and u into the memristor chaotic system NCSM. Use the fourth-order-fifth-order Runge-Kutta algorithm to calculate the numerical solution of the memristor chaotic system NCSM, and obtain 4 random chaotic sequences. Save the first two random chaotic sequences as the first random sequence A and the second random sequence C. The lengths of the first random sequence A and the second random sequence C are both m×n.

4. A generative image information steganography method according to claim 3, characterized in that: The step S4 comprises: S4.1: Preprocess the first random sequence A and the second random sequence C respectively to obtain a third random sequence D and a fourth random sequence J; S4.2: Sort the third random sequence D to obtain an index sequence I, and use the index sequence I to scramble and encrypt the three channels R, G and B respectively to obtain a scramble matrix S for the corresponding channel; S4.3: Expand the scrambled matrices S of the three channels R, G and B into a one-dimensional sequence F, and perform XOR operation on the elements of the one-dimensional sequence F and the elements in the fourth random sequence J to obtain a fifth random sequence E of the corresponding channel; S4.4: Restore the fifth random sequence E of the three channels R, G and B into m×n matrices respectively, and obtain encryption matrices R′, G′ and B′ of the corresponding channels.

5. A generative image information steganography method according to claim 4, characterized in that: In step S4.1, the following operations are performed on the first random sequence A to obtain a third random sequence D: Among them, A i is the i-th element in the first random sequence A; D i is the i-th element in the third random sequence D; The following operations are performed on the second random sequence C to obtain the fourth random sequence J: Among them, C i is the i-th element in the second random sequence C; J i is the i-th element in the fourth random sequence J.

6. A generative image information steganography method according to claim 4, characterized in that: In the step S4.2, the third random sequence D is sorted in descending order to obtain an index sequence I.

7. A generative image information steganography method according to claim 4, characterized in that: In step S4.3, the scrambled matrices S of the three channels R, G and B are expanded into a one-dimensional sequence F, and the elements of the one-dimensional sequence F are XORed with the elements in the fourth random sequence J to obtain the fifth random sequence E of the corresponding channel, specifically: Among them, E i is the i-th element in the fifth random sequence E; F i is the i-th element in the one-dimensional sequence F; J i is the i-th element in the fourth random sequence J.

8. A generative image information steganography method according to claim 1, characterized in that: In step S5, the type of the carrier image Q is the same as the image type of the training reversible generative model Glow, and the size is s×q×3, wherein p and q are the third and fourth positive integers, and p and q are divisible by 32; The sum of the sizes of the first, second, third and fourth latent variables ε1, ε2, ε3 and ε4 is equal to s×q×3; The sum of the sizes of the first, second and third latent variables ε1, ε2 and ε3 is greater than the sum of all elements of the ciphertext image G to ensure the complete hiding of the ciphertext information.

9. A generative image information steganography method according to claim 1, characterized in that: The step S6 includes: The elements of the ciphertext image G are sampled from a standard normal distribution, the domain of which is [-4, 4], and the domain is evenly divided into 256 intervals, each of which has a length of 0.03125; the intervals are sorted from left to right, with serial numbers ranging from 0 to 255; According to the value of the ciphertext image G element, the corresponding interval is found, and the probability density of the center point of the corresponding interval is taken as the probability density of the sampling. Finally, the probability density matrix L of the sampling is obtained, as shown below: L i =λ(-4+(1+G i )×0.03125-0.015625),i=1,2,…,m×n×3 Among them, G i is the i-th element in the ciphertext image G; L i is the i-th element in the sampled probability density matrix L; The elements in ε1, ε2 and ε3 are replaced with the elements in the probability density matrix L in sequence to obtain the first, second and third latent variables ε1′, ε2′ and ε3′ sampled.

10. A generative image information steganography method according to any one of claims 1 to 9, characterized in that: The method further comprises: S8: Use an asymmetric encryption algorithm to send the key and the generated stego-image H to the communication recipient.

Images