Generative industrial production image data steganography method

Abstract

The invention discloses a generative industrial production image data steganography method, which belongs to the technical field of information hiding, and comprises the following steps: S1, carrying out BCH channel coding preprocessing on industrial secret information, and generating an information sequence with redundancy check; s2, based on StyleGAN2, fusing the information sequence in the S1 with the injected noise to generate a carrier image; and S3, carrying out block matching and transformation rotation on the carrier image generated in the step S2 to realize reversible fuzzification and embedding authentication information. According to the generative industrial production image data steganography method, transmission interference is resisted through an error correction mechanism, the image is close to a natural state to avoid detection, the position of hidden information is fuzzified reversibly, authentication information is matched, and completeness and verifiability are guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of information hiding technology, specifically relating to a generative industrial production image data steganography method. Background Technology

[0002] Information hiding aims to provide a secure and covert image steganography solution. By embedding industrial secrets into the generated carrier image, it ensures the invisibility and imperceptibility of the secret information during image transmission. Its application scenarios are primarily in the industrial production field, suitable for scenarios requiring the protection of sensitive data transmission, such as information hiding in industrial images like production line monitoring diagrams, inspection images, and product quality inspection images. It can seamlessly embed high-value secret information such as production line information, safety tags, production batches, material tracking, and control commands into industrial images, meeting the information security transmission needs of quality inspection, fault diagnosis, and intelligent control processes.

[0003] However, existing technologies have the following shortcomings: traditional image steganography methods (such as LSB least significant bit substitution) are weak in resistance to compression and cropping, and lack the ability to resist steganalysis attacks, making it difficult to meet the high requirements of industrial images for stability and security; in generative steganography methods based on generative adversarial networks, the embedded secret information lacks robustness and is easily damaged by image processing such as compression and noise; there is a lack of secondary protection for the location and content of secret information, making it easy for attackers to maliciously tamper with it or cause information leakage.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a generative industrial production image data steganography method. This method resists transmission interference through error correction mechanism, makes the image close to its natural state to avoid detection, and reversibly blurs the location of hidden information. Combined with authentication information, it ensures completeness and verifiability.

[0006] To achieve the above objectives, the present invention provides a generative industrial production image data steganography method, comprising the following steps:

[0007] S1. Perform BCH channel coding preprocessing on industrial secret information to generate an information sequence with redundancy check;

[0008] S2. Based on StyleGAN2, the information sequence in S1 is fused with injected noise to generate a carrier image;

[0009] S3. The carrier image generated in S2 is reversibly blurred and the authentication information is embedded through block matching, transformation and rotation.

[0010] Preferably, in step S101, industrial secret information is collected and converted into a binary sequence. The bit stream is converted into a binary sequence of length k, represented as follows:

[0011] m = {m0, m1, m2, ..., m} k-1},m k ∈{0,1};

[0012] Where m represents the entire binary sequence of the original secret information; m k It represents the k-th binary value in the sequence;

[0013] S102. Construct the information polynomial corresponding to the original secret information, expressed as:

[0014] m(x) = m0 + m1x + m2x 2 +...+m 130 x 130 ;

[0015] Where m(x) is the polynomial corresponding to the original secret information bit stream of length 131 bits;

[0016] S103. Construct the generator polynomial of the BCH encoding, expressed as:

[0017] g(x)=g 124 x 124 +g 123 x 123 +…+g1x+g0;

[0018] Where g(x) is the constructive generator polynomial;

[0019] S104. Perform the BCH encoding operation to generate the complete encoded polynomial C(x);

[0020] S105. Convert the complete coding polynomial C(x) obtained in S104 into a binary codeword sequence of length n, represented as:

[0021] C = {c0, c1, c2, ..., c} 254},c i ∈{0,1};

[0022] Where C is the complete codeword sequence with a length of 255.

[0023] Preferably, S2 includes the following steps:

[0024] S201. Select the pre-trained StyleGAN2 as the image generator to generate carrier images;

[0025] S202. The secret information encoded by BCH is modulated with the injected noise of StyleGAN2 using distribution-preserving modulation. The modulated noise is then injected into the synthesis network to generate the carrier image. The information sequence C in S1 is reshaped into a binary secret information matrix c. The noise is then modulated according to the following formula:

[0026]

[0027] in, For element-wise Hadamard products; |N K | To preserve the noise amplitude distribution; c is the binary secret information matrix to be embedded; N′ k The original Gaussian noise injected into the k-th layer follows a normal distribution.

[0028] Preferably, in S3, block matching involves dividing the carrier image and the reference image into pairs of blocks of the same size, classifying them according to their standard deviation, and then matching them one by one.

[0029] Preferably, in S3, the transformation rotation includes linear transformation and rotation optimization. The calculation formula for the linear transformation is:

[0030]

[0031] y = μ B -x·μ A ;

[0032] Where, σ B Reference block B i Standard deviation; μ B Reference block B i The mean; σ A For carrier block A i Standard deviation; μ A For carrier block A i The mean; x is the scaling factor; y is the offset factor;

[0033] The formula for calculating rotation optimization is:

[0034] β * =argminRMSE β ;

[0035] Where β is the parameter variable to be optimized; β * The optimal parameters are: RMSE (root mean square error).

[0036] The present invention also provides a generative industrial production image data steganography system, comprising:

[0037] The encoding preprocessing module is used to perform BCH channel coding preprocessing on industrial secret information to generate an information sequence with redundancy check.

[0038] The steganography image generation module is connected to the encoding preprocessing module and fuses the information sequence with injected noise based on StyleGAN2 to generate a carrier image.

[0039] The reversible blurring and authentication embedding module is connected to the steganography image generation module and is used to reversibly blur the carrier image and embed authentication information by segmenting, matching, transforming and rotating the image.

[0040] Therefore, the present invention employs the above-mentioned generative industrial production image data steganography method, and compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0041] (1) This invention uses BCH channel coding to preprocess secret information and introduces an error correction mechanism to effectively resist noise, compression and other interference in industrial transmission. It solves the problem that secret information is easily destroyed and decoding fails in the prior art and adapts to the transmission needs of complex industrial environments.

[0042] (2) This invention fuses secret information with injected noise from StyleGAN2 using a distribution-preserving modulation method, making the generated image distribution highly approximate natural images. This solves the problem that existing generative steganographic images are easily detected due to their structural regularity and are difficult to be identified by traditional steganalysis tools.

[0043] (3) The present invention designs a reversible blurring process, which makes the steganographic image visually close to any reference image, destroys the predictability between pixels, and prevents attackers from locating the embedded information. At the same time, the authentication information is embedded through the LSB method to ensure the integrity and verifiability of the image, and solves the problems of the lack of secondary protection and easy tampering or leakage in the prior art.

[0044] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating an embodiment of a generative industrial production image data steganography method according to the present invention;

[0046] Figure 2 This is a specific architecture of the StyleGAN2 generative model, which is an embodiment of the generative industrial production image data steganography method of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0048] Example 1

[0049] like Figure 1 As shown, a generative industrial production image data steganography method of the present invention includes the following steps:

[0050] S1. Perform BCH channel coding preprocessing on industrial secret information to generate an information sequence with redundancy check; including the following steps:

[0051] S101. Collect industrial secret information and convert it into a binary sequence;

[0052] In this step, confidential information that needs to be protected during industrial production is collected, and its bit stream is converted into a binary sequence of length k, represented as:

[0053] m = {m0, m1, m2, ..., m} k-1},m k ∈{0,1};

[0054] Where m represents the entire binary sequence of the original secret information; m k It represents the k-th binary value in the sequence;

[0055] S102. Construct the information polynomial corresponding to the original secret information; construct the information polynomial based on the binary sequence obtained in S101, expressed as:

[0056] m(x) = m0 + m1x + m2x 2 +...+m 130 x 130 ;

[0057] Where m(x) is the polynomial corresponding to the original secret information bit stream of length 131 bits;

[0058] S103. Construct the generator polynomial for BCH encoding; for the selected BCH encoding scheme (such as BCH(255,131)), its coefficients can be automatically generated by consulting the BCH encoding standard specification or using the MATLAB general function bchgenpoly(255,131); in this embodiment, the generator polynomial is expressed as:

[0059] g(x)=g 124 x 124 +g 123 x 123 +…+g1x+g0;

[0060] Where g(x) is the constructive generator polynomial;

[0061] S104. Perform BCH encoding operation to generate a complete encoded polynomial;

[0062] In this step, the information polynomial is shifted left by 124 bits to obtain:

[0063] x 124 m(x) = m0x 124 +m1x 125 +...+m 130 x 254 ;

[0064] Use x 124 Taking the modulus of m(x) with respect to the generator polynomial g(x), we obtain the check polynomial, which is expressed as:

[0065] r(x)=(x 124 *m(x))modg(x);

[0066] Construct the complete encoded polynomial, represented as:

[0067] C(x)=x 124 *m(x)+r(x);

[0068] Where C(x) is the complete encoding polynomial;

[0069] S105. Output the binary codeword sequence after BCH encoding; convert the complete encoding polynomial C(x) obtained in S104 into a binary codeword sequence of length n, represented as:

[0070] C = {c0, c1, c2, ..., c} 254},c i ∈{0,1};

[0071] The complete codeword sequence C after BCH(255,131) encoding is obtained, with a length of 255;

[0072] S2, such as Figure 2 As shown, using a pre-trained StyleGAN2, the information sequence C in S1 is fused with injected noise through distribution-preserving modulation to generate a carrier image; the steps include:

[0073] S201. Select the pre-trained StyleGAN2 as the image generator to generate carrier images;

[0074] In this step, the pre-trained StyleGAN2 model is selected to start the image generation process. First, the input 1×512-dimensional standard normal distribution latent vector z is normalized by the Normalize layer to ensure distribution stability. Then, the normalized z is fed into the style encoder f, which consists of an 8-layer fully connected network, and mapped to a low-dimensional style vector w. Then, w is processed by a multi-layer independent FC network to output multiple sets of style vectors A, which correspond one-to-one with the AdaIN layers of each layer of the synthesis network g, and are used to control the local style and texture of the image.

[0075] S202. The secret information encoded by BCH is modulated with the injected noise of StyleGAN2 in a distribution-preserving manner, and the modulated noise is injected into the synthesis network to generate the carrier image.

[0076] In this step, each level of the synthetic network g is equipped with a noise injection node B, and the injected noise N is... k The Gaussian noise matrix, such as 4×4×512 or 8×8×256, is matched to the dimension of the hierarchical feature map and follows a standard normal distribution.

[0077] Reshape the information sequence C in S1 into a binary secret information matrix c; and modulate the noise according to the following formula:

[0078]

[0079] in, For element-wise Hadamard products; |N K | To preserve the noise amplitude distribution; c is the binary secret information matrix to be embedded; N′ k The original Gaussian noise injected into the k-th layer follows a normal distribution; 2*c-1 maps the secret information 0 / 1 to -1 / 1 to flip the noise symbol, ensuring that the modulated distribution is approximately consistent with the original noise;

[0080] During the noise injection and image generation stages, pure Gaussian noise N is directly injected into the non-final layer B nodes. K The AdaIN layer, 3×3 convolutional layer, and upsampling layer driven by style vector A work together to construct image features layer by layer. Finally, the B node injects modulated noise to replace the pure noise N′. k After fusing the secret information, the module generates a high-resolution carrier image such as 1024×1024, thus completing the steganographic embedding of the secret information.

[0081] S3. The generated carrier image is reversibly blurred and authentication information is embedded through block matching, transformation, and rotation; including the following steps:

[0082] S301. Divide the carrier image containing the secret information and the reference image used for feature matching into an equal number of non-overlapping / overlapping blocks of the same size to form a one-to-one block pair set, specifically:

[0083] Choose any reference image R, and divide the carrier image O and the reference image R into N blocks of the same size; where the blocks obtained by dividing the carrier image O are denoted as A. i The block obtained by segmenting the reference image R is denoted as B. i ; each A i With the corresponding B i Pairing up to form block pairs (A) i B i );

[0084] S302, Traverse all carrier blocks A i and reference block B i And calculate the pixel standard deviation for each block, using the following formula:

[0085]

[0086] Where σ is the standard deviation of pixels in the image patch; p i Let be the gray value of the i-th pixel in the image block;

[0087] And calculate the average pixel value of each block, using the following formula:

[0088]

[0089] Using the α% quantile of the standard deviation of all blocks as a threshold (α = 50% in this embodiment), the blocks are divided into two categories:

[0090] Class0 represents σ∈[0,N] α ]; where N α The α percentile of the standard deviation;

[0091] Class1 represents σ∈[N] α N 100 ];

[0092] In Class 0, the carrier block and the reference block are paired sequentially, and the same applies to Class 1, to ensure that blocks of the same complexity category are processed subsequently, and to avoid texture differences interfering with the transformation effect.

[0093] S303. Taking Class0 block as an example, Class1 logic is consistent. Calculate the carrier block A in Class0 respectively. i Reference block B i The mean and standard deviation are calculated and normalized; according to the formula:

[0094]

[0095] y = μ B -x·μ A ;

[0096] Where, σ B Reference block B i Standard deviation; μ B Reference block B i The mean; σ A For carrier block A i Standard deviation; μ A For carrier block A i The mean; x is the scaling factor; y is the offset factor;

[0097] The normalized data is modulated to make the statistical characteristics of the new block approximate those of the reference block. The calculation formula is as follows:

[0098]

[0099] in, For carrier block A i Modulated pixel values; These are the original pixel values ​​before the transformation;

[0100] Crop to the effective pixel range, such as 0-255, to avoid overflow. The calculation formula is:

[0101]

[0102] Then, a new carrier block is generated through inverse normalization. The calculation formula is as follows:

[0103]

[0104] in, For the final image pixels;

[0105] Then, the modulation parameters are saved for use in the reverse process of subsequent reversible blurring;

[0106] S304, Regarding the modulated new carrier block B i ', Perform rotation angle optimization, for each rotated block, calculate the angle relative to the corresponding reference block B. i The root mean square error (RMSE) is calculated using the following formula:

[0107]

[0108] Among them, R β (B) i,j For B i The rotated block; T i,jM×N represents the pixel value at the corresponding position of the target block in the reference image; M×N is the size of the image block.

[0109] Choose the rotation angle that minimizes RMSE, calculated as follows:

[0110] β * =argminRMSE β ;

[0111] Where β is the parameter variable to be optimized; β * The optimal parameters;

[0112] Will Replace the corresponding block in the reference image to construct the camouflaged image R′;

[0113] S305, Hash authentication information generation and LSB steganography embedding;

[0114] In this step, a fixed-length authentication digest is generated from the original secret information using a hash function, serving as an identifier for integrity verification; for each B... i The pixel value of ' is subjected to an LSB replacement operation, which embeds the k-th bit of the authentication information into the least significant bit (LSB) of the j-th pixel in the i-th block. The calculation formula is as follows:

[0115] p″ i,j =(p″ i,j &0xFE)|b k ;

[0116] Among them, b k The k-th bit of the authentication information; 0xFE is the hexadecimal mask; p″ i,j For the transformed carrier block B i The original value of the j-th pixel in ';

[0117] After the replacement is completed, B i The pixel values ​​store authentication information, and the embedded image serves as the authentication image for the recipient to verify authenticity.

[0118] Therefore, the present invention adopts the above-mentioned generative industrial production image data steganography method. This method resists transmission interference through error correction mechanism, makes the image close to the natural state to avoid detection, and reversibly blurs the hidden information location. Combined with authentication information, it ensures completeness and verifiability.

[0119] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A generative industrial production image data steganography method, characterized in that, Includes the following steps: S1. Perform BCH channel coding preprocessing on industrial secret information to generate an information sequence with redundancy check; S2. Based on StyleGAN2, the information sequence in S1 is fused with injected noise to generate a carrier image; S3. The carrier image generated in S2 is reversibly blurred and the authentication information is embedded through block matching, transformation and rotation.

2. The generative industrial production image data steganography method according to claim 1, characterized in that, S1 includes the following steps: S101. Collect industrial secret information and convert it into a binary sequence. Convert the bit stream into a binary sequence of length k, represented as: m={m0,m1,m2,…,m k-1 },m k ∈{0,1}; Where m represents the entire binary sequence of the original secret information; m k It represents the k-th binary value in the sequence; S102. Construct the information polynomial corresponding to the original secret information, expressed as: m(x)=m0+m1x+m2x 2 +...+m 130 x 130 ; Where m(x) is the polynomial corresponding to the original secret information bit stream of length 131 bits; S103. Construct the generator polynomial of the BCH encoding, expressed as: g(x)=g 124 x 124 +g 123 x 123 +…+g1x+g0; Where g(x) is the constructive generator polynomial; S104. Perform the BCH encoding operation to generate the complete encoded polynomial C(x); S105. Convert the complete coding polynomial C(x) obtained in S104 into a binary codeword sequence of length n, represented as: C={c0,c1,c2,…,c 254 },c i ∈{0,1}; Where C is the complete codeword sequence with a length of 255.

3. The generative industrial production image data steganography method according to claim 1, characterized in that, S2 includes the following steps: S201. Select the pre-trained StyleGAN2 as the image generator to generate carrier images; S202. The secret information encoded by BCH is modulated with the injected noise of StyleGAN2 using distribution-preserving modulation. The modulated noise is then injected into the synthesis network to generate the carrier image. The information sequence C in S1 is reshaped into a binary secret information matrix c. The noise is then modulated according to the following formula: in, For element-wise Hadamard products; |N K | To preserve the noise amplitude distribution; c is the binary secret information matrix to be embedded; N′ k The original Gaussian noise injected into the k-th layer follows a normal distribution.

4. The generative industrial production image data steganography method according to claim 1, characterized in that, In S3, block matching involves dividing the carrier image and the reference image into pairs of blocks of the same size, classifying them according to their standard deviation, and then matching them one by one.

5. The generative industrial production image data steganography method according to claim 1, characterized in that, In S3, the transformation rotation includes linear transformation and rotation optimization. The formula for calculating the linear transformation is: y=μ B -x*μ A ; Where, σ B Reference block B i Standard deviation; μ B Reference block B i The mean; σ A For carrier block A i Standard deviation; μ A For carrier block A i The mean; x is the scaling factor; y is the offset factor; The formula for calculating rotation optimization is: β * argminRMSE β ; Where β is the parameter variable to be optimized; β * The optimal parameters are: RMSE (root mean square error).

6. A generative industrial production image data steganography system, applied to the generative industrial production image data steganography method according to any one of claims 1-5, characterized in that, include: The encoding preprocessing module is used to perform BCH channel coding preprocessing on industrial secret information to generate an information sequence with redundancy check. The steganography image generation module is connected to the encoding preprocessing module and fuses the information sequence with injected noise based on StyleGAN2 to generate a carrier image. The reversible blurring and authentication embedding module is connected to the steganography image generation module and is used to reversibly blur the carrier image and embed authentication information by segmenting, matching, transforming and rotating the image.

Images