Adaptive medical image reversible information hiding method based on adversarial lifting
By adversarial training, adaptive positioning information embedded in location and combined with image repair, the problem of degradation in quality of traditional medical image information hiding algorithms is solved, and efficient, large-capacity and safe information hiding and recovery are achieved.
Patent Information
- Application Number
- CN202510614350.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-08-01
AI Technical Summary
After hiding information, traditional medical image information hiding algorithms will reduce the quality of the original medical image, and the image accuracy and information accuracy will become lower after the receiver extracts the hidden information, which will not effectively resist malicious attacks by third parties.
Adaptive medical image reversible information hiding method is adopted to train adaptive positioning information embedded in position through adversarial training, combined with image repair function, improve the capacity and image quality of hidden information, and recover hidden information losslessly at the receiver.
It realizes large-capacity information hiding while ensuring the integrity of medical images, improves the security and image quality of information hiding algorithms, and has the ability to resist attacks.
Smart Images
Figure CN120408681A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an adaptive reversible medical image information hiding method based on adversarial enhancement, belonging to the fields of computer medical image processing and information security. Background Art
[0002] Medical images transmitted in a Hospital Information System (HIS), a Picture Archiving and Communication System, and an Internet of Medical Things (IoMT) contain various important data such as patient sensitive information, medical examination device information, and hospital information, and are vulnerable to attacks by illegal organizations and individuals. When a medical imaging device scans a patient, the generated image is stored in a Picture Archiving and Communication System and then transmitted to a workstation. Doctors can comprehensively analyze them through other information in the Hospital Information System. These information can also be transmitted through the Internet of Medical Things and telemedicine in intelligent medicine for diagnosing, treating, and consulting the injured and sick in remote areas, islands, or ships with poor medical conditions, etc., and also provide the possibility for patients to view examination results and pathology reports outside the hospital. If the information is illegally steganographed, tampered with, or lost during transmission and storage, it will directly affect the consultation result, lead to the leakage of patient privacy, and trigger serious medical incidents. Information hiding technology is of great significance for the secure transmission of medical images.
[0003] The relationship between information detection and information hiding is similar to the relationship between a "spear" and a "shield". Information detection and information hiding technologies can optimize each other and improve through game theory. Medical image information hiding technology can hide information related to medical images in the redundant space of medical images while ensuring that this operation does not cause distortion of the structure and organization of medical images and affect the clinical treatment of doctors. The relevant information hidden by an excellent medical image information hiding algorithm cannot be discovered by third-party attackers, and the legitimate receiver can recover the hidden information losslessly when receiving the carrier image containing the secret information. At the same time, this information hiding technology can effectively resist malicious attacks and damage from third parties.
[0004] Traditional medical image information hiding algorithms will reduce the quality of the original medical image after hiding information, and both the image accuracy and information accuracy will become lower after the hidden information is extracted by the receiver of the medical image. The proposed adaptive reversible medical image information hiding method based on adversarial enhancement in the present invention uses an illegal hidden information detection algorithm as a discriminator to adversarially train and adaptively locate the position of information embedding in the original medical image, improving the capacity of the hidden information. In addition, the proposed method also introduces an image restoration function, which can improve the overall image quality and the signal-to-noise ratio of the lesion after key information is embedded, and can restore the hidden information and the original medical image losslessly at the receiver. Summary of the Invention
[0005] The purpose of the present invention is to provide an adversarial-enhanced adaptive reversible medical image information hiding method, which is used to achieve key information hiding and lossless recovery in the task of transmitting medical images in the IoMT telemedicine network with a relatively high hiding capacity.
[0006] The present invention adopts the following technical solutions to solve the above problems:
[0007] The present invention provides an adversarial-enhanced adaptive reversible medical image information hiding method, which can achieve key information hiding during the transmission of DICOM medical images with a relatively high hiding capacity. The method includes the following steps:
[0008] Step 1, import the original DICOM-format medical image. Specifically, import the DICOM-format medical image that needs information hiding. The medical image includes a standard file header for storing information and medical image data with a size of 512×512.
[0009] Step 2, extract the standard file header and medical image data from the DICOM-format medical image. Specifically, separate the standard file header and 512×512 medical image data contained in the DICOM file and evenly divide the medical image into low-resolution images.
[0010] Step 3, input the low-resolution image into a predictor to obtain a restored high-resolution image. Specifically, evenly divide the 512×512 medical image into lower-resolution medical images with a size of 256×256 and input them into the predictor to obtain a restored high-resolution image with a size of 512×512.
[0011] Step 4, input the medical image, the restored high-resolution image, and the standard file header into an embedder to generate a cipher-containing medical image. Specifically, the embedder receives the standard file header and 512×512 medical image output in Step 2 and the restored 512×512-size high-resolution image output in Step 3, determines the positions suitable for embedding information through a position selection network according to the interpolated and restored image, and encodes the information in the standard file header into a bit stream and embeds it into the cipher-containing medical image in combination with the method of prediction error expansion.
[0012] Step 5, synthesize the cipher-containing medical image and a false file header into a cipher-containing DICOM-format medical image. Specifically, the embedded cipher-containing medical image is combined with the false file header hiding the key information to generate a cipher-containing DICOM medical image, which hides the key information in the original DICOM image and does not affect the normal use of the image.
[0013] Among them, the network structure to be trained in step 3 is as follows: The predictor RM in step 3 includes three main modules: a feature extraction layer, a feature enhancement layer, and an upsampling layer, which can realize the restoration of medical image information hiding with any magnification factor. The feature extraction layer contains a convolutional kernel with a size of 3×3 and a stride of 2. The feature enhancement layer includes 64 feature enhancement modules FE and an RG residual group structure. Each layer of the feature enhancement layer FE contains 4 distillation modules and 1 spatial attention module. The distillation module consists of a convolutional kernel with a size of 3×3 and a stride of 2 and a Relu activation function layer. The spatial attention module consists of a max pooling layer, an average pooling layer, and a sigmoid activation layer. Finally, it passes through a convolutional kernel with a size of 7×7 and a stride of 2. Each RG residual group structure consists of a convolutional kernel with a size of 3×3 and a stride of 2, a ReLu activation function layer, a convolutional kernel with a size of 3×3 and a stride of 2, an average pooling layer, a convolutional kernel with a size of 3×3 and a stride of 2 and a ReLu activation function layer, a convolutional kernel with a size of 3×3 and a stride of 2, and a sigmoid activation layer. The upsampler uses a flow-based upsampling module to perform image upsampling by learning the offset grid, and the obtained deep image features are input into the upsampling module to obtain the final restored image.
[0014] Among them, the network structure to be trained in step 4 is as follows: The embedder in step 4 consists of 15 feature extraction blocks and 1 deconvolution layer. The first 8 feature extraction blocks are responsible for downsampling the image, and the last 7 feature extraction blocks are responsible for upsampling. The output of the i-th feature extraction block is connected to its corresponding input (the 16-i-th block) and then passed to the next feature extraction block. Each feature extraction block includes 1 convolutional layer, 2 normalization layers, and 2 LeakyReLU activation layers, and these layers are staggered to form the feature extraction block.
[0015] The network loss function L(θ) to be trained in step 3 is:
[0016]
[0017] where θ are the parameters included in the network, N is the total number of samples in the training set used, LR i and HR i are the i-th batch of data selected from the training set.
[0018] The network loss functions L(θ G ) and L(θ D ) for step 4 are
[0019]
[0020] The cross-entropy loss function crossEntropy is used to optimize the parameters of both the generator and the discriminator. Let the parameters of the generator be θ G , and the parameters of the discriminator be θ D . Among them, I B represents the non-encrypted image, I steo represents the encrypted image, and L out is the discrimination result obtained by the discriminator. I B represents the label of the non-encrypted image, and L steo represents the label of the encrypted image. The purpose of the generator is to generate images that the discriminator cannot distinguish, and the constructed label is exactly the opposite of the label of the input image. The purpose of the discriminator is to distinguish whether the input image is an encrypted image, and the constructed label is the same as the label of the input image.
[0021] Under the same prior conditions, the peak signal-to-noise ratio and structural similarity are used to measure the restoration effect of the restoration model on low-resolution images. The following formula shows:
[0022]
[0023] The peak signal-to-noise ratio (PSNR) is a commonly used metric for measuring image quality. It is usually used to compare the similarity between the original image and the processed image (such as compression, embedding, etc.), or to compare the similarity between two images. Among them, I represents the original image, and P represents the image after embedding information.
[0024] w and h represent the width and height of the beam image, I(i,j) and P(i,j) represent the pixel values at the position (i,j), MSE is the average of the squares of the differences in pixel values between the original image and the processed image, MAX is the maximum pixel value that can appear in a single-channel image, and the larger the value of PSNR, the higher the similarity between the image and the original image, the lower the distortion degree, and the better the quality.
[0025] The structural similarity (SSIM) is another metric for measuring image quality, aiming to simulate the perception of the human visual system for images. Among them, μ f and μ g represent the means of images f and g respectively, σ f and σ g are the corresponding standard deviations, and σ fg is the covariance between the two images; C1 and C2 are two constants: C1 = (0.01 × LG) 2 , C2 = (0.03 × LG) 2, LG is the gray value range of the image.
[0026] Under the same prior conditions, the embedding capacity is measured by the embedding rate, which is expressed by the following formula:
[0027]
[0028] If w and h represent the width and height of the image, and num is the total number of bits of the embedded information, the embedding rate R can be obtained.
[0029] In steps 3 and 4, the network model parameters to be trained are iteratively updated by the optimizer.
[0030] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the proposed adaptive reversible medical image information hiding method based on adversarial enhancement.
[0031] A computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, they implement the proposed adaptive reversible medical image information hiding method based on adversarial enhancement.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0033] 1. The proposed adaptive reversible medical image information hiding method based on adversarial enhancement in the present invention obtains a model with a large embedding capacity after training with a certain amount of data. Experiments prove that the present invention can embed a large amount of information in medical images while ensuring the integrity of medical images, and has a very significant advantage compared with information hiding algorithms.
[0034] 2. The predictor in the present invention uses the technology in the field of image inpainting. Experiments show that on a public dataset, the image generated by the inpainting model predictor has a better effect compared with the image generated by the traditional predictor, making the prediction error histogram for embedding more concentrated and steeper, thus improving the performance of the information hiding algorithm.
[0035] 3. The embedding part in the present invention introduces an adversarial training embedding network, enabling the network to learn how to find more suitable positions to better hide information, further enhancing the security of the information hiding algorithm. In the subsequent embedding stage, by introducing the embedding part enhanced by adversarial training, the embedding effect has been greatly improved compared with the traditional reversible hiding algorithm, and it can embed secret information in large quantities while retaining the details of medical images.
[0036] 4. After the encrypted image generated by the present invention is subjected to an attack test, it is verified that it has a certain anti-attack ability against common attack methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is the schematic diagram of the entire process of the present invention.
[0038] Figure 2 This is the schematic diagram of the composition of each unit mentioned in the network.
[0039] Figure 3 This is the flowchart of the reversible information hiding method for medical images based on adversarial enhancement proposed by the present invention. Specific embodiments
[0040] The following further clarifies the present invention in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art fall within the scope defined by the appended claims of this application.
[0041] Embodiment 1: A reversible information hiding method for medical images based on adversarial enhancement, and the specific implementation steps are as follows:
[0042] Step 1, import the DICOM image that needs to hide information as the input.
[0043] Step 2, extract the standard file header and medical image data from the DICOM format medical image. Specifically, separate the standard file header contained in the DICOM file from the 512×512 medical image data, and divide the image into low-resolution images of 256×256 size uniformly according to the formula when the magnification factor is 2.
[0044] Specifically, select an image with a width of w0 and a height of h0, and calculate the integer width w and height h of the image I obtained at different magnification factors according to the following formula. h of the image.
[0045]
[0046] Among them, round() is a rounding function used to round the decimal input to the nearest integer. floor() is a floor function used to round the decimal input down to the nearest integer. scale1 and scale2 correspond to the magnification factors of the images after width and height degradation, and step1 and step2 are the steps on the corresponding width and height. The calculation formula of step is as follows:
[0047]
[0048] Among them, abs() is an absolute value operation. After obtaining the corresponding size, crop the image I suitable for uniform sampling starting from the upper left corner of the original image. h . In Ih The uniform sampling operation is carried out as follows. First, calculate the height h of the size after degradation according to the selected magnification factors scale and scale2 l and the width w l . Then calculate I h and I l The differences in height and width H d and W d between the two images. The calculation formula is as follows:
[0049]
[0050] The two differences respectively represent the number of rows and columns that should be eliminated in the height and width directions of the image I h . Subsequently, calculate the specific row indices and column indices that should be selected according to the two differences. This process is the process of uniform sampling, which can be expressed by the formula as follows:
[0051]
[0052] Among them, Hsamples[i] represents the row index of the i-th pixel row that should be eliminated, and Wsamples[j] represents the column index of the j-th pixel column that should be eliminated in the width direction. start is selected to be 0, indicating uniform selection starting from index 0.
[0053] Compact the non-eliminated pixels in I h to obtain the degraded low-resolution image I l . This process can be expressed by the formula as:
[0054]
[0055] Among them, pixel1 represents the pixel after eliminating the entire row of pixels with the selected row index in I h , and pixel2 represents the remaining pixel after eliminating all the entire rows and columns of pixels where the selected row index and column index are located. Finally, first use the flatten operation to change the pixels contained in the retained set into a one-dimensional array, and finally, use the reshape operation to change it into a low-resolution image I l ×w l with the size of l .
[0056] Step 3, the medical image information prediction and repair module reads the 256×256 low-resolution image I l output in Step 2. Repair it into a 512×512 high-resolution image. Specifically, first use a convolutional layer Conv to extract the shallow image features F0 from LR, and the formula is as follows:
[0057] F0 = Conv(LR) (12)
[0058] Then, the shallow feature F0 is input into the deep feature extractor for feature extraction. The deep feature extraction layer includes k feature enhancement modules and an RG residual group structure. The residual group processes the feature information after feature enhancement and then inputs it to the next residual group for processing. This can effectively transmit gradients and reduce the problem of gradient disappearance during training. At the same time, it also helps to construct a deeper network to extract richer features. Let FE n be the nth feature enhancement group, and let RG n be the nth residual group, and F n-1 be the feature extracted from the (n - 1)th layer. Then, the feature F n extracted by the entire deep feature extraction layer can be expressed by the formula:
[0059] F n = Conv(RG i (FE i (…RG n (FE n (F n-1 ))…))) (13)
[0060] The stacked feature F n+1 input into the subsequent upsampling module can be expressed by the formula:
[0061] F n+1 = F n-1 + F n (14)
[0062] Finally, the obtained feature F n+1 is input into the upsampling module UP sample that can arbitrarily enlarge the size for medical image information hiding and restoration operations to obtain the finally generated high - resolution image HR. The specific process can be expressed by the formula
[0063] HR = Conv(UP sample (F n+1 )) (15)
[0064] The overall medical image information hiding and restoration module can be described by the formula:
[0065] HR = RM(LR) (16)
[0066] Step 4: Input the medical image, the restored high-resolution image, and the standard file header into the embedder to generate the encrypted medical image. Specifically, the embedder accepts the standard file header output in Step 2, the 512×512 medical image, and the restored 512×512 high-resolution image output in Step 3. The embedder determines the positions suitable for embedding information based on the interpolated and restored image through the position selection network and encodes the information in the standard file header as a bitstream and embeds it into the encrypted medical image in combination with the method of predicting error expansion. Specifically, the degraded image I in Step 2 l is enlarged through interpolation to obtain image I B , at this time I B has the same size as the image I before degradation h . This image is input into the position selection network (LSUnet) in the embedder to generate the embedding position map M unet . The restored image after hiding the medical image information, the bitstream after converting the information to be hidden, the degradation retention positions, and the embedding position map are input into the specially designed embedder S E to obtain the embedded image I steo . The whole process can be expressed by the formula:
[0067]
[0068] Effect evaluation:
[0069] The present invention proposes an adaptive reversible medical image information hiding method based on adversarial enhancement for hiding key information during the transmission of large-capacity medical images. Compared with existing methods, this method can ensure the integrity of medical images while guaranteeing a large embedding information volume.
[0070] Table 1. Restoration effects of the predictor of the present invention on different datasets in some datasets
[0071]
[0072] Table 2. PSNR values of the downloaded encrypted images with different embedding amounts
[0073]
[0074]
[0075] Table 3. SSIM values of the downloaded encrypted images with different embedding amounts
[0076]
[0077] It should be noted that the above embodiments are only the preferred embodiments of the present invention and do not limit the protection scope of the present invention. Any equivalent replacement or substitution made on the basis of the above technical solutions shall fall within the protection scope of the present invention.
Claims
1. An adaptive reversible information hiding method for medical images based on adversarial enhancement, characterized in that, The method includes the following steps: Step 1: Import the original DICOM-format medical image. Specifically, import the DICOM-format medical image that needs to hide information. The medical image includes a standard file header for storing information and medical image data of 512×512 size; Step 2: Extract the standard file header and medical image data from the DICOM-format medical image. Specifically, separate the standard file header and 512×512 medical image data contained in the DICOM file and evenly divide the medical image into low-resolution images; Step 3: Input the low-resolution image into the predictor to obtain the restored high-resolution image. Specifically, evenly divide the 512×512 medical image into lower-resolution medical images of 256×256 size and input them into the predictor to obtain the restored 512×512-sized high-resolution image; Step 4: Input the medical image, the restored high-resolution image, and the standard file header into the embedder to generate the encrypted medical image. Specifically, the embedder receives the standard file header and the 512×512 medical image output in Step 2 and the restored 512×512-sized high-resolution image output in Step 3, determines the positions suitable for embedding information through the position selection network based on the interpolated and restored image, and encodes the information in the standard file header as a bit stream and embeds it into the encrypted medical image in combination with the way of prediction error expansion; Step 5: Synthesize the encrypted medical image and the false file header into an encrypted DICOM-format medical image. Specifically, the encrypted medical image after embedding is combined with the false file header that hides the key information to generate the encrypted DICOM medical image, hiding the key information in the original DICOM image without affecting the normal use of the image.
2. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that The network structure to be trained in step 3 is as follows: The predictor RM in step 3 includes three main modules: a feature extraction layer, a feature enhancement layer, and an upsampling layer, which can achieve the post-repair of medical image information hiding with any magnification factor. The feature extraction layer contains a convolutional kernel with a size of 3×3 and a stride of 2. The feature enhancement layer includes 64 feature enhancement modules FE and an RG residual group structure. Each layer of the feature enhancement layer FE contains 4 distillation modules and 1 spatial attention module. The distillation module consists of a convolutional kernel with a size of 3×3 and a stride of 2 and a ReLu activation function layer. The spatial attention module consists of a max pooling layer, an average pooling layer, and a sigmoid activation layer. Finally, it passes through a convolutional kernel with a size of 7×7 and a stride of 2. Each RG residual group structure consists of a convolutional kernel with a size of 3×3 and a stride of 2, a ReLu activation function layer, a convolutional kernel with a size of 3×3 and a stride of 2, an average pooling layer, a convolutional kernel with a size of 3×3 and a stride of 2 and a ReLu activation function layer, a convolutional kernel with a size of 3×3 and a stride of 2, and a sigmoid activation layer. The upsampler uses a flow-based upsampling module to perform image upsampling by learning the offset grid, and the obtained deep image features are input into the upsampling module to obtain the final repaired image.
3. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that The network structure to be trained in step 4 is as follows: The embedder in step 4 consists of 15 feature extraction blocks and 1 deconvolution layer. The first 8 feature extraction blocks are responsible for downsampling the image, and the last 7 feature extraction blocks are responsible for upsampling. The output of the i-th feature extraction block is connected to its corresponding input (the 16-i-th block) and then passed to the next feature extraction block. Each feature extraction block includes 1 convolutional layer, 2 normalization layers, and 2 LeakyReLU activation layers, and these layers are arranged alternately to form the feature extraction block.
4. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that, The network loss function L(θ) to be trained in step 3 is: where θ is the parameter contained in the network, N is the total number of samples in the training set used, LR i and HR i is the i-th batch of data selected from the training set.
5. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that The network loss function L(θ G ) to be trained in step 4 and L(θ D ) are Both the generator and the discriminator use the cross - entropy loss function crossEntropy for parameter optimization. Let the parameters of the generator be θ G , and the parameters of the discriminator be θ D , where I B represents the non - encrypted image, I steo represents the encrypted image, L out is the discrimination result obtained by the discriminator, L B represents the label of the non - encrypted image, L steo represents the label of the encrypted image. The purpose of the generator is to generate images that the discriminator cannot distinguish, and the constructed label is exactly the opposite of the label of the input image. The purpose of the discriminator is to distinguish whether the input image is an encrypted image, and the constructed label is the same as the label of the input image.
6. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that, Under the same prior conditions, the peak signal-to-noise ratio and structural similarity are used to measure the repair effect of the repair model on low-resolution images, as shown in the following formula: The peak signal-to-noise ratio (PSNR) is a commonly used metric for measuring image quality. It is usually used to compare the similarity between the original image and the processed image, or to compare the similarity between two images. Here, I represents the original image, and P represents the image after embedding information. w and h represent the width and height of the image, I(i,j) and P(i,j) represent the pixel values at the position (i,j), MSE is the average of the squares of the differences in pixel values between the original image and the processed image, and MAX is the maximum pixel value that can occur in a single-channel image. The larger the value of PSNR, the higher the similarity between the image and the original image, the lower the distortion degree, and the better the quality. Among them, μ f and μ g represent the means of images f and g respectively, σ f and σ g are the corresponding standard deviations, and σ fg is the covariance between the two images; C1 and C2 are two constants: C1 = (0.01 × LG) 2 , C2 = (0.03 × LG) 2 , where LG is the gray value range of the image.
7. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that, Under the same prior conditions, the embedding capacity is measured using the embedding rate, as expressed by the following formula: If w and h represent the width and height of the image, and num is the total number of bits of the embedded information, the embedding rate R can be obtained.
8. An adaptive reversible medical image information hiding method based on adversarial enhancement according to claim 1, characterized in that, The network model parameters that need to be trained in Step 3 and Step 4 are iteratively updated by the optimizer.
9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements an adaptive reversible medical image information hiding method based on adversarial enhancement as described in any one of claims 1 to 8 above.
10. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the computer instruction is executed by the processor, it implements an adaptive reversible medical image information hiding method based on adversarial enhancement as described in any one of claims 1-8.