Structural perception adaptive steganography method for medical image privacy protection

Abstract

The invention discloses a medical image privacy protection-oriented structure perception adaptive steganography method, which comprises the following steps of: in a training stage, generating a diagnosis related region mask based on a pre-trained IMIS-Net, extracting an embeddable suppression mask, and obtaining an initial embedding probability under the minimum distortion constraint; and the self-adaptive adjustment of structure perception is carried out on the pixel-level embedding probability in combination with a reinforcement learning normal form, so that the ROI is protected preferentially and a high-risk region is inhibited on the premise of meeting the set capacity, and the steganalysis resistance and perception quality are improved. And after the training is completed, in a reasoning stage, the masks and probabilities are firstly generated or quoted, then action probabilities are obtained through a trained strategy network, then probability consistency is mapped into STC cost, real embedding is completed by the STC, and a secret-containing medical image is obtained. According to the invention, on the premise that the availability of image diagnosis is not reduced, safe embedding and hidden transmission of sensitive information are realized.

Description

Technical Field

[0001] This invention relates to a method for hiding and protecting medical information, belonging to the interdisciplinary field of medical image processing and computer vision, specifically a structure-aware adaptive steganography method for protecting the privacy of medical images. Background Technology

[0002] With the development of medical imaging technology, medical images have been widely used in clinical diagnosis and intelligent analysis, and often contain sensitive information such as patient identity and diagnostic results. Without effective protection during transmission and storage, privacy leaks are highly likely. Therefore, medical image steganography technology is increasingly being applied to medical data security protection. Its basic principle is to embed sensitive information into images without affecting diagnostic value, achieving covert transmission.

[0003] Existing medical image steganography methods mainly fall into two categories: one is based on spatial or transform domain embedding, such as hiding information by modifying pixels or transform coefficients; the other combines deep learning and structure awareness, using network models to automatically select embedding regions to enhance concealment and security. Meanwhile, some methods attempt to distinguish between diagnostic regions of interest (ROI) and non-diagnostic regions of non-interest (RONI), embedding information in non-critical regions to reduce interference with diagnostic content.

[0004] However, existing technologies still have significant shortcomings: First, the identification and protection of ROIs are often not precise enough, which can easily lead to mis-embedding or tampering of diagnostic regions; second, some methods that rely on training models require a large amount of labeled data, have limited generalization ability, and are difficult to adapt to different types of medical images; third, medical images have differences in pixel distribution, and if a large number of unstructured regions are directly used for embedding, it can easily cause statistical anomalies, thereby increasing the risk of detection; in addition, existing methods are difficult to balance between embedding capacity, image quality, and concealment, often resulting in a decline in image quality and obvious embedding traces, which is difficult to meet the requirements of medical applications for high security and high perceived quality. Summary of the Invention

[0005] Purpose of the invention: To address the problems and shortcomings of existing technologies, this invention provides a structure-aware adaptive steganography method for protecting the privacy of medical images.

[0006] The method of this invention achieves the following objectives: (1) Automatically locate and exclude embedded ROI regions, and explicitly suppress high-risk regions; (2) Under a given capacity, undetectability is significantly improved through adaptive probability optimization and adversarial feedback loop; (3) Achieve synergistic optimization and cross-scene generalization of capacity, concealment and image quality by taking ROI weighted quality as the guide.

[0007] (4) Overcome the shortcomings of medical image steganography technology in terms of insufficient protection of the diagnostic area, easy detection and difficulty in ensuring image quality.

[0008] Technical Solution: A structure-aware adaptive steganography method for medical image privacy protection. This method addresses the privacy leakage risks during the transmission and storage of medical images and the vulnerability of traditional steganography methods to detection in critical areas. It provides a highly secure steganography operation scheme while ensuring the diagnostic usability of medical images. The method comprises two stages: the first stage is the initial probability and mask generation stage, which automatically generates a structure-aware mask and initial embedding probabilities based on the input medical image, providing guidance for subsequent embedding operations; the second stage is the reinforcement learning-driven embedding adjustment stage, which dynamically optimizes the embedding strategy through the interactive feedback between the policy network and the environment network, controlling the spatial distribution and modification intensity of information embedding. This adaptively embeds information in non-critical areas while suppressing modifications in critical areas, ensuring that the steganographic image is both secure and usable for clinical diagnosis.

[0009] The method achieves adaptive probability control, dynamic feedback optimization, and joint optimization of multi-objective loss by jointly utilizing the zero-shot ROI localization model IMIS-Net, the embedding suppression mask module ESM, the initial steganography to generate embedding probabilities, and the reinforcement learning mechanisms of the policy network and the environment network. This enables steganography operations to balance security, embedding capacity, and visual quality in complex medical image scenarios.

[0010] The specific steps of the method are as follows: Step 101: Input a medical image, obtain the original image matrix X as input, and perform standardization processing on the image.

[0011] Step 102: Use the IMIS-Net model to perform zero-shot ROI localization on the standardized image and generate the ROI mask M. roi The IMIS-Net model identifies diagnostic-related regions in images and supports multimodal input prompts, including points, boxes, and text.

[0012] Step 103: The Embedding Suppression Mask (ESM) module is used to identify high-risk regions in the medical image. These high-risk regions are those with a detection probability higher than a set threshold during steganalysis, such as user-specified areas with a completely black background, uniform texture, or abnormal pixel distribution. The ESM module generates a suppression mask M. ESMThis is used to constrain subsequent information embedding operations, avoid modifications in high-risk areas, and thus reduce the likelihood that steganographic traces will be detected by the analyzer.

[0013] Step 104: Generate an initial embedding probability map P based on steganography (such as S-UNIWARD) to provide an initial embedding distribution reference for the policy network.

[0014] Step 105, mask the ROI M roi Suppression mask M ESM The initial embedded probability graph P is jointly input into the policy network to generate a temporary action probability matrix Q. θ .

[0015] Step 106, based on the temporary action probability matrix Q output by the policy network θ The modification probability at each pixel position is calculated, with the action set {-1, 0, +1}, corresponding to negative modification, no modification, and positive modification, respectively. The policy network generates an action (a) to be performed at pixel (i, j) given the original image matrix X. i,j probability distribution π θ (a i,j |X), this distribution is used to guide subsequent information embedding operations, thereby achieving concealment optimization while ensuring the stability of the diagnostic region.

[0016] Step 107: Using the STC embedding simulator, the final modified image M is generated through the action probability and noise matrix. The modification is then superimposed onto the original image matrix X to obtain the steganalytic image Y.

[0017] Step 108: Input the steganalyte image Y into the environment network (e.g., XuNet), and calculate the pixel-level reward signal r for each pixel. i,j It is used to measure the safety of modifications and guide policy network optimization.

[0018] Step 109: Train the policy network and the environment network to obtain the trained model; The policy network is trained using reinforcement learning based on feedback from the environment network. The parameters of the policy network are adjusted, and the environment network is iteratively updated until the steganographic image reaches the preset targets in terms of security, visual quality, and diagnostic usability. The final steganographic image Y is then output.

[0019] The next step is to apply the trained model to the general medical imaging stage: Step 201, input the medical image X to be protected. test The trained model is then standardized in preparation for generating steganalytic images.

[0020] Step 202: Use the IMIS-Net model to generate ROI masks, label the diagnostic-related regions of the medical image, and generate a suppression mask M.ESM Used for suppression in high-risk areas.

[0021] Step 203: Mask the ROI and suppress the mask M of the medical image. ESM The initial embedding probability map learned during the training phase is input into the trained policy network to directly generate the action probability matrix Q. θ .

[0022] Step 204: Calculate the modification probability of each pixel based on the action probability matrix to form an action probability distribution, which is used to control the actual embedding operation.

[0023] Step 205: Using the STC embedding simulator, the action probability distribution and encrypted doctor's diagnosis report information are combined to generate the final modified image M, which is then superimposed onto the test image X. test In the process, a dense image Y is obtained. test =X test +M.

[0024] In training the policy network, the temporary action probability matrix output by the policy network is combined with the ROI mask M. roi With suppression mask M ESM Constraints are imposed. The ROI represents the region of critical interest or region of interest, which is a region in the medical image directly related to clinical diagnosis, including but not limited to lesions, major organs, or other diagnostic target areas. Steganography embedding operations are prohibited within this region to avoid affecting the diagnostic validity of the medical image. Furthermore, the non-critical region (RONI) is a region in the medical image that does not affect diagnosis, including but not limited to background areas, homogeneous tissue, or areas without diagnostic significance. Information embedding is adaptively performed within this region, thereby achieving a balance between the covert transmission of privacy information and image quality while protecting the critical region.

[0025] During the training of the policy network and the environment network, a multi-objective joint loss function is adopted, including capacity loss, environment network reward loss, ROI protection loss, ESM constraint loss and modification rate constraint loss, to optimize the steganography strategy and achieve the best balance between security, image quality and embedding capacity in the embedding process.

[0026] During model training, the performance of stegogram images is evaluated, and security and image quality metrics are calculated, including the detection error rate P. E = (P FA +P MD ) / 2 and pixel-level reward signal Simultaneously, peak signal-to-noise ratio (PSNR) and structural similarity metric (SSIM) were calculated to verify the security, visual quality, and diagnostic usability of the stegtext images. Among these, P... FAP represents the false detection probability, indicating the probability of incorrectly identifying an image without embedded information as an image containing steganographic information. MD The false negative probability represents the probability of incorrectly classifying an image that actually contains steganographic information as an image without embedded information. Further, the pixel-level reward signal... These are the weighting coefficients. pixel position The mask value is used to indicate whether embedding of that pixel is allowed. denoted as the importance factor of a pixel. The reward signal calculated using the above formula is used to guide the embedding strategy to modify low-risk areas, thereby improving the overall steganography effect.

[0027] Repeat steps 101 to 109 for 52 rounds or more (e.g., 72 rounds) to complete the training of the policy network and the environment network, and obtain the trained model.

[0028] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the structure-aware adaptive steganography method for medical image privacy protection as described above.

[0029] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the structure-aware adaptive steganography method for medical image privacy protection as described above.

[0030] Beneficial Effects: This invention achieves the following beneficial effects by introducing a steganography method based on medical image segmentation and structure awareness, combined with a reinforcement learning optimization mechanism: ① Ensuring the safety of diagnostic regions: This invention utilizes the zero-shot medical segmentation model IMIS-Net to automatically identify regions of interest (ROIs) in medical images. During the embedding process, diagnostic-related regions are effectively protected, avoiding interference from information embedding with key diagnostic content, thereby ensuring the diagnostic usability of medical images.

[0031] ②Enhancing Steg Security: This invention automatically identifies and suppresses completely black areas and sparsely structured regions in medical images by embedding an suppression mask module (ESM), reducing the concentrated distribution of embedded information in high-risk detection areas and significantly enhancing resistance to advanced steganalysis detection.

[0032] ③ Achieving intelligent adaptive optimization: This invention introduces reinforcement learning into the embedding probability regulation process and combines it with multi-objective loss functions for joint optimization, enabling the model to learn autonomously during training and achieve a balance between embedding capacity, security and diagnostic integrity, with overall performance superior to traditional fixed-policy methods. Attached Figure Description

[0033] Figure 1 This is a flowchart of the network training process according to an embodiment of the present invention; Figure 2 This is an embedded flowchart of an embodiment of the present invention; Figure 3 This is a visual comparison diagram of specific image embedding in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0035] like Figure 1 As shown, a structure-aware adaptive steganography method (MARS-Net) for medical image privacy protection is presented. The network training process of this invention adopts a two-stage framework: Stage 1, initial probability and mask generation stage; Stage 2, reinforcement learning probability adjustment embedding stage. In Stage 1, diagnostic region-related information (ROI) masks are generated based on the pre-trained IMIS-Net, and embeddability suppression masks (ESM) are extracted, while initial embedding probabilities are obtained under minimum distortion constraints. In Stage 2, the pixel-level embedding probabilities are adaptively adjusted using a reinforcement learning paradigm, prioritizing the protection of ROIs and suppressing high-risk regions while meeting a given capacity, thereby improving anti-steganography analysis capability and perception quality. After training, the inference stage follows... Figure 2 The process described is as follows: First, the medical image is preprocessed to generate a mask and initial probabilities. Next, action probabilities are obtained through a trained network model and mapped to distortion costs. Then, sensitive information such as diagnostic reports requiring protection is embedded using STC (Self-Containment Technology) combined with the distortion costs to achieve true embedding, resulting in a encrypted medical image. Through this approach, the present invention achieves secure embedding and covert transmission of sensitive information without reducing the usability of image diagnosis; when needed, the diagnostic report can also be extracted and recovered using STC.

[0036] Network training process: Figure 1 The diagram shows the complete training framework of the network described in this invention. The MARS-Net proposed in this invention is a "two-stage" medical image steganalysis training framework: the first stage generates initial embedding priors and structural masks to achieve preprocessing of medical images; the second stage performs adaptive adjustment of the embedding probability under the reinforcement learning paradigm to achieve both protection of medical diagnosis-related regions (ROIs) and suppression of high-risk regions, and maximizes steganalysis resistance and perception quality.

[0037] Phase 1: Initial Probability and Structure Mask Generation Step 1101: Extract the region of interest from the original medical image. Specifically, extract the region of interest from the medical image X=(x i,j ) H×W Input to the pre-trained IMIS-Net model with frozen weights, x i,j H represents the pixel value at position (i,j) in the image, and H and W represent the height and width of the image, respectively. Combining information such as points, bounding boxes, or text prompts, the IMIS-Net model automatically outputs the region of interest mask M. roi =((m roi ) i,j ) H×W , where m roi The mask value is used to indicate whether the corresponding pixel (i,j) belongs to the region of interest. When m roi =1 indicates that the pixel belongs to the ROI, and when m roi =0 indicates that the pixel belongs to a non-critical area.

[0038] M roi =IMIS-Net(X,Prompt) (1) Prompt represents an input prompt, which can be any of a point, box, or text. The IMIS-Net model has the capability for zero-shot / low-interaction fine-grained segmentation in medical scenarios. It is a segmentation module that achieves high-quality boundary and small structure recognition through large-scale interactive data and multi-round mask decoders.

[0039] Step 1102: Extract embeddability suppression masks from medical images. Design an embedding suppression mask module (ESM). The embedding suppression mask module automatically scans each medical image and extracts all locations with a pixel value of 0 (regions with a pixel value of 0 are regions that need to be suppressed. In medical images such as CT or MRI, regions with a pixel value of 0 usually correspond to air or background regions unrelated to diagnosis. Embedding information in these regions is easily detected by steganalysis tools). This forms a suppression mask M. ESM =((m ESM ) i,j ) H×W , where (m ESM ) i,j This represents the mask value at pixel position (i,j), when (m ESM ) i,j When = 0, it indicates that the pixel belongs to a high-risk area and should not be modified during subsequent embedding; when (m ESM ) i,j When =1, it indicates that the pixel belongs to a low-risk area and can be used for information embedding.

[0040] (2) Mask M ESM It is used as an explicit penalty term in the loss function of the policy network training to suppress embedding into the aforementioned high-risk regions.

[0041] Step 1103: Initial Embedding Probability Estimation. An initial probability generation function S is defined using a symmetric steganographic prior (such as SUNIWARD). init (.), resulting in the initial embedding probability map P=( ) H×W = S init (X), where ∈[0,1] represents the probability that a pixel is modified during the information embedding process.

[0042] Within the framework of "minimum distortion", the standard form of probability and capacity entropy constraints are given: (3) in, The expression represents the desired distortion, where H and W represent the height and width of the image, respectively, and i and j are pixel indices. , For a set of actions, Let represent the probability of choosing action m at pixel position (i,j). This indicates the corresponding cost of distortion. The capacity entropy of the probability distribution. This represents the capacity constraint value. Furthermore, the standard form of the probability distribution is defined as: (4) in These are Lagrange multipliers used to balance capacity constraints and distortion minimization; their values ​​are determined through a binary search. The action is represented by the temporary variable m, which is used to calculate the total probability.

[0043] Step 1104: M produced in Phase 1 roi =((m roi ) i,j ) H×W M ESM =((m ESM ) i,j ) H×W P=( ) H×W They will be used as inputs to the Phase 2 policy network.

[0044] Phase Two: Probability-Adjusted Embedding Phase Based on Reinforcement Learning Step 1201: Place M roi MESM A trainable policy network with input parameters θ is cascaded with P. This policy network outputs a temporary action probability matrix. ,in H represents the temporary modification probability of a pixel (i,j), and H×W are the length and width of the image. The pixel-by-pixel action a... i,j ∈{-1,0,+1} represents a ternary modification of "decrease by one / remain unchanged / increase by one", and the policy network provides the corresponding embedding action probability: (5) Thus, the action probability matrix is ​​obtained. .

[0045] Step 1202: STC Embedding Simulation and Stegature Sample Generation. To seamlessly simulate the real embedding process during training, a uniform noise matrix N=(n) at the same scale as the image is first sampled. i,j ) H×W , where n i,j This represents the noise value at pixel position (i,j), which follows a uniform distribution and is used to provide a randomized selection signal during the embedding process. (Based on action probability) Generate the final modified graph M={ using the "STC-like" rule. }: (6) Modify the image The steganographic image Y = X + M is obtained by adding it to the cover image. Then, X and Y are input together into the environment network to obtain a pixel-by-pixel reward signal; the environment network then feeds the reward back to the policy network. This invention uses the deep steganalysis analyzer XuNet as the environment network.

[0046] Step 1203: Multi-objective loss and optimization. Following step 1202, the following loss function is calculated for optimization.

[0047] Environmental network loss (cross-entropy): (7) Where the base of log is e, N is the total number of samples, and x i For the input image, z i Its true label (carrier image z) i =0, steganalysis z i =1), E(x) i ) represents the probability that the environment network identifies the image as a steganalysis.

[0048] Policy network loss: (8) in These are the weighting coefficients.

[0049] (a) Countering reward loss Used to encourage policy networks to generate more secure embedding probability distributions: (9) The relationship between pixel-level reward and gradient is as follows: (10) In the formula, X represents the input medical image; For the policy network, apply action a to pixel (i, j). i,j The probability of ∈{-1,0,+1} (which is defined by formula (5)) (Mapped) This is an indicator function that only counts for bonuses when a non-zero modification occurs.

[0050] (b) Capacity constraint loss This is used to make the embedding entropy of the entire graph consistent with the target capacity C (in bpp): (11) (c) ROI repulsion loss Used for embedding tendency within the ROI region: (12) Where m roi This is generated for step 1101. N ROI The number of pixels in the ROI; This represents the probability of a non-zero modification occurring at that pixel (see the correspondence defined in formula (5)).

[0051] (d) ESM suppresses loss (Mask loss), used to constrain the embedding probability within high-risk regions (ESM mask): (13) in For the ESM mask generated in step 1102, N ESM This represents the number of pixels in that area.

[0052] (e) Modification rate loss This is used to reduce the overall modification rate and improve the quality of steganalog images. (14) Where N = H × W.

[0053] Step 1204: Joint network optimization until convergence.

[0054] Embedding / Privacy Protection Process After Network Training like Figure 2 As shown, after the network training is completed, the actual embedding (privacy protection) process is presented. The inference phase no longer participates in the training and uses STC to perform real steganography; the variables and formulas used are consistent with those in the previous section, and no new technical settings are introduced.

[0055] Step 2101: Input and parameter settings.

[0056] Input the medical image X to be protected, set the target embedding capacity C, and provide the diagnostic report to be embedded; the image size and intensity normalization preprocessing is performed according to the unified rules of the aforementioned training phase.

[0057] Step 2102: Prior Mask and Initial Probability Generation. Following the process described in Stage 1, the diagnostic-related region mask M is generated. roi (Generated by IMIS-Net with given prompts), Embeddability suppression mask M ESM And the initial embedding probability map P=S under the minimum distortion framework. init (X) (its solution satisfies the capacity entropy constraint, (Determined through binary search), the three factors, along with X, are used as subsequent inputs.

[0058] Step 2103: Policy Network Forward: Obtain pixel-level modification probabilities. [X, M] roi M ESM [P] is input into the pre-trained policy network to obtain pixel-level output Q. θ And obtain the ternary action probability per pixel according to the aforementioned action probability mapping relationship. .

[0059] Step 2104: Construct the weights of STC from probability to cost. To achieve a true embedding with "minimum total cost" using STC during the inference phase, based on the aforementioned "probability-distortion consistency" principle, the weights are... The equivalent mapping to the pixel-action cost of STC The correspondence of "smaller probability, higher cost" can be achieved using monotonically logarithmic forms, since M... roi With M ESM Non-zero actions have been suppressed during training, and their corresponding costs are significantly increased under this mapping, thus being naturally avoided in subsequent STC solutions.

[0060] Step 2105: STC Constraint Construction. A complex is generated based on the parity check matrix of the bitstream and the STC, and the result obtained in Step 2104 is used... As a pixel-wise cost input to STC; motion components that are not allowed at the intensity saturation boundary are masked according to the aforementioned boundary rules (equivalent to assigning infinite cost).

[0061] Step 2106: Solve the modified graph for minimizing the total cost using STC. Within the feasible set satisfying the STC linear constraints, solve the modified graph for minimizing the total cost. ,in, The adjustment cost for selecting embedding action m at pixel position (i,j) is used to characterize the overall overhead of different embedding actions in terms of concealment and distortion. If the obtained embeddable length deviates from the target capacity, the probability distribution is binary-tuned according to the capacity entropy constraint method to scale the overall cost, and the above steps are repeated until the capacity requirement is met. If the obtained embeddable length deviates from the target capacity, the aforementioned capacity entropy constraint method is used to adjust the cost. Perform a binary fine-tuning to scale the overall probability (and thus scale the cost equivalently), and repeat this step until the capacity requirement is met.

[0062] Step 2107: Generation of the steganographic medical image. Add the optimal modified image to the medical image to obtain the steganographic result Y=X+M. * Output steganographic images for privacy protection scenarios during subsequent transmission and storage.

[0063] A visual comparison of the modifications to different steganography methods before and after introducing ESM constraints was presented. For each group, the original carrier image, ROI mask, superposition of carrier image and mask were given sequentially, along with the modifications of the three steganography methods under unconstrained conditions and under the constraints of the proposed method (MARS-Net). To visually verify the effectiveness of the proposed MARS-Net framework in avoiding the embedding of sensitive region information, a comparative visualization of the modifications to three representative steganography methods (DBT-RL, S-UNIWARD, and WOW) before and after introducing embedding constraints was also provided.

[0064] It is evident that in the unrestricted modified images, all three steganography methods exhibit significant embedding traces within the ROI mask region, some even concentrated in completely black areas of the image (such as the air background in CT scans), which are high-risk areas where embedding should not occur. This indiscriminate embedding behavior not only potentially interferes with the diagnostic usability of medical images but also significantly increases the risk of detection.

[0065] In contrast, the constrained modification map after embedding adjustment using this method exhibits significantly different distribution characteristics—almost no embedding occurs in ROI regions and structurally simple regions. Embedding traces are effectively guided to non-critical regions in the image that do not affect diagnosis but possess a certain degree of texture complexity, thereby improving embedding security while ensuring image quality. This enables the policy network to effectively learn and avoid high-detection-risk regions during training, achieving structure-aware control over embedding locations.

[0066] Table 1. Evaluation of image steganalysis using the method of this invention on the MSD pancreatic tumor dataset: Image quality metrics (WPSNR, PSNR, SSIM) at different embedding rates. Table 1 shows the WPSNR, conventional PSNR, and Structural Similarity Index (SSIM) evaluation results for medical cryptic images generated using existing steganography methods and those optimized using the method of this invention on the MSD-Pancreas dataset (SUNIWARD, WOW, and DBT-RL represent the existing initial steganography methods; MARS-Net(WOW), MARS-Net(SUNIWARD), and MARS-Net(DBT-RL) represent the optimized methods after this invention). It can be seen that MARS-Net based on the method of this invention significantly improves WPSNR while maintaining the overall quality of conventional images, indicating its superior performance in preserving image quality in key diagnostic regions.

[0067] Table 2 shows the detection error rate (%) of the method of the present invention against three steganalysts on the MSD pancreatic tumor dataset. Table 2: Experimental comparisons were conducted on the MSD-Pancreas dataset of three common steganalysis methods (SUNIWARD, WOW, and DBT-RL) and their combinations with this method, and detection was performed using three steganalysis analyzers, including a traditional analyzer SRM and two CNN-based deep learning analyzers CovNet and LWENet.

[0068] The performance of this method is comparable to the original method under the traditional SRM analyzer, with slight improvements in certain scenarios (such as an improvement of about 6% in DBT-RL). However, under both deep learning analyzers, the method of this invention significantly improves the anti-detection capability of all steganography methods. The significant improvement compared to the original method means that the image is almost impossible for the analyzer to effectively distinguish, successfully achieving a qualitative leap from "fully detectable" to "almost undetectable".

[0069] at the same time Figure 3 The embedding visualization shown illustrates that, for medical images, existing methods still modify the region of interest (ROI) during the embedding process (e.g., using DBT-RL, SUNIWARD, and WOW steganography methods as the initial embedding), which can easily interfere with medical diagnosis. In contrast, the method of this invention, under the constraint of the ROI mask, explicitly assigns the embedding location to non-critical regions, thus protecting the ROI and ensuring that the steganographic image maintains diagnostic usability while preserving its concealment.

[0070] In summary, this invention revolves around the complete process of "stage one prior and structure mask generation — stage two reinforcement learning adaptive adjustment — inference stage STC true embedding," and adopts... Figure 1 , Figure 2 The technical approach described achieves priority protection of ROIs in medical images and effective suppression of high-risk areas, while striking a balance between capacity, concealment, and image quality. The embodiments described in this specification are used to illustrate the technical solution and working mechanism of the present invention and do not constitute a limitation on the scope of protection. Without departing from the core ideas and technical effects of the present invention, modifications to the step order, parameter settings, loss weights, and related aspects are permitted. Figure 1 / Figure 2 Equivalent substitutions or formal transformations made by the corresponding module implementations should all fall within the protection scope of this invention.

[0071] It is obvious to those skilled in the art that the steps of the structure-aware adaptive steganography method for medical image privacy protection described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, which can then be stored in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A structure-aware adaptive steganography method for medical image privacy protection, characterized in that, Automatically generate structure-aware masks and initial embedding probabilities based on the input medical images; Based on reinforcement learning, the embedding policy is dynamically optimized through the interaction and feedback between the policy network and the environment network, controlling the spatial distribution and modification intensity of information embedding; including the following steps: Step 101: Input a medical image, obtain the original image matrix X as input, and perform standardization processing on the image; Step 102: Use the IMIS-Net model to perform zero-shot ROI localization on the standardized image and generate the ROI mask M. roi ; Step 103: Use the embedded suppression mask module to identify high-risk regions in medical images. High-risk regions are those that need to be suppressed and modified. The embedded suppression mask module generates a suppression mask M. ESM This is used to constrain subsequent information embedding operations and prevent modifications from being made in high-risk areas; Step 104: Use steganography to generate an initial embedding probability map P to provide an initial embedding distribution reference for the policy network; Step 105, mask the ROI M roi Suppression mask M ESM The initial embedded probability graph P is jointly input into the policy network to generate a temporary action probability matrix Q. θ ; Step 106, based on the temporary action probability matrix Q output by the policy network θ Calculate the probability of modification at each pixel position; Step 107: Using the STC embedding simulator, the final modified image M is generated through the action probability and noise matrix. The modification is then superimposed on the original image matrix X to obtain the steganalytic image Y. Step 108: Input the steganalytic image Y into the environment network, and calculate the pixel-level reward signal r for each pixel. i,j This is used to measure the safety of modifications and guide policy network optimization; Step 109: Train the policy network and the environment network to obtain the trained model; The policy network is trained by reinforcement learning based on the feedback from the environment network. The parameters of the policy network are adjusted, and the environment network is iteratively updated until the stegana image reaches the preset target in terms of security, visual quality and diagnostic usability. The final stegana image Y is then output. Adaptive steganography of medical images is achieved using a trained model, and the steganographic image is output.

2. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, Applying the trained model to medical images to achieve adaptive steganography of medical images includes the following steps: Step 201, input the medical image X to be protected. test The trained model is then standardized. Step 202: Use the IMIS-Net model to generate ROI masks, label the diagnosis-related regions of the medical image, and generate suppression masks; Step 203: Input the ROI mask, suppression mask, and initial embedding probability map learned during the training phase of the medical image into the trained policy network to directly generate the action probability matrix; Step 204: Calculate the modification probability of each pixel based on the action probability matrix to form an action probability distribution, which is used to control the actual embedding operation; Step 205: Using the STC embedding simulator, the action probability distribution and the encrypted doctor's diagnosis report are embedded together to generate the final modified map M, and the modified map M is superimposed on the medical image X. test In the process, a dense image Y is obtained. test =X test +M.

3. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, The IMIS-Net model supports multimodal prompt input; the multimodal prompt input includes point, box, and text prompt input.

4. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, In step 106, the action set is set to {-1, 0, +1}, corresponding to negative modification, no modification, and positive modification, respectively; the policy network generates an action to be performed at pixel (i, j) based on the original image matrix X. i,j probability distribution π θ (a i,j |X), this distribution is used to guide subsequent information embedding operations.

5. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, In training the policy network, the temporary action probability matrix output by the policy network is constrained by combining the ROI mask and the suppression mask; During the training of the policy network and the environment network, a multi-objective joint loss function is used to optimize the steganography strategy. The loss function includes capacity loss, environment network reward loss, ROI protection loss, ESM constraint loss and modification rate constraint loss. During model training, the performance of stegograms is evaluated, and security and image quality metrics are calculated, including the detection error rate P. E = (P FA +P MD ) / 2 and pixel-level reward signal Simultaneously, peak signal-to-noise ratio and structural similarity metrics are calculated to verify the security, visual quality, and diagnostic usability of stegographic images; Among them, P FA P represents the false detection probability, indicating the probability of incorrectly identifying an image without embedded information as an image containing steganographic information. MD The false negative probability represents the probability of incorrectly classifying an image containing steganographic information as an image not embedded; the pixel-level reward signal in These are the weighting coefficients. pixel position The mask value, For pixels Importance factors.

6. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, Using steganography, the initial probability generation function S is defined prior. init (.), resulting in the initial embedding probability map P=( ) H×W = S init (X), where ∈[0,1] represents the probability that a pixel is modified during the information embedding process; Within the framework of minimum distortion, the standard form of probability and capacity entropy constraints are given: (3) in, This represents the expected distortion, where H and W represent the height and width of the image, respectively, and i and j are pixel indexes. The action... , For a set of actions, Let represent the probability of choosing action m at pixel position (i,j). This indicates the corresponding cost of distortion. The capacity entropy of the probability distribution. The capacity constraint value; the standard form of the probability distribution is defined as: (4) in These are Lagrange multipliers used to balance capacity constraints and distortion minimization; their values ​​are determined through a binary search. Indicates an action.

7. The structure-aware adaptive steganography method for medical image privacy protection according to claim 1, characterized in that, Mask the ROI M roi Suppression mask M ESM The initial embedded probability map P is cascaded with a policy network with input parameters θ, and the policy network outputs a temporary action probability matrix. ,in H represents the temporary modification probability of a pixel (i,j), where H×W is the length and width of the image; and the pixel-wise action a... i,j ∈{-1,0,+1} represents a ternary modification of "decrease by one / remain unchanged / increase by one", and the policy network provides the corresponding embedding action probability: (5) Thus, the action probability matrix is ​​obtained. .

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the structure-aware adaptive steganography method for medical image privacy protection as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the structure-aware adaptive steganography method for medical image privacy protection as described in any one of claims 1-5.

Images