Medical data dynamic access control system based on multi-modal feature fusion authentication

Abstract

The application discloses a medical data dynamic access control system based on multi-modal feature fusion authentication, and relates to the technical field of information security. The system comprises a multi-modal acquisition and preprocessing module, which is used for acquiring finger vein, iris and voiceprint features and generating a desensitization vector; a feature-level fusion authentication module extracts maximum correlation features by using a multi-class correlation analysis model and a deep neural network, generates a biological feature fusion fingerprint, and prevents forgery attacks by combining a multi-dimensional liveness detection mechanism; a TEE privacy computing module performs feature comparison in a hardware isolation environment; a dynamic access control module calculates a dynamic trust value based on a blockchain smart contract, biological authentication credibility, environmental risk factors and historical operation behaviors, and realizes dynamic granting of permissions; and an audit module ensures that operations are tamper-proof by means of hash chaining. The application effectively solves the problems of single features being easily broken, privacy leakage and static permissions being rigid, and realizes high-security dynamic access control of medical data.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a dynamic access control system for medical data based on multimodal feature fusion authentication. Background Technology

[0002] With the rapid development of medical informatization, hospital information systems (HIS) and electronic medical record systems (EMR) have become the core infrastructure of medical institutions, storing massive amounts of patient privacy data and sensitive medical records. Therefore, how to ensure the security of access to medical data and prevent data leakage and unauthorized access is a key issue in the current construction of medical informatization.

[0003] Current medical information systems mainly rely on traditional account + password or single biometric authentication for data access control. In terms of access control architecture, they generally adopt a static role-based access control (RBAC) model, which means that once medical staff are authorized for a certain role, they have fixed operating permissions within the validity period.

[0004] However, the aforementioned existing technologies have many defects and safety risks in actual medical applications: 1. Single-feature authentication has security bottlenecks: Single biometric features are easily forged, such as attacking facial recognition with high-definition photos or attacking fingerprint recognition with 3D-printed fingerprint films, posing a high risk of fraud. Although some studies have attempted to introduce multimodal biometrics, existing solutions are mostly simple feature splicing or weighted averaging, without fully considering the inherent correlation and complementarity between different modal features, resulting in limited improvement in recognition accuracy and anti-attack capabilities.

[0005] 2. Lack of continuous security assessment of the operation process: Existing systems usually adopt the "one-time authentication, full trust" mechanism, which only verifies identity at the time of login. If the terminal device is impersonated by others while the screen is unlocked, or if the operator behaves abnormally after logging in, such as exporting a large amount of data in a short period of time, the system often cannot identify and block it in time, which can easily lead to the leakage of sensitive medical data.

[0006] 3. Centralized storage of privacy data poses high risks: Traditional biometric templates are usually stored in a centralized database in plaintext or with simple encryption. Since biometrics are immutable, once the centralized server is attacked or the database is leaked, the user's biometric information will be permanently leaked, causing irreversible security risks.

[0007] 4. Static access control lacks flexibility and fine granularity: Static RBAC models cannot perceive changes in environmental context. In emergency scenarios such as medical emergencies, doctors may need to temporarily exceed their privileges or attempt to access the system in abnormal environments such as remote locations or late at night. Static policies cannot dynamically adjust access permissions based on real-time environmental risks and historical behavioral habits, resulting in insufficient permissions when emergency care is needed, and failing to achieve a balance between security and convenience.

[0008] In summary, there is an urgent need for a highly secure medical data access control system that can integrate multimodal biometrics, possess anti-fraud capabilities, protect biometric privacy, and dynamically adjust access permissions based on environment and behavior. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention proposes a dynamic access control system for medical data based on multimodal feature fusion authentication. By constructing a feature-level fusion authentication model based on MCCA and DNN, a hardware-isolated TEE privacy computing architecture, and a dynamic trust value evaluation mechanism based on blockchain smart contracts, this invention solves core technical problems such as the ease with which single biometric features can be forged, the risk of centralized storage of privacy data, and the lack of fine-grained adaptability of static permissions. This achieves joint optimization of the accuracy of identity authentication, data privacy security, and flexibility of access control in medical data access.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A dynamic access control system for medical data based on multimodal feature fusion authentication includes: The multimodal acquisition and preprocessing module is used to acquire the user's multimodal biometric data, and to perform quality assessment and desensitization processing on the multimodal biometric data to generate corresponding desensitized feature vectors; The feature-level fusion authentication module is used to receive the desensitized feature vector, extract the maximum correlation feature using a multi-canonical correlation analysis model, and map the heterogeneous features to a unified high-dimensional space through a deep neural network to generate a unique biometric fusion fingerprint. At the same time, it performs liveness verification on the user through a liveness detection mechanism. The TEE privacy computing module is used to compare the biometric fused fingerprint with a pre-stored template in a hardware-isolated trusted execution environment, and output the comparison result and biometric authentication confidence score. The dynamic access control module runs on the blockchain network and is used to calculate the access trust value and compare the access trust value with the security threshold to perform dynamic permission determination and execution. Audit and Traceability Module: Key data from the entire system operation process are hashed and stored on the blockchain to achieve tamper-proof and traceable auditing of operation records.

[0011] Specifically, the multimodal biometric data includes finger vein features, iris features, and voiceprint features, which are collected through a client.

[0012] Specifically, the steps for generating desensitized feature vectors in the multimodal acquisition and preprocessing module include: S11. The image variance of finger vein features and iris features is calculated using the Laplacian operator to evaluate sharpness, and the illumination uniformity is detected using grayscale histogram. If the sharpness or illumination uniformity does not meet the expected value, the client re-collects the corresponding data to finally obtain data that meets the quality assessment. S12. For data that has passed the quality assessment, perform Region of Interest (ROI) cropping: For finger vein features, use the Sober operator to locate the finger edge boundary and establish a normalized coordinate system to crop the dense vein area near the interphalangeal joint; for iris features, use the Hough transform to locate the inner circle of the pupil and the outer circle of the sclera, and use the Daugman rubber sheet model to unfold the annular iris region into a rectangular polar coordinate image; for voiceprint features, use the speech activity detection algorithm to remove the static segments at the beginning and end of the audio. S13. The original feature codes extracted from finger vein features, iris features, and voiceprint features after ROI cropping are concatenated with the hardware unique identifier of the client device to serve as salt values. A one-way hash digest is generated using the SHA-256 or SM3 algorithm, which serves as the desensitized feature vector for each modality feature.

[0013] Specifically, the steps of the feature-level fusion authentication module in generating a biometric fusion fingerprint include: S21. Construct the objective function of the Multiple Canonical Correlation Analysis (MCCA) model, and obtain the optimal projection vectors of each modality feature by solving the generalized eigenvalue problem. The objective function satisfies: ; in Indicates the first The autocovariance matrix of a biometric modality is used to describe the feature distribution within that modality. Representing modes With mode The cross-covariance matrix is ​​used to quantify the linear correlation between different modes. These are the optimal projection vectors corresponding to finger vein features, iris features, and voiceprint features, respectively. S22. Using the obtained optimal projection vector, project and stitch the features to generate a fused feature vector. The generation logic is as follows: ; in, These are the desensitization feature vectors corresponding to finger veins, iris, and voiceprint, respectively. S23, Fusing feature vectors The input is a deep neural network, which includes several fully connected layers. Through several layers of nonlinear spatial transformation, heterogeneous biological features are mapped to a unified high-dimensional space. The transformation logic of each layer is as follows: ; in, It is the ReLU activation function. and Here are the weight matrix and bias terms for this layer. For the first Output characteristics of layered networks; The output layer of the deep neural network is then regularized to generate a biometric fusion fingerprint. ; in, For output layer features, for The Middle One element, is the feature dimension of the output layer.

[0014] Specifically, the liveness detection mechanism of the feature-level fusion authentication module is as follows: (1) The user’s finger vein is detected by dual-wavelength near-infrared detection technology. The preset first near-infrared band and second near-infrared band infrared light are emitted. The absorption difference of hemoglobin to the first near-infrared band and the second near-infrared band infrared light is used to capture the pulsation rhythm of the subcutaneous blood flow of the user’s finger in real time by combining photoelectric volume plethysmography. If no periodic blood flow pulsation rhythm that meets the characteristics of a living body is detected, the finger vein liveness verification is deemed to have failed. (2) Obtain the user's audio signal by means of random text challenge response, analyze the spectral entropy, instantaneous frequency change and cepstral residual characteristics of the audio signal, and if the loss of high-frequency details and hardware sound field distortion of the recording playback attack are identified, the voiceprint liveness verification is determined to fail.

[0015] Specifically, the TEE privacy computing module performs the following steps: S31. Pre-store encrypted user biometric templates in a hardware-isolated memory region of the TEE trusted execution environment. The hardware-isolated memory region is isolated from external operating systems, processes, and network environments; S32. Receive the biometric fusion fingerprint generated by the feature-level fusion authentication module via the TEE dedicated security interface. Within a hardware-isolated memory region, a preset comparison algorithm is invoked to fuse biometric features into fingerprints. With pre-stored user biometric templates Perform similarity measurement and map the calculation results to The interval is used to obtain the biometric authentication confidence score. ; S33, Biometric authentication confidence score Real-time determination based on preset safety threshold: If If the result exceeds the preset security threshold, the identity matching result will be output as True, and the biometric authentication confidence score will be output simultaneously. Otherwise, output the identity matching result as False.

[0016] Specifically, the preset comparison algorithm mentioned in step S32 includes a cosine similarity algorithm and a normalized Euclidean distance algorithm. The formula for calculating similarity using the cosine similarity algorithm is as follows: ; The formula for calculating similarity using the normalized Euclidean distance algorithm is: ; in, This is the preset maximum Euclidean distance threshold.

[0017] Specifically, the steps for calculating the access trust value in the dynamic access control module are as follows: S41. Collect and quantify the full set of parameters used to calculate the access trust value. The parameters are defined as follows: Biometric authentication confidence score, used to quantify the consistency between the current visitor's biometric features and their pre-stored biometric templates; The environmental risk factor components include quantitative assignments for preset scenarios, including remote access. Access outside of working hours Other environmental scenarios are assigned values ​​according to preset rules; The weighting coefficients for each environmental risk factor component are preset according to the risk level of the medical scenario, satisfying the following requirements. ; The behavioral characteristic value of the current operation. The mean and standard deviation of user's historical behavior features; This is a role-level adjustment coefficient, mapped to a preset trust base based on the medical personnel's professional title and departmental jurisdiction. The mapping rule is that the higher the professional title, the greater the jurisdiction. The larger the assigned value; S42. Calculate the initial access trust value. The calculation formula is: ; in, , , To preset the global weighting coefficients, satisfying , Characterizing the core weights of biometric authentication Deduction weights that characterize environmental risk Auxiliary weights that represent historical behavior; S43, calculate the results Character level correction coefficient After weighted adjustment, the final access trust value is obtained. .

[0018] Specifically, the rules for dynamic permission determination and execution in the dynamic access control module are as follows: [The text abruptly shifts to a different topic] ...access trust value... The system compares the data in real time with the preset two-level security thresholds, triggering the corresponding permission policy. If the value exceeds the first security threshold, then grant the visitor full read and write permissions to the medical data. If the data falls between the second and first security thresholds, only read-only access to the de-identified medical data will be granted to the visitor, while triggering secondary authentication. If the value is less than the second security threshold, the access request will be rejected directly, and an alarm log will be generated.

[0019] In summary, the beneficial technical effects of the present invention are as follows: 1. Extremely high anti-fraud and environmental adaptability, solving the problem of single features being easily compromised and limited: This invention adopts a feature-level fusion algorithm based on MCCA (Multi-Canonical Correlation Analysis) and DNN (Deep Neural Network) to map three heterogeneous biometric features—finger vein, iris, and voiceprint—to a unified high-dimensional space to generate a unique fused fingerprint. Compared with simple feature splicing, this deep fusion fully utilizes the complementarity between modalities, significantly reducing the recognition error rate. At the same time, combined with a multi-dimensional liveness detection mechanism using dual-wavelength near-infrared spectroscopy and audio spectrum entropy analysis, it can effectively defend against spoofing attacks such as high-definition photos, 3D-printed fingerprint films, and audio playback.

[0020] 2. This invention achieves fine-grained access control based on dynamic trust values, breaking the rigid limitations of the static RBAC model: It innovatively constructs a dynamic access control mechanism based on blockchain smart contracts, no longer relying on a single static role authorization. The system introduces environmental risk factors (such as remote location, non-working hours) and historical operational behavior characteristics, using a weighted algorithm to calculate the user's access trust value in real time. When the trust value fluctuates, the system can automatically trigger permission downgrade or secondary verification mechanisms. This not only meets the flexibility requirements for data access in medical emergency scenarios but also promptly blocks abnormal behavior, effectively solving the security vulnerabilities caused by "one-time authentication, full trust," and achieving a dynamic balance between security and convenience.

[0021] 3. A hardware-level privacy protection system has been constructed, eliminating the risk of leakage from centralized storage: This invention utilizes TEE (Trusted Execution Environment) technology to forcibly restrict the biometric comparison process to the hardware isolation area of ​​the CPU, preventing external operating systems from spying on the comparison logic and feature data. At the same time, combined with the desensitization processing method of salted hashing, it ensures that even if the backend database is compromised by hackers, the leaked data cannot be reversed to restore the original biometric features. This hardware and software combined architecture completely solves the irreversible privacy leakage crisis caused by the centralized plaintext storage of biometric features in traditional medical information systems.

[0022] 4. End-to-end trusted auditing, meeting the stringent regulatory requirements of the medical industry: This invention utilizes the immutability of blockchain technology to hash and store all access requests, authentication results, policy determination processes, and alarm logs on the blockchain. This distributed auditing mechanism prevents administrators or internal personnel from maliciously tampering with or deleting logs, providing credible legal evidence for accountability and tracing of medical data breaches, and fully complies with the stringent requirements of the medical industry for data security compliance. Attached Figure Description

[0023] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0024] To make the technical means, creative features, objectives and effects of this invention clearer and easier to understand, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0025] Example like Figure 1 As shown, the medical data dynamic access control system based on multimodal feature fusion authentication provided by this invention specifically includes: The multimodal acquisition and preprocessing module is used to collect users' multimodal biometric data through the client, including finger vein features, iris features and voiceprint features, and to perform quality assessment and desensitization processing on the multimodal biometric data to generate corresponding desensitized feature vectors; The specific steps for generating the desensitized feature vector include: S11. The image variance of finger vein features and iris features is calculated using the Laplacian operator to evaluate sharpness, and the illumination uniformity is detected using grayscale histogram. If the sharpness or illumination uniformity does not meet the expected value, the client re-collects the corresponding data to finally obtain data that meets the quality assessment. S12. For data that has passed the quality assessment, perform Region of Interest (ROI) cropping: For finger vein features, use the Sober operator to locate the finger edge boundary and establish a normalized coordinate system to crop the dense vein area near the interphalangeal joint; for iris features, use the Hough transform to locate the inner circle of the pupil and the outer circle of the sclera, and use the Daugman rubber sheet model to unfold the annular iris region into a rectangular polar coordinate image; for voiceprint features, use the speech activity detection algorithm to remove the static segments at the beginning and end of the audio. S13. The original feature codes extracted from finger vein features, iris features, and voiceprint features after ROI cropping are concatenated with the hardware unique identifier of the client device to serve as salt values. A one-way hash digest is generated using the SHA-256 or SM3 algorithm, which serves as the desensitized feature vector for each modality feature.

[0026] The feature-level fusion authentication module is used to receive the desensitized feature vector, extract the maximum correlation feature using a multi-canonical correlation analysis model, and map the heterogeneous features to a unified high-dimensional space through a deep neural network to generate a unique biometric fusion fingerprint. This step specifically includes: S21. Construct the objective function of the Multiple Canonical Correlation Analysis (MCCA) model. Obtain the optimal projection vectors of each modality feature by solving the generalized eigenvalue problem, maximizing the sum of the canonical correlation coefficients of each modality feature within the shared subspace. The objective function satisfies: ; in Indicates the first The autocovariance matrix of a biometric modality is used to describe the feature distribution within that modality. Representing modes With mode The cross-covariance matrix is ​​used to quantify the linear correlation between different modes. These are the optimal projection vectors corresponding to finger vein features, iris features, and voiceprint features, respectively. S22. Extract the top-K most relevant typical features from the obtained optimal projection vector and concatenate them. While eliminating redundant information between modes, retain the synergistic features with strong complementarity to complete the extraction of the most relevant features and generate a fused feature vector. The generation logic is as follows: ; in, These are the desensitization feature vectors corresponding to finger veins, iris, and voiceprint, respectively. S23, Fusing feature vectors The input is a deep neural network, which includes 3-5 fully connected layers, including a ReLU activation function and a Batch Normalization layer. Through several layers of nonlinear spatial transformation, heterogeneous biological features are mapped to a unified high-dimensional space. The transformation logic of each layer is as follows: ; in, It is the ReLU activation function. and Here are the weight matrix and bias terms for this layer. For the first Output characteristics of layered networks; The output layer of the deep neural network is then regularized to generate a fixed-length (e.g., 512 bits) and highly unique unit vector, which is the biometric fingerprint. : ; in, For output layer features, for The Middle One element, is the feature dimension of the output layer.

[0027] Meanwhile, the feature-level fusion authentication module performs liveness verification on users through a liveness detection mechanism, which specifically includes: (1) The user’s finger vein is detected by dual-wavelength near-infrared detection technology. The preset first near-infrared band and second near-infrared band infrared light are emitted. The absorption difference of hemoglobin to the first near-infrared band and the second near-infrared band infrared light is used to capture the pulsation rhythm of the subcutaneous blood flow of the user’s finger in real time by combining photoelectric volume plethysmography. If no periodic blood flow pulsation rhythm that meets the characteristics of a living body is detected, the finger vein liveness verification is deemed to have failed. (2) Obtain the user's audio signal by means of random text challenge response, analyze the spectral entropy, instantaneous frequency change and cepstral residual characteristics of the audio signal, and if the loss of high-frequency details and hardware sound field distortion of the recording playback attack are identified, the voiceprint liveness verification is determined to fail.

[0028] The TEE privacy computing module is used to compare biometric fused fingerprints with pre-stored templates in a hardware-isolated trusted execution environment, and output the comparison results and biometric authentication confidence scores. The module specifically performs the following steps: S31. Pre-store encrypted user biometric templates in a hardware-isolated memory region of the TEE trusted execution environment. The hardware-isolated memory region is isolated from external operating systems, processes, and network environments; S32. Receive the biometric fusion fingerprint generated by the feature-level fusion authentication module via the TEE dedicated security interface. Within a hardware-isolated memory region, a preset comparison algorithm is invoked to fuse biometric features into fingerprints. With pre-stored user biometric templates Perform similarity measurement and map the calculation results to The interval is used to obtain the biometric authentication confidence score. ; The preset comparison algorithms include cosine similarity algorithm and normalized Euclidean distance algorithm. The formula for calculating similarity using the cosine similarity algorithm is as follows: ; The formula for calculating similarity using the normalized Euclidean distance algorithm is: ; in, This is the preset maximum Euclidean distance threshold.

[0029] S33, Biometric authentication confidence score Real-time determination based on preset safety threshold: If If the result exceeds the preset security threshold, the identity matching result will be output as True, and the biometric authentication confidence score will be output simultaneously. Otherwise, output the identity matching result as False.

[0030] The dynamic access control module adopts a smart contract-based dynamic access control strategy to replace the traditional gateway. The smart contract on the consortium blockchain decides whether to allow data. The dynamic strategy logic adopts the ABAC extended model, and the trust value R is the sole logical basis for determining whether to allow, de-identify, or block data.

[0031] This module runs on the blockchain network. It first calculates an access trust value and then compares this value with a security threshold to perform dynamic permission determination and execution. The steps for calculating the access trust value are as follows: S41. Collect and quantify the full set of parameters used to calculate the access trust value. The parameters are defined as follows: Biometric authentication confidence score, used to quantify the consistency between the current visitor's biometric features and their pre-stored biometric templates; The environmental risk factor components include quantitative assignments for preset scenarios, including remote access. Access outside of working hours Other environmental scenarios are assigned values ​​according to preset rules; The weighting coefficients for each environmental risk factor component are preset according to the risk level of the medical scenario, satisfying the following requirements. ; The behavioral characteristic value of the current operation. The mean and standard deviation of user's historical behavior features; This is a role-level adjustment coefficient, mapped to a preset trust base based on the medical personnel's professional title and departmental jurisdiction. The mapping rule is that the higher the professional title, the greater the jurisdiction. The larger the assigned value; S42. Calculate the initial access trust value. The calculation formula is: ; in, , , To preset the global weighting coefficients, satisfying , Characterizing the core weights of biometric authentication Deduction weights that characterize environmental risk Auxiliary weights that represent historical behavior; S43, calculate the results Character level correction coefficient After weighted adjustment, the final access trust value is obtained. .

[0032] Furthermore, access trust value The system compares the data in real time with the preset two-level security thresholds, triggering the corresponding permission policy. If the value exceeds the first security threshold, then grant the visitor full read and write permissions to the medical data. If the data falls between the second and first security thresholds, only read-only access to the de-identified medical data will be granted to the visitor, while triggering secondary authentication. If the value is less than the second security threshold, the access request will be rejected directly, and an alarm log will be generated.

[0033] Audit and Traceability Module: Key data from the entire system operation process are hashed and stored on the blockchain to achieve tamper-proof and traceable auditing of operation records.

[0034] Therefore, the medical data dynamic access control system based on multimodal feature fusion authentication provided by this invention establishes a high-precision anti-fraud defense by using a feature-level fusion mechanism combining MCCA and DNN to map heterogeneous biometrics into a unified high-dimensional fingerprint, utilizing a privacy computing architecture based on TEE hardware isolation to perform "usable but invisible" security measurement of biometric templates, and finally constructing a core process based on blockchain smart contracts and multidimensional environmental risk factors. This realizes a closed-loop security control process of "multimodal deep fusion authentication - hardware-level privacy protection - context-aware dynamic authorization," solving the problems of identity forgery risks caused by relying on single features or simple splicing in existing medical information systems, as well as the pain points of rigid permissions and data leakage caused by the lack of continuous evaluation mechanism in static RBAC models. It improves the accuracy of identity recognition, the security of privacy data, and the flexibility of permission management in complex medical scenarios, significantly suppresses the risks of spoofing attacks and unauthorized access, and enhances the model's ability to perceive and adapt to abnormal operation behavior and emergency rescue environments. It provides an anti-fraud, adaptive, and auditable solution for information systems requiring high-security protection in digital hospitals and sensitive medical data management environments.

[0035] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic access control system for medical data based on multimodal feature fusion authentication, characterized in that, include: The multimodal acquisition and preprocessing module is used to acquire the user's multimodal biometric data, and to perform quality assessment and desensitization processing on the multimodal biometric data to generate corresponding desensitized feature vectors; The feature-level fusion authentication module is used to receive the desensitized feature vector, extract the maximum correlation feature using a multi-canonical correlation analysis model, and map the heterogeneous features to a unified high-dimensional space through a deep neural network to generate a unique biometric fusion fingerprint. At the same time, it performs liveness verification on the user through a liveness detection mechanism. The TEE privacy computing module is used to compare the biometric fused fingerprint with a pre-stored template in a hardware-isolated trusted execution environment, and output the comparison result and biometric authentication confidence score. The dynamic access control module runs on the blockchain network and is used to calculate the access trust value and compare the access trust value with the security threshold to perform dynamic permission determination and execution. Audit and Traceability Module: Key data from the entire system operation process are hashed and stored on the blockchain to achieve tamper-proof and traceable auditing of operation records.

2. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 1, characterized in that: The multimodal biometric data includes finger vein features, iris features, and voiceprint features, which are collected through a client.

3. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 2, characterized in that, The steps for generating desensitized feature vectors in the multimodal acquisition and preprocessing module include: S11. The image variance of finger vein features and iris features is calculated using the Laplacian operator to evaluate sharpness, and the illumination uniformity is detected using grayscale histogram. If the sharpness or illumination uniformity does not meet the expected value, the client re-collects the corresponding data to finally obtain data that meets the quality assessment. S12. For data that has passed the quality assessment, perform Region of Interest (ROI) cropping: For finger vein features, use the Sober operator to locate the finger edge boundary and establish a normalized coordinate system to crop the dense vein area near the interphalangeal joint; for iris features, use the Hough transform to locate the inner circle of the pupil and the outer circle of the sclera, and use the Daugman rubber sheet model to unfold the annular iris region into a rectangular polar coordinate image; for voiceprint features, use the speech activity detection algorithm to remove the static segments at the beginning and end of the audio. S13. The original feature codes extracted from finger vein features, iris features, and voiceprint features after ROI cropping are concatenated with the hardware unique identifier of the client device to serve as salt values. A one-way hash digest is generated using the SHA-256 or SM3 algorithm, which serves as the desensitized feature vector for each modality feature.

4. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 1, characterized in that, The steps of generating a biometric fusion fingerprint by the feature-level fusion authentication module include: S21. Construct the objective function of the Multiple Canonical Correlation Analysis (MCCA) model, and obtain the optimal projection vectors of each modality feature by solving the generalized eigenvalue problem. The objective function satisfies: ； in Indicates the first The autocovariance matrix of a biometric modality is used to describe the feature distribution within that modality. Representing modes With mode The cross-covariance matrix is ​​used to quantify the linear correlation between different modes. These are the optimal projection vectors corresponding to finger vein features, iris features, and voiceprint features, respectively. S22. Using the obtained optimal projection vector, project and stitch the features to generate a fused feature vector. The generation logic is as follows: ； in, These are the desensitization feature vectors corresponding to finger veins, iris, and voiceprint, respectively. S23, Fusing feature vectors The input is a deep neural network, which includes several fully connected layers. Through several layers of nonlinear spatial transformation, heterogeneous biological features are mapped to a unified high-dimensional space. The transformation logic of each layer is as follows: ； in, It is the ReLU activation function. and Here are the weight matrix and bias terms for this layer. For the first Output characteristics of layered networks; The output layer of the deep neural network is then regularized to generate a biometric fusion fingerprint. ； in, For output layer features, for The Middle One element, is the feature dimension of the output layer.

5. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 1, characterized in that, The liveness detection mechanism of the feature-level fusion authentication module is as follows: (1) The user’s finger vein is detected by dual-wavelength near-infrared detection technology. The preset first near-infrared band and second near-infrared band infrared light are emitted. The absorption difference of hemoglobin to the first near-infrared band and the second near-infrared band infrared light is used to capture the pulsation rhythm of the subcutaneous blood flow of the user’s finger in real time by combining photoelectric volume plethysmography. If no periodic blood flow pulsation rhythm that meets the characteristics of a living body is detected, the finger vein liveness verification is deemed to have failed. (2) Obtain the user's audio signal by means of random text challenge response, analyze the spectral entropy, instantaneous frequency change and cepstral residual characteristics of the audio signal, and if the loss of high-frequency details and hardware sound field distortion of the recording playback attack are identified, the voiceprint liveness verification is determined to fail.

6. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 1, characterized in that, The TEE privacy computing module specifically performs the following steps: S31. Pre-store encrypted user biometric templates in a hardware-isolated memory region of the TEE trusted execution environment. The hardware-isolated memory region is isolated from external operating systems, processes, and network environments; S32. Receive the biometric fusion fingerprint generated by the feature-level fusion authentication module via the TEE dedicated security interface. Within a hardware-isolated memory region, a preset comparison algorithm is invoked to fuse biometric features into fingerprints. With pre-stored user biometric templates Perform similarity measurement and map the calculation results to The interval is used to obtain the biometric authentication confidence score. ; S33, Biometric authentication confidence score Real-time determination based on preset safety threshold: If If the result exceeds the preset security threshold, the identity matching result will be output as True, and the biometric authentication confidence score will be output simultaneously. Otherwise, output the identity matching result as False.

7. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 6, characterized in that, The preset comparison algorithm mentioned in step S32 includes a cosine similarity algorithm and a normalized Euclidean distance algorithm. The formula for calculating similarity using the cosine similarity algorithm is: ； The formula for calculating similarity using the normalized Euclidean distance algorithm is: ； in, This is the preset maximum Euclidean distance threshold.

8. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 7, characterized in that, The steps for calculating the access trust value in the dynamic access control module are as follows: S41. Collect and quantify the full set of parameters used to calculate the access trust value. The parameters are defined as follows: Biometric authentication confidence score, used to quantify the consistency between the current visitor's biometric features and their pre-stored biometric templates; The environmental risk factor components include quantitative assignments for preset scenarios, including remote access. Access outside of working hours Other environmental scenarios are assigned values ​​according to preset rules; The weighting coefficients for each environmental risk factor component are preset according to the risk level of the medical scenario, satisfying the following requirements. ; The behavioral characteristic value of the current operation. The mean and standard deviation of user's historical behavior features; This is a role-level adjustment coefficient, mapped to a preset trust base based on the medical personnel's professional title and departmental jurisdiction. The mapping rule is that the higher the professional title, the greater the jurisdiction. The larger the assigned value; S42. Calculate the initial access trust value. The calculation formula is: ； in, , , To preset the global weighting coefficients, satisfying , Characterizing the core weights of biometric authentication Deduction weights that characterize environmental risk Auxiliary weights that represent historical behavior; S43, calculate the results Character level correction coefficient After weighted adjustment, the final access trust value is obtained. .

9. The medical data dynamic access control system based on multimodal feature fusion authentication according to claim 8, characterized in that, The rules for dynamic permission determination and execution in the dynamic access control module are as follows: Access trust value... The system compares the data in real time with the preset two-level security thresholds, triggering the corresponding permission policy. If the value exceeds the first security threshold, then grant the visitor full read and write permissions to the medical data. If the data falls between the second and first security thresholds, only read-only access to the de-identified medical data will be granted to the visitor, while triggering secondary authentication. If the value is less than the second security threshold, the access request will be rejected directly, and an alarm log will be generated.

Images