Private data protection method and system based on homomorphic encryption and federated learning

Abstract

The invention relates to the technical field of privacy data protection, and discloses a privacy data protection method and system based on homomorphic encryption and federated learning, and the method comprises the steps: processing original health data collected by Internet of Things equipment of a plurality of pension institutions, and obtaining a preprocessing data set and a privacy risk assessment result; encrypting to obtain a double encrypted data set; constructing a decentralized federated learning framework, and calculating global training parameters; performing noise injection processing and model iteration training to obtain a privacy protection model; performing trusted execution environment deployment on the privacy protection model, and performing identity verification through role-based multi-level access control and zero-knowledge proof to obtain a security access strategy; edge node distributed deployment is conducted on the privacy protection model, cross-mechanism health data analysis is conducted through a secure multi-party computing protocol, a collaborative analysis result is obtained, various privacy protection technologies are adopted, and efficient utilization of data is achieved while data privacy is protected.

Description

Technical Field

[0001] The present invention relates to the field of privacy data protection technology, and in particular to a privacy data protection method and system based on homomorphic encryption and federated learning. Background Art

[0002] As the global aging process accelerates, the health management of the elderly population has become one of the major challenges facing the medical systems of various countries. Chronic diseases such as diabetes, hypertension, and heart disease are common health problems among the elderly. These diseases require long-term and continuous monitoring and management to ensure that the patient's condition is effectively controlled. The traditional health management model mainly relies on regular hospital visits and physical examinations, but this method is inefficient and cannot monitor the patient's health status in real time.

[0003] The smart health management platform based on the Internet of Things technology can collect patients' physiological data in real time by connecting smart health devices to the Internet, and upload these data to the cloud platform for analysis and monitoring. Doctors and health managers can use the remote monitoring system to keep abreast of patients' health dynamics and provide personalized health advice and intervention measures. However, when these platforms collect, transmit and store a large amount of personal sensitive data, once the data is leaked or abused, it may pose a serious threat to patients' privacy. Although the existing health management system adopts basic security measures such as encryption technology and data access control, these measures are powerless in the face of complex and changeable network attacks and data abuse threats. At the same time, the traditional centralized data storage and processing mode also brings a high risk of data leakage. In order to meet the high standards of privacy protection, it is urgent to introduce new privacy protection technologies such as differential privacy, homomorphic encryption, and federated learning into the smart health management system. Summary of the invention

[0004] The present invention provides a privacy data protection method and system based on homomorphic encryption and federated learning. The present invention adopts a variety of privacy protection technologies to achieve efficient use of data while protecting data privacy.

[0005] In a first aspect, the present invention provides a privacy data protection method based on homomorphic encryption and federated learning, and the privacy data protection method based on homomorphic encryption and federated learning includes: The original health data collected by IoT devices in multiple nursing homes were processed by Laplace noise addition and data standardization to obtain preprocessed data sets and privacy risk assessment results; According to the privacy risk assessment result, the numerical data in the preprocessed data set is subjected to BFV homomorphic encryption processing, and the neural network calculation data is subjected to CKKS homomorphic encryption processing to obtain a double encrypted data set; A decentralized federated learning framework is constructed based on the double encrypted data set, and training weights are calculated through data heterogeneity analysis and a federated optimization algorithm to obtain global training parameters; The global training parameters are input into the gradient clipping module for noise injection processing, and the model is iteratively trained based on the differential privacy budget constraint to obtain a privacy protection model; Deploy the privacy protection model in a trusted execution environment, perform identity authentication through role-based multi-level access control and zero-knowledge proof, and obtain a secure access policy; Based on the secure access strategy, the privacy protection model is deployed in a distributed manner on edge nodes, and cross-institutional health data analysis is performed through a secure multi-party computing protocol to obtain collaborative analysis results.

[0006] In a second aspect, the present invention provides a privacy data protection system based on homomorphic encryption and federated learning, and the privacy data protection system based on homomorphic encryption and federated learning includes: The standardization module is used to add Laplace noise and perform data standardization on the raw health data collected by IoT devices in multiple nursing homes to obtain preprocessed data sets and privacy risk assessment results; An encryption module, used to perform BFV homomorphic encryption processing on the numerical data in the preprocessed data set and perform CKKS homomorphic encryption processing on the neural network calculation data according to the privacy risk assessment result, to obtain a double encrypted data set; A computing module, used to construct a decentralized federated learning framework based on the double encrypted data set, calculate training weights through data heterogeneity analysis and a federated optimization algorithm, and obtain global training parameters; A training module, used for inputting the global training parameters into a gradient clipping module for noise injection processing, performing model iterative training based on a differential privacy budget constraint, and obtaining a privacy protection model; A verification module, used to deploy the privacy protection model in a trusted execution environment, perform identity authentication through role-based multi-level access control and zero-knowledge proof, and obtain a secure access policy; The analysis module is used to perform distributed edge node deployment of the privacy protection model based on the security access policy, perform cross-institutional health data analysis through a secure multi-party computing protocol, and obtain collaborative analysis results.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned privacy data protection method based on homomorphic encryption and federated learning.

[0008] The fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned privacy data protection method based on homomorphic encryption and federated learning.

[0009] In the technical solution provided by the present invention, by constructing a multi-level privacy protection architecture, the security protection of privacy data is achieved in the whole process of data collection, transmission, storage and analysis. Specifically, data preprocessing is achieved through Laplace noise addition and data standardization, and BFV and CKKS dual homomorphic encryption are used to ensure data transmission security; data collaborative analysis among multiple institutions is achieved based on a decentralized federated learning framework, and the model training process is protected by differential privacy budget constraints and gradient clipping; the Intel SGX and ARM TrustZone dual trusted execution environment deployment model is used, combined with role-based multi-level access control and zero-knowledge proof for identity authentication; finally, cross-institutional collaborative analysis of health data is achieved through edge computing and secure multi-party computing. The present invention not only effectively solves the problem of data privacy leakage in traditional health management systems, but also realizes efficient use of data while protecting data privacy by innovatively integrating multiple privacy protection technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of the steps of a privacy data protection method based on homomorphic encryption and federated learning in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a privacy data protection system based on homomorphic encryption and federated learning in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a privacy data protection method and system based on homomorphic encryption and federated learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the privacy data protection method based on homomorphic encryption and federated learning in an embodiment of the present invention includes: Step S1, performing Laplace noise addition and data standardization processing on the original health data collected by IoT devices of multiple nursing homes to obtain a preprocessed data set and privacy risk assessment results; It is understandable that the execution subject of the present invention can be a privacy data protection system based on homomorphic encryption and federated learning, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0014] Specifically, the original health data collected by IoT devices in multiple nursing homes are collected. The health data monitored by IoT devices mainly include chronic disease monitoring data such as blood pressure, blood sugar, heart rate and blood oxygen. The temporal integrity of these original data is verified. The verification process can identify the situation of missing data or abnormal timestamps and obtain the validity index of the original data. Based on the validity index of the original data, the original health data is evaluated in multiple dimensions. The evaluation includes the accuracy, completeness and consistency of the data. By analyzing the correlation between each data dimension and other data, outliers are found, and corresponding correction measures are taken to ensure the reliability of the data, and the preliminary cleaned data is obtained. The cleaned data is input into the Intel SGX trusted execution environment for classification and privacy level classification. As a hardware-level secure execution environment, Intel SGX ensures the security of the data processing process and helps to protect privacy without exposing the data. Through this process, the data is divided into different security levels according to its privacy risk, and the multimodal security classification results are obtained. According to the multimodal security classification results, a three-layer evaluation matrix is ​​constructed. The first layer evaluates the sensitivity of the data and calculates the sensitivity score of each data item; the second layer analyzes the exposure scope of the data and obtains the exposure score of each data item in different scenarios; the third layer performs statistics based on the frequency of data use to obtain the frequency of use score. The data sensitivity score matrix is ​​generated by weighted combination of the three-layer scores. According to the results of the data sensitivity score matrix, the distribution parameters of Laplace noise are determined. The addition of Laplace noise is a differential privacy technology that prevents sensitive information leakage by injecting noise into the data. Through directional noise injection, the data is ensured to be used for analysis and modeling without exposing sensitive information, and a privacy-protected data set is obtained. The noise-processed data is grouped according to the type of physiological indicators of the elderly and standardized so that different types of data can be compared and analyzed on a unified scale to obtain a standardized preprocessed data set. Deep feature extraction and correlation analysis are performed on the preprocessed data set. Meaningful features are extracted from the data and the correlation between each data item is identified. Through deep feature extraction, the structure and rules of the data are understood at a higher level to obtain data feature mapping information. A risk assessment model is constructed based on the data feature mapping information and the data sensitivity score matrix. Through the constructed risk assessment model, multiple threshold judgments are made to determine the privacy risk level of the data, quantify the potential risk of data leakage, and take corresponding privacy protection measures based on the risk assessment results. Obtain the privacy risk assessment results.

[0015] Step S2: According to the privacy risk assessment results, the numerical data in the preprocessed data set is subjected to BFV homomorphic encryption, and the neural network calculation data is subjected to CKKS homomorphic encryption to obtain a double encrypted data set; Specifically, the preprocessed data set is classified according to the privacy risk assessment results. Basic physiological indicators such as blood pressure and blood sugar are divided into numerical health data subsets, and complex physiological data such as electrocardiograms are divided into neural network calculation data subsets. Through classification, suitable encryption methods can be effectively applied to different types of data to improve the efficiency and accuracy of privacy protection. For chronic disease monitoring data in the numerical health data subset, basic operation requirements analysis is performed, the complexity of basic addition and multiplication operations required for these data is analyzed, and their numerical calculation parameters are evaluated to help determine the key generation strategy and encryption parameters required for subsequent encryption calculations. According to the numerical calculation parameters, a BFV homomorphic encryption key generation strategy is constructed, and the corresponding encryption key pair is generated using the Intel SGX trusted execution environment. The BFV homomorphic encryption operation is performed on the numerical health data subset to obtain the first encrypted data set. By adopting the BFV homomorphic encryption technology, it is ensured that the data remains encrypted during the calculation process, and addition and multiplication operations can be performed without leaking the data itself. At the same time, the neural network calculation data subset is analyzed, and the multi-layer perceptron algorithm is used to balance the accuracy requirements and computational overhead of the data to determine the appropriate CKKS encryption parameters. CKKS encryption is suitable for neural network data, which can handle higher precision floating point operations and maintain data privacy during the calculation process. The neural network optimization configuration is input into the ARM TrustZone trusted execution environment, and a CKKS encryption key pair is generated in this environment to ensure that the key generation process is also protected. Perform homomorphic encryption operations on the subset of neural network calculation data to obtain the second encrypted data set. Distributed storage node planning is performed for the first encrypted data set and the second encrypted data set. Blockchain technology is used to record data distribution and access logs to ensure that the data storage and access process is transparent and traceable. Based on the location information of data storage, a secure data sharing channel is constructed between nursing homes. By using the secret sharing mechanism, the secure transmission of encrypted data between multiple nursing homes is ensured. This mechanism can divide the data into multiple ciphertext parts. Only with the cooperation of all participants can the original data be restored, thereby preventing the risk of single point leakage. Based on the distribution results of encrypted data, the Yao encryption circuit protocol is designed, through which the encrypted data synchronization calculation between nodes of nursing homes is realized. A double encrypted data set is obtained.

[0016] Step S3: construct a decentralized federated learning framework based on the double encrypted data set, calculate the training weights through data heterogeneity analysis and federated optimization algorithm, and obtain global training parameters; Specifically, multi-level feature extraction is performed on the chronic disease monitoring data of the elderly in the double encrypted data set. The first layer extracts the time series features of the data to obtain the time series feature matrix, capturing the law of the change of the health status of the elderly over time; the second layer extracts the category features of the data and generates the category feature matrix to help identify the category attributes of different health states; the third layer extracts the association features of the data and generates the association feature matrix to reveal the correlation between various health indicators. Through the fusion operation of these features, the data distribution feature vector is obtained. The degree of difference in data distribution between different nursing homes is calculated based on the data distribution feature vector. The combined calculation method of cosine similarity and Euclidean distance is used. Cosine similarity measures the similarity of different data distributions, while Euclidean distance quantifies the absolute value of the difference in data distribution. Through calculation, the heterogeneity matrix between institutions is obtained, reflecting the differences in health data distribution among various nursing homes. The heterogeneity matrix is ​​input into the federated optimization algorithm module, which can adjust the parameters of the training strategy according to the data heterogeneity characteristics of different nursing homes, ensuring that in the federated learning process, the model update of each institution can adapt to its specific data distribution and obtain appropriate federated learning configuration information. The FedAvg algorithm framework is constructed based on the federated learning configuration information. The framework uses the model updates on the local devices of multiple nursing homes to aggregate the update results of each local model through weighted average operation to obtain the global model aggregation parameters. This process enables the local models of each nursing home to be updated synchronously without centralizing all data to the cloud, effectively protecting data privacy. The model aggregation parameters are calculated by secret sharing. The model aggregation parameters are divided into different shares and distributed to each nursing home. Through the secret sharing mechanism, it is ensured that each institution only has part of the parameters to prevent the risk of single point failure or data leakage. The parameters distributed to each nursing home are securely synchronized during the cross-institutional calculation process, and the cross-institutional parameter aggregation is performed using a secure multi-party computing model to obtain the globally updated training parameters. The globally updated parameters are input into the differential privacy protection module for privacy protection processing. In this module, gradient clipping and noise addition techniques are used for privacy protection. Gradient clipping is used to control the gradient change amplitude during model training to prevent gradient information from leaking sensitive data; noise addition adds noise to the gradient, so that even if the data is obtained during model training, the original data cannot be accurately restored, effectively preventing data leakage. After privacy protection processing, the leakage-proof training parameters are obtained. After completing the privacy protection, the anti-leakage training parameters are used to verify the model performance and evaluate the privacy protection. In order to ensure that the model has high accuracy while protecting privacy, the model is continuously improved through multiple rounds of iterative optimization, and the privacy protection effect is evaluated through the threshold judgment method.Multiple rounds of iterative optimization gradually improve the model performance while maintaining its privacy protection effect; and threshold judgment is used to determine whether privacy protection meets the predetermined standards, ensuring that the privacy protection measures are effective and will not significantly affect the training effect of the model. Through these steps, the global training parameters are finally obtained.

[0017] Step S4: input the global training parameters into the gradient clipping module for noise injection processing, perform model iterative training based on the differential privacy budget constraint, and obtain a privacy protection model; Specifically, the gradient sensitivity analysis of the physiological index characteristics of the elderly in the global training parameters is performed to evaluate the sensitivity of the gradient change to the model training, so as to ensure that sensitive information will not be leaked during the model training process. By using the Intel SGX trusted execution environment, the L2 norm of the gradient vector is securely calculated to obtain the gradient clipping threshold. The L2 norm of the gradient is a common method to measure the gradient size, which can help the system identify the maximum range of gradient change, so as to set a reasonable clipping threshold. Ensure that there is no excessive change when the gradient is updated, and avoid the risk of privacy leakage caused by excessive gradients. Constrain the parameter update of the neural network model based on the gradient clipping threshold. The gradient projection method is used to limit the gradient size of each dimension. The gradient projection method projects the gradient vector to a specific subspace to ensure that the update of each dimension does not exceed the predetermined limit, and obtains the clipped parameter matrix. The multi-layer perceptron algorithm analysis is performed on the clipped parameter matrix. By analyzing the balance between the model's data accuracy and privacy protection requirements, the appropriate Laplace noise distribution parameters are determined, and the noise injection strategy is obtained. During the model training process, the noise injection strategy is applied to the gradient update of each round of iteration. Random noise is added to each round of update through the differential privacy mechanism to obtain the perturbed training parameters. The privacy budget consumption of the perturbed training parameters is calculated. The privacy budget is a way to measure the cost of privacy protection of the differential privacy mechanism. In order to ensure that privacy protection does not exceed the predetermined budget limit, the budget allocation ratio is dynamically adjusted according to the privacy risk assessment mechanism. Dynamic adjustment can flexibly allocate the budget according to the privacy protection requirements of each round of training to ensure that the expected privacy protection goals are achieved without affecting the model training effect. According to the budget constraint scheme, a multi-round iterative training framework is constructed, and the local model parameters are aggregated according to the budget limit to obtain the global model update sequence. The global model update sequence is input into the ARM TrustZone trusted execution environment for security verification. ARM TrustZone provides a hardware-level security mechanism that can ensure that no data leakage occurs during the model update process. Security verification is performed through the Yao encrypted circuit protocol to strictly evaluate the security of the model. The Yao encrypted circuit protocol is a secure multi-party computing protocol that can verify the correctness and security of model parameters without exposing any original data. In this embodiment, it is ensured that during the federated learning process, the calculation results of multiple participants will not leak any sensitive information. A comprehensive evaluation is performed on the security score and training accuracy of the model to ensure that the model can maintain a high prediction accuracy while protecting privacy. In the evaluation process, not only the training accuracy of the model is paid attention to, but also whether the privacy protection measures have excessively affected the model performance. Through comprehensive evaluation, the relationship between privacy protection and model performance is balanced to obtain a privacy protection model.

[0018] Step S5: deploy the privacy protection model in a trusted execution environment, perform identity authentication through role-based multi-level access control and zero-knowledge proof, and obtain a secure access policy; Specifically, the access rights of monitoring data such as blood pressure, blood sugar, heart rate and blood oxygen in the privacy protection model are hierarchically processed. By building a role-based access control model (RBAC), users with different roles (such as caregivers, family members, doctors, etc.) are assigned to different access levels, and a multi-level access control matrix is ​​generated accordingly. The matrix shows which roles access which data and sets the permissions for different roles to access data to achieve fine-grained control. A zero-knowledge proof protocol is built based on the multi-level access control matrix to ensure that the authentication of access requests does not require the disclosure of any sensitive information of the user. Zero-knowledge proof allows a party to prove that it knows a certain information through a mathematical algorithm without revealing the information itself. In the access rights verification process, zero-knowledge proof is used to verify the access rights of roles such as caregivers and family members, and an identity authentication protocol is formed by generating zero-knowledge proof parameters. The identity authentication protocol is deployed in the Intel SGX trusted execution environment. IntelSG provides a hardware-level secure execution environment. Through TEE (trusted execution environment) hardware isolation technology, it ensures that data and calculations during model execution are not affected by external attacks, and a secure execution area is obtained. Edge computing nodes are configured for the secure execution area. Edge computing nodes can perform calculations and analysis at the location where data is generated, thereby reducing data transmission delays and improving data privacy protection. Through the ARM TrustZone trusted execution environment, a model access control interface is built to strengthen the model's access control capabilities. ARM TrustZone provides hardware-level security protection for IoT devices, which can effectively isolate the device's main processor and security processor to ensure security during the computing process. By configuring the edge security environment, it is ensured that edge devices do not expose sensitive information when performing model reasoning and data access. Connect the edge computing environment to the secure multi-party computing network. Data between multiple nursing homes is collaboratively calculated without exposing the original data. Through the secret sharing mechanism, access requests are dispersed to multiple nodes for verification, ensuring that each request is securely and distributedly verified. The secret sharing mechanism divides the data into several copies and distributes them to different nodes. Only when all nodes work together can the complete information be restored. Effectively avoid single point failures and data leakage risks and improve the security of the system. After successfully completing the access verification, the blockchain technology is used to record the data access between the nodes of the nursing home. Blockchain provides a distributed ledger technology that ensures that every data access is accurately recorded and can trace historical access records. Through the data access recorded by the blockchain, an access tracing mechanism is established, which can provide behavioral audit data for future security audits. Differential privacy analysis is performed on behavioral audit data. Differential privacy is a technology that ensures that individual privacy is not leaked when statistical analysis is performed on data.The Laplace mechanism is used to desensitize the access behavior characteristics, cover up the sensitive information of the access data, avoid the leakage of any individual information, and obtain a privacy protection audit report. Based on the privacy protection audit report, a dynamic access control strategy is constructed. Through this strategy, the system can dynamically adjust access rights and strategies according to the real-time privacy protection requirements and the results of the audit report. In order to ensure the security and effectiveness of the access control strategy, the Yao encryption circuit is used to update the security strategy to obtain a secure access strategy.

[0019] Step S6: Based on the secure access policy, the privacy protection model is distributedly deployed on edge nodes, and cross-institutional health data analysis is performed through a secure multi-party computing protocol to obtain collaborative analysis results.

[0020] Specifically, the edge node performs a preliminary analysis of the chronic disease monitoring data of the elderly collected by IoT devices according to the security access policy. The information is uploaded to the cloud only when an abnormality is detected, minimizing the risk of data exposure and obtaining local anomaly detection results. The local anomaly detection results are processed by BFV homomorphic encryption, and a data security computing environment is built using the Intel SGX trusted execution environment to ensure that the processing of encrypted data will not be affected by external attacks, and the first layer of encrypted data is obtained to ensure that the data remains encrypted during transmission and calculation. The first layer of encrypted data is input into the CKKS homomorphic encryption module for encryption processing, and the ARM TrustZone technology is used to provide hardware-level protection to ensure that the encryption process is not affected by malware, and double encrypted analysis data is obtained. The double encrypted analysis data is used to build a group intelligence federated learning framework between the nodes of the nursing home. Under this framework, the FedAvg algorithm is used for collaborative training, and the model is updated on the local devices of each nursing home, and then the global model is aggregated by weighted average. Multiple nursing homes can share useful information in the data while protecting data privacy, thereby optimizing the global prediction model. Through a decentralized learning method, a distributed prediction model is obtained to effectively predict changes in the health status of the elderly and help identify health risks in a timely manner. The distributed prediction model is calculated by secret sharing. By dividing the model parameters into multiple shares and distributing them to various nursing homes, it is ensured that each institution can only see part of the data and cannot obtain the complete model parameters. Only when all participating institutions cooperate can the complete model information be restored. In this process, the secret sharing of model parameters protects data privacy and prevents the possibility of data abuse by any single institution. The parameters of the joint analysis are input into the Yao encrypted circuit protocol to ensure secure model parameter aggregation among multiple nursing homes. The Yao encrypted circuit protocol is a secure multi-party computing protocol that allows all participants to perform collaborative computing without exposing any sensitive data to obtain cross-institutional analysis data. Differential privacy processing is performed on cross-institutional analysis data. By adopting the Laplace mechanism, sensitive information in the analysis results is desensitized. Through differential privacy technology, even if the analysis results are maliciously accessed, the sensitive information of the data cannot be recovered, protecting the privacy of users. Health trend statistical analysis is performed based on the privacy protection analysis results to help the system identify the laws of group health changes and take timely intervention measures. The analysis process and results are recorded through blockchain technology. The distributed ledger feature of blockchain ensures the transparency and traceability of data, so that each step of the data analysis process can be audited and traced, and a collaborative analysis result that has been securely verified by multiple parties is obtained.

[0021] In the embodiment of the present invention, by constructing a multi-level privacy protection architecture, the security protection of privacy data is achieved in the whole process of data collection, transmission, storage and analysis. Specifically, data preprocessing is achieved through Laplace noise addition and data standardization, and BFV and CKKS dual homomorphic encryption are used to ensure data transmission security; data collaborative analysis among multiple institutions is achieved based on a decentralized federated learning framework, and the model training process is protected by differential privacy budget constraints and gradient clipping; the Intel SGX and ARM TrustZone dual trusted execution environment deployment model is used, combined with role-based multi-level access control and zero-knowledge proof for identity authentication; finally, cross-institutional collaborative analysis of health data is achieved through edge computing and secure multi-party computing. The present invention not only effectively solves the problem of data privacy leakage in traditional health management systems, but also realizes efficient use of data while protecting data privacy by innovatively integrating multiple privacy protection technologies.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Collect the original health data corresponding to the IoT devices of multiple nursing homes, and verify the time series integrity of the chronic disease monitoring data of blood pressure, blood sugar, heart rate, and blood oxygen collected by the IoT devices to obtain the validity index of the original data; Based on the original data validity index, the original health data is evaluated in multiple dimensions and outliers are corrected to obtain preliminary cleaned data. The preliminary cleaned data is then input into the Intel SGX trusted execution environment for data classification and privacy level division to obtain multimodal security classification results. A three-layer evaluation matrix is ​​constructed for the multimodal security classification results. The first layer calculates data sensitivity to obtain a sensitivity score. The second layer analyzes the data exposure range to obtain an exposure score. The third layer performs data usage frequency statistics to obtain a frequency score. The data sensitivity score matrix is ​​obtained through a weighted combination of the three-layer scores. The Laplace noise distribution parameters are determined according to the data sensitivity score matrix, and the preliminary cleaned data is subjected to directional noise injection to obtain a privacy-preserving dataset. The privacy-preserving dataset is then grouped and standardized according to the physiological indicator types of the elderly to obtain a preprocessed dataset. Deep feature extraction and correlation analysis are performed on the preprocessed data set to obtain data feature mapping information. A risk assessment model is built based on the data feature mapping information and the data sensitivity scoring matrix, and the privacy risk assessment results are obtained through multiple threshold judgments.

[0023] Specifically, the original health data corresponding to the IoT devices of multiple nursing homes are collected, and the chronic disease monitoring data of blood pressure, blood sugar, heart rate, and blood oxygen collected by the IoT devices are verified for temporal integrity to ensure that the data is correctly recorded and stored in chronological order and is not lost or tampered with during transmission. When performing temporal integrity verification, the validity indicators of the data at each time point are calculated, including the accuracy of the timestamp, the continuity of the data, and the abnormal fluctuations of the data. For example, for blood sugar data, if there is an abrupt and large fluctuation within the normal time window (such as a sudden increase in blood sugar levels to an abnormally high value within 30 minutes), it is necessary to further check whether there is a problem with the data source or sensor. Based on the validity indicators, the original health data is evaluated for quality in multiple dimensions and outliers are corrected. During the quality assessment process, statistical analysis methods, such as Z-score or box plot analysis, are used to detect outliers in the data. Z-score is a way to measure the distance between a data point and the mean, and the calculation formula is: in, represents the i-th data point, represents the mean of the data set, is the standard deviation of the data set. If the Z value is greater than a certain set threshold (such as 3), the data point is considered an outlier and needs to be corrected or deleted. After outlier detection and correction, preliminary cleaned data is obtained. The cleaned data is input into the Intel SGX trusted execution environment for data classification and privacy level classification. Intel SGX is a hardware-level trusted execution environment that ensures that data remains encrypted during processing and cannot be accessed by external attackers even during execution. By running in Intel SGX, data is strictly protected for privacy and classified according to the privacy level of different data. The privacy level of data is determined by the sensitivity of the data. For example, blood sugar and heart rate data belong to high privacy level data, while blood pressure data belongs to medium privacy level data. After this step, a multimodal security classification result is obtained, in which each data category is labeled with a privacy level. The multimodal security classification result is evaluated by constructing a three-layer evaluation matrix to comprehensively analyze the sensitivity, exposure risk and frequency of use of the data. In the first layer, the sensitivity of the data is calculated to obtain a sensitivity score. By analyzing the content and usage scenarios of the data, a sensitivity score is assigned to each data item using a weighted algorithm. The sensitivity score is calculated as follows: ;in, Indicates The probability of a role or device accessing data, Indicates the risk level of the role or device, and m is the total number of roles or devices. In the third layer, the frequency of data use is counted to obtain a frequency score. The frequency of data use within a certain period of time and the possibility of frequent access are evaluated. The frequency score is calculated as follows: ;in, represents the access frequency of data in the kth time period, is the total number of time periods. By weighted combination of these three levels of scores, a comprehensive data sensitivity score matrix is ​​obtained, which can comprehensively evaluate the privacy risk and usage value of each data. On this basis, the distribution parameters of Laplace noise are determined according to the data sensitivity score matrix. Laplace noise is a mechanism for differential privacy that protects the privacy of data and prevents the leakage of sensitive information by adding noise to the data. The distribution parameters of Laplace noise are dynamically adjusted according to the sensitivity score of the data. The formula for noise injection is as follows: ; in, represents the sensitivity of the function, It's the privacy budget. and is the parameter of the Laplace distribution. According to the calculated noise distribution parameters, the preliminary cleaned data is subjected to directional noise injection to obtain a privacy-preserving data set. The privacy-preserving preprocessed data set is grouped and standardized according to the types of physiological indicators of the elderly. For example, data such as blood sugar, blood pressure, and heart rate are standardized separately so that each type of data has the same scale. The standardized data is input into the deep feature extraction and correlation analysis module, and the data feature mapping information is obtained by extracting the high-order features of the data and analyzing the correlation between the various data indicators. A risk assessment model is constructed based on the data feature mapping information and the data sensitivity scoring matrix. The privacy risk assessment results are obtained by using the multiple threshold judgment method.

[0024] In a specific embodiment, the execution step constructs a three-layer evaluation matrix for the multimodal security classification results, the first layer performs data sensitivity calculation to obtain a sensitivity score, the second layer performs data exposure range analysis to obtain an exposure score, and the third layer performs data usage frequency statistics to obtain a frequency score. The process of obtaining a data sensitivity score matrix by weighted combination of the three-layer scores can specifically include the following steps: Map the data items of different types of health data in the multimodal security classification results, encode the data in layers according to the privacy attributes, obtain the initial data sensitivity matrix, and establish a sensitivity evaluation index system for each type of health data based on the initial data sensitivity matrix, and calculate the first-level sensitivity score; The multimodal security classification results are graded according to the data access scope, and the data exposure probability is calculated through the access control matrix to obtain the data exposure risk vector. The data exposure risk vector is normalized to obtain the second-level exposure score. Perform data access frequency statistics on the multimodal security classification results based on the preset time window, construct a frequency distribution histogram, obtain the original frequency characteristics, and perform time series analysis on the original frequency characteristics. Calculate the usage frequency trend through the exponential smoothing method to obtain the third-level frequency score; The first layer sensitivity score, the second layer exposure score and the third layer frequency score are input into the multi-layer neural network, and the weight parameters are optimized by the back propagation algorithm to obtain the score fusion weight; Based on the score fusion weight, the first-layer sensitivity score, the second-layer exposure score, and the third-layer frequency score are weighted and summed to obtain the data sensitivity score matrix.

[0025] Specifically, data items of different types of health data are mapped and hierarchically encoded according to privacy attributes. For different health data, such as blood pressure, blood sugar, heart rate, and blood oxygen, each data item is classified according to its impact on personal privacy. Based on the classification, each type of data item is hierarchically encoded and a preliminary data sensitivity matrix is ​​formed, in which each data item corresponds to a sensitivity score, indicating the privacy risk and confidentiality of the data. Based on the initial data sensitivity matrix, a sensitivity evaluation index system is constructed according to the privacy attributes of each type of health data. Data is evaluated according to the type of data, collection method, and usage scenario. The calculation formula for the sensitivity score is as follows: ;in, represents the weight of the i-th feature, Indicates the privacy sensitivity level of the feature, is the total number of data items. Through weighted calculation, the first-level sensitivity score is obtained, which reflects the privacy risk of each health data. Taking into account the access control of the data, the data is graded according to the multimodal security classification results. Different health data have different access scopes and usage permissions. An access control matrix is ​​defined for each data item. The exposure probability of the data is calculated through the matrix to evaluate the exposure risk of the data. The calculation of the exposure probability is based on factors such as the access frequency of the data, the access permissions of the visitor, and the access channel. The calculation formula for the exposure probability is: ;in, Indicates The probability of visiting a role, Indicates the risk level of the role. is the total number of access roles. By calculating the exposure probability, we get the exposure risk vector, that is, the exposure risk of each data item under different access roles. The exposure risk vector is normalized to get the second-level exposure score, which measures the risk of each data exposure. Based on the preset time window, the multimodal security classification results are statistically analyzed for data access frequency to identify the access patterns of data in different time periods, and then determine which data is frequently accessed. For each data item, count the number of accesses in each time period and construct a frequency distribution histogram. Perform time series analysis on the original frequency features, and calculate the frequency trend of data usage by exponential smoothing. The exponential smoothing method smoothes the fluctuations of data by giving higher weights to the most recent time points to obtain a smooth trend forecast. The calculation formula is: ;in, represents the predicted value at time t, represents the actual data at time t, represents the predicted value at the previous moment, is a smoothing coefficient, which usually takes a value between 0 and 1. Through this method, the third-layer frequency score is obtained, which reflects the usage frequency trend of the data in the time series. The first-layer sensitivity score, the second-layer exposure score, and the third-layer frequency score are input into the multi-layer neural network for optimization. The neural network optimizes the weights through the back-propagation algorithm to find an optimal score fusion solution. During the training process, the neural network adjusts its weight parameters according to the importance of each score to the final output result to improve the accuracy of score fusion. The training formula of the neural network is: ;in, represents the loss function, Indicates the actual label, Represents the predicted value of the neural network, and n is the number of samples. Through the back propagation algorithm, the parameters in the neural network are adjusted to minimize the loss function and obtain the optimal score fusion weight. Based on the optimized score fusion weight, the first-layer sensitivity score, the second-layer exposure score, and the third-layer frequency score are weighted and summed to obtain the data sensitivity score matrix. The calculation formula for the weighted summation is: ;in, Represents the sensitivity score matrix of the data, They represent the weights of sensitivity score, exposure score and frequency score respectively, and They represent the sensitivity score, exposure score, and frequency score, respectively. Through weighted combination, a sensitivity score matrix is ​​obtained to comprehensively evaluate the privacy risk of each health data.

[0026] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Based on the privacy risk assessment results, the preprocessed data set is classified, and the basic physiological indicators corresponding to blood pressure and blood sugar are divided into numerical health data subsets, and the complex physiological data corresponding to electrocardiogram is divided into neural network calculation data subsets; The basic operation requirements of the elderly chronic disease monitoring data in the numerical health data subset were analyzed, and the numerical calculation parameters were obtained through the complexity evaluation of addition and multiplication operations; Construct a BFV homomorphic encryption key generation strategy based on numerical calculation parameters, use the Intel SGX trusted execution environment to generate an encryption key pair, perform BFV encryption operations on the numerical health data subset, and obtain a first encrypted data set; Perform multi-layer perceptron algorithm analysis on the neural network calculation data subset, determine the CKKS encryption parameters based on the data accuracy requirements and the balance between computational overhead, and obtain the optimal configuration of the neural network; Input the neural network optimization configuration into the ARM TrustZone trusted execution environment, generate a CKKS encryption key pair, perform homomorphic encryption calculations on the neural network calculation data subset, and obtain a second encrypted data set; Perform distributed storage node planning for the first encrypted data set and the second encrypted data set, record data distribution and access logs based on blockchain technology, and obtain data storage location information; A secure data sharing channel is built between nursing homes based on the data storage location information. The ciphertext data is distributed through a secret sharing mechanism to obtain the encrypted data distribution result. Based on the encrypted data distribution result, the Yao encryption circuit protocol is designed to perform secure and synchronous calculations on the encrypted data between the nodes of the nursing homes to obtain a doubly encrypted data set.

[0027] Specifically, the preprocessed data set is classified based on the privacy risk assessment results. Basic physiological indicators such as blood pressure and blood sugar are divided into numerical health data subsets, while complex physiological data such as electrocardiograms are divided into neural network calculation data subsets. For the chronic disease monitoring data of the elderly in the numerical health data subset, the basic operation requirements are analyzed. The complexity of addition and multiplication operations is evaluated. The complexity of addition and multiplication operations is determined by evaluating the impact of each operation on the amount of data. On this basis, the relevant parameters of numerical calculation are obtained. According to the numerical calculation parameters, a BFV homomorphic encryption key generation strategy is constructed for the numerical health data subset. BFV homomorphic encryption is an encryption scheme that supports homomorphic calculations of addition and multiplication, which is suitable for processing numerical data such as blood sugar and blood pressure. When constructing the key, the corresponding public key, private key and decryption key are generated according to the encryption requirements of the data. The Intel SGX (Software Guard Extension) trusted execution environment is used in the key generation process to ensure that the encryption operation is performed in a secure hardware environment to prevent key leakage. After the encryption key is generated, the BFV encryption operation is performed on the numerical health data subset to obtain the first encrypted data set. By processing data in an encrypted state, data is always encrypted during storage and calculation, protecting data privacy. Multilayer perceptron algorithm analysis is performed on a subset of neural network calculation data. Multilayer perceptron is a basic neural network structure used to process time series data or structured data. When analyzing these data, the CKKS encryption parameters are determined based on the balance between the accuracy requirements of the data and the computational overhead. The CKKS homomorphic encryption scheme handles encrypted calculations with real data and is suitable for floating calculations in neural networks. The calculation process is optimized by adjusting encryption parameters (such as encryption depth and noise level) to ensure that data can still be effectively trained in an encrypted state. Based on the parameter settings, the optimized configuration of the neural network is obtained. The optimized configuration is input into the ARM TrustZone trusted execution environment to generate the corresponding CKKS encryption key pair. In ARM TrustZone, computing and data storage are strictly separated to protect data from unauthorized access. The generated key pair is used to perform encryption calculations on a subset of neural network calculation data to obtain a second encrypted data set. After the encryption operation is completed, the data storage phase is entered. Distributed storage nodes are planned for the first encrypted data set and the second encrypted data set. The distribution of data and access logs are recorded through blockchain technology. The distributed ledger technology of blockchain provides transparency and traceability for data storage, ensuring that each data access request can be tracked and recorded. This information includes who accessed the data, the time of access, and the content of the access. In this embodiment, data storage location information is obtained. Based on the data storage location information, a secure data sharing channel between nursing homes is constructed. This channel ensures the secure sharing of data between multiple nursing homes by using a secret sharing mechanism.The secret sharing mechanism protects data by dividing it into multiple parts and distributing these parts to different participants. The data can only be restored when all participants collaborate. This method avoids the risk of a single institution mastering the complete data, increases the difficulty of data leakage, and thus ensures privacy security. After the distribution of the ciphertext data, the Yao encryption circuit protocol is designed based on the encrypted data distribution results. Through the Yao encryption circuit protocol, participating nursing homes can collaborate on data calculations without exposing the original data. The Yao encryption circuit uses encryption circuits to simulate the data processing process to ensure that the privacy of the data is still protected even when multiple parties are involved. Through this protocol, the encrypted data between the nodes of the nursing home is securely and synchronously calculated to obtain a double encrypted data set.

[0028] In a specific embodiment, the process of executing step S3 may specifically include the following steps: A multi-level feature extraction network is constructed for the chronic disease monitoring data of the elderly in the double encrypted data set. The first layer extracts the data time series features to obtain the time series feature matrix, the second layer extracts the data category features to obtain the category feature matrix, and the third layer extracts the data association features to obtain the association feature matrix. The data distribution feature vector is obtained through feature fusion operation. The degree of data distribution difference between different nursing homes is calculated based on the data distribution feature vector. The heterogeneity matrix between institutions is obtained through the combined calculation of cosine similarity and Euclidean distance. The heterogeneity matrix between institutions is input into the federated optimization algorithm module. The training strategy parameters are adjusted based on the heterogeneity characteristics of the nursing home data to obtain the federated learning configuration information. The FedAvg algorithm framework is constructed based on the federated learning configuration information, and the model aggregation parameters are obtained through model updates and weighted average operations on local devices of multiple nursing homes; The model aggregation parameters are calculated by secret sharing. The parameters are divided into different shares and distributed to various nursing homes to obtain distributed training weights. Based on the distributed training weights, a secure multi-party computing model is built to aggregate parameters across institutions to obtain global update parameters. The global update parameters are input into the differential privacy protection module, and privacy protection processing is performed through gradient clipping and noise addition to obtain anti-leakage training parameters. The model performance and privacy protection evaluation of the anti-leakage training parameters are then performed. Through multiple rounds of iterative optimization and threshold judgment, the global training parameters are obtained.

[0029] Specifically, a multi-level feature extraction network is constructed for the chronic disease monitoring data of the elderly in the double encrypted data set. Different types of features are extracted through a three-layer structure. The first layer extracts data time series features, the second layer extracts data category features, and the third layer extracts data association features. The first layer of feature extraction is completed by sliding window processing of historical data points and calculating the statistical features of each window (such as maximum, minimum, mean, etc.) to generate a time series feature matrix, which represents the changing trend of health status in different time periods. For example, suppose the blood sugar data at a series of time points , calculate its mean and standard deviation through the sliding window, and get the time series feature matrix , where each Indicated in The statistical features of the data within a time window. In the second layer, the data category features are extracted. Category features reflect the characteristics of different types of data. For example, numerical data such as blood sugar and blood pressure are classified by setting thresholds and divided into "normal" and "abnormal" categories. A category feature matrix is ​​constructed by counting the frequency of occurrence of each type of data or through other feature engineering methods. Each row of the matrix represents a sample (for example, an elderly person), and each column represents the performance of the sample under different categories. Assuming that blood sugar levels are divided into "normal" and "abnormal", the category feature of each data point may be 0, 1 (normal or abnormal). In this embodiment, the category feature matrix of the data is obtained. , where each Indicates the distribution of a sample under different categories. The third-level feature extraction focuses on the correlation features of the data, and reveals their mutual relationship by calculating the correlation between various health monitoring indicators. For example, there is a certain correlation between blood sugar and blood pressure. In order to extract the correlation features, the correlation coefficient between each two data items is calculated, such as the Pearson correlation coefficient. Assume and If is a data set of two variables, their correlation coefficient is calculated by the following formula: ;in, and Indicates data points, and They are and By this method, the correlation feature matrix between health monitoring indicators is calculated. , each Indicates the degree of correlation between different indicators. Perform feature fusion operations on time series features, category features, and association features to obtain data distribution feature vectors. Feature fusion combines multiple feature matrices into a unified feature representation through weighted summation or principal component analysis. For example, if the time series feature matrix, category feature matrix, and association feature matrix are respectively , and Indicates that the fused feature vector is obtained in the following way: ;in, is the weight of each feature matrix, controlling the importance of each feature in the final fusion. Through this step, a feature vector representing the data distribution is obtained . Based on the data distribution feature vector, the data distribution differences between different nursing homes are calculated. In order to evaluate the data heterogeneity between different nursing homes, a combined calculation method of cosine similarity and Euclidean distance is used. Cosine similarity measures the angle between two vectors in space, thereby reflecting their similarity. The formula is as follows: ;in, and Represent the data distribution feature vectors of the two nursing homes respectively, and is the modulus of the vector. The Euclidean distance measures the absolute distance between two vectors, and its formula is: ; By combining these two calculation methods, the inter-institutional heterogeneity matrix is ​​obtained, which represents the degree of difference in data distribution between different nursing homes. The heterogeneity matrix is ​​input into the federated optimization algorithm module, and the training strategy parameters are adjusted by analyzing the heterogeneous characteristics of the data of each nursing home. This process helps the system determine how to conduct model training between different institutions, and adjusts the learning weight of each participant according to the heterogeneity, ensuring that the final model can perform more robustly on heterogeneous data. Based on the federated learning configuration information, the FedAvg algorithm framework is constructed. The FedAvg algorithm trains the model on the local devices of each nursing home and updates the local model, and then aggregates the updated results of all local models through weighted average to obtain a global model. This process is expressed by the following formula: ;in, Indicates The local model parameters of the nursing home, is the total number of participants. In this embodiment, the federated learning framework is able to share knowledge between different nursing homes without exposing the private data of any institution. In order to protect privacy, the model aggregation parameters are calculated by secret sharing. By dividing the model parameters into different shares and distributing them to each nursing home, it is ensured that each institution can only access part of the parameters. Through the secret sharing mechanism, all participants can only restore the complete parameters when they cooperate, thereby enhancing privacy protection. During the calculation process, the "Secure Multi-Party Computation (SMPC)" model is used for cross-institutional parameter aggregation. The global update parameters are input into the differential privacy protection module, and gradient clipping and noise addition techniques are used to prevent privacy leakage. Gradient clipping avoids the leakage of sensitive information by controlling the gradient range of each model update. The formula for gradient clipping is: ;in, is the gradient, is the preset maximum threshold, is the clipped gradient. Noise addition achieves privacy protection by adding random noise to each update. The formula is: ;in, It means that it has a mean of 0 and a variance of Gaussian noise. Through multiple rounds of iterative optimization and threshold judgment, the anti-leakage training parameters are obtained, and the model performance verification and privacy protection evaluation are performed to ensure that the model can still provide sufficiently high performance while protecting privacy. Get the global training parameters.

[0030] In a specific embodiment, the execution step constructs a multi-level feature extraction network for the elderly chronic disease monitoring data in the double encrypted data set, the first layer extracts the data time series features to obtain a time series feature matrix, the second layer extracts the data category features to obtain a category feature matrix, and the third layer extracts the data association features to obtain an association feature matrix. The process of obtaining the data distribution feature vector through feature fusion operation can specifically include the following steps: The monitoring data of blood pressure, blood sugar, heart rate, and blood oxygen in the double encrypted data set are divided into time windows, and a time series feature extraction layer is constructed according to the preset time scale to obtain data sequences of different time granularities; Based on the data sequences of different time granularities, time series statistical operations are performed, and the time series feature matrix is ​​obtained by calculating the maximum value, minimum value, mean value, standard deviation and abnormal fluctuation amplitude. The normal value range and risk level threshold are set for the monitoring data of blood pressure, blood sugar, heart rate and blood oxygen. The classification model is constructed through numerical interval mapping to obtain the data category label; The monitoring data is divided into risk levels according to the data category labels. The category feature matrix is ​​obtained by clustering and grouping the data of the same risk level. The monitoring indicators of blood pressure, blood sugar, heart rate, and blood oxygen are paired and combined. The correlation between the indicators is calculated to obtain the indicator correlation measurement value. Based on the indicator association measurement value, a physiological indicator association network is constructed. By extracting the topological structure features between indicator nodes, the association feature matrix is ​​obtained. The time series feature matrix, category feature matrix and association feature matrix are transformed by feature mapping. The feature fusion parameters are obtained by calculating the importance weights of different features. Based on the feature fusion parameters, the time series feature matrix, category feature matrix and association feature matrix are normalized and weighted combined, and the data distribution feature vector is obtained through dimension conversion and feature selection.

[0031] Specifically, the monitoring data of blood pressure, blood sugar, heart rate, and blood oxygen in the double encrypted data set are divided into time windows. The continuous health monitoring data is divided into multiple time periods so that the data of each time period can be analyzed. Assume that there is a set of blood sugar monitoring data: , each data point represents the blood glucose level at a certain time point, and the interval between each time point is (For example, blood sugar is measured once every minute). According to the preset time scale, the data is divided into multiple time windows, each window contains a certain number of data points. For example, each time window is set to contain 10 data points, and the range of the first time window is , the second time window is , and so on. Perform time series statistical operations on data sequences of different time granularities. Extract meaningful statistical features from the data in each time window. For example, for blood glucose data, calculate the maximum value, minimum value, mean value, standard deviation, and abnormal fluctuation range of each time window. Through time series statistical operations, a time series feature matrix is ​​obtained, which contains the statistical information of each time window. For example, assuming there are 5 time windows, and the maximum value, minimum value, mean value, standard deviation, and fluctuation range are calculated for each window, the time series feature matrix is: ; Each row represents the statistical features of a time window, and each column represents a feature type, such as maximum, minimum, mean, etc. According to the normal value range and risk level threshold of the monitoring data, a grading model is constructed, and a data category label is assigned to each data point. The health status of the data is judged according to the different value ranges of the health data. For example, blood glucose data is divided into risk levels by setting the normal value range and threshold. The data category labels after grading are 0 (normal), 1 (low risk), and 2 (high risk), and each data point is assigned to a different risk level. According to the data category label, the monitoring data is divided into risk levels. By clustering and grouping data with the same risk level, a category feature matrix is ​​formed. Samples are classified according to the health risk of the data. The indicator association metric value is obtained by calculating the change correlation between the indicators. For each pair of health monitoring indicators (such as blood glucose and blood pressure), the correlation coefficient between them is calculated. In this embodiment, the correlation between different health monitoring indicators is quantified to obtain an association metric matrix. Using the association metric value, a physiological indicator association network is constructed. Each node in the network represents a health monitoring indicator, and the edges between the nodes represent the correlation between them. For example, if the correlation coefficient between blood sugar and blood pressure is high, the connection strength between them in the physiological index association network is large. By calculating the topological structure characteristics between the nodes in the network, the association feature matrix is ​​obtained, which reflects the relationship between the various physiological indicators and their position in the entire system. The time series feature matrix, the category feature matrix, and the association feature matrix are transformed by feature mapping to fuse different types of features into a unified space. Using a weighted combination method, the different parts of each feature matrix are fused according to their importance weights. The time series feature matrix, the category feature matrix, and the association feature matrix are normalized and weighted combined. Features of different scales are converted to a unified scale for use in the model. Through dimensionality conversion and feature selection, the most representative features are selected and the representation of the data is optimized. These matrices are weighted combined and standardized based on the feature fusion parameters to obtain the data distribution feature vector.

[0032] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Perform gradient sensitivity analysis on the physiological index characteristics of the elderly in the global training parameters, calculate the L2 norm of the gradient vector through the Intel SGX trusted execution environment, and obtain the gradient clipping threshold; Based on the gradient clipping threshold, the neural network model is constrained to update parameters, and the gradient size of each dimension is limited by the gradient projection method to obtain the clipped parameter matrix; Perform multi-layer perceptron algorithm analysis on the pruned parameter matrix, determine the Laplace noise distribution parameters based on data accuracy and privacy protection requirements, and obtain the noise injection strategy; The noise injection strategy is applied to the model training process. Random noise is added to the gradient update of each iteration through the differential privacy mechanism to obtain the perturbed training parameters. The privacy budget consumption of the perturbation training parameters is calculated, and the budget allocation ratio is dynamically adjusted according to the privacy risk assessment mechanism to obtain the budget constraint scheme. A multi-round iterative training framework is constructed based on the budget constraint scheme to aggregate local model parameters and obtain the global model update sequence. The global model update sequence is input into the ARM TrustZone trusted execution environment, and security verification is performed through the Yao encrypted circuit protocol to obtain the model security score. A comprehensive evaluation is then performed on the model security score and training accuracy, and model parameters are protected based on the secret sharing mechanism to obtain a privacy protection model.

[0033] Specifically, the gradient sensitivity analysis is performed on the physiological index characteristics of the elderly in the global training parameters to identify the sensitivity of each parameter during the training process, especially in the data model training of different nursing homes, how to adjust the model parameters to prevent privacy leakage. The L2 norm is obtained by summing the squares of all gradient elements and taking the square root. The calculation of the L2 norm helps the system identify which parameters change greatly during the training process, so as to determine what constraints to apply to these parameters. On this basis, the gradient clipping threshold is calculated according to the L2 norm, that is, when the L2 norm of the gradient exceeds this threshold, the model will clip the gradient to avoid instability and overfitting of the model caused by excessive gradient updates. The parameter update of the neural network model is constrained based on the gradient clipping threshold. The gradient size of each dimension is limited by the gradient projection method. The purpose of the gradient projection method is to ensure that the gradient does not exceed the clipping threshold each time the model is updated. If the gradient of a parameter exceeds the preset clipping threshold, it is "projected" back to a restricted range. The deep learning model training is performed on the clipped parameters using the multi-layer perceptron algorithm to efficiently classify or regress the data. The multi-layer perceptron algorithm can effectively learn complex patterns in data through nonlinear transformations of multiple hidden layers. During the training process, ensure the balance between data accuracy and privacy protection. At this stage, the determination of the parameters of the Laplace noise distribution is very critical, and the size of the noise directly affects the trade-off between the privacy protection effect and the training accuracy. Assume that noise is added to each round of gradient updates, and use the differential privacy mechanism to add Laplace noise to each gradient update. In this embodiment, each gradient update is accompanied by noise, so that even if the attacker obtains the model parameters, the information of each data point cannot be accurately inferred. Perform a privacy budget consumption calculation to determine the consumption of the privacy budget by noise injection. The privacy budget is a limited resource that is gradually consumed as the training process proceeds. During the calculation process, each gradient update consumes a certain privacy budget, and the budget allocation ratio is dynamically adjusted according to the privacy risk assessment mechanism. For example, in the early stage, more privacy budget is allowed to improve the accuracy of the model, while more budget is used to protect privacy in the later stage. In this embodiment, ensure a reasonable balance between privacy protection and model accuracy. According to the privacy budget consumption and privacy risk assessment mechanism, a multi-round iterative training framework is constructed, which can dynamically adjust the privacy protection strategy of each round according to the budget allocation plan. In each round of training, local model parameter aggregation is performed, and each participating organization updates the local model parameters by weighted average, and the updated parameters are summarized as a global model update sequence. In this embodiment, multiple organizations jointly train a global model without exposing their respective data. The globally updated parameters are input into the differential privacy protection module for privacy protection. In order to ensure the protection effect, gradient clipping and noise addition are performed to ensure that the privacy risk of each update is within an acceptable range.After processing, the anti-leakage training parameters are obtained, and these parameters are used for model performance verification and privacy protection evaluation. During the evaluation process, the balance between privacy protection and model accuracy is dynamically adjusted through multiple rounds of iterative optimization and threshold judgment. The specific threshold judgment is achieved through the error of the verification set. If the error is too large, the privacy protection strength is reduced. If the risk of privacy leakage is high, the protection strength is increased. In the verified anti-leakage training parameters, the Yao encrypted circuit protocol is used to verify the model security. The Yao encrypted circuit protocol verifies the model updates provided by each participant through encrypted calculations to ensure that no private data is leaked between the participants. Through this process, the model security score is obtained, and the reliability and privacy protection effect of the model are comprehensively evaluated based on the score and training accuracy. Through the secret sharing mechanism, the parameters of the model are protected to ensure that each participant can only access part of the parameters it owns, but cannot access the entire model, and obtain a privacy-preserving model.

[0034] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The access rights of the monitoring data of blood pressure, blood sugar, heart rate and blood oxygen in the privacy protection model are graded, and a multi-level access control matrix is ​​obtained by constructing a role-based access control model RBAC; A zero-knowledge proof protocol is constructed based on a multi-level access control matrix. The zero-knowledge proof parameters are generated by verifying the access rights of caregivers and family members, and an identity authentication protocol is obtained. Deploy the identity authentication protocol to the Intel SGX trusted execution environment, and use TEE hardware isolation technology to protect the model execution space and obtain a secure execution area. Configure edge computing nodes in the secure execution area, build a model access control interface through the ARM TrustZone trusted execution environment, obtain an edge security environment, and connect the edge security environment to the secure multi-party computing network. Perform distributed verification of access requests through a secret sharing mechanism to obtain access verification results. Based on the access verification results, the data access between the nodes of the nursing home is recorded on the blockchain, and the access tracing mechanism is established through the distributed ledger technology to obtain the behavior audit data; Perform differential privacy analysis on behavioral audit data, desensitize the access behavior characteristics through the Laplace mechanism to obtain a privacy-preserving audit report, and build a dynamic access control strategy based on the privacy-preserving audit report. Update the security policy through the Yao encryption circuit to obtain a secure access strategy.

[0035] Specifically, the access rights of the monitoring data of blood pressure, blood sugar, heart rate and blood oxygen in the privacy protection model are graded. The role-based access control model (RBAC) is adopted to assign different access roles to each user through the RBAC model, thereby limiting the scope and type of data they can access. In this model, roles such as "nurse", "family member" and "doctor" will have different permissions. By assigning different access levels according to each role, a multi-level access control matrix is ​​obtained, in which each row represents a different user role, each column represents a different data type (such as blood sugar, blood pressure, etc.), and the value in the matrix represents the access level of a certain role to a certain data. For example, the doctor has access to all data, while the family members can only access blood sugar and blood pressure data, and can only view low-risk data. Based on the role-based access control model, a zero-knowledge proof protocol is constructed to verify the identity of the user and ensure the legitimacy of data access. Zero-knowledge proof is a cryptographic protocol that allows the prover to prove to the verifier that a proposition is correct without revealing any information other than the correctness of the proposition. In this application scenario, the zero-knowledge proof protocol verifies the access rights of the roles of "caregiver" and "family member", generates zero-knowledge proof parameters, and uses these parameters to generate the authentication protocol. The authentication protocol is deployed in the Intel SGX trusted execution environment, and the hardware isolation technology of SGX is used to ensure that identity authentication and data processing are performed in a secure execution environment. Intel SGX provides a hardware-level protection that can ensure that sensitive data in the computing process is not accessible to external attackers, even under the control of the operating system or virtual machine hypervisor. The execution space of the SGX environment model is securely protected to generate a secure execution area, making the entire data access control process more secure and reliable. Edge computing nodes are configured in the secure execution area to improve the efficiency of data processing, and the computing is dispersed from the data source to the place close to the data, reducing transmission delays and enhancing data privacy protection. At this stage, ARM TrustZone technology is used to create the access control interface of the model. ARM TrustZone protects the security of data and applications by hardware isolating the computing environment to ensure that only authorized devices access specific parts of the model. By connecting the edge computing environment with a secure multi-party computing network, it is possible to utilize the computing power between multiple institutions without leaking any sensitive data. Data access requests in the network are distributedly verified through a secret sharing mechanism. Information is divided into several parts, and each participant can only obtain a part of it. Only when all participants cooperate can the original information be restored. Data access is recorded on the blockchain. The distributed ledger of blockchain technology provides an unalterable record for data access, so that every access request can be traced. By storing each access record in the blockchain as a hash value, the security of the access record is guaranteed, and the access tracing mechanism is implemented.When analyzing behavioral audit data, differential privacy technology is used to protect sensitive information of data access. Differential privacy adds noise to the data to ensure that even if the data is analyzed or queried, no information about the individual is leaked. In behavioral auditing, the Laplace mechanism is used to add noise to protect the privacy of users. In this embodiment, sensitive information in the behavioral audit data is effectively desensitized to ensure that there is no risk of leakage. Based on the privacy protection audit report, a dynamic access control policy is constructed. By evaluating the effectiveness of privacy protection and the needs of access control, access rights and policies are adjusted in real time so that the system can respond flexibly according to different security requirements. The Yao encrypted circuit protocol is used to update security policies. The Yao encrypted circuit protocol ensures that in multi-party computing, all participants can obtain correct results without exposing any intermediate data through encrypted calculations. Through this encrypted circuit protocol, the access control policy is encrypted and protected during the update process, while ensuring privacy security.

[0036] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Based on the security access strategy, the edge node performs preliminary analysis on the elderly chronic disease monitoring data collected by IoT devices, and only uploads the information to the cloud when an abnormality is detected to obtain local anomaly detection results; Perform BFV homomorphic encryption on the local anomaly detection results, build a data security computing environment through the Intel SGX trusted execution environment, obtain the first layer of encrypted data, and input the first layer of encrypted data into the CKKS homomorphic encryption module, perform secondary encryption protection through ARM TrustZone technology, and obtain double encrypted analysis data; Based on the double encrypted analysis data, a group intelligence federated learning framework is built between the nodes of the nursing home, and the FedAvg algorithm is used for collaborative training to obtain a distributed prediction model; Perform secret sharing calculations on the distributed prediction model, divide the model parameters into different shares and distribute them to each nursing home to obtain joint analysis parameters, and input the joint analysis parameters into the Yao encryption circuit protocol to perform secure analysis on the health data between nursing homes through secure multi-party computing to obtain cross-institutional analysis data; Differential privacy processing is performed on cross-institutional analysis data, and sensitive information in the analysis results is desensitized through the Laplace mechanism to obtain privacy-preserving analysis results. Health trend statistical analysis is performed based on the privacy-preserving analysis results, and the analysis process and results are recorded through blockchain technology to obtain collaborative analysis results.

[0037] Specifically, the data collected by the sensor is initially analyzed at the edge node, and data processing is performed directly on the IoT device or local device. Only when an abnormal situation is monitored, such as a sudden increase in blood sugar or abnormal fluctuations in heart rate, the relevant data is uploaded to the cloud. Once an anomaly is detected, the data is uploaded and processed in the cloud. At this stage, the local anomaly detection results are processed by BFV homomorphic encryption. BFV homomorphic encryption can support addition and multiplication operations on encrypted data. With BFV encryption, addition and multiplication calculations are performed on the data without decryption. For each data point , encrypted data Encrypted by BFV: ;in, Indicates that it is based on public key Encryption, is the input health data, It is encrypted data. The encrypted data is input into the Intel SGX trusted execution environment, which provides a secure computing environment through the hardware security module to prevent unauthorized access to the data during the processing process. When data is calculated in this environment, all intermediate calculation results are also kept encrypted. Through hardware security assurance, it is ensured that even if the execution environment is attacked, the data cannot be leaked. This process generates the first layer of encrypted data, which is transferred to the CKKS homomorphic encryption module for encryption. CKKS homomorphic encryption is suitable for encryption operations of floating values. ARM TrustZone technology is used to perform secondary encryption protection using its hardware isolation characteristics. Through multi-layer encryption protection, the privacy of the data is ensured not to be leaked during the processing process. Based on the double encryption analysis data, a group intelligence federated learning framework between the nodes of the nursing home is constructed. Federated learning allows multiple distributed nodes (such as different nursing homes) to collaboratively train models through local computing without sharing the original data. Under this framework, each nursing home can process data locally and perform global aggregation based on the training results of the local model to obtain the global model parameters. By using the FedAvg algorithm, the model updates of multiple nursing homes are weighted averaged to obtain a global model. For example, suppose the model parameters trained locally in each nursing home are , then the aggregation parameter of the global model It is expressed as: ; in, is the number of participating nursing homes. In this embodiment, each nursing home collaborates to train a global prediction model while ensuring data privacy. Secret sharing calculations are performed on the distributed prediction model to enhance data privacy protection. Through the secret sharing mechanism, each model parameter is divided into multiple shares and distributed to different nursing homes. Each institution can only access part of the model parameters it owns, and the complete model can only be restored when multiple institutions collaborate. Use the Yao encrypted circuit protocol for secure multi-party computing. The Yao encrypted circuit protocol uses encrypted computing to allow multiple institutions to jointly calculate the parameters of the model without exposing the original data to obtain the final analysis results. Through this protocol, each participant can only see part of the data during the calculation process and cannot access the data or intermediate calculation results of other parties. For example, if multiple nursing homes perform collaborative computing through encrypted circuits, the final calculation results It is expressed by the following formula: ;in, Represents encrypted data held by different elderly care institutions, represents the computational function performed by the encryption circuit, is the final calculation result. The cross-institutional analysis results are processed using differential privacy technology. Differential privacy adds noise to the data to ensure that even if the data is analyzed or queried, only the overall statistical information is leaked, not individual information. The analysis results are desensitized using the Laplace mechanism, and Laplace noise is added to each query result to ensure privacy protection. For example, suppose the query result is , then add noise to the query results using the following formula: ;in, The mean is 0 and the standard deviation is Gaussian noise, It is the query result after noise processing. Blockchain technology is used to record the analysis process and results to establish an access traceability mechanism. Blockchain ensures the immutability and traceability of each analysis process and result through its distributed ledger technology. For example, the data analysis process is stored in the blockchain through hash values, and each hash value contains the timestamp of the analysis, the participating institution information, and the analysis results.

[0038] The above describes the privacy data protection method based on homomorphic encryption and federated learning in the embodiment of the present invention. The following describes the privacy data protection system based on homomorphic encryption and federated learning in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a privacy data protection system based on homomorphic encryption and federated learning includes: The standardization module is used to add Laplace noise and perform data standardization on the raw health data collected by IoT devices in multiple nursing homes to obtain preprocessed data sets and privacy risk assessment results; The encryption module is used to perform BFV homomorphic encryption on the numerical data in the preprocessed data set and CKKS homomorphic encryption on the neural network calculation data according to the privacy risk assessment results to obtain a double encrypted data set; The computing module is used to build a decentralized federated learning framework based on a double-encrypted data set, calculate the training weights through data heterogeneity analysis and federated optimization algorithm, and obtain global training parameters; The training module is used to input the global training parameters into the gradient clipping module for noise injection processing, and perform iterative model training based on the differential privacy budget constraint to obtain a privacy-preserving model; The verification module is used to deploy the privacy-preserving model in a trusted execution environment and obtain a secure access policy through role-based multi-level access control and zero-knowledge proof for identity authentication. The analysis module is used to perform distributed deployment of edge nodes for privacy protection models based on secure access policies, and to perform cross-institutional health data analysis through secure multi-party computing protocols to obtain collaborative analysis results.

[0039] Through the collaboration of the above components and the construction of a multi-level privacy protection architecture, the security protection of privacy data is achieved in the whole process of data collection, transmission, storage and analysis. Specifically, data preprocessing is achieved through Laplace noise addition and data standardization, and BFV and CKKS dual homomorphic encryption are used to ensure data transmission security; data collaborative analysis among multiple institutions is achieved based on a decentralized federated learning framework, and the model training process is protected by differential privacy budget constraints and gradient clipping; the Intel SGX and ARM TrustZone dual trusted execution environment deployment model is used, combined with role-based multi-level access control and zero-knowledge proof for identity authentication; finally, cross-institutional collaborative analysis of health data is achieved through edge computing and secure multi-party computing. The present invention not only effectively solves the problem of data privacy leakage in traditional health management systems, but also realizes efficient use of data while protecting data privacy by innovatively integrating multiple privacy protection technologies.

[0040] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a display screen, an input system, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0041] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0042] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0043] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0044] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0045] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0046] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A privacy data protection method based on homomorphic encryption and federated learning, characterized in that: The method comprises: The original health data collected by IoT devices in multiple nursing homes were processed by Laplace noise addition and data standardization to obtain preprocessed data sets and privacy risk assessment results; According to the privacy risk assessment result, the numerical data in the preprocessed data set is subjected to BFV homomorphic encryption processing, and the neural network calculation data is subjected to CKKS homomorphic encryption processing to obtain a double encrypted data set; A decentralized federated learning framework is constructed based on the double encrypted data set, and training weights are calculated through data heterogeneity analysis and a federated optimization algorithm to obtain global training parameters; The global training parameters are input into the gradient clipping module for noise injection processing, and the model is iteratively trained based on the differential privacy budget constraint to obtain a privacy protection model; Deploy the privacy protection model in a trusted execution environment, perform identity authentication through role-based multi-level access control and zero-knowledge proof, and obtain a secure access policy; Based on the secure access strategy, the privacy protection model is deployed in a distributed manner on edge nodes, and cross-institutional health data analysis is performed through a secure multi-party computing protocol to obtain collaborative analysis results.

2. The privacy data protection method based on homomorphic encryption and federated learning according to claim 1 is characterized in that: The raw health data collected by IoT devices of multiple nursing homes are subjected to Laplace noise addition and data standardization processing to obtain a preprocessed data set and privacy risk assessment results, including: Collect the original health data corresponding to the IoT devices of multiple nursing homes, and verify the time series integrity of the chronic disease monitoring data of blood pressure, blood sugar, heart rate, and blood oxygen collected by the IoT devices to obtain the validity index of the original data; Based on the raw data validity index, the raw health data is subjected to multi-dimensional quality assessment and outlier correction to obtain preliminary cleansed data, and the preliminary cleansed data is input into the Intel SGX trusted execution environment for data classification and privacy level classification to obtain a multimodal security classification result; A three-layer evaluation matrix is ​​constructed for the multimodal security classification results. The first layer performs data sensitivity calculation to obtain a sensitivity score, the second layer performs data exposure range analysis to obtain an exposure score, and the third layer performs data usage frequency statistics to obtain a frequency score. The data sensitivity score matrix is ​​obtained by weighted combination of the three-layer scores. Determine Laplace noise distribution parameters according to the data sensitivity score matrix, perform directional noise injection on the preliminary cleaned data to obtain a privacy-preserving data set, and group and standardize the privacy-preserving data set according to the type of physiological indicators of the elderly to obtain a preprocessed data set; Deep feature extraction and correlation analysis are performed on the preprocessed data set to obtain data feature mapping information, and a risk assessment model is constructed based on the data feature mapping information and the data sensitivity scoring matrix, and a privacy risk assessment result is obtained through multiple threshold determination.

3. The privacy data protection method based on homomorphic encryption and federated learning according to claim 2 is characterized in that: The three-layer evaluation matrix is ​​constructed for the multimodal security classification results. The first layer performs data sensitivity calculation to obtain a sensitivity score, the second layer performs data exposure range analysis to obtain an exposure score, and the third layer performs data usage frequency statistics to obtain a frequency score. The data sensitivity score matrix is ​​obtained by weighted combination of the three-layer scores, including: Mapping different types of health data in the multimodal security classification results, hierarchically encoding the data according to privacy attributes, obtaining an initial data sensitivity matrix, and establishing a sensitivity evaluation index system for each type of health data based on the initial data sensitivity matrix, and calculating a first-level sensitivity score; The multimodal security classification results are classified according to the data access scope, the data exposure probability is calculated through the access control matrix to obtain a data exposure risk vector, and the data exposure risk vector is normalized to obtain a second-level exposure score; Performing data access frequency statistics on the multimodal security classification results based on a preset time window, constructing a frequency distribution histogram, obtaining original frequency features, performing time series analysis on the original frequency features, calculating the usage frequency trend through an exponential smoothing method, and obtaining a third-level frequency score; Inputting the first layer sensitivity score, the second layer exposure score and the third layer frequency score into a multi-layer neural network, optimizing weight parameters by a back propagation algorithm, and obtaining a score fusion weight; Based on the score fusion weight, a weighted sum operation is performed on the first layer sensitivity score, the second layer exposure score, and the third layer frequency score to obtain a data sensitivity score matrix.

4. The privacy data protection method based on homomorphic encryption and federated learning according to claim 2 is characterized in that: According to the privacy risk assessment result, the numerical data in the preprocessed data set is subjected to BFV homomorphic encryption processing, and the neural network calculation data is subjected to CKKS homomorphic encryption processing to obtain a double encrypted data set, including: Classifying the preprocessed data set based on the privacy risk assessment result, dividing the basic physiological indicators corresponding to blood pressure and blood sugar into numerical health data subsets, and dividing the complex physiological data corresponding to the electrocardiogram into a neural network calculation data subset; Performing basic computational demand analysis on the elderly chronic disease monitoring data in the numerical health data subset, and obtaining numerical calculation parameters through addition and multiplication computational complexity evaluation; Constructing a BFV homomorphic encryption key generation strategy according to the numerical calculation parameters, generating an encryption key pair using the Intel SGX trusted execution environment, and performing a BFV encryption operation on the numerical health data subset to obtain a first encrypted data set; Performing multi-layer perceptron algorithm analysis on the neural network calculation data subset, determining CKKS encryption parameters based on data accuracy requirements and calculation overhead balance, and obtaining the neural network optimization configuration; Input the neural network optimization configuration into the ARM TrustZone trusted execution environment, generate a CKKS encryption key pair, perform homomorphic encryption calculation on the neural network calculation data subset, and obtain a second encrypted data set; Performing distributed storage node planning for the first encrypted data set and the second encrypted data set, recording data distribution and access logs based on blockchain technology, and obtaining data storage location information; A secure data sharing channel is established between nursing homes based on the data storage location information, and ciphertext data is distributed through a secret sharing mechanism to obtain encrypted data distribution results. The Yao encryption circuit protocol is designed based on the encrypted data distribution results, and the encrypted data between the nodes of the nursing homes are securely and synchronously calculated to obtain a doubly encrypted data set.

5. The privacy data protection method based on homomorphic encryption and federated learning according to claim 4 is characterized in that: The decentralized federated learning framework is constructed based on the double encrypted data set, and the training weights are calculated through data heterogeneity analysis and federated optimization algorithm to obtain global training parameters, including: A multi-level feature extraction network is constructed for the elderly chronic disease monitoring data in the double encrypted data set. The first layer extracts the data time series features to obtain a time series feature matrix, the second layer extracts the data category features to obtain a category feature matrix, and the third layer extracts the data association features to obtain an association feature matrix. The data distribution feature vector is obtained through feature fusion operation. Based on the data distribution feature vector, the degree of difference in data distribution between different nursing homes is calculated, and the heterogeneity matrix between institutions is obtained by combining cosine similarity and Euclidean distance. The heterogeneity matrix between institutions is input into the federated optimization algorithm module, and the training strategy parameters are adjusted based on the heterogeneity characteristics of the nursing home data to obtain the federated learning configuration information. Constructing a FedAvg algorithm framework according to the federated learning configuration information, and obtaining model aggregation parameters through model updating and weighted average operation on local devices of multiple nursing homes; Perform secret sharing calculation on the aggregation parameters of the model, divide the parameters into different shares and distribute them to various nursing homes to obtain distributed training weights, and build a secure multi-party computing model based on the distributed training weights to aggregate parameters across institutions to obtain global update parameters; The global update parameters are input into the differential privacy protection module, and privacy protection processing is performed through gradient clipping and noise addition to obtain anti-leakage training parameters. Model performance verification and privacy protection evaluation are performed on the anti-leakage training parameters, and global training parameters are obtained through multiple rounds of iterative optimization and threshold judgment.

6. The privacy data protection method based on homomorphic encryption and federated learning according to claim 5 is characterized in that: The multi-level feature extraction network is constructed for the elderly chronic disease monitoring data in the double encrypted data set. The first layer extracts the data time series features to obtain a time series feature matrix, the second layer extracts the data category features to obtain a category feature matrix, and the third layer extracts the data association features to obtain an association feature matrix. The data distribution feature vector is obtained through feature fusion operation, including: The monitoring data of blood pressure, blood sugar, heart rate and blood oxygen in the double encrypted data set are divided into time windows, and a time series feature extraction layer is constructed according to a preset time scale to obtain data sequences of different time granularities; Based on the data sequences of different time granularities, time series statistical operations are performed to obtain a time series feature matrix by calculating the maximum value, minimum value, mean value, standard deviation and abnormal fluctuation amplitude, and a normal value range and a risk level threshold are set for the monitoring data of blood pressure, blood sugar, heart rate and blood oxygen, and a grading model is constructed by mapping the numerical interval to obtain a data category label; The monitoring data is divided into risk levels according to the data category labels, and the category feature matrix is ​​obtained by clustering and grouping the data of the same risk level, and the monitoring indicators of blood pressure, blood sugar, heart rate, and blood oxygen are paired and combined in pairs, and the correlation degree of changes between the indicators is calculated to obtain the indicator association measurement value; Based on the indicator association metric values, a physiological indicator association network is constructed, and an association feature matrix is ​​obtained by extracting the topological structure features between indicator nodes, and the time series feature matrix, the category feature matrix and the association feature matrix are subjected to feature mapping transformation, and feature fusion parameters are obtained by calculating the importance weights of different features; The time series feature matrix, the category feature matrix and the association feature matrix are normalized and weighted combined based on the feature fusion parameters, and a data distribution feature vector is obtained through dimension conversion and feature selection.

7. The privacy data protection method based on homomorphic encryption and federated learning according to claim 6 is characterized in that: The global training parameters are input into the gradient clipping module for noise injection processing, and the model is iteratively trained based on the differential privacy budget constraint to obtain a privacy protection model, including: Performing gradient sensitivity analysis on the physiological indicator characteristics of the elderly in the global training parameters, calculating the L2 norm of the gradient vector through the Intel SGX trusted execution environment, and obtaining the gradient clipping threshold; Based on the gradient clipping threshold, the neural network model is constrained to update parameters, and the gradient size of each dimension is limited by the gradient projection method to obtain a clipped parameter matrix; Performing a multi-layer perceptron algorithm analysis on the trimmed parameter matrix, determining Laplace noise distribution parameters based on data accuracy and privacy protection requirements, and obtaining a noise injection strategy; The noise injection strategy is applied to the model training process, and random noise is added to the gradient update of each iteration through the differential privacy mechanism to obtain the perturbed training parameters; Calculate the privacy budget consumption of the perturbation training parameters, dynamically adjust the budget allocation ratio according to the privacy risk assessment mechanism, obtain a budget constraint scheme, and build a multi-round iterative training framework according to the budget constraint scheme to aggregate local model parameters and obtain a global model update sequence; The global model update sequence is input into the ARM TrustZone trusted execution environment, and security verification is performed through the Yao encrypted circuit protocol to obtain a model security score. A comprehensive evaluation is performed on the model security score and training accuracy, and model parameter protection is performed based on a secret sharing mechanism to obtain the privacy protection model.

8. The privacy data protection method based on homomorphic encryption and federated learning according to claim 7 is characterized in that: The privacy protection model is deployed in a trusted execution environment, and identity authentication is performed through role-based multi-level access control and zero-knowledge proof to obtain a secure access policy, including: The access rights of the monitoring data of blood pressure, blood sugar, heart rate and blood oxygen in the privacy protection model are graded, and a multi-level access control matrix is ​​obtained by constructing a role-based access control model RBAC; A zero-knowledge proof protocol is constructed based on the multi-level access control matrix, and zero-knowledge proof parameters are generated by verifying the access rights of caregivers and family members to obtain an identity authentication protocol; Deploy the authentication protocol to the Intel SGX trusted execution environment, and use TEE hardware isolation technology to securely protect the model execution space to obtain a secure execution area; Configure edge computing nodes for the secure execution area, build a model access control interface through the ARM TrustZone trusted execution environment, obtain an edge security environment, connect the edge security environment to a secure multi-party computing network, perform distributed verification on access requests through a secret sharing mechanism, and obtain access verification results; Based on the access verification results, data access between nodes of the nursing home is recorded on the blockchain, and an access tracing mechanism is established through distributed ledger technology to obtain behavioral audit data; A differential privacy analysis is performed on the behavior audit data, and the access behavior characteristics are desensitized through the Laplace mechanism to obtain a privacy protection audit report. A dynamic access control strategy is constructed based on the privacy protection audit report, and the security strategy is updated through the Yao encryption circuit to obtain the security access strategy.

9. The privacy data protection method based on homomorphic encryption and federated learning according to claim 8 is characterized in that: Based on the secure access policy, the privacy protection model is distributedly deployed on edge nodes, and cross-institutional health data analysis is performed through a secure multi-party computing protocol to obtain collaborative analysis results, including: Based on the security access strategy, the edge node performs preliminary analysis on the elderly chronic disease monitoring data collected by the IoT device, and only uploads the information to the cloud when an abnormality is detected to obtain a local abnormality detection result; Perform BFV homomorphic encryption processing on the local anomaly detection result, build a data security computing environment through the Intel SGX trusted execution environment, obtain the first layer of encrypted data, and input the first layer of encrypted data into the CKKS homomorphic encryption module, perform secondary encryption protection through ARM TrustZone technology, and obtain double encrypted analysis data; A swarm intelligence federated learning framework is constructed among the nodes of the nursing home based on the double encrypted analysis data, and a distributed prediction model is obtained by collaborative training through the FedAvg algorithm; Perform secret sharing calculation on the distributed prediction model, divide the model parameters into different shares and distribute them to each nursing home to obtain joint analysis parameters, input the joint analysis parameters into the Yao encryption circuit protocol, perform secure analysis on the health data between nursing homes through secure multi-party computing, and obtain cross-institution analysis data; The cross-institutional analysis data is differentially privacy processed, and the sensitive information of the analysis results is desensitized through the Laplace mechanism to obtain privacy-preserving analysis results. Health trend statistical analysis is performed based on the privacy-preserving analysis results, and the analysis process and results are recorded through blockchain technology to obtain the collaborative analysis results.

10. A privacy data protection system based on homomorphic encryption and federated learning, characterized in that: Used to execute the privacy data protection method based on homomorphic encryption and federated learning as described in any one of claims 1 to 9, the privacy data protection system based on homomorphic encryption and federated learning includes: The standardization module is used to add Laplace noise and perform data standardization on the raw health data collected by IoT devices in multiple nursing homes to obtain preprocessed data sets and privacy risk assessment results; An encryption module, used to perform BFV homomorphic encryption processing on the numerical data in the preprocessed data set and perform CKKS homomorphic encryption processing on the neural network calculation data according to the privacy risk assessment result, to obtain a double encrypted data set; A computing module, used to construct a decentralized federated learning framework based on the double encrypted data set, calculate training weights through data heterogeneity analysis and a federated optimization algorithm, and obtain global training parameters; A training module, used for inputting the global training parameters into a gradient clipping module for noise injection processing, performing model iterative training based on a differential privacy budget constraint, and obtaining a privacy protection model; A verification module, used to deploy the privacy protection model in a trusted execution environment, perform identity authentication through role-based multi-level access control and zero-knowledge proof, and obtain a secure access policy; The analysis module is used to perform distributed edge node deployment of the privacy protection model based on the security access policy, perform cross-institutional health data analysis through a secure multi-party computing protocol, and obtain collaborative analysis results.

Images