Differential privacy protection method and device for user medical data

Abstract

The invention provides a differential privacy protection method and device for user medical data, the method and device are applied to a first server deployed at an edge node in a data processing system, the data processing system further comprises user equipment and a cloud second server, and the method comprises the following steps: receiving a preprocessed medical data set transmitted by the user equipment; identifying attribute categories and counting the total number; calculating a Shapley value and an information entropy of each attribute category, and after standardization processing, dividing privacy leakage levels by combining skyline calculation; distributing the privacy budget corresponding to the inverse proportion according to the grade; and injecting Laplacian noise into each attribute category to generate a synthetic data set, and transmitting the synthetic data set to a second server for storage. According to the method, differential privacy protection of medical data attributes is realized, data privacy security and data availability are effectively balanced, the leakage risk of the medical data is reduced, the leakage risk of sensitive data in a network transmission process is reduced through an edge computing architecture, and the data processing efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of differential privacy technology, and in particular to a differential privacy protection method and apparatus for user medical data. Background Technology

[0002] In the field of smart healthcare applications, the privacy protection of medical data has become a core bottleneck restricting data sharing and value mining. Differential privacy technology, due to its mathematical provability and robustness, has become the mainstream privacy protection solution in this field.

[0003] However, traditional differential privacy protection schemes generally adopt a uniform "one-size-fits-all" noise-adding mode, lacking effective means to distinguish the differentiated features of different attributes in medical datasets. This limitation results in insufficient protection of core attributes with high user identity relevance and significant data utility (such as peak heart rate and key blood glucose indicators), while excessive noise addition to low-sensitivity attributes leads to data distortion, severely impacting data usability in medical research, disease diagnosis, and other scenarios. Furthermore, traditional solutions often rely on centralized cloud processing, exposing raw medical data to the risk of leakage during long-distance transmission. They also lack scientific attribute classification and privacy budget allocation mechanisms, making it difficult to balance the strength of privacy protection with the value of data utilization, and failing to meet the dual practical needs of high sensitivity and high availability in medical data. Summary of the Invention

[0004] This application proposes a differential privacy protection method and apparatus for user medical data, aiming to reduce the risk of leakage of sensitive data transmission through local processing at edge nodes. It combines Shapley value and information entropy to quantify attribute utility and privacy risks, achieves accurate classification through skyline calculation, and then allocates differentiated privacy budgets according to the level and independently injects Laplace noise. This achieves accurate privacy protection at the attribute level of medical data, maximizes data availability and reduces distortion, meets differential privacy compliance requirements, and balances privacy security and data value mining.

[0005] In a first aspect, embodiments of this application provide a differential privacy protection method for user medical data, applied to a first server in a data processing system. The data processing system further includes a user device and a second server, wherein the first server is deployed at an edge node within a preset distance from the user device, and the second server is deployed in the cloud; the method includes: Receive a set of pre-processed user medical data transmitted from user equipment; Identify the attribute categories in the user's medical data set and count the total number of attributes corresponding to each attribute category; Based on the attribute category and the total number of attributes, calculate the Shapley value for each attribute category. The Shapley value is used to quantify the marginal contribution of each attribute category to the data utility of the user's medical dataset. Also, based on the attribute category and the user's medical dataset, calculate the information entropy for each attribute category. The information entropy is used to quantify the information uncertainty of each attribute category. The Shapley value and information entropy are standardized to obtain standardized Shapley value and standardized information entropy respectively; and the reciprocal of standardized Shapley value and standardized information entropy are calculated and analyzed through the skyline to classify the privacy leakage level of each attribute category. Based on the privacy leakage level, a corresponding privacy budget is allocated to the corresponding attribute category. The privacy leakage level is negatively correlated with the corresponding privacy budget value. The difference in privacy budget values ​​between any two attribute categories is determined by the difference in privacy leakage levels and a preset step size. Based on each privacy budget, Laplace noise is injected into the corresponding attribute category to generate a synthetic dataset; The synthesized data set is transmitted and stored to a second server.

[0006] Secondly, embodiments of this application provide a differential privacy protection device for user medical data, applied to a first server in a data processing system. The data processing system further includes a user device and a second server, wherein the first server is deployed at an edge node within a preset distance from the user device, and the second server is deployed in the cloud; the device includes: The data receiving unit is used to receive the pre-processed set of user medical data transmitted by the user equipment. The attribute recognition unit is used to identify the attribute categories in the user's medical data set and count the total number of attributes corresponding to the attribute categories; The numerical calculation unit is used to calculate the Shapley value for each attribute category based on the attribute category and the total number of attributes. The Shapley value is used to quantify the marginal contribution of each attribute category to the data utility of the user's medical data set. It is also used to calculate the information entropy for each attribute category based on the attribute category and the user's medical data set. The information entropy is used to quantify the information uncertainty of each attribute category. The grading unit is used to standardize the Shapley value and information entropy respectively to obtain the standardized Shapley value and standardized information entropy; and to calculate and analyze the reciprocal of the standardized Shapley value and standardized information entropy through the skyline to classify the privacy leakage level of each attribute category. The budget allocation unit is used to allocate a corresponding privacy budget to the corresponding attribute category according to the privacy leakage level. The privacy leakage level is negatively correlated with the corresponding privacy budget value. The difference in privacy budget values ​​between any two attribute categories is determined by the difference in privacy leakage levels and a preset step size. A noise injection unit is used to inject Laplace noise into the corresponding attribute category according to each privacy budget to generate a synthetic dataset; A data storage unit is used to transmit and store the synthesized data set to a second server.

[0007] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and one or more programs, the one or more programs being stored in the memory and configured to be executed by the processor, the programs including instructions for performing the steps as described in the first aspect of embodiments of this application.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps in the first aspect of embodiments of this application.

[0009] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement some or all of the steps described in the first aspect of embodiments of this application.

[0010] As can be seen in this embodiment, the first server receives and processes user medical data locally through edge nodes. It combines Shapley values ​​and information entropy to quantify attribute utility and privacy risks, and uses Skyline calculation to achieve precise classification of attribute privacy leaks. Then, it allocates differentiated privacy budgets according to the classification and independently injects Laplace noise. This method eliminates the need for centralized cloud processing, has strong anti-interference capabilities (through effective feature filtering and standardization), and automates attribute-level privacy protection. It not only reduces the risk of sensitive data transmission leakage and improves processing efficiency but also maximizes data availability, balancing privacy security with the need for medical data value mining, and meeting differential privacy compliance requirements. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a differential privacy protection method for user medical data provided in an embodiment of this application; Figure 3 This is a flowchart illustrating the calculation process of Shapley value for attribute category provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the execution of operational functions by a second server, as provided in an embodiment of this application. Figure 5 This is a functional unit block diagram of a differential privacy protection device for user medical data provided in an embodiment of this application; Figure 6 This is a functional unit block diagram of another differential privacy protection device for user medical data provided in this application embodiment; Figure 7 This is a structural block diagram of a server provided in an embodiment of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0014] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0015] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0016] Please see Figure 1 , Figure 1This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application. For example... Figure 1 As shown, the data processing system 100 includes a first server 110, a user device 120, and a second server 130. The user device 120 is a communication device with medical data acquisition and preprocessing functions, including wearable devices (such as smart bracelets, portable vital sign collectors, medical-grade smart clothing, etc.), mobile terminal devices, and computer terminal devices. These terminal devices can collect real-time medical data such as users' heart rate, blood pressure, and sleep data.

[0017] Specifically, the first server 110 is deployed at an edge node (such as a regional data processing center for a user's mobile phone) within a preset distance from the user device 120, and has local data processing capabilities; the second server 130 is deployed in the cloud and is responsible for the centralized storage and shared management of the synthesized data. To achieve a balance between medical data privacy protection and data availability, the first server 110, the user device 120, and the second server 130 establish a communication connection through the network, forming a closed-loop system of "collection-local processing-cloud storage".

[0018] In the data processing system 100, user equipment 120 is responsible for collecting medical data and performing preliminary preprocessing (data cleaning, noise reduction, screening for outliers / missing values, and preliminary standardization), and then transmitting the preprocessed user medical data set to the first server 110. The first server 110, as the core processing unit, is responsible for performing key operations such as attribute identification, feature filtering, numerical calculation, grading, privacy budget allocation, and noise injection. The second server 130 is responsible for receiving and storing the synthetic data generated by the first server 110, which can then be legally accessed and used by authorized parties such as hospitals and research teams.

[0019] Furthermore, after collecting user medical data according to a preset cycle, user equipment 120 first performs local preprocessing to eliminate data heterogeneity, and then uploads the preprocessed medical data set to the first server 110. After receiving the data, the first server 110 first loads a preset data meta-model, extracts independent feature items and classifies them into multiple attribute categories, and removes invalid, duplicate and redundant feature items to retain valid feature items; then it calculates the Shapley value (quantifying the marginal contribution of data utility) and information entropy (quantifying information uncertainty) of each attribute category, and after standardization, it divides the privacy leakage level through the skyline calculation; then it allocates a differentiated privacy budget according to the level, and independently injects noise into each attribute category based on the Laplace noise mechanism to generate a synthetic data set that meets the requirements of differential privacy; finally, the synthetic data set is transmitted to the second server 130 for storage.

[0020] Typically, a data processing system 100 may include one or more first servers 110, each first server 110 may connect to one or more user devices 120, and each data processing system may include a second server 130, each second server 130 may store the synthetic data set uploaded by one or more first servers 110, thereby enabling parallel processing of multi-user medical data, which not only ensures data privacy and security, but also meets the needs of subsequent value mining of medical data.

[0021] Based on this, this application provides a differential privacy protection method for user medical data. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0022] Example 1: The following describes the framework of the differential privacy protection method for user medical data in this application.

[0023] Please see Figure 2 , Figure 2 This is a flowchart illustrating a differential privacy protection method for user medical data provided in an embodiment of this application. The method is applied to a first server 110 in a data processing system 100. The data processing system 100 also includes a user device 120 and a second server 130. The first server 110 is deployed at an edge node within a preset distance from the user device 120, and the second server 130 is deployed in the cloud. The method includes: Step S201: Receive the set of user medical data transmitted by the user equipment after preprocessing.

[0024] Among them, the user devices at the device layer are directly facing the user's daily health monitoring and clinical assistance scenarios. They accurately collect multi-dimensional highly sensitive medical data such as the user's basic information, real-time physiological indicators and related examination and test results (integrated into a user medical data set), and continuously capture the user's personal data information according to a preset cycle to ensure the timeliness and integrity of the data.

[0025] Furthermore, user devices at the device layer perform basic preprocessing on the data collected from users, eliminating heterogeneity through initial standardization and integration. Specifically, before uploading the dataset to the edge layer, the user devices perform preprocessing operations based on medical data industry standards. These operations include data cleaning (removing missing or abrupt values ​​from sensors installed on the user device), noise reduction (filtering high-frequency noise using a moving average method), and format unification (e.g., converting heterogeneous data from different acquisition devices on the user device into a unified standard format). The aim is to eliminate data heterogeneity and provide high-quality input for subsequent attribute identification and quantification calculations. Data transmission can use the AES-256 encryption protocol (AES-256 is the 256-bit key version of the Advanced Encryption Standard (AES), a globally used symmetric encryption protocol primarily used for the encryption and protection of sensitive data, widely adaptable to data transmission and storage scenarios). Combined with the short-range communication advantages of edge nodes, this reduces the risk of leakage during long-distance transmission of raw sensitive data, complying with relevant regulations regarding the security requirements for sensitive data transmission.

[0026] Furthermore, such as Figure 1 The data processing system shown includes a device layer, an edge layer, and a cloud layer. All core computations (calculating Shapley values, classifying privacy levels, and adding noise) are completed at the edge layer (regional data processing center), avoiding data leakage due to multiple transmissions. In this embodiment of the method, the first server is deployed at an edge node (i.e., the edge layer) within a preset distance from the user's device. Therefore, the first server performs the differential privacy protection data processing for the user's medical dataset. Its high reliability and low latency can compensate for the shortcomings of traditional cloud computing (where the second server performs the differential privacy protection data processing) in handling real-time, high-precision data analysis, thus helping to reduce the risk of data leakage.

[0027] Before receiving the user's medical data set transmitted by the user device, the first server first performs a hash operation on the pre-processed complete dataset to generate a unique checksum. The dataset and checksum are then bound together and transmitted via an encrypted communication channel (such as encrypted Bluetooth / network link). Upon receiving the data, the first server first separates the original dataset and checksum according to a preset protocol. Then, it performs the same hash operation on the received dataset to obtain a new checksum. The server verifies data integrity by comparing the new checksum byte-by-byte with the received checksum. If the first server determines that the comparison is consistent, it determines that the data has not been lost or tampered with due to network fluctuations or device disconnection, and proceeds normally to the subsequent attribute recognition process. If the first server determines that the comparison is inconsistent, it immediately sends a data transmission anomaly command to the user device, triggering a breakpoint resume or full retransmission mechanism, and records the anomaly time, transmission link, and other logs to ensure the dataset is complete before proceeding with subsequent processing.

[0028] Step S202: Identify the attribute categories in the user's medical data set and count the total number of attributes corresponding to each attribute category.

[0029] The data processing performed in steps S202-S206 aims to address the vulnerability of edge infrastructure to attacks and prevent the leakage of highly sensitive medical and health data acquired by the primary server, thereby protecting users from risks such as information theft and fraud. Because the data with different attributes collected by user devices contribute differently to the overall information, the requirements for privacy protection also vary.

[0030] For example, taking diabetes data as an example, the data information it contains includes two aspects: first, the analysis of multiple blood indicators in the blood glucose dataset; and second, the matching of user identity information based on blood glucose characteristics and blood indicators. Based on the sensitivity of different attribute types, a differentiated strategy should be adopted for the protection of user medical data sets collected by user devices. Therefore, in step S202, the attribute categories of the medical data in the user medical data set will be distinguished first, so that different levels of privacy protection can be applied to different attribute categories subsequently.

[0031] In one possible embodiment, identifying attribute categories in a user's medical data set and counting the total number of attributes corresponding to each attribute category includes: loading a preset data meta-model and extracting independent feature items representing single data features from the user's medical data set; classifying the independent feature items according to the classification rules built into the preset data meta-model to obtain multiple attribute categories; and, for each attribute category, removing invalid and duplicate attribute items from at least one corresponding independent feature item to retain valid feature items corresponding to the attribute category; and counting the number of attribute categories after classification to obtain the total number of attributes.

[0032] Each attribute category corresponds to at least one independent feature item, and any two attribute categories are different from each other.

[0033] The pre-defined data element model incorporates core classification rules such as "physiological indicators," "identity identifiers," and "collection context." Each classification rule corresponds to a clear feature matching logic (e.g., "heart rate, blood pressure" matches the "physiological indicators" category, "device number, user unique identifier" matches the "identity identifier" category), ensuring the accuracy and consistency of attribute classification. The system identifies user medical data sets based on the pre-defined data element model. The attribute categories in the dataset include identity identifiers, various physiological indicators, etc. The number of attributes in the dataset is precisely counted and denoted as... .

[0034] As can be seen, in this example, the accurate classification of independent feature items is achieved by using a pre-set medical data meta-model. Combined with the removal of invalid and duplicate attribute items, this not only ensures the uniqueness and validity of the attribute categories, but also provides high-quality data input focused on core features for the subsequent calculation of Shapley value and information entropy. This avoids redundant data interfering with the accuracy of the quantification results and improves the pertinence of the privacy protection strategy.

[0035] In one possible embodiment, for each attribute category, invalid and duplicate attribute items are removed from at least one independent feature item to retain the valid feature items corresponding to the attribute category. This includes: performing a uniqueness check on at least one independent feature item corresponding to the attribute category; if duplicate feature items representing the same data feature exist, performing a deduplication operation on the duplicate feature items to update at least one independent feature item corresponding to the attribute category; and obtaining a preset medical data threshold and preset format specification for each attribute category; and removing meaningless and redundant items from the updated at least one independent feature item according to the preset medical data threshold and preset format specification to retain the valid feature items.

[0036] Among them, meaningless items are divided into abnormal items and erroneous items according to the removal conditions. Abnormal items are independent feature items whose feature values ​​exceed the corresponding preset medical data thresholds, and erroneous items are independent feature items whose data formats do not conform to the corresponding preset format specifications. Redundant items are independent feature items that cannot match the classification rules.

[0037] Among them, meaningless items are divided into abnormal items and erroneous items according to the removal conditions. Abnormal items are independent feature items whose value exceeds the corresponding preset medical data threshold, and erroneous items are independent feature items whose data format does not conform to the corresponding preset format specification. Redundant items are independent feature items that cannot match the classification rules. The uniqueness verification is based on the unique identifier of the feature item (such as "heart rate_data collection date_XXXX") and performs a hash comparison. The preset medical data threshold is derived from medical clinical standards (such as the adult systolic blood pressure threshold of 90-140 mmHg). The preset format specification includes the number of digits retained in the value (such as the heart rate retains an integer number of digits) and the date format (YYYY-MM-DDHH:MM:SS), etc.

[0038] As can be seen, in this example, the three-layer screening mechanism of uniqueness verification for deduplication, medical threshold verification, and format specification verification effectively filters out duplicate data, abnormal data, and data with incorrect format. At the same time, it removes redundant data that cannot match the classification rules, significantly improving the purity and usability of the feature items corresponding to the attribute categories. This lays a reliable data foundation for subsequent quantitative calculations and privacy level classification, and avoids misjudgments of privacy protection strategies caused by invalid data.

[0039] Step S203: Calculate the Shapley value for each attribute category based on the attribute category and the total number of attributes; and calculate the information entropy for each attribute category based on the attribute category and the user medical data set.

[0040] Among them, the Shapley value is used to quantify the marginal contribution of each attribute category to the data utility of the user's medical dataset, and the information entropy is used to quantify the information uncertainty of each attribute category.

[0041] Among them, the Shapley value is used to quantify the marginal contribution of each attribute category to the data utility of the user's medical data set. Its calculation is based on cooperative game theory. The core logic is to traverse all subsets that do not contain the target attribute, compare the model utility difference between "containing the target attribute" and "not containing the target attribute", and combine the weight factor corresponding to the subset size to obtain the average marginal contribution of the target attribute. Information entropy is used to quantify the information uncertainty of each attribute category. Based on the principle of information theory, the more uniform the distribution of attribute values, the higher the information entropy, the greater the difficulty for attackers to infer the user's identity through this attribute, and the lower the risk of privacy leakage. The two together constitute an attribute privacy-utility dual-dimensional evaluation system, which solves the problem that traditional differential privacy cannot distinguish the differentiated features of attributes.

[0042] Example 2: The following describes the framework of the calculation process for the Shapley value of the attribute category in the embodiments of this application.

[0043] Please see Figure 3 , Figure 3This is a flowchart illustrating the calculation process of Shapley value for an attribute category according to an embodiment of this application. The method is applied to the first server 110 in the data processing system 100, and includes: Step S301: Generate all attribute subsets for all attribute categories and initialize the Shapley value accumulator for each attribute category.

[0044] Among them, for the dataset In One attribute, If the value is greater than 1, generate all possible subsets of attributes. Any two attribute subsets must be distinct. The generation of attribute subsets is based on the principles of combinatorics, traversing all non-empty subsets and empty sets that do not contain the target attribute. For example, for three attributes (A, B, C), the subsets for the target attribute A are {}, {B}, {C}, and {B,C}, ensuring that the marginal contribution scenarios of the target attribute in all possible combinations are covered.

[0045] The first server also needs to initialize each attribute. The Shapley accumulator is set to 0, and 64-bit floating-point storage is used to avoid loss of precision during the calculation process.

[0046] Step S302: Load the utility evaluation model adapted to the user's medical data set and obtain the corresponding validation set.

[0047] The validation set contains sample data that match known real labels.

[0048] The utility evaluation model uses a random forest model adapted to medical data classification tasks (exemplarily, 100 decision trees with a maximum depth of 8 layers), which is pre-trained on data-rich labeled medical data. The validation set is derived from desensitized samples of historical medical data (including disease risk labels, physiological indicator labels, etc.) to ensure the objectivity and accuracy of the model's utility evaluation.

[0049] Step S303: Traverse the attribute subsets that do not contain the target attribute category, and train the first subset model that contains the target attribute category and the second subset model that does not contain the target attribute category respectively.

[0050] The target attribute category is the attribute category currently being processed. Iterate through each attribute subset. For each attribute Training "containing" separately "Subset model" and "not containing" "Subset model".

[0051] The model training employs cross-validation, iterating repeatedly until the loss function converges. The training parameters of the first subset model and the second subset model are completely identical, ensuring that the difference in utility is determined solely by the presence or absence of the target attribute.

[0052] Step S304: Input the sample data into the first subset model and the second subset model respectively to obtain the first prediction result and the second prediction result.

[0053] Among them, the prediction results are evaluated based on model accuracy, precision, and recall. Accuracy is preferred as the basis for quantifying utility because it can comprehensively reflect the model's ability to identify positive and negative samples, ensuring the comprehensiveness of the marginal contribution calculation.

[0054] Step S305: Input the first prediction result, the second prediction result, and the true labels corresponding to the sample data into the utility evaluation model to output the marginal contribution of the target attribute category to the model utility.

[0055] The marginal contribution is the difference between the utility of the first subset of the model that includes the target attribute category and the utility of the second subset of the model that does not include the target attribute category.

[0056] Among them, the utility metric is the prediction accuracy of the model on the validation set, which is used here. Indicates, calculation Add attribute subset The marginal contribution after that is calculated using the following formula: .

[0057] Furthermore, the utility evaluation model directly outputs the marginal contribution value by calculating the difference in matching degree between the two sets of prediction results and the true labels. For example, if the accuracy of the first subset model is 82% and the accuracy of the second subset model is 65%, then the marginal contribution is 17%.

[0058] Step S306: Multiply the marginal contribution by the weight factor corresponding to the attribute subset and accumulate it to the Shapley value accumulator to determine the Shapley value of the target attribute category.

[0059] The weighting factor is determined by the size of the subset of attributes.

[0060] Among them, the weighting factor is calculated based on the subset size, for subsets containing A subset of attributes The formula for calculating the weight is as follows: , Here, "!" is the factorial operator in mathematics, and "n!" means "n factorial".

[0061] Furthermore, each Multiply by the corresponding weight, and the result is added to the attribute. The Shapley value accumulator ultimately yields the attribute. Shapley value The Shapley value is a fair allocation method that calculates the average marginal contribution of each participant across all possible sets of participants. It is often used to describe the fair allocation of the total payoff of a game based on the participants' contributions.

[0062] As can be seen, in this embodiment, the refined calculation process of full subset traversal, dual-model comparison, and weighted factor accumulation accurately quantifies the marginal contribution of each attribute category to the utility of medical data, solving the problem that traditional methods cannot objectively distinguish the importance of attributes. This provides a scientific quantitative basis for subsequent differentiated privacy budget allocation, ensuring that the most useful attributes obtain a reasonable level of privacy protection, and balancing data privacy and usability.

[0063] In one possible embodiment, the information entropy corresponding to each attribute category is calculated based on the attribute category and the user medical data set, including: obtaining the value range of each attribute category; determining the frequency of occurrence of each attribute value in the fixed value set in the user medical data set; dividing the frequency of occurrence by the total number of records corresponding to the user medical data set to determine the probability of occurrence of each attribute value; multiplying the probability of occurrence of each attribute value in the currently processed attribute category by the logarithm of the probability of occurrence, summing the results, and taking the negative number to obtain the information entropy of the currently processed attribute category.

[0064] Among them, the value range indicates the fixed set of values ​​corresponding to the attribute category, and the total number of records indicates the number of data records in the collection period corresponding to the user's medical data set. Information entropy is used to measure the information uncertainty of an attribute (for example, "gender" only has male / female, with low uncertainty and low entropy; "blood pressure value" has a wide range, with high uncertainty and high entropy), that is, the higher the information entropy (stronger uncertainty), the lower the need for privacy protection.

[0065] Among them, for the dataset Each attribute in Determine its range Statistical analysis of this dataset Each value Number of times Calculate each value probability , ); .

[0066] Furthermore, attributes Information entropy The calculation formula is as follows: ; in, Indicates the range middle The probability of occurrence Then the value of this item is 0.

[0067] As can be seen, in this example, the standardized calculation process of determining the value range, frequency statistics, probability calculation, and logarithmic summation and inversion accurately quantifies the information uncertainty of each attribute category, providing an objective basis for the risk assessment of attribute privacy leakage. This complements the utility quantification of Shapley values, jointly supporting the accurate classification of attribute privacy levels and ensuring that privacy protection strategies are both consistent with data characteristics and meet compliance requirements.

[0068] Step S204: Standardize the Shapley value and information entropy to obtain standardized Shapley value and standardized information entropy; and calculate and analyze the reciprocals of the standardized Shapley value and standardized information entropy through the skyline to classify the privacy leakage level of each attribute category.

[0069] The process involves standardizing the Shapley value and information entropy to obtain standardized Shapley values ​​and standardized information entropy. This includes: using a fractional standardization method, mapping the Shapley value of each attribute category to the [0,1] interval based on the mean and standard deviation of the Shapley values ​​corresponding to all attribute categories in the user's medical dataset to obtain standardized Shapley values; and mapping the information entropy of each attribute category to the [0,1] interval based on the mean and standard deviation of the information entropy of all attribute categories to obtain standardized information entropy.

[0070] Among them, the standardization of Shapley value and information entropy adopts... Fractional standardization restricts the range of Shapley values ​​and information entropy values ​​to within a certain range.

[0071] ; ; ; ; The above formulas are the Shapley value, the standardized formula for information entropy, and the standard deviation formula, respectively. It is a dataset The mean of Shapley values ​​and information entropy for all attributes. It is a dataset The standard deviation of all Shapley values ​​and information entropy.

[0072] In one possible embodiment, the privacy leakage level of each attribute category is determined by calculating and analyzing the standardized Shapley value and the reciprocal of the standardized information entropy using the skyline calculation method. This includes: constructing a two-dimensional data space based on the standardized Shapley value and the reciprocal of the standardized information entropy corresponding to each attribute category; performing the following operations on the two-dimensional data space to determine multiple privacy leakage levels and the attribute category corresponding to each privacy leakage level, where each attribute category corresponds to a data point in the two-dimensional data space; performing a skyline calculation operation in the two-dimensional data space to determine a target data point and constructing a current skyline set based on the target data point; classifying the attribute categories corresponding to the current skyline set into the current privacy leakage level; removing the attribute categories that have been classified into the current privacy leakage level, and repeating the skyline calculation operation on the remaining unclassified attribute categories to determine the next skyline set and classify the next privacy leakage level based on the next skyline set, repeating the above operations until the privacy leakage level corresponding to each attribute category is determined.

[0073] The target data point is a data point in the two-dimensional data space that is not dominated by other data points. The dominance condition is that the values ​​of other data points in both dimensions of the two-dimensional data space are not lower than those of the target data point, and at least one dimension has a value higher than that of the target data point. The privacy leakage risk corresponding to the next privacy leakage level is lower than the privacy leakage risk corresponding to the current privacy leakage level.

[0074] Furthermore, the first dimension of the two-dimensional data space is the standardized Shapley value, and the second dimension is the reciprocal of the standardized information entropy. Therefore, each attribute category corresponds to a static data point in the two-dimensional data space. The dominance relationship of data points in the two-dimensional data space based on the dominance condition is as follows: For any two data points corresponding to different attribute categories in the two-dimensional data space, namely the first data point and the second data point, if the value of the first data point in the first dimension is not less than the value of the second data point in the first dimension, and the value of the first data point in the second dimension is not less than the value of the second data point in the second dimension, and at the same time, the value of the first data point in at least one dimension is greater than the value of the second data point in the corresponding dimension, then it is determined that the first data point dominates the second data point.

[0075] Among them, the reciprocal of information entropy The calculation formula is A skyline, or skyline, in a multidimensional dataset refers to the set of points in the dataset that are not dominated by any other data points. Skyline calculation involves finding all the data points corresponding to each skyline from the aggregated dataset.

[0076] Furthermore, when performing a grading operation on the two-dimensional data space to determine multiple privacy leakage levels and the attribute category corresponding to each privacy leakage level, the grading operation performed by the first server has the lowest probability of user privacy leakage for Skyline's first privacy leakage, denoted as... Delete the points on the first Skyline, and calculate the second Skyline from the remaining points, denoted as . And so on. The smaller the Shapley value and the reciprocal of information entropy, the lower the privacy requirement of the attribute, and the less likely privacy is to be leaked. The smaller the value, the less likely privacy is to be leaked, the lower the requirement for privacy protection level, and the higher the privacy budget value.

[0077] As can be seen in this example, by eliminating the dimensional differences of different indicators through standardization, the skyline calculation based on the dominance relationship realizes the hierarchical division of attribute privacy levels, avoiding the traditional "one-size-fits-all" classification method. This ensures that attributes with high Shapley values ​​(high utility) and low information entropy reciprocals (high privacy risks) are classified as high-level and receive stronger privacy protection, while low-utility and low-risk attributes retain higher availability, effectively balancing the privacy and security of medical data with the needs of subsequent value mining.

[0078] Step S205: Allocate the corresponding privacy budget to the corresponding attribute category according to the privacy leakage level.

[0079] Among them, the privacy breach level is negatively correlated with the corresponding privacy budget value, and the difference in privacy budget value between any two attribute categories is determined by the difference in privacy breach levels and the preset step size.

[0080] The user medical data set includes There are several attributes, therefore they need to be divided. Each privacy protection level. For each level middle The larger the value, the smaller the corresponding privacy budget. The initial minimum value of the privacy budget is set to 1 (to meet the basic requirements of differential privacy), and the preset step size is 0.1 (based on the need for a balance between privacy and utility in medical data). For example, Level 1 (high risk) is allocated a budget of 1.0, Level 2 is allocated 1.1, Level 3 is allocated 1.2, and so on. The privacy budget configuration conforms to the privacy parallel combination theorem, that is, the overall privacy budget of the dataset is equal to the sum of the individual privacy budgets of all attributes in the dataset. The privacy budget of the cloud-integrated global dataset is equal to the sum of the privacy budgets of the datasets corresponding to all user devices. This ensures that after each attribute is independently annoyed, the overall dataset still satisfies the mathematical provability of differential privacy.

[0081] Step S206: Based on each privacy budget, inject Laplace noise into the corresponding attribute category to generate a synthetic dataset.

[0082] Laplace noise is the core noise mechanism of differential privacy. Its generation and injection process strictly follow the mathematical principles of differential privacy to ensure that the synthesized data hides real privacy information while retaining the statistical distribution characteristics of the data.

[0083] In one possible embodiment, Laplace noise is injected into the corresponding attribute category according to each privacy budget to generate a synthetic dataset, including: determining the query function corresponding to each attribute category; calculating the global sensitivity based on the user medical data set and the query function; determining the scale parameter of the corresponding Laplace noise according to the ratio of the privacy budget to the global sensitivity of the attribute category; generating noise data based on the scale parameter; and performing noise injection operations on the effective feature terms corresponding to the attribute category based on the noise data to obtain the noise-injected attribute category data; and integrating the attribute category data corresponding to each attribute category into a synthetic dataset, wherein the noise injection operations for any two attribute categories are independent of each other.

[0084] The query function is a mapping function that performs preset statistical calculations on the valid feature items of the attribute category, such as "mean calculation" and "extreme value statistics" for physiological indicators, and "frequency statistics" for identity identification. The query result is a statistical quantification result.

[0085] Among them, dataset and They are adjacent if and only if One of the records and The difference is that adjacent datasets refer to two reference medical datasets that differ by a single data record, as determined by the user's medical dataset.

[0086] The global sensitivity is determined by the maximum difference in the query results calculated using a query function for any pair of adjacent datasets. and Query function f The global sensitivity is expressed as and with Norm measurement, where the formula for calculating global sensitivity is as follows: .

[0087] Among them, the scale parameter of Laplace noise is λ=Δf / ε (ε is the privacy budget). The smaller ε is (the higher the privacy level), the smaller λ is, and the greater the noise intensity. The noise injection adopts the method of "true value + noise value". For numerical features, noise is directly superimposed. For categorical features, the values ​​are adjusted according to the noise probability to ensure that the data after injection still conforms to the logic of medical data (such as the heart rate is still within a reasonable range after noise).

[0088] Furthermore, after completing the noise injection operation for the attribute category data, the noise-injected attribute category data is validated against the preset medical data threshold and clinical rules corresponding to that attribute category. If outliers exceeding the preset medical data threshold exist, these outliers are corrected to a reasonable range corresponding to the preset medical data threshold. The noise injection and outlier correction for all attribute categories are completed sequentially using the above steps. The noise injection operations for each attribute category are independent of each other. Finally, all processed attribute category data are integrated to form a synthetic dataset that meets the application requirements of medical scenarios.

[0089] The probability ratio of query results after injecting noise into adjacent datasets satisfies the differential privacy constraint and conforms to the parallel combination theorem, ensuring that the user medical dataset as a whole meets the differential privacy requirements. As can be seen, in this example, the entire process of defining query functions, calculating global sensitivity, determining noise scale parameters, and injecting independent noise achieves attribute-level differentiated noise addition. This satisfies the mathematical provability of differential privacy and minimizes the damage of noise to data usability by precisely controlling noise intensity. This allows the synthesized data to still support applications such as population statistics and trend analysis in medical research, solving the pain point of "excessive noise causing data invalidation" in traditional differential privacy.

[0090] Step S207: The synthesized data set is transmitted and stored to the second server.

[0091] Among them, the entire data processing process from step S201 to step S207 satisfies the mathematical provability of differential privacy.

[0092] Furthermore, differential privacy verification is performed for individual attributes. For any attribute in the user's medical data set, a "neighboring dataset" is defined as two datasets that differ by only one data record (e.g., one dataset contains a user's heart rate record, and the other dataset excludes that record, remaining identical). For this attribute, the global sensitivity of the query function is first calculated, which is the maximum difference in query results obtained by any pair of neighboring datasets using the same query function (e.g., calculating the attribute mean or frequency). This difference is the core indicator for measuring the sensitivity of the query function to changes in data records. Then, an independent privacy budget is allocated to this attribute, and the scale parameter of the Laplace noise is determined based on the ratio of "global sensitivity to privacy budget." The injected noise follows the Laplace distribution corresponding to this parameter. The core characteristic of this noise mechanism is that the probability ratio of query results between neighboring datasets is strictly limited to a reasonable range of "the privacy budget power of the natural constant e" (where e is the natural constant; the larger the privacy budget, the looser the limit and the lower the privacy protection strength; conversely, the smaller the budget, the stricter the limit and the higher the protection strength). In simple terms, attackers cannot infer the existence of a specific data record in the dataset with a high probability by the difference in query results. Therefore, the noise injection operation of a single attribute meets the core requirements of differential privacy.

[0093] Furthermore, differential privacy verification is performed on the entire dataset. Since the noise injection process for each attribute is independent (i.e., noise generation and injection for one attribute does not affect other attributes), overall verification can be performed using the "parallel combination theorem" of differential privacy. The core logic of this theorem is: when multiple mechanisms satisfying differential privacy operate on mutually independent datasets (or data attributes), the new mechanism formed by combining these mechanisms still satisfies differential privacy, and the overall privacy budget is equal to the sum of the privacy budgets of each mechanism. In this application, the overall privacy budget of the dataset is the sum of the privacy budgets of all attributes. Given that a single attribute already satisfies differential privacy, it can be deduced that the noise injection operation on the entire user medical data set also satisfies differential privacy. Furthermore, from a system-wide perspective, the processing of medical data sets for each user device is independent (processing one user's dataset does not affect other users' datasets). Based on the transitivity of the aforementioned "parallel combination theorem," the datasets of all user devices satisfy differential privacy, and their processing is independent. Therefore, the global synthetic dataset (the sum of synthetic data from all users) integrated by the second server in the cloud also satisfies differential privacy. The privacy budget of the global dataset is equal to the sum of the privacy budgets of all user device datasets—that is, first summing the privacy budgets of all attributes in each user's dataset, then further summing the results of all users to form the global privacy budget. This mechanism ensures that attackers cannot infer the specific medical information of a single user through the global data, achieving end-to-end differential privacy protection from the attribute level to the global level.

[0094] Example 3: The following describes the framework for the second server to perform its operational functions in this embodiment of the application.

[0095] Please see Figure 4 , Figure 4 This is a flowchart illustrating the execution of operational functions by a second server according to an embodiment of this application. The method is applied to a second server 130 in a data processing system 100, and includes: Step S401: Receive the composite data set transmitted by the first server and perform a double verification operation.

[0096] The second server receives the synthesized data through an encrypted communication protocol. It first uses a checksum comparison and a data integrity verification algorithm to confirm that the data has not been lost, tampered with, or damaged during transmission.

[0097] Furthermore, based on the differential privacy compliance verification rules, the compliance of noise injection into the synthetic data is verified (such as verifying the independence of noise injection into each attribute category and the consistency of privacy budget allocation with level matching). If the verification fails, the specific reasons for the verification failure and the correction instructions are fed back to the first server; if the verification passes, proceed to the next step.

[0098] Step S402: Perform categorized storage and structured index construction on the synthetic data.

[0099] The data is stored in partitions according to its application scenario (such as clinical diagnosis reference, chronic disease management analysis, and medical research statistics) and data type (such as physiological indicator data and examination and testing data), and encrypted storage is used to ensure data security. The second server builds a structured index for each piece of synthetic data. The index information includes the data generation timestamp, the corresponding attribute category privacy leakage level, the first server identifier, and the data availability label (such as "suitable for group statistics" and "suitable for individual trend analysis") to support subsequent fast and accurate retrieval.

[0100] Step S403: Receive the access request from the authorized party and perform multi-level permission verification.

[0101] The authorizing party (such as medical institutions or research institutions) submits an access request through the real-name authentication interface, along with the purpose of access, the scope of data use, and authorization credentials. For example, the second server first verifies the legitimacy of the authorized party's identity (by comparing it with the preset authorized list and verifying the validity of the authorization certificate), and then matches the corresponding data access permission level according to the purpose of access (e.g., for scientific research statistics access, only aggregated data is allowed, while for clinical diagnosis access, high-precision de-identified synthetic data is allowed). Furthermore, for access requests involving highly sensitive medical data types, an additional compliance audit is performed (such as verifying whether they comply with medical ethics requirements). Once the audit is passed, the data adaptation process begins.

[0102] Step S404: Adapt data as needed and grant access, and record usage logs in real time.

[0103] Specifically, based on the access permissions and usage scenarios of the authorized party, the synthesized data is adapted to specific scenarios (such as providing statistical dimension aggregated data for research teams and single-user trend data for clinical institutions); data access permissions are granted to authorized parties through temporary access channels, with limited access duration and data download volume to prevent data abuse; and a full log of data usage is recorded in real time, including access time, authorized party identification, purpose of use, data access volume, data export records, etc., and the log retention period complies with relevant regulatory requirements.

[0104] Step S405: Receive the synthetic data update notification from the first server and perform data synchronization and old data archiving.

[0105] Specifically, when the first server generates updated synthetic data based on the user's latest medical data, the second server receives the update notification and the new synthetic data, repeating the double verification in step S401. After successful verification, the corresponding old data is replaced with the new synthetic data, and the old synthetic data is archived according to a preset period (e.g., archived and retained for 1 year, after which it is encrypted and destroyed), ensuring the timeliness and security of the stored data. In addition, the second server synchronously updates the structured index to ensure that authorized parties access the latest compliant data.

[0106] Step S406: Regularly generate compliance audit reports and feed them back to the first server optimization strategy.

[0107] Here, "regularly" can refer to a monthly cycle for statistically analyzing indicators such as the frequency of access to synthesized data, the compliance rate of authorized users, and data adaptation satisfaction, and generating an audit report based on the privacy compliance verification results. Furthermore, the audit report is transmitted to the corresponding primary server, providing data support for the primary server to optimize the privacy budget allocation ratio, adjust noise injection scale parameters, and optimize attribute level classification rules, thereby continuously improving the balance between data privacy protection and usability.

[0108] As can be seen, in this embodiment, through the layered verification, classified storage, hierarchical authorization and full-process auditing of the second server, this embodiment not only achieves the legal and compliant sharing of synthetic data, but also ensures the privacy and security of medical data. At the same time, the feedback mechanism promotes the continuous optimization of edge processing strategies and improves the closed loop of the entire process of "edge processing - cloud sharing - strategy iteration".

[0109] The following are embodiments of the apparatus of this application. These embodiments of the apparatus and the embodiments of the method of this application belong to the same concept and are used to execute the methods described in the embodiments of this application. For ease of explanation, only the parts related to the apparatus embodiments of this application are shown in the embodiments of this application. For specific technical details not disclosed, please refer to the description of the embodiments of the method of this application, which will not be repeated here.

[0110] This application provides a differential privacy protection device for user medical data, applied to a first server in a data processing system. Specifically, the differential privacy protection device for user medical data is used to execute the steps performed by the first server in the above-described differential privacy protection method for user medical data. The differential privacy protection device for user medical data in this application may include modules corresponding to the respective steps.

[0111] This application embodiment can divide the differential privacy protection device for user medical data into functional modules based on the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. The module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0112] When dividing each function into modules according to its corresponding function. Figure 5 This is a functional unit block diagram of a differential privacy protection device for user medical data provided in an embodiment of this application; the differential privacy protection device for user medical data is applied to... Figure 1The data processing system 100 shown includes a first server 110, a user device 120, and a second server 130. The first server 110 is deployed at an edge node within a preset distance from the user device, and the second server 120 is deployed in the cloud. The device 50 includes: a collection receiving unit 501 for receiving a pre-processed set of user medical data transmitted from the user device; an attribute identification unit 502 for identifying attribute categories in the user medical data set and counting the total number of attributes corresponding to each category; a numerical calculation unit 503 for calculating the Shapley value for each attribute category based on the attribute category and the total number of attributes, where the Shapley value quantifies the marginal contribution of each attribute category to the data utility of the user medical data set; and calculating the information entropy for each attribute category based on the attribute category and the user medical data set, where the information entropy is used to... The system quantifies the information uncertainty of each attribute category; a grading unit 504 is used to standardize the Shapley value and information entropy to obtain standardized Shapley value and standardized information entropy; and to calculate and analyze the reciprocal of the standardized Shapley value and standardized information entropy through the skyline to classify the privacy leakage level of each attribute category; a budget allocation unit 505 is used to allocate a corresponding privacy budget to the corresponding attribute category according to the privacy leakage level, the privacy leakage level and the corresponding privacy budget value are negatively correlated, and the difference in privacy budget values ​​between any two attribute categories is determined by the difference in privacy leakage levels and a preset step size; a noise injection unit 506 is used to inject Laplace noise into the corresponding attribute category according to each privacy budget to generate a synthetic data set; and a data storage unit 507 is used to transmit and store the synthetic data set to a second server.

[0113] In one possible embodiment, in identifying attribute categories in a user's medical data set and counting the total number of attributes corresponding to each attribute category, the attribute identification unit 502 is specifically configured to: load a preset data meta-model and extract independent feature items representing single data features from the user's medical data set; classify the independent feature items according to the classification rules built into the preset data meta-model to obtain multiple attribute categories, each attribute category corresponding to at least one independent feature item, and any two attribute categories being distinct; and, for each attribute category, remove invalid and duplicate attribute items from the corresponding at least one independent feature item to retain the valid feature items corresponding to the attribute category; and, count the number of attribute categories after classification to obtain the total number of attributes.

[0114] In one possible embodiment, regarding the removal of invalid and duplicate attribute items from at least one independent feature item for each attribute category to retain valid feature items corresponding to the attribute category, the attribute identification unit 502 is specifically configured to: perform a uniqueness check operation on at least one independent feature item corresponding to the attribute category; if there are duplicate feature items characterized as the same data feature, perform a deduplication operation on the duplicate feature items to update at least one independent feature item corresponding to the attribute category; and obtain a preset medical data threshold and preset format specification corresponding to each attribute category; according to the preset medical data threshold and preset format specification, remove meaningless and redundant items from the updated at least one independent feature item to retain valid feature items. Meaningless items are divided into abnormal items and erroneous items according to the removal conditions. Abnormal items are independent feature items whose feature item values ​​exceed the corresponding preset medical data threshold, and erroneous items are independent feature items whose data format does not conform to the corresponding preset format specification. Redundant items are independent feature items that cannot match the classification rules.

[0115] In one possible embodiment, in calculating the Shapley value corresponding to each attribute category based on the attribute category and the total number of attributes, the numerical calculation unit 503 is specifically used for: generating all attribute subsets for all attribute categories and initializing the Shapley value accumulator for each attribute category, wherein any two attribute subsets are distinct; loading a utility evaluation model adapted to the user's medical data set and obtaining the corresponding validation set, the validation set containing sample data matching known real labels; traversing attribute subsets that do not contain the target attribute category, and training a first subset model containing the target attribute category and a second subset model that does not contain the target attribute category, respectively. The current attribute category is defined. Sample data is input into the first subset model and the second subset model respectively to obtain the first prediction result and the second prediction result. The first prediction result, the second prediction result, and the true labels corresponding to the sample data are input into the utility evaluation model to output the marginal contribution of the target attribute category to the model utility. The marginal contribution is the difference between the utility of the first subset model containing the target attribute category and the utility of the second subset model not containing the target attribute category. The marginal contribution is multiplied by the weight factor corresponding to the attribute subset and accumulated into the Shapley value accumulator to determine the Shapley value of the target attribute category. The weight factor is determined by the subset size of the attribute subset.

[0116] In one possible embodiment, in calculating the information entropy corresponding to each attribute category based on the attribute category and the user medical data set, the level division unit 504 is specifically used to: obtain the value range of each attribute category, the value range being used to indicate the fixed value set corresponding to the attribute category; determine the frequency of occurrence of each attribute value in the fixed value set in the user medical data set; divide the frequency of occurrence by the total number of records corresponding to the user medical data set to determine the probability of occurrence of each attribute value, the total number of records being used to indicate the number of data records in the collection period corresponding to the user medical data set; and multiply the probability of occurrence of each attribute value in the currently processed attribute category by the logarithm of the probability of occurrence, sum them up, and take the negative number to obtain the information entropy of the currently processed attribute category.

[0117] In one possible embodiment, in calculating and analyzing the standardized Shapley value and the reciprocal of the standardized information entropy through the skyline to classify the privacy leakage level of each attribute category, the budget allocation unit 505 is specifically used to: construct a two-dimensional data space based on the standardized Shapley value and the reciprocal of the standardized information entropy corresponding to each attribute category; for the two-dimensional data space, perform the following operations to determine multiple privacy leakage levels and the attribute category corresponding to each privacy leakage level, each attribute category corresponding to a data point in the two-dimensional data space; perform a skyline calculation operation in the two-dimensional data space to determine target data points, and construct the current processing skyline set based on the target data points, wherein the target data points are the points in the two-dimensional data space... Data points that are not dominated by other data points are identified by the following conditions: the dominance condition is that the values ​​of other data points in both dimensions of the two-dimensional data space are not lower than the target data point, and at least one dimension has a value higher than the target data point; the attribute categories corresponding to the currently processed skyline set are classified into the current privacy leakage level; the attribute categories that have been classified into the current privacy leakage level are removed, and the skyline calculation operation is repeated for the remaining unclassified attribute categories to determine the next skyline set and to classify the next privacy leakage level based on the next skyline set. The above operation is repeated until the privacy leakage level corresponding to each attribute category is determined, and the privacy leakage risk corresponding to the next privacy leakage level is lower than the privacy leakage risk corresponding to the current privacy leakage level.

[0118] In one possible embodiment, in injecting Laplace noise into the corresponding attribute category according to each privacy budget to generate a synthetic dataset, the noise injection unit 506 is specifically configured to: determine the query function corresponding to each attribute category, wherein the query function is a mapping function that performs a preset statistical calculation on the effective feature items of the attribute category; calculate the global sensitivity based on the user medical dataset and the query function, wherein the global sensitivity is determined by the maximum difference value of the query results obtained by the query function for any pair of adjacent datasets, and the query result is a statistical quantification result, wherein the adjacent datasets refer to two reference medical datasets that differ by a single data record as determined by the user medical dataset; determine the scale parameter of the corresponding Laplace noise according to the ratio of the privacy budget of the attribute category to the global sensitivity; generate noise data based on the scale parameter, and perform noise injection operation on the effective feature items corresponding to the attribute category according to the noise data to obtain the attribute category data after noise injection; and integrate the attribute category data corresponding to each attribute category into a synthetic dataset, wherein the noise injection operations corresponding to any two attribute categories are independent of each other.

[0119] When using integrated units, such as Figure 6 As shown, Figure 6 This is a functional unit block diagram of another differential privacy protection device for user medical data provided in this application embodiment. Figure 6 The differential privacy protection device 60 for user medical data includes a processing module 602 and a communication module 601. The processing module 602 controls and manages the operation of the differential privacy protection device 60, including, for example, the steps of the collection receiving unit 501, attribute identification unit 502, numerical calculation unit 503, level division unit 504, budget allocation unit 505, noise injection unit 506, and data storage unit 507, and / or other processes for performing the techniques described herein. The communication module 601 supports interaction between the differential privacy protection device and other devices. Figure 6 As shown, the differential privacy protection device for user medical data may include a storage module 603, which is used to store the program code and data of the differential privacy protection device for user medical data.

[0120] The processing module 602 may be a processor or processing module, such as a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication module 601 may be a transceiver, RF circuitry, or a communication interface, etc. The storage module 603 may be a memory.

[0121] All relevant content in each scenario involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here. The differential privacy protection device 60 for user medical data described above can all perform the above... Figure 2 The example shown is a differential privacy protection method for user medical data.

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0123] Figure 7 This is a structural block diagram of a server provided in an embodiment of this application. For example... Figure 7As shown, server 70 may include one or more of the following components: processor 701, memory 702 coupled to processor 701, wherein memory 702 may store one or more computer programs 703, and the one or more computer programs 703 may be configured to implement the methods described in the above embodiments when executed by one or more processors 701. Server 70 here is either the first server 110 or the second server 130 in the above embodiments.

[0124] Processor 701 may include one or more processing cores. Processor 701 connects to various parts of the server 70 using various interfaces and lines, and performs various functions and processes data of the server 70 by running or executing instructions, programs, code sets, or instruction sets stored in memory 702, and by calling data stored in memory 702. Optionally, processor 701 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 701 may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 701 and may be implemented separately using a communication chip.

[0125] The memory 702 may include random access memory (RAM) or read-only memory (ROM). The memory 702 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 702 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described above. The data storage area may also store data created by the server 70 during use.

[0126] It is understood that server 70 may include more or fewer structural elements than those shown in the above block diagram, and this is not limited here.

[0127] This application also provides a computer storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements some or all of the steps of any of the methods described in the above method embodiments.

[0128] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments.

[0129] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0130] In the several embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, and systems can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and there may be other division methods in actual implementation; for example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can be physically comprised separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware or in the form of hardware plus software functional units.

[0133] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute partial steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, volatile memory, or non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM), etc., which are various media capable of storing program code.

[0134] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions without departing from the spirit and scope of the present invention, and various modifications and alterations can be made, including combinations of the different functions and implementation steps described above, as well as software and hardware implementation methods, all of which are within the protection scope of the present invention.

Claims

1. A method for differential privacy protection of user medical data, characterized in that, A first server applied to a data processing system, the data processing system further comprising a user device and a second server, wherein the first server is deployed at an edge node within a preset distance from the user device, and the second server is deployed in the cloud; the method comprises: receiving a set of pre-processed user medical data transmitted by the user device; identifying attribute categories in the set of user medical data and counting the total number of attributes corresponding to the attribute categories; calculating a Shapley value corresponding to each attribute category according to the attribute categories and the total number of attributes, the Shapley value being used to quantify the marginal contribution of each attribute category to the data utility of the set of user medical data; and calculating an information entropy corresponding to each attribute category according to the attribute categories and the set of user medical data, the information entropy being used to quantify the information uncertainty of each attribute category; standardizing the Shapley value and the information entropy respectively to obtain a standardized Shapley value and a standardized information entropy; and calculating and analyzing the reciprocals of the standardized Shapley value and the standardized information entropy by skyline to divide the privacy leakage level of each attribute category; allocating a corresponding privacy budget to the corresponding attribute category according to the privacy leakage level, the privacy leakage level being negatively correlated with the value of the corresponding privacy budget, and the value difference of the privacy budget between any two attribute categories being determined by the level difference of the privacy leakage level and a preset step length; injecting Laplace noise into the corresponding attribute category according to each privacy budget to generate a set of synthetic data; transmitting and storing the set of synthetic data to the second server.

2. The method of claim 1, wherein, The identification of attribute categories in the set of user medical data and the counting of the total number of attributes corresponding to the attribute categories comprises: loading a preset data meta-model to extract independent feature items representing single data features from the set of user medical data; classifying the independent feature items according to the classification rules built-in the preset data meta-model to obtain a plurality of attribute categories, each attribute category corresponding to at least one independent feature item, and any two attribute categories being different from each other; and for each attribute category, eliminating invalid attribute items and duplicate attribute items in the corresponding at least one independent feature item to retain valid feature items corresponding to the attribute category; and counting the number of classified attribute categories to obtain the total number of attributes.

3. The method of claim 2, wherein, The elimination of invalid attribute items and duplicate attribute items in the corresponding at least one independent feature item for each attribute category to retain valid feature items corresponding to the attribute category comprises: performing a uniqueness verification operation on the at least one independent feature item corresponding to the attribute category, and if there are duplicate feature items representing the same data feature, performing a deduplication operation on the duplicate feature items to update the at least one independent feature item corresponding to the attribute category; and obtaining a preset medical data threshold and a preset format specification corresponding to each attribute category; According to the preset medical data threshold and the preset format specification, the meaningless items and the redundant items in the updated at least one independent feature item are removed to retain the effective feature items, the meaningless items are divided into abnormal items and error items according to the removal condition, the abnormal items are independent feature items with feature item values exceeding the corresponding preset medical data threshold, and the error items are independent feature items with data formats not conforming to the corresponding preset format specification, and the redundant items are independent feature items that cannot match the classification rule.

4. The method according to any one of claims 1 to 3, characterized in that, The calculation of the Shapley value corresponding to each attribute category according to the attribute category and the total number of attributes includes: Generating all attribute subsets for all attribute categories, and initializing the Shapley value accumulator of each attribute category, any two attribute subsets being different from each other; Loading an utility evaluation model adapted to the user medical data set, and obtaining a corresponding verification set, the verification set containing sample data matching known true labels; Traverse the attribute subset not containing the target attribute category, respectively train the first subset model containing the target attribute category and the second subset model not containing the target attribute category, and the target attribute category is the currently processed attribute category; Input the sample data into the first subset model and the second subset model respectively to obtain the first prediction result and the second prediction result; Input the first prediction result, the second prediction result and the true label corresponding to the sample data into the utility evaluation model to output the marginal contribution of the target attribute category to the model utility, and the marginal contribution is the difference between the utility of the first subset model containing the target attribute category and the utility of the second subset model not containing the target attribute category; Multiply the marginal contribution by the weight factor corresponding to the attribute subset, and accumulate it to the Shapley value accumulator to determine the Shapley value of the target attribute category, and the weight factor is determined by the subset size of the attribute subset.

5. The method of claim 4, wherein, The calculation of the information entropy corresponding to each attribute category according to the attribute category and the user medical data set includes: Obtaining the value domain of each attribute category, the value domain being used to indicate the fixed value set corresponding to the corresponding attribute category; Determine the occurrence frequency of each attribute value in the fixed value set in the user medical data set; Divide the occurrence frequency by the total number of records corresponding to the user medical data set to determine the occurrence probability of each attribute value, and the total number of records is used to indicate the number of data records in the collection period corresponding to the user medical data set; Multiply the occurrence probability of each attribute value in the currently processed attribute category by the logarithm of the occurrence probability, sum and take the opposite number to obtain the information entropy of the currently processed attribute category.

6. The method of claim 5, wherein, The skyline is calculated and analyzed to divide the privacy leakage level of each attribute category, including: constructing a two-dimensional data space according to the normalized Shapley value and the reciprocal of the normalized information entropy corresponding to each attribute category; determining a plurality of privacy leakage levels and an attribute category corresponding to each privacy leakage level, each attribute category corresponding to a data point in the two-dimensional data space, by performing the following operations on the two-dimensional data space; performing a skyline calculation operation in the two-dimensional data space to determine a target data point, and forming a current skyline set being processed according to the target data point, the target data point being a data point in the two-dimensional data space that is not dominated by other data points, the domination condition being that the two-dimensional data space values of the other data points are not lower than those of the target data point, and at least one dimension value is higher than that of the target data point; dividing the attribute category corresponding to the current skyline set being processed into a current privacy leakage level; repeating the skyline calculation operation on the remaining attribute categories of unclassified levels to determine a next skyline set, and dividing a next privacy leakage level based on the next skyline set, and repeating the above operations until the privacy leakage level corresponding to each attribute category is determined, the privacy leakage risk corresponding to the next privacy leakage level being lower than that corresponding to the current privacy leakage level.

7. The method of claim 6, wherein, the injecting of Laplace noise into the corresponding attribute category according to each privacy budget to generate a synthetic data set, comprising: determining a query function corresponding to each attribute category, the query function being a mapping function that performs a predetermined statistical calculation on the effective feature items of the attribute category; calculating a global sensitivity based on the user medical data set and the query function, the global sensitivity being determined by the maximum difference value of the query results calculated by the query function on any pair of adjacent data sets, the query result being a statistical quantitative result, and the adjacent data sets being two reference medical data sets determined by the user medical data set and differing by a single data record; determining a scale parameter of the corresponding Laplace noise according to the ratio of the privacy budget of the attribute category to the global sensitivity; generating noise data based on the scale parameter, and performing noise injection operation on the effective feature items corresponding to the attribute category according to the noise data to obtain attribute category data after noise injection; integrating the attribute category data corresponding to each attribute category into a synthetic data set, and the noise injection operations corresponding to any two attribute categories being independent of each other.

8. An apparatus for differential privacy protection of user medical data, characterized in that, A first server applied to a data processing system, the data processing system further comprising a user device and a second server, wherein the first server is deployed on an edge node within a preset distance from the user device, and the second server is deployed in the cloud; the device comprises: a set receiving unit configured to receive a preprocessed user medical data set transmitted by the user device; An attribute identifying unit is configured to identify attribute categories in the user medical data set and count total numbers of attributes corresponding to the attribute categories. A numerical calculation unit is configured to calculate Shapley values corresponding to each of the attribute categories according to the attribute categories and the total numbers of attributes, the Shapley values being used to quantify marginal contributions of each of the attribute categories to data utility of the user medical data set, and calculate information entropies corresponding to each of the attribute categories according to the attribute categories and the user medical data set, the information entropies being used to quantify information uncertainty of each of the attribute categories. A grade dividing unit is configured to perform standardization on the Shapley values and the information entropies respectively to obtain standardized Shapley values and standardized information entropies, and calculate and analyze reciprocals of the standardized Shapley values and the standardized information entropies by using a skyline to divide a privacy leakage grade of each of the attribute categories. A budget allocating unit is configured to allocate a corresponding privacy budget to a corresponding attribute category according to the privacy leakage grade, the privacy leakage grade and a value of the corresponding privacy budget being negatively correlated, and a numerical difference of the privacy budget between any two of the attribute categories being determined by a grade difference of the privacy leakage grade and a preset step length. A noise injecting unit is configured to inject Laplace noise into a corresponding attribute category according to each of the privacy budgets to generate a synthetic data set. A data storage unit is configured to transmit and store the synthetic data set to the second server.

9. A server, characterized by A computer program for electronic data exchange, wherein the computer program causes a computer to perform the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program for electronic data exchange, wherein the computer program causes a computer to perform the method of any one of claims 1-7.

Images