A method for secure sharing of heterogeneous data based on differential privacy budgeting
By unifying the modeling and cross-task allocation of heterogeneous data, the problems of insufficient data accuracy and insufficient task priority allocation in differential privacy budget allocation are solved, and efficient and secure data sharing in different scenarios is achieved.
Patent Information
- Application Number
- CN202611340594.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, differential privacy budget allocation strategies fail to effectively consider the differences in the value of data under different use scenarios, resulting in insufficient data accuracy in high-value scenarios. Furthermore, the lack of elastic allocation capabilities across multiple concurrent data sharing tasks leads to impaired data availability for high-priority tasks.
A heterogeneous data security sharing method based on differential privacy budget is adopted. By uniformly modeling the data to be shared, calculating the availability value score, and combining the sensitivity level, the basic privacy budget is solved using multi-objective optimization. Online convex optimization is used for cross-task elastic allocation, and finally differential noise addition processing is performed.
It enables differentiated budget allocation for data in different scenarios, ensures data accuracy in high-value scenarios, and ensures data quality for high-priority tasks through cross-task allocation, thus avoiding damage to data availability.
Smart Images

Figure CN123069435A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data security and privacy protection technology, and in particular to a method for secure sharing of heterogeneous data based on differential privacy budgeting. Background Technology
[0002] With the increasing demand for data sharing, differential privacy technology is widely used to provide privacy protection during data release. Existing methods typically employ a globally uniform privacy budget allocation strategy, or simply apply static hierarchical noise based on data sensitivity.
[0003] This type of approach has two main problems. First, the value differences of the same data in different use cases are not included in the budget allocation, and uniform noise addition leads to insufficient data accuracy in high-value scenarios. Second, there is a lack of flexible budget allocation capabilities across multiple concurrent data-sharing tasks. When the total system budget is insufficient, it cannot be reasonably allocated according to task priority, resulting in impaired data availability for high-priority tasks. Summary of the Invention
[0004] The purpose of this invention is to propose a method for secure sharing of heterogeneous data based on differential privacy budget adaptation in order to solve the above-mentioned problems.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A heterogeneous data secure sharing method based on differential privacy budget adaptation includes:
[0007] To model shared heterogeneous data in a unified way, each data item is represented as a triplet structure containing data entity, attribute, and sensitivity level;
[0008] Receive a data sharing request, obtain the usage scenario identifier, requester trust level and data timestamp carried in the request, encode and convert the usage scenario identifier to obtain a scenario weight value, normalize the requester trust level, and perform weighted summation based on the processed scenario weight value, the normalized requester trust level value and the data time decay factor to calculate the usability value score of each data item.
[0009] Combining the usability value score and the sensitivity level, with the dual objectives of maximizing data usability and minimizing the risk of privacy leakage, and with the upper limit of the total privacy budget of the system as a constraint, the basic privacy budget amount of each data item is solved through multi-objective optimization.
[0010] Obtain the concurrent task queue. When the total privacy budget required by each task exceeds the system's total privacy budget limit, based on the urgency, value, and noise tolerance of each task, perform cross-task elastic allocation of the basic privacy budget amount through online convex optimization to obtain the actual allocated budget for each data item.
[0011] Based on the actual allocated budget and the data type of the data items, perform differentiated noise addition processing on each data item to generate noisy data and respond to the sharing request.
[0012] Preferably, the usability value score is calculated as follows:
[0013] ;
[0014] in, For the first Usability value score for each data item For the first The usage scenario weight value corresponding to each data item For the first The normalized requester trust level value corresponding to each data item For the first Data time decay factor for each data item , , The preset weighting coefficients satisfy the following conditions: .
[0015] Preferably, the data time decay factor is calculated as follows:
[0016] ;
[0017] in, Represents an exponential function. The decay rate constant is For the first The difference between the generation time of each data item and the current time. The range of values is .
[0018] Preferably, the objective function for solving the basic privacy budget for each data item through multi-objective optimization is:
[0019] ;
[0020] ;
[0021] in, The total number of data items. For the first Basic privacy budget for each data item, This is the upper limit of the system's total privacy budget. For the first Sensitivity coefficient for each data item Weighting for privacy risk penalties.
[0022] Preferably, the sensitivity level is divided into multiple discrete levels, and each level corresponds to a sensitivity coefficient. Among them, personal identification information corresponds to the highest sensitivity level, behavioral trajectory data corresponds to the medium sensitivity level, and aggregated statistical data corresponds to the low sensitivity level. The larger the value, the greater the privacy leakage risk penalty term in the objective function for the corresponding data item.
[0023] Preferably, the objective function for the cross-task elastic allocation of the basic privacy budget through online convex optimization is:
[0024] ;
[0025] ;
[0026] in, The total number of concurrent tasks. For the first Normalized urgency of each task For the first The normalized task value of each task. For the first Normalized noise tolerance for each task, For the first The set of data item numbers contained in each task For set Data item number in For the first The actual allocation budget for each data item For the first The basic privacy budget for each data item.
[0027] Preferably, when the task queue is idle, the released budget is replenished to the data items corresponding to the previously compressed tasks according to the historical compression ratio, so that the basic privacy budget of each data item is restored to the initial level.
[0028] Preferably, the differential noise addition process for each data item includes:
[0029] For numerical data, Laplace noise is applied, with a noise scale of [value missing]. ,in To ensure global sensitivity of the query function; for categorical data, exponential noise is used, and the output category is selected probabilistically based on the utility function score; for image data and high-dimensional data, differential privacy stochastic gradient descent is used, which prunes the gradient and adds noise when updating the model parameters; for time series data, a sliding window method is used to add noise window by window, which satisfies differential privacy requirements within the window while preserving the temporal correlation between adjacent windows.
[0030] Preferably, the method further includes a policy update step:
[0031] After each data sharing task is completed, the data availability score from the requester and the privacy leak event records detected by the privacy audit module are collected. The data characteristics, actual budget allocation, noise addition method, data availability score, and privacy leak event records of each data item are stored as samples in the feedback database. The samples in the feedback database are statistically analyzed periodically. If the data availability score of a certain type of data is higher than the preset threshold under a low budget allocation, the base budget for that type of data in subsequent tasks is reduced. If the data availability score of a certain type of data is still lower than the preset threshold under a high budget allocation, its base budget is increased or its noise addition method is switched to an alternative method with less loss of accuracy for that data type.
[0032] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0033] 1. This invention comprehensively considers the usage scenario, the trust level of the requester, and the time decay factor through usability value scoring. Each factor is normalized and then participates in the weighted calculation, so that the same data can obtain different budget amounts in different scenarios, and the data accuracy in high-value scenarios is guaranteed.
[0034] 2. This invention dynamically allocates budgets among concurrent tasks through online convex optimization by elastically allocating resources across tasks. High-priority tasks do not suffer from reduced data quality due to budget contention, and compressed tasks are restored to their initial level by budget replenishment when the queue is idle. Attached Figure Description
[0035] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:
[0036] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0037] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.
[0038] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.
[0039] Example 1
[0040] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.
[0041] In this embodiment, it includes:
[0042] In the field of differential privacy data sharing, existing methods mostly adopt a globally uniform privacy budget allocation strategy or simply apply static, graded noise based on data sensitivity. This leads to two prominent problems: First, the value difference of the same data in different use cases is not considered. For example, the availability requirement of medical data in emergency scientific research scenarios is much higher than that in regular statistical scenarios, and uniform noise application will result in insufficient data accuracy in high-value scenarios. Second, there is competition for privacy budgets among multiple concurrent data sharing tasks, and there is a lack of flexible allocation capabilities across tasks. When the total budget is insufficient, it is impossible to allocate it reasonably according to task priority.
[0043] A differential privacy data sharing method according to an embodiment of the present invention includes the following steps:
[0044] Step 1: Perform a unified modeling of the heterogeneous data to be shared, and obtain the triple representation of each data item.
[0045] All types of heterogeneous data to be shared are uniformly represented as a triple structure of "data entity, attribute, and sensitivity". Here, the data entity is the unique identifier of the data, the attribute is the information of each field carried by the data entity, and the sensitivity is the inherent sensitivity level preset according to the data category. Structured data, text data, image data, and time-series data are all registered and stored according to this triple structure.
[0046] Step 2: Calculate the availability value score for each data item based on the context information of the current sharing request.
[0047] Upon receiving a data sharing request, the system retrieves the usage scenario identifier, requester trust level, and data timestamp carried in the request. Data preprocessing is then performed on these three types of raw input: the usage scenario identifier is encoded and mapped to its corresponding scenario weight value; the requester trust level is normalized using a mean-based range-based method to scale it to the same dimension as the scenario weights, eliminating dimensional differences when directly weighting indicators from different sources. Based on these three preprocessed factors, the usability value score of each data item in the current sharing task is calculated.
[0048] The usability value score is calculated as follows:
[0049] ;
[0050] in, For the first Usability value score for each data item For the first The usage scenario weight value corresponding to each data item For the first The normalized requester trust level value corresponding to each data item For the first Data time decay factor for each data item , , The preset weighting coefficients satisfy the following conditions: , This is the number of the data item.
[0051] It should be noted that the data time decay factor The calculation is based on the interval between the data generation time and the current time, specifically:
[0052] ;
[0053] in, Represents an exponential function. The decay rate constant is For the first The difference between the generation time and the current time for each data item, expressed in units of 1 and 2. The time unit corresponds to ensure the exponent is dimensionless. The further back in time the data was generated, the greater the dimension of the exponent. The smaller the value, the less valuable the historical data becomes over time. Because... The range of values is ,and , After normalization, the range of values is consistent, and the three can be directly weighted and summed.
[0054] Step 3: Combine usability value score and inherent sensitivity level to generate basic privacy budget for each data item.
[0055] The usability value score is jointly calculated with the sensitivity level in the triple to generate the basic privacy budget for each data item. The allocation principle is as follows: Allocate more budget to data items with high sensitivity and low sensitivity. Allocate less budget to data items that are low in sensitivity and high in price.
[0056] For data items where value scoring and sensitivity are correlated, the allocation value is solved through multi-objective optimization. With the dual objectives of maximizing data availability and minimizing privacy leakage risk, and constrained by the system's total privacy budget ceiling, the basic privacy budget amount for each data item is determined. The objective function for multi-objective optimization is:
[0057] ;
[0058] ;
[0059] in, The total number of data items. This is the upper limit of the system's total privacy budget. For the first Sensitivity coefficient for each data item This serves as a privacy risk penalty weight, used to balance usability gains with privacy breach risks. Because... It has been normalized. The range of values and Since their dimensions are consistent, they can directly participate in the weighted operation of the same objective function.
[0060] It should be noted that the inherent sensitivity levels are divided into multiple discrete levels. Personal identification information corresponds to the highest sensitivity level, behavioral trajectory data corresponds to the medium sensitivity level, and aggregated statistical data corresponds to the low sensitivity level. Each level corresponds to a sensitivity coefficient. , The larger the value, the higher the sensitivity, and the greater the weight of the privacy leakage risk penalty item in multi-objective optimization.
[0061] Step 4: Based on the attributes of each task in the concurrent task queue, perform cross-task elastic allocation of the basic privacy budget to obtain the actual allocated budget for each data item.
[0062] Retrieve a task queue consisting of all concurrent data-sharing tasks in the current system. When the total privacy budget required by all tasks exceeds the system's total privacy budget limit, allocate the privacy budget among multiple concurrent tasks based on the urgency, value, and sensitivity of each task to data accuracy.
[0063] The allocation strategy is as follows: for tasks with high urgency and high value, allocate the full budget required; for tasks with low urgency and low value, reduce their budget proportionally. The reduction ratio is positively correlated with the task's tolerance to noise; tasks with higher tolerance will undergo a larger reduction.
[0064] The allocation process is solved in real time using an online convex optimization algorithm. The urgency level is considered. Task value Noise tolerance Before being used in calculations, mean normalization based on the range was applied. , , To eliminate dimensional differences when directly weighting different attributes, the scale is uniformly scaled to the same range. The objective function of the online convex optimization algorithm is to maximize the weighted availability of all tasks, with the constraint that the sum of the budgets allocated to each task does not exceed the system's total budget limit.
[0065] ;
[0066] ;
[0067] in, The total number of concurrent tasks. Number the task. For the first Normalized urgency of each task For the first The normalized task value of each task. For the first Normalized noise tolerance for each task, For the first The set of data item numbers contained in each task For set Data item number in For the first The actual allocation budget for each data item For the first The basic privacy budget for each data item, constraining The higher the tolerance level, the greater the compression allowed.
[0068] In this embodiment, to prevent low-priority tasks from being continuously compressed and becoming unusable for extended periods, a budget rollback process is introduced. When the task queue is idle, the released budget is rolled back to the data items corresponding to the previously compressed tasks according to the historical compression ratio, restoring the basic privacy budget of each data item to its initial level.
[0069] Step 5: Based on the actual budget allocation and data type, perform differential noise enhancement on each data item, generate the noise-enhanced data, and respond to the sharing request.
[0070] according to Based on the data type of the data item, select the corresponding noise-adding method to add noise to the data item.
[0071] For numerical data, a Laplace mechanism is used to add noise, with a noise scale of [value missing]. ,in To optimize the global sensitivity of the query function, an exponential mechanism is used for categorical data, selecting the output category probabilistically based on the utility function score. For image and high-dimensional data, differential privacy stochastic gradient descent is employed, pruning gradients and adding noise during model parameter updates. For time-series data, a sliding window approach is used to add noise window by window, preserving temporal correlations between adjacent windows while satisfying differential privacy requirements within each window.
[0072] The data with added noise is sent to the requester through a secure channel, and the response includes the privacy budget consumption of each data item.
[0073] Step 6: Collect feedback data on the sharing effect and update the privacy budget allocation strategy based on the feedback data.
[0074] After each data sharing task is completed, feedback data is collected from two dimensions. Feedback from the requester's side includes data availability scores and task completion status. Feedback from the privacy side includes records of privacy breaches detected by the privacy audit module.
[0075] The data characteristics of each data item, The noise addition method, data availability score, and privacy breach event record are stored as a complete sample in the feedback database. Statistical analysis is performed on the samples in the feedback database periodically, and the calculation parameters for the basic privacy budget are updated based on the analysis results.
[0076] The specific update rules are as follows: if the data availability score of a certain type of data is higher than the preset threshold under a low budget allocation, the base budget for that type of data in subsequent tasks will be reduced; if the data availability score of a certain type of data is still lower than the preset threshold under a high budget allocation, its base budget will be increased or its noise addition method will be switched to an alternative method with less loss of accuracy for that data type.
[0077] Through a continuous cycle of allocation, use, feedback, and calibration, the privacy budget allocation strategy gradually converges to a balance between the strength of privacy protection and data availability.
[0078] Example 2
[0079] A medical data sharing platform undertook multiple concurrent research data sharing tasks in a certain year. The platform's database stores heterogeneous data, including structured medical records, text-based diagnostic records, medical imaging data, and vital sign time-series data. On a certain day, the platform simultaneously received three data sharing requests: Task 1 from an emergency infectious disease research team at a university, Task 2 from a pharmaceutical company's routine drug efficacy statistical analysis team, and Task 3 from a public health institution's retrospective epidemiological research team. What is the system's total privacy budget limit? If the total budget requirement for the three tasks exceeds the upper limit, cross-task elastic allocation needs to be triggered.
[0080] The platform uses a unified model to model four types of data to be shared, registering each data item as a triple of "data entity, attribute, and sensitivity". Structured medical record data contains personal identification information and corresponds to the highest sensitivity level; text diagnostic records contain behavioral and medical history information and correspond to the medium sensitivity level; medical imaging data contains biometric information and corresponds to the highest sensitivity level; and vital sign time-series data consists of aggregated and associative records and corresponds to the low sensitivity level, as shown in Table 1 below.
[0081] Table 1 Data Items: Triple Registration Information
[0082]
[0083] The platform receives the sharing request from Task 1 and extracts the context information: the usage scenario is identified as "urgent scientific research," the requester's trust level is "institutional certification level A," and the time interval between the generation of each data item and the current time is different. The usage scenario identifier is encoded and converted, with "urgent scientific research" mapped to a scenario weight. The trust level of the requester is normalized using mean normalization based on the range. After normalization, level A... Calculate the time decay factor for each data item separately. ,in , The unit is monthly. Weighting coefficients are used. , , ,according to The usability value score for each data item is calculated, as shown in Table 2 below.
[0084] Table 2. Calculation of Usability Value Score for Each Data Item in Task 1 (Table 2)
[0085]
[0086] Each data item and Substituting into the multi-objective optimization objective function, To constrain this, a weighted approach is taken to account for privacy risk penalties. Solve for the basic privacy budget for each data item. Data item 1 and data item 3 High but It is also high, belonging to the case where value and sensitivity are in the same direction, and the two are balanced through multi-objective optimization; data item 4 higher and Low, allocated a larger budget; data item 2 is centered, as shown in Table 3 below.
[0087] Table 3: Allocation Results of Basic Privacy Budget for Each Data Item
[0088]
[0089] The total basic budget requirement for the three concurrent tasks for the above four data items is: ,Exceed This triggers cross-task elastic allocation.
[0090] Retrieve the attributes of three concurrent tasks, including their urgency. Task value Noise tolerance Each method employs mean normalization based on the range, and is then uniformly scaled to... Range. The actual allocation budget for each data item is solved using an online convex optimization algorithm. The lower limit for compression of each task is constrained to be... As shown in Tables 4 and 5 below.
[0091] Table 4. Normalization Results of Concurrent Task Attributes
[0092]
[0093] Table 5: Actual Allocation Budget for Each Data Item After Cross-Task Flexible Allocation
[0094]
[0095] Task 1, due to its highest urgency and value, received a budget allocation close to the base amount, with minimal reduction. Task 2 had the highest noise tolerance and withstood the largest reduction. Task 3 received a moderate reduction. The sum of the actual budgets for the six tasks was... This just meets the total budget constraint.
[0096] Based on each data item Based on data type, differential noise processing was performed on the four data categories. Medical record records (data item 1) are structured numerical data, and the Laplace algorithm was used with a noise scale of [value missing]. Diagnostic text (data item 2) is categorical data, and the output category is selected according to the utility function score probability using an exponential mechanism; medical images (data item 3) are high-dimensional data, and the gradient is clipped and noise is added using differential privacy stochastic gradient descent; vital signs time series (data item 4) uses a sliding window method to add noise window by window, while preserving the temporal correlation between adjacent windows, as shown in Table 6 below.
[0097] Table 6: Noise Addition Methods and Budget Consumption Records for Each Data Item
[0098]
[0099] The data with added noise is sent to each task requester through a secure channel, with the privacy budget consumption of each data item included in the response.
[0100] After the three tasks are completed, the platform collects feedback data from both the requester and privacy perspectives. For Task 1, the medical records and imaging data both scored above the preset threshold under a higher budget, so the strategy parameters remained unchanged. For Task 2, the vital sign time-series data scored below the preset threshold under a compressed budget, triggering an upward adjustment of the baseline budget. For Task 2, the diagnostic text still scored above the threshold under a compressed budget, triggering a downward adjustment of the baseline budget. These samples are stored in the feedback database, and the baseline budget parameters for subsequent tasks are updated periodically after statistical analysis, as shown in Table 7 below.
[0101] Table 7 Summary of Feedback Data and Strategy Update Results
[0102]
[0103] The basic privacy budget corresponding to data item 4 that was compressed in Task 2 is restored to its initial level through budget replenishment when the task queue is idle, so as to avoid the data item being in a low-precision state for a long time in subsequent tasks.
[0104] Throughout the implementation process, data originates from four types of heterogeneous raw data and undergoes triple registration to obtain unified sensitivity labeling. Combined with the context information from Task 1, the scenario weights, normalized confidence levels, and time decay factors of each data item are aggregated into an usability value score. This value score, along with the sensitivity coefficient, drives multi-objective optimization, generating a basic privacy budget that reflects the principle of "high value, low sensitivity, and high allocation." Faced with budget competition among three concurrent tasks, online convex optimization flexibly allocates resources across tasks based on their urgency, value, and noise tolerance. High-priority tasks receive sufficient budget, while high-tolerance tasks bear the main burden of compression. The noise-adding stage selects the appropriate mechanism based on data type, transforming budget differences into accuracy differences. The feedback stage feeds back the actual usability score to the parameter library, driving the baseline budget to converge towards the accuracy level truly required for each type of data. The output data of each step serves as direct input for the next step, maintaining complete logical continuity in the data flow within the same medical research sharing scenario.
[0105] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
[0106] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A heterogeneous data secure sharing method based on differential privacy budget adaptive approach, characterized in that, include: To create a unified model for shared heterogeneous data, each data item is represented as a triplet structure containing a data entity, attribute, and sensitivity level. Receive a data sharing request, obtain the usage scenario identifier, requester trust level and data timestamp carried in the request, encode and convert the usage scenario identifier to obtain a scenario weight value, normalize the requester trust level, and perform weighted summation based on the processed scenario weight value, the normalized requester trust level value and the data time decay factor to calculate the usability value score of each data item. Combining the usability value score and the sensitivity level, with the dual objectives of maximizing data usability and minimizing the risk of privacy leakage, and with the upper limit of the total privacy budget of the system as a constraint, the basic privacy budget amount of each data item is solved through multi-objective optimization. Obtain the concurrent task queue. When the total privacy budget required by each task exceeds the system's total privacy budget limit, based on the urgency, value, and noise tolerance of each task, perform cross-task elastic allocation of the basic privacy budget amount through online convex optimization to obtain the actual allocated budget for each data item. Based on the actual allocated budget and the data type of the data items, perform differentiated noise addition processing on each data item to generate noisy data and respond to the sharing request.
2. The heterogeneous data secure sharing method based on differential privacy budget adaptive method according to claim 1, characterized in that, The usability value score is calculated as follows: ; in, For the first Usability value score for each data item For the first The usage scenario weight value corresponding to each data item For the first The normalized requester trust level value corresponding to each data item For the first Data time decay factor for each data item , , The preset weighting coefficients satisfy the following conditions: .
3. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, The data time decay factor is calculated as follows: ; in, Represents an exponential function. The decay rate constant is For the first The difference between the generation time of each data item and the current time. The range of values is .
4. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, The basic privacy budget for each data item is determined through multi-objective optimization, with the objective function being: ; ; in, The total number of data items. For the first Basic privacy budget for each data item This is the upper limit of the system's total privacy budget. For the first Sensitivity coefficient for each data item Weighting for privacy risk penalties.
5. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, Sensitivity levels are divided into multiple discrete levels, each level corresponding to a sensitivity coefficient. Among them, personal identification information corresponds to the highest sensitivity level, behavioral trajectory data corresponds to the medium sensitivity level, and aggregated statistical data corresponds to the low sensitivity level. The larger the value, the greater the privacy leakage risk penalty term in the objective function for the corresponding data item.
6. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, The objective function for elastically allocating the basic privacy budget across tasks through online convex optimization is: ; ; in, The total number of concurrent tasks. For the first Normalized urgency of each task For the first The normalized task value of each task. For the first Normalized noise tolerance for each task, For the first The set of data item numbers contained in each task For set Data item number in For the first The actual allocation budget for each data item For the first The basic privacy budget for each data item.
7. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 6, characterized in that, When the task queue is idle, the released budget will be replenished to the data items corresponding to the previously compressed tasks according to the historical compression ratio, so that the basic privacy budget of each data item is restored to the initial level.
8. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, Differential noise addition processing is performed on each data item, including: For numerical data, Laplace noise is applied, with a noise scale of [value missing]. ,in To ensure global sensitivity of the query function; for categorical data, exponential noise is used, and the output category is selected probabilistically based on the utility function score; for image data and high-dimensional data, differential privacy stochastic gradient descent is used, which prunes the gradient and adds noise when updating the model parameters; for time series data, a sliding window method is used to add noise window by window, which satisfies differential privacy requirements within the window while preserving the time series correlation between adjacent windows.
9. The heterogeneous data secure sharing method based on differential privacy budget adaptation according to claim 1, characterized in that, It also includes the policy update step: After each data sharing task is completed, the data availability score from the requester and the privacy leak event records detected by the privacy audit module are collected. The data characteristics, actual budget allocation, noise addition method, data availability score, and privacy leak event records of each data item are stored as samples in the feedback database. The samples in the feedback database are statistically analyzed periodically. If the data availability score of a certain type of data is higher than the preset threshold under a low budget allocation, the base budget for that type of data in subsequent tasks is reduced. If the data availability score of a certain type of data is still lower than the preset threshold under a high budget allocation, its base budget is increased or its noise addition method is switched to an alternative method with less loss of accuracy for that data type.
