A method and system for dynamic allocation of edge device privacy budget for chronic disease monitoring

Abstract

The application relates to an edge device privacy budget dynamic allocation method and system for chronic disease monitoring, which comprises the following steps: S1, collecting user physiological data flow and device resource state of an edge device, extracting device resource features and user physiological features, and setting a global security event flag; S2, constructing a privacy protection strength function, a data utility function and a resource efficiency function; S3, firstly judging the value of the global security event flag, if the value is true, outputting a final privacy budget as a preset maximum safety value; if the value is false, executing a decision target optimization process to obtain a candidate solution set of the privacy budget; S4, performing Pareto front search fast non-dominated sorting on the candidate solution set to output a first Pareto front; and S5, based on the first Pareto front, adopting a preset multi-factor priority strategy to output a final privacy budget. Compared with the traditional technology, the application can solve the problem that a privacy budget allocation strategy cannot systematically balance multi-target conflicts.

Description

Technical Field

[0001] This invention relates to the field of data privacy technology, and more specifically, to a method and system for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring. Background Technology

[0002] With the increasing prevalence of edge computing in chronic disease management, real-time physiological data collection presents significant privacy challenges. Differential privacy protects data by adding noise, and its privacy budget is a key parameter for balancing privacy and data utility. Existing technologies mostly employ fixed budgets or dynamic adjustment strategies based on weighted summation. These methods have the following limitations: 1. Privacy, utility, and device resource consumption are three conflicting objectives. Existing methods, using linear weighting or rule-based decision-making, cannot systematically explore the optimal balance among these three factors, often resulting in suboptimal solutions. 2. Weighting coefficients or rule thresholds heavily rely on domain expert experience for setting, lacking a theoretical basis for adaptation in complex dynamic scenarios, leading to weak model generalization ability. 3. A single output value cannot reflect all possible solutions in the current state, and the decision-making process lacks a global insight into the decision space.

[0003] Therefore, considering the above requirements and the shortcomings of existing privacy budget allocation strategies in systematically balancing conflicts among multiple objectives, this application proposes a method and system for dynamic allocation of privacy budget for edge devices used in chronic disease monitoring. Summary of the Invention

[0004] This invention provides a method and system for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring. By quantifying three core objectives—privacy protection strength, data utility, and device resource efficiency—a three-dimensional decision space is constructed on the edge device. A lightweight Pareto frontier search algorithm is then used to find a set of optimal privacy budget candidate solutions in real time, thereby achieving a Pareto optimal balance among privacy, utility, and resources in long-term chronic disease monitoring.

[0005] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows: The first aspect of this invention provides a method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring. This method includes the following steps: S1. Collect user physiological data stream and edge device resource status, extract device resource features and user physiological features through preprocessing, and set global security event flags based on the user physiological features.

[0006] S2. Maximizing privacy protection strength, data utility, and resource efficiency are taken as the decision objectives for privacy budget allocation, and privacy protection strength function, data utility function, and resource efficiency function are constructed respectively.

[0007] S3. First, determine the value of the global security event flag. If it is true, interrupt the decision target optimization process, set the privacy budget value to the preset maximum security value, and output the final privacy budget. If it is false, execute the decision target optimization process, input the device resource characteristics and user physiological characteristics into the privacy protection strength function, data utility function, and resource efficiency function to obtain the candidate solution set of the privacy budget.

[0008] S4. Perform Pareto Front Search and Fast Non-Dominated Sort on the candidate solution set to output the first Pareto Front. The first Pareto Front includes a vector consisting of all non-dominated privacy budget candidate solution sets and their corresponding objective function values.

[0009] S5. Based on the first Pareto front, determine the value of the privacy budget using a preset multi-factor priority strategy, and output the final privacy budget.

[0010] A second aspect of the present invention provides a dynamic allocation system for privacy budgets of edge devices for chronic disease monitoring. This system is used in the aforementioned method for dynamic allocation of privacy budgets of edge devices for chronic disease monitoring, and includes: a data acquisition module, a privacy budget calculation module, a privacy budget sorting module, and a privacy budget allocation module.

[0011] The data acquisition module collects user physiological data streams and edge device resource status, extracts device resource features and user physiological features through preprocessing, and sets a global security event flag based on the user physiological features. The privacy budget calculation module takes maximizing privacy protection strength, data utility, and resource efficiency as the decision objectives for privacy budget allocation, constructs a privacy protection strength function, a data utility function, and a resource efficiency function, respectively, and receives the device resource features, user physiological features, and global security event flags output by the data acquisition module. First, it judges the value of the global security event flag. If it is true, the decision objective optimization process is interrupted, the privacy budget value is set to the preset maximum security value, and the final privacy budget is output. If it is false, the decision objective optimization process is executed, and the device resource features and user physiological features are input into the privacy protection strength function, the data utility function, and the resource efficiency function to obtain a candidate solution set for the privacy budget. The privacy budget sorting module performs Pareto front search fast non-dominated sorting on the candidate solution set output by the privacy budget calculation module, and outputs a first Pareto front. The first Pareto front includes a vector composed of all non-dominated privacy budget candidate solution sets and their corresponding objective function values.

[0012] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention provides a method and system for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring. By modeling privacy budget allocation as a multi-objective optimization problem and solving the Pareto front in real time at the edge, it achieves a dynamic optimal balance between privacy protection strength, data utility, and device resource efficiency, effectively overcoming the limitations of traditional fixed or heuristic allocation strategies. While ensuring user privacy, it adaptively adjusts the privacy budget based on real-time device battery level, processor load, and user physiological risk scenarios. It prioritizes device battery life and stability when resources are scarce, and prioritizes data utility to support clinical intervention when high-risk physiological events are detected. Furthermore, through a cloud-edge co-evolutionary mechanism, it continuously optimizes the local overhead prediction model, enabling the entire system to continuously improve itself as usage data accumulates and the device environment changes. Ultimately, it achieves more accurate and context-aware dynamic privacy protection in resource-constrained edge computing environments. Attached Figure Description

[0013] Figure 1 This is a flowchart of a method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to the present invention; Figure 2 This is a flowchart of Pareto optimization in one embodiment of the present invention; Figure 3 This is a schematic diagram of Pareto frontier decision-making in one embodiment of the present invention; Figure 4 This is a diagram of an edge-cloud collaborative optimization system in one embodiment of the present invention; Figure 5 This is a schematic diagram of a dynamic allocation system for privacy budget of edge devices for chronic disease monitoring according to the present invention. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0016] Example 1 like Figure 1 As shown, this invention provides a method for dynamically allocating the privacy budget of edge devices for chronic disease monitoring. This method includes the following steps: S1. Collect user physiological data stream and edge device resource status, extract device resource features and user physiological features through preprocessing, and set global security event flags based on the user physiological features.

[0017] S2. Maximizing privacy protection strength, data utility, and resource efficiency are taken as the decision objectives for privacy budget allocation, and privacy protection strength function, data utility function, and resource efficiency function are constructed respectively.

[0018] S3. First, determine the value of the global security event flag. If it is true, interrupt the decision target optimization process, set the privacy budget value to the preset maximum security value, and output the final privacy budget. If it is false, execute the decision target optimization process, input the device resource characteristics and user physiological characteristics into the privacy protection strength function, data utility function, and resource efficiency function to obtain the candidate solution set of the privacy budget.

[0019] S4. Perform Pareto Front Search and Fast Non-Dominated Sort on the candidate solution set to output the first Pareto Front. The first Pareto Front includes a vector consisting of all non-dominated privacy budget candidate solution sets and their corresponding objective function values.

[0020] S5. Based on the first Pareto front, determine the value of the privacy budget using a preset multi-factor priority strategy, and output the final privacy budget.

[0021] Based on the aforementioned technical features, this invention models privacy budget decision-making as a multi-objective optimization problem. By quantifying three core objectives—privacy protection strength, data utility, and device resource efficiency—a three-dimensional decision space is constructed on edge devices. A lightweight Pareto front search algorithm is employed to find a set of optimal privacy budget candidate solutions in real time. The Pareto front of the problem is solved in real time at the edge to guide the decision-making process. Finally, by combining a security arbitration mechanism and user context, the final budget value is selected from the set of optimal solutions, thereby achieving a Pareto optimal balance among privacy, utility, and resources in long-term chronic disease monitoring.

[0022] like Figure 1 As shown, step S1 performs multi-dimensional state perception. The process of collecting user physiological data stream includes: continuously collecting user physiological data stream using sensors built into the edge device, and extracting user physiological features from the physiological data stream using a sliding window approach. The user's physiological characteristics Input into the pre-trained risk assessment model Model_Risk, and output a risk score. Specifically:

[0023] in For the sigmoid function, and All model parameters are obtained by training the pre-trained risk assessment model Model_Risk using historical data, and are based on user physiological characteristics. When the value exceeds the preset threshold range, the global security event flag Safety_Event_Flag will be set to true; The process of collecting device resource status data for edge devices includes: real-time collection of the edge device's battery level (BL) and the processor's average load (CL) over the past minute via the operating system API; for the battery level (BL), a piecewise linear normalization method is used to calculate its normalized value, which is then mapped to... Interval:

[0024] in, and As configurable parameters set based on device hardware characteristics, in this embodiment, based on the low power protection threshold of common mobile devices, the system may enter power-saving mode when the battery level is below 20%. Therefore, Set to 20, Set to 100% for full charge; for average load CL, the value itself has a range of values. The interval is given, therefore its normalized value is:

[0025] Final output device resource characteristics .

[0026] It should be noted that the risk assessment model Model_Risk is trained using historical data, and the specific process is as follows: (1) Data Collection and Labeling: Public datasets or anonymized data from partner hospitals were used. The data format was time series, including timestamps and physiological values. In this embodiment, blood glucose values ​​in mmol / L were used as the application object. Labeling was based on clinical guidelines. For each time window, if the blood glucose value was continuously below 3.9 mmol / L for more than 5 minutes, it was marked as a hypoglycemic event, and the label y was set to 1; if the blood glucose value was continuously above 11.1 mmol / L for more than 10 minutes, it was marked as a hyperglycemic event, and the label y was set to 2; otherwise, it was marked as a normal event, and the label y was set to 0.

[0027] The data preprocessing process includes using linear interpolation to impute missing values ​​and using a moving average filter to remove noise.

[0028] (2) Feature extraction: User physiological features are extracted from the raw physiological data stream using a sliding window. The sliding window size is set to the most recent 30 readings, i.e., 2.5 hours of data, specifically including: The arithmetic mean of the data within the sliding window is calculated using the following formula:

[0029] Then calculate the variance of the data within the sliding window, using the following formula:

[0030] The slope of the data change over time within the sliding window is fitted using linear regression, and the calculation formula is as follows:

[0031] in For time indexing, .

[0032] Outliers based on the Z-score are calculated using local statistics within the same sliding window. For each reading within the window... The following abnormal detection indicators exist:

[0033] in and This represents the mean and standard deviation of all readings within this window. If If a value is found to be an outlier, it is marked as such. The characteristic value obtained is the proportion of outliers within the window, or the number of outliers divided by the window size.

[0034] Based on the above anomaly detection metrics, including the maximum length of a continuous outlier sequence and the exponentially weighted moving average (EWMA) residual, the maximum length of the outlier sequence is set to 3 in this embodiment, representing three consecutive outlier sequences. Indicates a potential event.

[0035] In EWMA residual calculation, the initial value of EWMA... Set as the first reading in the window For subsequent readings The residual is:

[0036]

[0037] in , representing a commonly used smoothing factor in time series analysis. The eigenvalue is the average of the residuals within the window.

[0038] User physiological characteristics The final structure is a five-dimensional vector: [mean] ,variance trend slope Outlier ratio, EWMA residuals (mean).

[0039] (3) Model training: Using a logistic regression model, the loss function is binary cross-entropy:

[0040] Where M is the number of training samples. The optimization algorithm uses stochastic gradient descent (SGD) with a learning rate of 0.01, which is determined in the range [0.001, 0.1] through grid search. The algorithm iterates for 1000 rounds with a batch size of 32.

[0041] The dataset was split into training and test sets in a 70:30 ratio, and parameters were tuned using 5-fold cross-validation. After training on a specific dataset, typical parameter values ​​were obtained: w = [0.12, 0.28, 0.38, 0.15, 0.07], b = -0.48. These specific parameters vary depending on the training dataset, but the feature weights remain relative to the user's physiological characteristics. The order must remain consistent.

[0042] The trained parameters w and b are embedded into the edge device firmware and updated monthly or quarterly via the cloud.

[0043] A parallel, high-frequency monitoring process is responsible for detecting safety events such as persistent hypoglycemia and setting a global safety event flag, Safety_Event_Flag. Safety event detection is based on threshold rules. In this embodiment, if the blood glucose level remains below 3.9 mmol / L for more than 5 minutes, Safety_Event_Flag is set to true. The threshold can be remotely configured and adjusted via the cloud.

[0044] In step S2, the privacy protection strength function is specifically as follows:

[0045] in A larger value indicates a higher level of privacy protection, and the optimization direction of the privacy protection strength function is to maximize... Value, i.e., finding a smaller value. Value; the data utility function is specifically:

[0046] Data utility The higher the value, the less the added noise degrades the data quality, and the stronger the data usability. The query function sensitivity is represented by the width of the data range. During the calculation and sorting process of the same physiological data, the query function sensitivity is the same relative to all candidate solutions. In this embodiment, the width of the blood glucose meter's data range... In subsequent Pareto sorting, due to the constant factor The fact that all candidate solutions are identical does not affect the relative ranking of solutions; therefore, in practical calculations, it can be simplified to... To reduce computational overhead, but it is necessary to ensure that within the same application scenario ( (Fixed) Use this simplification.

[0047] The optimization direction of the data utility function is to maximize data utility. Data utility measures the ability of data, after differential privacy protection, to retain the true information of the original data and support correct analysis in subsequent data analysis, model training, or clinical decision-making tasks.

[0048] It should be noted that in the differential privacy framework, a query function refers to an information extraction operation performed on the dataset. In the chronic disease monitoring scenario of this invention, a typical query object is the publication of a single user's current instantaneous physiological data point. Therefore, the query function f can be defined as f(dataset) = current blood glucose value. Its sensitivity... Defined as: for any two adjacent datasets that differ by only one user and one record, the maximum possible change in the result of this query function. For single-point numerical publications, Typically, the width of the entire possible range of values ​​for this physiological parameter is taken. In this embodiment, the blood glucose range is assumed to be 0-20 mmol / L. =20.0 mmol / L Therefore, constant factor This is fixed for specific types of physiological data. In actual calculations, since we only need to compare different... The relative magnitude of utility under different values, and within the same application (e.g., all publishing blood glucose values). It is constant, therefore it can be simplified to To reduce the computing overhead of edge devices.

[0049] The resource efficiency function is specifically as follows:

[0050] in To account for the additional system overhead associated with privacy protection, the additional system overhead is calculated using the following linear regression model:

[0051] in , , , , All of these are model parameters obtained through pre-training. This characterizes the positive impact of processor load on overhead; the higher the load, the greater the base overhead. Used to characterize the amplification effect of low power state on overhead. The lower the power, the more likely the system is to be in energy-saving or frequency-reducing state, and the actual time or energy consumption to complete the same calculation may increase. Used to characterize the increase under high load conditions. The additional overhead amplification effect of the value means that noise with smaller variance needs to be generated, which may involve more complex random number calculations.

[0052] The input for the additional system overhead is the privacy budget candidate value. and device resource feature vector The output is a scalar parameter. The optimization direction of the resource efficiency function is to maximize resource efficiency. That is, minimizing additional system overhead. .

[0053] Resource efficiency measures how efficiently limited computing resources are utilized on edge devices when performing privacy-preserving operations. Higher efficiency indicates less additional system overhead incurred in achieving privacy protection.

[0054] In this embodiment, the process of training the cost prediction model is as follows: In the experimental environment, edge devices were simulated to perform privacy-preserving operations and collect data, including feature variables and label variables. The feature variables included: Value and equipment resource characteristics ,in Value at A uniform sampling of 50 points was performed within the range, based on consensus from differential privacy literature and preliminary experiments. The data is too inefficient. The privacy protection is too weak. The labeled variables include actual resource consumption, CPU time in milliseconds, measured using the Linux system tool perf.

[0055] The collected dataset includes 5,000 samples, covering different combinations of ε and resource states.

[0056] Construct the characteristic matrix X, which includes a bias term consisting of all 1 columns and four characteristic terms: , , , All features are Z-score normalized before training to avoid dimensional differences.

[0057] The model training process specifically involves minimizing the mean squared error (MSE) using multiple linear regression.

[0058] Then, the parameter estimates are solved:

[0059] Where X is the normalized feature matrix and y is the label variable.

[0060] Typical parameters obtained from training are: , , , , In actual deployment, model parameters to The actual hardware performance of the target edge device needs to be re-determined through offline training.

[0061] Finally, calculations were performed on the test set. To ensure prediction accuracy, model validation is completed. Model parameters are verified through methods such as... Figure 4 The cloud-based collaborative evolution mechanism shown has been updated.

[0062] like Figure 3 As shown, in step S3, the first step is to determine the value of the global security event flag. If it is true, the decision target optimization process is interrupted, the privacy budget value is set to the preset maximum security value, and the final privacy budget is output. Specifically, before executing the optimization process, the value of the global security event flag Safety_Event_Flag is first determined to be true, i.e., the extracted user physiological characteristics are considered true. If the value exceeds a preset threshold, output the final privacy budget. And end the dynamic allocation process of the privacy budget:

[0063] in, This represents the maximum privacy budget security value preset while ensuring the complete transmission of information. In this embodiment, In the event of a security incident, priority should be given to ensuring the utility of data to support medical intervention, while Version 5.0 still provides basic privacy protection.

[0064] When the global security event flag is false, such as Figure 3 The execution of the decision objective optimization process shown to obtain the candidate solution set S for privacy budget includes: Privacy Budget Feasible range Discretize into The candidate values ​​constitute the candidate solution set. In this embodiment That is, the step size = (5.0-0.1) / 20 = 0.245.

[0065] for Each candidate solution , Substitute them into the privacy protection strength function respectively. Data utility function and resource efficiency function The corresponding three objective function values ​​were calculated. Each candidate solution and its three corresponding objective function values To form a complete solution.

[0066] In step S4, Pareto front search and fast non-dominated sorting are performed on the candidate solution set to output the first Pareto front F1, specifically including: First, a first Pareto front F1 is constructed and initialized as an empty set. Then, the three objective function values ​​calculated for all solutions in the candidate solution set S are normalized using Min-Max to obtain the normalized objective values. , and Then, traverse the candidate solution set S, for each solution... Check if there is another solution. This makes it possible to achieve the following in the normalized target space: , and At least one inequality holds; if no such inequality exists... Then determine It is a non-dominated solution, and will the Joining the first Pareto frontier F1.

[0067] The Min-Max normalization process is as follows:

[0068]

[0069]

[0070] in, , , , , , These are the minimum and maximum values ​​of the objective function in the current candidate solution set S.

[0071] like Figure 3 As shown, step S5 specifically involves: if the global security event flag is true, then security events are given the highest priority, and the final privacy budget is calculated accordingly. Otherwise, prioritize high-risk scenarios and calculate the user's physiological risk factors. ,like Risk exceeding the preset threshold Then select the data utility function from F1. The candidate solution with the largest value is used as the final privacy budget. Among them, the preset risk threshold .

[0072] like The preset risk threshold is not met. Then focus on resource-scarce scenarios, check equipment resource status, and if the power normalization value... Less than or equal to the preset power threshold Or processor load normalization value Greater than the preset load threshold Then, select the resource efficiency function from F1. The candidate solution with the largest value is used as the final privacy budget. Among them, the preset power threshold Preset load threshold When the battery level is below 20% or the load exceeds 80%, the system performance and stability are significantly reduced, posing a high risk.

[0073] If the user has provided custom privacy settings If the above conditions are not met, then the user's needs will be the primary consideration, and the option selected from F1 will be the one that best meets those needs. The candidate solution with the smallest absolute difference, i.e. .

[0074] If none of the above priority strategies are met, the default settings will be executed, and the function that increases the privacy protection strength will be selected from F1. The candidate solution with the largest value is used as the final privacy budget. ,Right now .

[0075] Output the final privacy budget in step S5 This also includes the application of a privacy budget, specifically: based on the determined final privacy budget. Values ​​that add noise to the collected user physiological data stream; the noise originates from the scale parameter. Laplace distribution The random sampling was obtained, among which To determine the sensitivity of the query function, the generated noise is added to the original data to obtain privacy-preserving data.

[0076] like Figure 4 As shown, after the privacy budget application, it also includes cloud-edge collaborative evolution, specifically, after obtaining authorization, generating data including timestamps and device resource characteristics. User's physiological risk score First Pareto Frontier F1 and Ultimate Privacy Budget Anonymous decision summaries are generated and uploaded to the cloud. In this embodiment, the decision summaries include feature variables and label variables. The label variables include total operation time and energy consumption estimate. The total operation time is obtained by sampling before and after noisy operations using the Linux system call clock_gettime(). The energy consumption estimate is a relative energy consumption score calculated based on CPU time and load = CPU time × (1 + ... ).

[0077] After the cloud server aggregates the summary data, feature variables are extracted. The linear regression model, which calculates additional system overhead, is retrained with respect to the label variables. Its purpose is to serve as a more powerful and accurate mentor model for analyzing privacy budget ε and device resource status on real-world, complex, large-scale data. The potential non-linear relationship between the model and actual resource overhead. It is not directly deployed to resource-constrained edge devices. Using aggregated anonymized data, an XGBoost regression model is trained to achieve higher prediction accuracy than the simple linear models currently used on edge devices. The specific hyperparameters of the XGBoost model are set as follows: maximum depth (max_depth) = 6, learning rate (learning_rate) = 0.1, number of iterations (n_estimators) = 100, and the remaining parameters use default values.

[0078] If the predictive performance improvement of the linear regression model located in the cloud compared to the linear regression model currently used by the edge device exceeds the expected threshold, then feature combinations are extracted from the linear regression model located in the cloud, and these are used to reconstruct the linear regression model for the edge device to account for the additional computational overhead. The parameters. In this embodiment, the desired threshold is 0.1. Finally, the reconstructed new model parameters are sent to the edge devices for them to update in the next computing cycle.

[0079] Example 2 Based on the above embodiment 1, this embodiment takes continuous glucose monitoring (CGM) for diabetic patients as an example.

[0080] like Figure 2 As shown, firstly, multi-dimensional state perception is performed to obtain device resource characteristics. The surface charge is low and the load is high.

[0081] Obtaining user physiological characteristics And calculate risk score The surface indicates low risk, so Safety_Event_Flag = False.

[0082] User settings: .

[0083] Perform Pareto front search, discretize There are 21 candidate values, ranging from 0.1 to 5.0, with a step size of 0.245.

[0084] Calculate each objective function value And normalized, the calculation in this embodiment , , They were obtained respectively. , and ; , and ; , and After performing a non-dominated sort, the result is... .

[0085] Finally, the final decision was made, but due to limited equipment resources... And does not meet the high-risk scenario. The system triggers a resource-saving priority strategy.

[0086] exist In the middle, calculate their respective R values, assuming When R reaches its maximum value of 0.35, the final output will be... .

[0087] Example 3 like Figure 5 As shown, the present invention also provides a dynamic allocation system for privacy budget of edge devices for chronic disease monitoring. The system is used in the aforementioned method for dynamic allocation of privacy budget of edge devices for chronic disease monitoring, and includes: a data acquisition module, a privacy budget calculation module, a privacy budget sorting module, and a privacy budget allocation module.

[0088] The data acquisition module collects user physiological data streams and edge device resource status, extracts device resource features and user physiological features through preprocessing, and sets a global security event flag based on the user physiological features. The privacy budget calculation module takes maximizing privacy protection strength, data utility, and resource efficiency as the decision objectives for privacy budget allocation, constructs a privacy protection strength function, a data utility function, and a resource efficiency function, respectively, and receives the device resource features, user physiological features, and global security event flags output by the data acquisition module. First, it judges the value of the global security event flag. If it is true, the decision objective optimization process is interrupted, the privacy budget value is set to the preset maximum security value, and the final privacy budget is output. If it is false, the decision objective optimization process is executed, and the device resource features and user physiological features are input into the privacy protection strength function, the data utility function, and the resource efficiency function to obtain a candidate solution set for the privacy budget. The privacy budget sorting module performs Pareto front search fast non-dominated sorting on the candidate solution set output by the privacy budget calculation module, and outputs a first Pareto front. The first Pareto front includes a vector composed of all non-dominated privacy budget candidate solution sets and their corresponding objective function values.

[0089] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] Alternatively, if the above embodiments of the present invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0091] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. The icons depicting structural positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring, characterized in that, Includes the following steps: S1. Collect user physiological data stream and edge device resource status, extract device resource features and user physiological features through preprocessing, and set global security event flags based on the user physiological features; S2. Taking the maximization of privacy protection strength, data utility, and resource efficiency as the decision objectives for privacy budget allocation, construct privacy protection strength function, data utility function, and resource efficiency function respectively; S3. First, determine the value of the global security event flag. If it is true, interrupt the decision target optimization process, set the privacy budget value to the preset maximum security value, and output the final privacy budget. If it is false, execute the decision target optimization process, input the device resource characteristics and user physiological characteristics into the privacy protection strength function, data utility function, and resource efficiency function to obtain the candidate solution set of the privacy budget. S4. Perform Pareto front search and fast non-dominated sorting on the candidate solution set to output the first Pareto front. The first Pareto front includes a vector consisting of all non-dominated privacy budget candidate solution sets and their corresponding objective function values. S5. Based on the first Pareto front, determine the value of the privacy budget using a preset multi-factor priority strategy, and output the final privacy budget.

2. The method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 1, characterized in that, In step S1, the process of collecting user physiological data stream includes: continuously collecting user physiological data stream using sensors built into the edge device, and extracting user physiological features from the physiological data stream using a sliding window approach. The user's physiological characteristics Input into the pre-trained risk assessment model Model_Risk, and output a risk score. Specifically: in For the sigmoid function, and All model parameters are obtained by training the pre-trained risk assessment model Model_Risk using historical data, and are based on user physiological characteristics. When the value exceeds the preset threshold range, the global security event flag Safety_Event_Flag will be set to true; The process of collecting device resource status data for edge devices includes: real-time collection of the edge device's battery level (BL) and the processor's average load (CL) over the past minute via the operating system API; for the battery level (BL), a piecewise linear normalization method is used to calculate its normalized value, which is then mapped to... Interval: in, and These are configurable parameters set based on the device's hardware characteristics; for the average load CL, since the value itself has a range of values... The interval is given, therefore its normalized value is: Final output device resource characteristics .

3. The method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 2, characterized in that, In step S2, the privacy protection strength function is specifically as follows: in A larger value indicates a higher level of privacy protection, and the optimization direction of the privacy protection strength function is to maximize... Value, i.e., finding a smaller value. Value; the data utility function is specifically: Data utility The higher the value, the less the added noise degrades the data quality, and the stronger the data usability. The query function sensitivity is represented by the width of the data range. During the calculation and sorting process of the same physiological data, the query function sensitivity is the same relative to all candidate solutions. The optimization direction of the data utility function is to maximize data utility. The resource efficiency function is specifically as follows: in To account for the additional system overhead associated with privacy protection, the additional system overhead is calculated using the following linear regression model: in , , , , All parameters are model parameters obtained through pre-training, and the input to the additional system overhead is the privacy budget candidate value. and device resource feature vector The output is a scalar parameter. The optimization direction of the resource efficiency function is to maximize resource efficiency. That is, minimizing additional system overhead. .

4. The method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 1, characterized in that, In step S3, the first step is to determine the value of the global security event flag. If the value is true, the decision target optimization process is interrupted, the privacy budget value is set to the preset maximum security value, and the final privacy budget is output. Specifically, before executing the optimization process, the value of the global security event flag Safety_Event_Flag is first determined to be true, i.e., the extracted user physiological characteristics. If the value exceeds a preset threshold, output the final privacy budget. And end the dynamic allocation process of the privacy budget: in, This represents the maximum privacy budget security value preset under the condition of ensuring the complete transmission of information.

5. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 4, characterized in that, In step S3, when the global security event flag is false, the decision objective optimization process is executed to obtain a candidate solution set S for the privacy budget, specifically including: Privacy Budget Feasible range Discretize into The candidate values ​​constitute the candidate solution set. ;for Each candidate solution , Substitute them into the privacy protection strength function respectively. Data utility function and resource efficiency function The corresponding three objective function values ​​were calculated. Each candidate solution and its three corresponding objective function values To form a complete solution.

6. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 5, characterized in that, In step S4, Pareto front search and fast non-dominated sorting are performed on the candidate solution set to output the first Pareto front F1, specifically including: First, a first Pareto front F1 is constructed and initialized as an empty set. Then, the three objective function values ​​calculated for all solutions in the candidate solution set S are normalized using Min-Max to obtain the normalized objective values. , and Then, traverse the candidate solution set S, for each solution... Check if there is another solution. This makes it possible for the target space to be normalized. , and At least one inequality holds; if no such inequality exists... Then determine It is a non-dominated solution, and will the Joining the first Pareto frontier in F1; The Min-Max normalization process is as follows: in, , , , , , These are the minimum and maximum values ​​of the objective function in the current candidate solution set S.

7. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 6, characterized in that, Step S5 specifically involves: if the global security event flag is true, then the final privacy budget... ; Otherwise, calculate the user's physiological risk factors. ,like Risk exceeding the preset threshold Then select the data utility function from F1. The candidate solution with the largest value is used as the final privacy budget. ;like The preset risk threshold is not met. Then check the device resource status; if the power normalization value is... Less than or equal to the preset power threshold Or processor load normalization value Greater than the preset load threshold Then, select the resource efficiency function from F1. The candidate solution with the largest value is used as the final privacy budget. If the user provides custom privacy settings If the above conditions are not met, then select from F1 the option that matches... The candidate solution with the smallest absolute difference, i.e. If none of the above priority strategies are met, then select the privacy protection strength function from F1. The candidate solution with the largest value is used as the final privacy budget. ,Right now .

8. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 7, characterized in that, Output the final privacy budget in step S5 This also includes the application of a privacy budget, specifically: based on the determined final privacy budget. This value adds noise to the collected user physiological data stream; The noise comes from the scale parameter Laplace distribution The random sampling was obtained, among which To determine the sensitivity of the query function, the generated noise is added to the original data to obtain privacy-preserving data.

9. A method for dynamically allocating privacy budgets for edge devices used in chronic disease monitoring according to claim 8, characterized in that, Following the application of privacy budget, it also includes cloud-edge collaborative evolution, specifically, after obtaining authorization, generating data including timestamps and device resource characteristics. User's physiological risk score First Pareto Frontier F1 and Ultimate Privacy Budget Anonymous decision summaries are generated and uploaded to the cloud; After the cloud server aggregates the summary data, feature variables are extracted. The linear regression model for calculating additional system overhead is retrained using the label variables. If the predictive performance of the cloud-based linear regression model is improved by a greater than expected threshold compared to the linear regression model currently used on the edge device, feature combinations are extracted from the cloud-based linear regression model and used to reconstruct the linear regression model for calculating additional system overhead on the edge device. The parameters are then determined; finally, the reconstructed new model parameters are sent to the edge devices for use in the next computing cycle.

10. A dynamic privacy budget allocation system for edge devices used in chronic disease monitoring, the system being used in the dynamic privacy budget allocation method for edge devices used in chronic disease monitoring as described in any one of claims 1-9, characterized in that, It includes: a data collection module, a privacy budget calculation module, a privacy budget sorting module, and a privacy budget allocation module; The data acquisition module collects user physiological data streams and edge device resource status, extracts device resource features and user physiological features through preprocessing, and sets a global security event flag based on the user physiological features. The privacy budget calculation module takes maximizing privacy protection strength, data utility, and resource efficiency as the decision objectives for privacy budget allocation, constructs a privacy protection strength function, a data utility function, and a resource efficiency function, respectively, and receives the device resource features, user physiological features, and global security event flags output by the data acquisition module. First, it judges the value of the global security event flag. If it is true, the decision objective optimization process is interrupted, the privacy budget value is set to the preset maximum security value, and the final privacy budget is output. If it is false, the decision objective optimization process is executed, and the device resource features and user physiological features are input into the privacy protection strength function, the data utility function, and the resource efficiency function to obtain a candidate solution set for the privacy budget. The privacy budget sorting module performs Pareto front search fast non-dominated sorting on the candidate solution set output by the privacy budget calculation module, and outputs a first Pareto front. The first Pareto front includes a vector composed of all non-dominated privacy budget candidate solution sets and their corresponding objective function values.

Images