An Adaptive Feature Preservation Analysis Method for Sensitive Information

Abstract

This invention relates to the field of information protection technology, specifically to an adaptive feature retention analysis method for sensitive information, comprising: constructing a privacy risk-information utility dual evaluation model including a privacy risk analysis unit and an information utility evaluation unit to evaluate sensitive information and generate a dual evaluation result matrix; based on the dual evaluation result matrix, analyzing the interaction relationship of different feature retention strategies when used in combination through a strategy compatibility analysis model to obtain a strategy compatibility matrix; according to the dual evaluation result matrix and the strategy compatibility matrix, generating multiple candidate feature retention strategy combination schemes through a hierarchical combination search model and performing conflict detection to generate a conflict-free optimal strategy combination; evaluating the impact of the optimal strategy combination on information availability and privacy protection through a performance-privacy trade-off mechanism, determining the optimal feature retention scheme within a preset time, retaining features of the sensitive information, and obtaining feature-retained information.

Description

Technical Field

[0001] This invention relates to the field of information protection technology, specifically to an adaptive feature retention analysis method for sensitive information. Background Technology

[0002] With the rapid development of information protection technologies and the widespread application of data-driven decision-making, the contradiction between protecting sensitive information and mining data value is becoming increasingly prominent. Traditional information protection technologies mainly include methods such as data anonymization, differential privacy, homomorphic encryption, and secure multi-party computation. However, these technologies face the problem of balancing protection effectiveness and data usability in practical applications.

[0003] Existing information protection technologies suffer from three main shortcomings: First, they lack a comprehensive evaluation mechanism for feature sensitivity and analytical importance. Traditional methods typically employ a uniform transformation strategy to process all features, failing to differentiate them based on their sensitivity and analytical importance. Second, privacy budget allocation strategies are overly simplistic. Existing differential privacy technologies often use static budget allocation methods, such as uniform allocation or allocation based on simple heuristic rules, ignoring the differences in the contribution of different features to the analytical task. Third, they lack adaptive adjustment capabilities. Existing technologies often use static sensitivity definitions based on expert knowledge, failing to dynamically optimize based on data context, application scenarios, and time-varying characteristics.

[0004] Therefore, how to assess the sensitivity and analytical importance of data features in real time, adaptively select the optimal feature transformation strategy based on the dual assessment results, and dynamically allocate protection resources under limited privacy budget constraints to maximize overall data availability has become a core challenge facing the field of big data analytics.

[0005] To address this, an adaptive feature retention analysis method for sensitive information is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive feature preservation analysis method for sensitive information. By evaluating the sensitivity and analytical importance of data features in real time, the optimal feature transformation strategy is adaptively selected based on the dual evaluation results, and protection resources are dynamically allocated under the constraint of a limited privacy budget to maximize the overall data availability.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An adaptive feature-preserving analysis method for sensitive information includes: A dual assessment model of privacy risk and information utility is constructed to evaluate sensitive information and generate a dual assessment result matrix. The dual assessment model of privacy risk and information utility includes: a privacy risk analysis unit, which performs privacy risk analysis in parallel at different granularities; and an information utility assessment unit, which evaluates information utility through different combinations of multiple assessment modes. Based on the privacy risks and information utility of each sensitive information in the dual evaluation result matrix, the interaction relationship of different feature retention strategies when used in combination is analyzed through a strategy compatibility analysis model to obtain a strategy compatibility matrix. Based on the dual evaluation result matrix and the policy compatibility matrix, multiple candidate feature preservation policy combination schemes are generated through a hierarchical combination search model, and conflict detection is performed to generate the conflict-free optimal policy combination. The impact of the optimal strategy combination on information availability and privacy protection is evaluated through a performance-privacy trade-off mechanism. Combining the dual evaluation result matrix, the optimal feature retention scheme is determined within a preset time using heuristic search and local optimization, and the sensitive information is feature-retained to obtain feature-retained information.

[0008] Preferably, the privacy risk analysis unit includes: a privacy risk analysis layer, which performs privacy risk analysis at the record level, field level, value range level, and pattern level to obtain privacy risk analysis results; and a granularity switching layer, which adaptively switches the privacy risk analysis granularity according to preset granularity switching conditions. The information utility evaluation unit includes: an information utility evaluation layer, which evaluates information utility using statistical importance evaluation mode, functional importance evaluation mode, reliability importance evaluation mode, dynamic importance evaluation mode, and dependency importance evaluation mode respectively; and a scenario feature matching layer, which dynamically selects a combination of evaluation modes based on the current analysis scenario type and user needs to generate information utility evaluation results. The privacy risk-information utility dual assessment model also includes a cross-validation unit to verify the consistency between the privacy risk analysis results and the information utility assessment results.

[0009] Preferably, the preset granularity switching conditions include: automatically switching from field level to pattern level when the data association density exceeds the association density threshold; automatically switching from value range level to field level when the data missing rate is higher than the data missing threshold; and triggering multi-granularity parallel analysis when the data distribution skewness is abnormal.

[0010] Preferably, the strategy compatibility analysis model includes: The privacy risk-information utility mapping unit establishes the compatibility relationship between sensitive information and feature retention strategies based on the privacy risk and information utility information of each sensitive information in the dual evaluation result matrix; the strategy interaction relationship identification unit identifies the positive synergistic effect, negative conflict effect, and neutral independence effect of different feature retention strategies when used together; the compatibility score calculation unit obtains the compatibility score between feature retention strategies based on the overlap of strategy functions, the degree of parameter conflict, and the degree of resource competition; and the strategy compatibility matrix generation unit generates a strategy compatibility matrix describing the compatibility relationship between each feature retention strategy based on the strategy interaction identification results and the compatibility score.

[0011] Preferably, the hierarchical combined search model includes: The compatibility filtering layer filters out a set of mutually compatible candidate strategies based on the strategy compatibility matrix; the complementarity selection layer selects a combination of strategies with functional complementarity from the set of candidate strategies, giving priority to strategies with positive synergistic effects; and the efficiency optimization layer optimizes the execution order of the strategies in the strategy combination according to computational complexity and resource consumption, generating multiple candidate feature-preserving strategy combination schemes.

[0012] Preferably, the specific process for generating conflict-free optimal strategy combinations is as follows: Conflict detection is performed on candidate feature retention strategy combinations to identify conflict types and locations, including parameter conflicts, logical conflicts, and resource conflicts. Detected conflicts are resolved based on a dual evaluation result matrix and a policy compatibility matrix to generate conflict-free optimal policy combinations. When parameter conflicts are detected, the value range and execution priority of conflicting parameters are dynamically adjusted based on the privacy risk weight and information utility weight in the dual evaluation result matrix. When logical conflicts are detected, the policy execution order and combination method are reconstructed based on the policy compatibility matrix. When resource conflicts are detected, the allocation scheme of computing and storage resources is optimized based on the dual evaluation result matrix.

[0013] Preferably, the performance-privacy trade-off mechanism includes: The system comprises the following components: a preliminary feature retention unit, which retains features based on the optimal strategy combination to obtain preliminary feature retention information; an information availability evaluation unit, which calculates the availability retention degree of the preliminary feature retention information in the analysis task; a privacy protection strength evaluation unit, which quantifies the privacy protection strength of the preliminary feature retention information in the analysis task based on k-anonymity, l-diversity, and differential privacy theory; a performance-privacy tradeoff function construction unit, which establishes a multi-objective optimization function with the objectives of maximizing information availability and maximizing privacy protection strength as the performance-privacy tradeoff function; and a Pareto front search unit, which searches for the Pareto optimal solution set of the performance-privacy tradeoff function under constraints, including computational resource constraints, time constraints, and privacy budget constraints.

[0014] Preferably, the specific process for determining the optimal protection scheme is as follows: combining the privacy risk weight vector and information utility weight vector in the dual evaluation result matrix, a comprehensive score is given to each scheme in the Pareto optimal solution set; a heuristic search algorithm is used to quickly locate the candidate region of the solution with a high comprehensive score in the solution space, thus narrowing the search range; a local optimization algorithm is used to perform a fine search within the candidate region to find an approximate optimal solution within a preset time, which is then used as the optimal feature retention scheme.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a dual assessment model of privacy risk and information utility, enabling simultaneous quantification of the privacy leakage risk and analytical application value of sensitive information. This overcomes the problems of existing technologies where either over-protection leads to data losing its analytical value, or insufficient protection results in privacy leakage risks. Through multi-granularity privacy risk analysis and multi-mode information utility assessment, this invention can formulate differentiated protection strategies based on the specific characteristics of different sensitive information, achieving precise privacy protection. The cross-validation mechanism in the dual assessment model ensures the consistency and reliability of the privacy risk assessment and information utility assessment results, avoiding the bias that may arise from single-dimensional assessment. The granularity switching mechanism adaptively adjusts the assessment precision according to data characteristics, ensuring both accuracy and computational efficiency, providing a scientific and reliable decision-making basis for subsequent feature preservation strategy selection.

[0016] 2. This invention effectively solves the problem of system performance degradation or even processing failure caused by conflicts and coordination issues when multiple strategies are used in parallel, significantly improving the stability and efficiency of feature preservation processing. The strategy compatibility analysis model effectively avoids conflicts and resource waste when multiple strategies are used in parallel by identifying positive synergistic effects, negative conflict effects, and neutral independence effects among strategies. The hierarchical combination search model ensures the optimality and executability of strategy combinations through a three-layer screening mechanism of compatibility filtering, complementarity selection, and efficiency optimization. The conflict detection and resolution mechanism can prevent conflicts before strategy execution and quickly mediate when conflicts occur, greatly reducing system operational risks. This intelligent strategy combination management method not only improves processing efficiency but also provides reliable technical support for complex and sensitive information processing scenarios.

[0017] 3. This invention effectively solves the challenge of balancing data availability and privacy protection in sensitive information processing through a performance-privacy trade-off mechanism and adaptive optimization algorithm, significantly improving the system's practicality and applicability. The algorithm rapidly evaluates the impact of different transformation strategies on data availability and privacy protection, quickly finding Pareto optimal solutions in complex multi-objective optimization problems, meeting the stringent efficiency requirements of practical applications. The solution method combining heuristic search and local optimization ensures both the quality of the solution and controls computation time, enabling the system to find an approximate optimal solution within a finite time. The real-time optimization capability of this invention allows the system to adjust protection strategies in real time according to changes in data flow and user needs, adapting to the processing requirements of dynamic data environments. This provides technical support for large-scale sensitive information processing, enabling the system to maximize the analytical value of data while ensuring privacy protection, providing reliable technical assurance for data sharing and collaborative applications. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an adaptive feature retention analysis method for sensitive information provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the privacy risk-information utility dual assessment model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the strategy compatibility analysis model provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention collects and analyzes user information with access rights to a certain extent on the basis of ensuring legality and compliance. All acquisition, use and analysis of user information involved have been explicitly authorized by the users.

[0020] This invention proposes an adaptive feature preservation analysis method for sensitive information. It can adaptively select the optimal feature transformation strategy based on the dual evaluation results by real-time assessing the sensitivity and analytical importance of data features, and dynamically allocate protection resources under limited privacy budget constraints to maximize overall data availability. The effectiveness of this invention will be illustrated below with two embodiments.

[0021] Example 1: In this embodiment, the method proposed in this invention is used to perform adaptive feature retention analysis on sensitive patient information in medical institutions. This medical data includes multi-dimensional sensitive information such as the patient's basic information (name, ID number, age, gender), medical information (medical history, diagnosis results, treatment plan), and examination data (laboratory indicators, imaging data). Medical institutions need to use this data for disease research, drug development, and medical quality analysis while protecting patient privacy. Figure 1 The specific flowchart of the method of this invention includes: constructing a privacy risk-information utility dual evaluation model containing a privacy risk analysis unit and an information utility evaluation unit to evaluate sensitive information and generate a dual evaluation result matrix; evaluating information utility through different combinations of various evaluation modes; based on the privacy risk and information utility of each sensitive information in the dual evaluation result matrix, analyzing the interaction relationship of different feature retention strategies when used in combination through a strategy compatibility analysis model to obtain a strategy compatibility matrix; generating multiple candidate feature retention strategy combination schemes through a hierarchical combination search model and performing conflict detection according to the dual evaluation result matrix and the strategy compatibility matrix, generating the optimal strategy combination without conflict; evaluating the impact of the optimal strategy combination on information availability and privacy protection through a performance-privacy trade-off mechanism, and determining the optimal feature retention scheme within a preset time using heuristic search and local optimization in conjunction with the dual evaluation result matrix, retaining features of sensitive information, and obtaining feature-retained information. The following is based on... Figure 1 The following explanation is provided regarding the content: A dual assessment model of privacy risk and information utility is constructed to evaluate sensitive information and generate a dual assessment result matrix. The dual assessment model of privacy risk and information utility includes: a privacy risk analysis unit, which performs privacy risk analysis in parallel at different granularities; and an information utility assessment unit, which evaluates information utility through different combinations of multiple assessment modes. refer to Figure 2 The privacy risk analysis unit includes: a privacy risk analysis layer, which performs privacy risk analysis at the record level, field level, value range level, and pattern level to obtain privacy risk analysis results; the record level performs hierarchical privacy risk assessment for each record; the field level performs privacy risk analysis for each field; the value range level performs privacy risk assessment for a preset value range; the pattern level performs privacy risk analysis for data patterns and relationships; and a granularity switching layer, which adaptively switches the privacy risk analysis granularity according to preset granularity switching conditions; the preset granularity switching conditions include: automatically switching from the field level to the pattern level when the data association density exceeds the association density threshold; automatically switching from the value range level to the field level when the data missing rate is higher than the data missing threshold; and triggering multi-granularity parallel analysis when the data distribution skewness is abnormal. The information utility evaluation unit includes: an information utility evaluation layer, which evaluates information utility using statistical importance evaluation mode based on information gain, functional importance evaluation mode based on business objectives, reliability importance evaluation mode based on data quality, dynamic importance evaluation mode based on timeliness, and dependency importance evaluation mode based on correlation; and a scenario feature matching layer, which dynamically selects a combination of evaluation modes based on the current analysis scenario type and user needs to generate information utility evaluation results. The privacy risk-information utility dual assessment model also includes a cross-validation unit to verify the consistency between the privacy risk analysis results and the information utility assessment results.

[0022] Specifically, record-level privacy risk assessment involves a hierarchical privacy risk assessment for each patient record. Record-level sensitivity is obtained by calculating the weighted average of record integrity score, uniqueness score, and relevance score. The record integrity score is calculated as the proportion of sensitive fields in the patient record relative to the total number of sensitive fields; the more identifying fields included, the higher the score. The uniqueness score is obtained by calculating the average similarity between the patient record and other records, subtracting the similarity value from 1. The relevance score is calculated based on the proportion of fields in the patient record that can be linked to external data sources.

[0023] Field-level privacy risk assessment: Privacy risk analysis is performed on each field in the medical data. Field-level sensitivity is obtained through a comprehensive calculation of basic sensitivity, distribution risk, and labeling strength. Basic sensitivity is assigned a fixed score based on the field type; for example, 1.0 for identifiable fields, 0.6 for quasi-identifiable fields, and 0.3 for ordinary attribute fields. Distribution risk is obtained by calculating the reciprocal of the field's value range and the reciprocal of its distribution uniformity. Labeling strength is obtained by calculating the reciprocal of the field's information entropy.

[0024] Value-range-level privacy risk assessment: This involves evaluating preset threshold ranges of privacy risks and obtaining a value-range-level sensitivity score through a weighted average of rarity, identifiability, and inference risk scores. Specifically, the rarity score is calculated as the reciprocal of the proportion of values ​​within the high-risk threshold range to the total number of values ​​in that field; the identifiability score is calculated based on the uniqueness of values ​​within that value range; and the inference risk score is obtained by analyzing the correlation strength between that value range and other sensitive information.

[0025] Pattern-level privacy risk assessment: This involves analyzing data patterns and relationships within medical data to determine privacy risks. The analysis yields pattern-level sensitivity through a comprehensive calculation of association pattern risk, sequence pattern risk, and pattern frequency risk. Association pattern risk is calculated by averaging the confidence scores of association rules involving the relevant feature; sequence pattern risk is calculated based on the traceability score of time-series patterns containing the relevant feature; and pattern frequency risk is calculated as the reciprocal of the pattern's support, with lower-frequency patterns carrying higher risk.

[0026] The granularity switching layer adaptively switches the privacy risk analysis granularity according to preset granularity switching conditions.

[0027] Switching Condition 1: When the correlation density of medical data exceeds the correlation density threshold, the system automatically switches from field-level analysis to schema-level analysis. Data correlation density is determined by calculating the correlation coefficient matrix between fields; the correlation density threshold is when the proportion of correlation coefficients with an absolute value greater than 0.5 in the correlation coefficient matrix exceeds 60% of the total logarithms of correlation coefficients, indicating that the data has a complex internal structure and that schema-level analysis is needed to identify potential privacy risks.

[0028] Switching Condition 2: When the data missing rate exceeds the data missing threshold, automatically switch from value range-level analysis to field-level analysis. The data missing rate is calculated by statistically analyzing the proportion of missing values ​​in each field; the data missing threshold is set when the proportion of missing values ​​in a single field exceeds 20% or the average missing rate of the entire dataset exceeds 15%, at which point the reliability of value range-level analysis decreases, and field-level analysis is used.

[0029] Switching condition three: When the data distribution skewness is abnormal, trigger multi-granularity parallel analysis. The data distribution skewness is determined by calculating the third central moment of the data distribution; when the absolute value of the skewness coefficient of the data distribution exceeds 1.0, it indicates that there are outliers or special patterns in the data, and multiple granularities need to be analyzed simultaneously to obtain a comprehensive risk assessment.

[0030] The information utility assessment layer evaluates information utility using five different assessment models, each tailored to different application needs and data characteristics: The statistical importance assessment model based on information gain calculates the information contribution of each feature to the target analysis task. In this embodiment, the target analysis task is disease diagnosis, and the information gain value of each medical feature (symptoms, examination indicators, etc.) is calculated using a decision tree model.

[0031] The functional importance assessment model based on business objectives evaluates the importance of features according to specific medical business objectives. It determines the feature importance weights for different business scenarios through an expert knowledge base and a business rule engine. For drug efficacy studies in this embodiment, features such as patient genetic information, medication history, and treatment response have high functional importance; for epidemiological studies, features such as patient geographical location, contact history, and onset time are more important.

[0032] The reliability importance assessment model based on data quality evaluates the importance of features from a data quality perspective. Data quality includes: the proportion of non-missing values ​​in features, the reliability of data assessed through outlier detection and consistency checks, and the consistency of feature distribution across different subsets of data.

[0033] The timeliness-based dynamic importance assessment model considers the timeliness characteristics of medical data. The timeliness weight of features is calculated using a time decay function based on an exponential decay model. In this embodiment, recent examination results are more important than historical data, and diagnostic information for acute diseases is more time-sensitive than information for chronic diseases.

[0034] The dependency importance assessment model based on correlation analyzes the dependencies between features to evaluate their importance. By constructing a feature correlation graph, graph algorithms are used to calculate the network centrality index of each feature, including degree centrality, betweenness centrality, and eigenvector centrality, to comprehensively determine the dependency importance of the features. For example, in medical data, some features may not be highly important individually, but they are highly correlated with other important features, thus possessing significant dependency value.

[0035] The scenario feature matching layer dynamically selects and adjusts the combination of evaluation modes based on the current analysis scenario type and user needs. It identifies the current analysis scenario through task description parsing, user preference settings, and data feature analysis. Scenario types include diagnostic assistance, drug development, epidemiological surveys, and medical quality assessment. Each scenario corresponds to a different emphasis on feature importance. Based on the scenario identification results, the most suitable combination is selected from five evaluation modes. For example, for diagnostic assistance scenarios, statistical importance and functional importance modes are mainly selected; for scientific research analysis scenarios, reliability importance and dependency importance modes are emphasized; and for real-time monitoring scenarios, dynamic importance modes have higher weights. A combination of the analytic hierarchy process (AHP) and an expert system is used to adjust the combined weights of each mode. The AHP calculates the relative importance of each mode by constructing a judgment matrix, while the expert system provides weight adjustment suggestions based on domain knowledge and historical experience.

[0036] The information utility fusion layer employs a weighted fusion algorithm to linearly combine the results of multiple evaluation models. To address the dimensional differences in the evaluation results from different models, the results of each model are standardized, and then a weighted average method is used to calculate the comprehensive information utility score.

[0037] The cross-validation unit performs consistency checks on the privacy risk analysis results and information utility assessment results, including correlation analysis and conflict detection. Correlation analysis calculates the correlation between privacy risk and information utility; theoretically, features with high privacy risk often have high information utility (such as identity information). The system uses correlation analysis to verify whether the assessment results meet expectations. Conflict detection identifies potential conflicts in the assessment results, such as a feature being assessed as low privacy risk but high information utility, or high privacy risk but low information utility. These situations require further analysis and adjustment. Consistency adjustment is used to adjust the assessment results when inconsistencies are detected, by re-analyzing feature attributes, adjusting assessment parameters, or introducing expert intervention.

[0038] The privacy risk-information utility dual assessment model achieves precise dual assessment of sensitive information. The four-layer granularity parallel analysis and adaptive switching mechanism solves the problems of insufficient accuracy or computational redundancy caused by the fixed granularity analysis of existing technologies, and realizes adaptive processing to dynamically adjust the analysis accuracy according to data characteristics. At the same time, the dynamic combination mechanism of five assessment modes overcomes the technical limitations of static assessment methods that cannot adapt to different application scenarios, and significantly improves the accuracy and applicability of sensitivity and importance assessment.

[0039] Furthermore, based on the privacy risks and information utility of each sensitive information in the dual evaluation result matrix, the interaction relationship of different feature retention strategies when used in combination is analyzed through a strategy compatibility analysis model to obtain a strategy compatibility matrix; Reference Figure 3 The strategy compatibility analysis model includes: The privacy risk-information utility mapping unit establishes the compatibility relationship between sensitive information and feature retention strategies based on the privacy risk and information utility information of each sensitive information in the dual evaluation result matrix; the strategy interaction relationship identification unit identifies the positive synergistic effect, negative conflict effect, and neutral independence effect of different feature retention strategies when used together; the compatibility score calculation unit obtains the compatibility score between feature retention strategies based on the overlap of strategy functions, the degree of parameter conflict, and the degree of resource competition; and the strategy compatibility matrix generation unit generates a strategy compatibility matrix describing the compatibility relationship between each feature retention strategy based on the strategy interaction identification results and the compatibility score.

[0040] Specifically, the privacy risk-information utility mapping unit first acquires a feature preservation strategy library containing various feature preservation strategies. It then determines a suitable protection strategy based on the privacy risk and information utility of the features. For example, features with high privacy risk and low information utility are suitable for data suppression strategies; features with high privacy risk and high information utility are suitable for differential privacy processing; and features with low privacy risk and high information utility can be treated with mild data generalization. A feature-policy mapping matrix is ​​constructed, where rows represent features, columns represent policies, and matrix elements represent the fit score between the feature and the policy. The fit score comprehensively considers both the protection effect and the utility loss.

[0041] Strategy interaction relationship identification unit: Positive synergistic effects are obtained by analyzing how combinations of strategies can produce enhanced protection. For example, simultaneously applying data generalization and noise addition to relevant features can provide stronger privacy protection; applying data suppression to identifiable features and data generalization to quasi-identifiable features can form multi-layered protection.

[0042] Negative conflict effects are identified by recognizing potential conflicts between strategies. For example, applying both data generalization and precise computation to the same feature simultaneously can create logical conflicts; using incompatible processing methods for related features can lead to data inconsistencies.

[0043] Neutral independence effects are obtained by identifying situations where strategies are independent of each other and have no significant interaction. For example, applying different strategies to unrelated features usually has an independent effect.

[0044] The strategy functional overlap in the compatibility scoring unit is calculated by analyzing the functional similarity of two strategies; the parameter conflict degree checks whether there are conflicts in the parameter settings of the strategies; the resource contention degree evaluates the strategy's requirements for computing resources, storage resources, and time resources; and a weighted summation method is used to calculate the comprehensive compatibility score.

[0045] The generated policy compatibility matrix is ​​a symmetric matrix, where rows and columns represent policies, and matrix elements represent the compatibility scores between two policies.

[0046] By constructing a strategy compatibility analysis model, intelligent identification and quantitative evaluation of the interaction relationships between feature preservation strategies were achieved, solving the technical problem that existing technologies cannot handle conflicts arising when multiple strategies are used in synergy. An intelligent adaptation mechanism for sensitive information and protection strategies was established, accurately identifying the synergistic, conflicting, and independent effects between strategies. Through multi-dimensional compatibility scoring and dynamic matrix updates, conflicts and resource waste when multiple strategies are used in parallel were effectively avoided, significantly improving the stability, coordination, and execution efficiency of strategy combinations.

[0047] Furthermore, based on the dual evaluation result matrix and the policy compatibility matrix, multiple candidate feature retention policy combination schemes are generated through a hierarchical combination search model, and conflict detection is performed to generate the conflict-free optimal policy combination. The hierarchical combined search model includes: The compatibility filtering layer filters out a set of mutually compatible candidate strategies based on the strategy compatibility matrix; the complementarity selection layer selects a combination of strategies with functional complementarity from the set of candidate strategies, giving priority to strategies with positive synergistic effects; and the efficiency optimization layer optimizes the execution order of the strategies in the strategy combination according to computational complexity and resource consumption, generating multiple candidate feature-preserving strategy combination schemes.

[0048] Specifically, the compatibility filtering layer first identifies policy pairs with negative scores in the compatibility matrix, indicating conflicting effects that prevent simultaneous use. A conflict graph is constructed based on these policy pairs, and a graph coloring algorithm is used to determine the largest set of conflict-free policies. A minimum threshold for compatibility scores is set, and only policy pairs with compatibility scores above this threshold are retained in the candidate set. The threshold setting considers both policy effectiveness and computational efficiency. After filtering, multiple candidate policy sets are generated, with each set containing policies exhibiting good compatibility.

[0049] Complementary Selection Layer: Analyze the protection coverage of each strategy after compatibility filtering, including the data types protected, protection strength, and applicable scenarios, and select strategy combinations that can comprehensively cover all protection needs; furthermore, prioritize strategy combinations with positive synergistic effects among these combinations, which can reduce resource consumption or improve processing efficiency while maintaining or improving protection effectiveness.

[0050] Efficiency Optimization Layer: By analyzing the dependencies between strategy combinations obtained from the Complementarity Selection Layer using a directed acyclic graph, this layer determines which strategies must be executed in a specific order and which strategies can be executed in parallel. It evaluates the computational complexity, memory requirements, storage requirements, and execution time of each strategy to provide a basis for resource scheduling. The layer performs topological sorting on the strategy dependency graph, prioritizing strategies with lower resource requirements or shorter execution times while satisfying dependencies, thus determining a reasonable execution order. Finally, it analyzes the possibility of parallel execution of strategies, and for strategies without dependencies, executes them in parallel to improve efficiency.

[0051] By designing a hierarchical combinatorial search model, this paper solves the technical problem of excessive computational complexity in traditional exhaustive search within a large-scale strategy combination space, achieving efficient search and intelligent optimization of strategy combinations. Specifically, the compatibility filtering layer effectively eliminates conflicting strategy combinations and narrows the search space; the complementarity selection layer ensures the integrity of functional coverage and maximizes synergistic effects; and the efficiency optimization layer significantly improves processing efficiency through execution order optimization and parallelism analysis, rapidly finding high-quality candidate solutions in the vast combinatorial space, greatly improving the search efficiency and solution quality of strategy combinations. This three-layer progressive search achieves efficient solution in an exponential combinatorial space, reducing search complexity from exponential to polynomial level while ensuring the quality and diversity of search results.

[0052] Furthermore, the specific process for generating conflict-free optimal strategy combinations is as follows: Conflict detection is performed on candidate feature retention strategy combinations to identify conflict types and locations, including parameter conflicts, logical conflicts, and resource conflicts. Detected conflicts are resolved based on a dual evaluation result matrix and a policy compatibility matrix to generate conflict-free optimal policy combinations. When parameter conflicts are detected, the value range and execution priority of conflicting parameters are dynamically adjusted based on the privacy risk weight and information utility weight in the dual evaluation result matrix. When logical conflicts are detected, the policy execution order and combination method are reconstructed based on the policy compatibility matrix. When resource conflicts are detected, the allocation scheme of computing and storage resources is optimized based on the dual evaluation result matrix.

[0053] Specifically, parameter conflict detection involves checking parameter type, value range, and precision requirements to identify conflicts in parameter settings between strategies. When a conflict is detected, the impact weights of the conflicting parameters on privacy protection and information utility are calculated. The parameter adjustment scheme with the least impact on the overall goal is selected, including parameter value negotiation, priority reordering, and alternative parameter selection. The privacy protection impact weight is the ratio of the change in privacy protection strength to the current total privacy protection score, and the information utility impact weight is the ratio of the change in information utility retention to the current total information utility score.

[0054] Logical conflict detection: Through data flow analysis, operation sequence checks, and dependency verification, this function analyzes whether there are contradictions in the execution logic of the strategy. When a logical conflict is detected, the dependencies of the conflicting strategy are analyzed, and the conflict is resolved by adjusting the execution order, modifying the strategy logic, or replacing the conflicting strategy.

[0055] Resource conflict detection: By calculating resource conflicts (CPU and memory requirements exceeding limits), storage resource conflicts (insufficient storage space), and time resource conflicts (execution time exceeding limits), the system assesses whether the policy's demand for system resources exceeds available resources. When a resource conflict is detected, based on the privacy risk weight and information utility weight in the dual evaluation result matrix, resource reallocation, policy degradation, and parallelism adjustment methods are used to allocate more resources to the processing strategies for important features, while adopting less resource-intensive strategies for secondary features.

[0056] By establishing an adaptive conflict resolution mechanism, the system achieves automatic identification and intelligent processing of policy combination conflicts, comprehensively covering the detection of parameter conflicts, logical conflicts, and resource conflicts. Based on the dual evaluation results and policy compatibility matrix, it can automatically resolve various conflicts in policy combinations while maintaining system performance, realizing intelligent negotiation, logical reconstruction, and optimized resource allocation for conflicts. This effectively prevents system failures and performance degradation caused by conflicts, ensures the stable execution of policy combinations, and significantly improves the reliability and robustness of the system.

[0057] Furthermore, the impact of the optimal strategy combination on information availability and privacy protection is evaluated through a performance-privacy trade-off mechanism. Combining the dual evaluation result matrix, the optimal feature retention scheme is determined within a preset time using heuristic search and local optimization, and the sensitive information is feature-retained to obtain feature-retained information.

[0058] The performance-privacy trade-off mechanism includes: The system comprises the following components: a preliminary feature retention unit, which retains features based on the optimal strategy combination to obtain preliminary feature retention information; an information availability evaluation unit, which calculates the availability retention degree of the preliminary feature retention information in the analysis task; a privacy protection strength evaluation unit, which quantifies the privacy protection strength of the preliminary feature retention information in the analysis task based on k-anonymity, l-diversity, and differential privacy theory; a performance-privacy tradeoff function construction unit, which establishes a multi-objective optimization function with the objectives of maximizing information availability and maximizing privacy protection strength as the performance-privacy tradeoff function; and a Pareto front search unit, which searches for the Pareto optimal solution set of the performance-privacy tradeoff function under constraints, including computational resource constraints, time constraints, and privacy budget constraints.

[0059] Specifically, the preliminary feature preservation unit performs feature preservation processing on medically sensitive information according to a determined combination of strategies and execution order to obtain preliminary feature-preserved information. For example, a pseudonym strategy is used for patient names, a range generalization strategy is used for ages, and differential privacy noise is added to diagnostic results.

[0060] Information Usability Assessment Unit: This unit assesses the usability of retained preliminary feature information based on the specific medical analysis task. For example, for disease prediction tasks, it assesses the impact of retained information on the accuracy of the prediction model; for drug efficacy analysis, it assesses the impact of retained information on the statistical analysis results.

[0061] Privacy Protection Strength Assessment Unit: Based on k-anonymity, l-diversity, and differential privacy theory, this unit quantifies the privacy protection strength of the preliminary feature retention information in the analysis task. Performance-Privacy Trade-off Function Building Block: Construct a bi-objective optimization function. The first objective is to maximize information availability, and the second objective is to maximize the strength of privacy protection. Since the two objectives are often in conflict, a Pareto optimal solution needs to be found.

[0062] Pareto Front Search Unit: Transforms the constraints in practical applications into optimization constraints, including computational resource constraints (processing time does not exceed a preset threshold), storage constraints (storage space does not exceed available capacity), and privacy budget constraints (noise addition does not exceed budget limits); Pareto front search is performed using a non-dominated sorting genetic algorithm.

[0063] By constructing a performance-privacy trade-off mechanism, we have achieved quantitative evaluation and multi-objective optimization of information availability and privacy protection strength. We have established a multi-dimensional information availability evaluation system and a comprehensive privacy protection strength evaluation mechanism. Through the construction of multi-objective optimization functions and Pareto front search, we have found the optimal trade-off solution under constraints such as computing resources, time and privacy budget, so that the system can maximize data availability while meeting privacy protection requirements.

[0064] Furthermore, the specific process for determining the optimal feature preservation scheme is as follows: Combining the privacy risk weight vector and information utility weight vector in the dual evaluation result matrix, a comprehensive score is given to each scheme in the Pareto optimal solution set; a heuristic search algorithm is used to quickly locate the candidate region of the solution with a high comprehensive score in the solution space, narrowing the search range; a local optimization algorithm is used to perform a fine search within the candidate region to find an approximate optimal solution within a preset time, which is then used as the optimal feature retention scheme.

[0065] Specifically, the privacy risk weight vector is obtained by calculating a weighted average after normalizing the four granularity sensitivity scores of each sensitive information; the information utility weight vector is obtained by calculating a weighted average after normalizing the importance scores of the five evaluation modes of each sensitive information.

[0066] Taking into account the privacy protection effect, information availability retention, and corresponding weighting coefficients, a comprehensive score is calculated for each Pareto optimal solution. Heuristic rules are designed based on medical data characteristics and privacy protection experience to guide the search process. A heuristic search algorithm is used to analyze the distribution characteristics of the Pareto optimal solution set, identifying regions with high-scoring solutions and marking them as key search areas. Simulated annealing is used for local search, with time limits and quality thresholds set as convergence conditions. The search stops when the search time reaches the preset time limit or the improvement in solution quality is less than the threshold.

[0067] By combining heuristic search with local optimization, a hybrid algorithm is developed to address the technical problem of slow convergence or easy getting trapped in local optima in complex constrained multi-objective problems by traditional optimization algorithms. By combining a scheme scoring mechanism with dual evaluation weight information, a near-optimal solution to a high-dimensional complex optimization problem can be found in a finite time. This significantly improves the convergence performance and solution quality of the algorithm and meets the time requirements of real-time applications.

[0068] Example 2: In Example 1, the method proposed in this invention successfully assesses the sensitivity and analytical importance of data features in real time, adaptively selects the optimal feature transformation strategy based on the dual assessment results, and dynamically allocates protection resources under limited privacy budget constraints to maximize overall data availability. To further verify the effectiveness of this invention, adaptive feature retention analysis is also performed on sensitive patient information from another medical institution in this embodiment.

[0069] A dual assessment model of privacy risk and information utility is constructed to evaluate sensitive information and generate a dual assessment result matrix. The dual assessment model of privacy risk and information utility includes: a privacy risk analysis unit, which performs privacy risk analysis in parallel at different granularities; and an information utility assessment unit, which evaluates information utility through different combinations of multiple assessment modes. The privacy risk analysis unit includes: a privacy risk analysis layer, which performs privacy risk analysis at the record level, field level, value range level and pattern level to obtain privacy risk analysis results; and a granularity switching layer, which adaptively switches the privacy risk analysis granularity according to preset granularity switching conditions. The information utility evaluation unit includes: an information utility evaluation layer, which evaluates information utility using statistical importance evaluation mode based on information gain, functional importance evaluation mode based on business objectives, reliability importance evaluation mode based on data quality, dynamic importance evaluation mode based on timeliness, and dependency importance evaluation mode based on correlation; and a scenario feature matching layer, which dynamically selects a combination of evaluation modes based on the current analysis scenario type and user needs to generate information utility evaluation results. The privacy risk-information utility dual assessment model also includes a cross-validation unit to verify the consistency between the privacy risk analysis results and the information utility assessment results.

[0070] The preset granularity switching conditions include: automatically switching from field level to pattern level when the data association density exceeds the association density threshold; automatically switching from value range level to field level when the data missing rate is higher than the data missing threshold; and triggering multi-granularity parallel analysis when the data distribution skewness is abnormal. Furthermore, the scene feature matching layer includes: a task description parsing module, which parses the user-input task description text using natural language processing technology, extracts analysis objectives, data processing requirements, and business scenario keywords, and generates a task feature vector; a user preference setting module, which obtains the user's preference weight settings in terms of privacy protection strength, processing speed, analysis accuracy, and resource consumption, and generates a user preference vector; a data feature analysis module, which automatically identifies the scale, dimension, type, completeness, distribution characteristics, and association patterns of the input data through statistical analysis methods, and generates a data feature vector; and a scene type analysis module, which calculates the scene comprehensive feature value based on the weighted fusion of the task feature vector, user preference vector, and data feature vector, and determines the current analysis scene type by matching the similarity with scene features in the historical scene database.

[0071] By constructing a scene recognition mechanism that includes a task description parsing module, a user preference setting module, and a data feature analysis module, the subjectivity and inaccuracy issues caused by the reliance on manual judgment in existing scene recognition technologies are resolved. This achieves intelligent scene recognition based on natural language processing, user preference acquisition, and automatic data feature analysis. Through weighted fusion calculation of task feature vectors, user preference vectors, and data feature vectors, and similarity matching of historical scene databases, the current analysis scene type can be accurately identified and the most suitable evaluation mode combination can be automatically selected, significantly improving the accuracy of scene recognition and the relevance of evaluation mode selection.

[0072] Furthermore, based on the privacy risks and information utility of each sensitive information in the dual evaluation result matrix, the interaction relationship of different feature retention strategies when used in combination is analyzed through a strategy compatibility analysis model to obtain a strategy compatibility matrix; The strategy compatibility analysis model includes: The privacy risk-information utility mapping unit establishes the compatibility relationship between sensitive information and feature retention strategies based on the privacy risk and information utility information of each sensitive information in the dual evaluation result matrix; the strategy interaction relationship identification unit identifies the positive synergistic effect, negative conflict effect, and neutral independence effect of different feature retention strategies when used together; the compatibility score calculation unit obtains the compatibility score between feature retention strategies based on the overlap of strategy functions, the degree of parameter conflict, and the degree of resource competition; and the strategy compatibility matrix generation unit generates a strategy compatibility matrix describing the compatibility relationship between each feature retention strategy based on the strategy interaction identification results and the compatibility score.

[0073] Furthermore, the specific calculation method of the compatibility score calculation unit is as follows: The overlap of the policy functions is determined by calculating the ratio of the intersection to the union of the target sets protected by the two policies. The higher the ratio, the greater the overlap. The conflict of parameters is determined by analyzing the range of values ​​and constraints of the policy parameters. The conflict is at its maximum when the range of values ​​of the parameters have no intersection or the constraints are contradictory. The conflict is zero when the parameters are completely compatible. The resource contention is determined by calculating the ratio of the policy's demand for CPU, memory, and storage resources to the amount of available resources. The contention is at its maximum when the total demand exceeds the available resources.

[0074] By designing methods for calculating functional overlap based on set intersection-union ratio, parameter conflict based on value range and constraint analysis, and resource competition based on the ratio of resource demand to available quantity, this paper solves the technical problem that existing technologies cannot quantify the compatibility relationship between strategies. It realizes accurate mathematical modeling and quantitative evaluation of strategy compatibility. Through clear calculation rules and boundary condition definitions, it can objectively and accurately evaluate the compatibility degree of any two strategies, providing a reliable quantitative basis for strategy combination selection and avoiding the subjectivity and ambiguity of compatibility evaluation.

[0075] Furthermore, based on the dual evaluation result matrix and the policy compatibility matrix, multiple candidate feature retention policy combination schemes are generated through a hierarchical combination search model, and conflict detection is performed to generate the conflict-free optimal policy combination. The hierarchical combined search model includes: The compatibility filtering layer filters out a set of mutually compatible candidate strategies based on the strategy compatibility matrix; the complementarity selection layer selects a combination of strategies with functional complementarity from the set of candidate strategies, giving priority to strategies with positive synergistic effects; and the efficiency optimization layer optimizes the execution order of the strategies in the strategy combination according to computational complexity and resource consumption, generating multiple candidate feature-preserving strategy combination schemes.

[0076] Furthermore, the specific process for generating conflict-free optimal strategy combinations is as follows: Conflict detection is performed on candidate feature retention strategy combinations to identify conflict types and locations, including parameter conflicts, logical conflicts, and resource conflicts. Detected conflicts are resolved based on a dual evaluation result matrix and a policy compatibility matrix to generate conflict-free optimal policy combinations. When parameter conflicts are detected, the value range and execution priority of conflicting parameters are dynamically adjusted based on the privacy risk weight and information utility weight in the dual evaluation result matrix. When logical conflicts are detected, the policy execution order and combination method are reconstructed based on the policy compatibility matrix. When resource conflicts are detected, the allocation scheme of computing and storage resources is optimized based on the dual evaluation result matrix.

[0077] Furthermore, the impact of the optimal strategy combination on information availability and privacy protection is evaluated through a performance-privacy trade-off mechanism. Combining the dual evaluation result matrix, the optimal feature retention scheme is determined within a preset time using heuristic search and local optimization, and the sensitive information is feature-retained to obtain feature-retained information.

[0078] The performance-privacy trade-off mechanism includes: The system comprises the following components: a preliminary feature retention unit, which retains features based on the optimal strategy combination to obtain preliminary feature retention information; an information availability evaluation unit, which calculates the availability retention degree of the preliminary feature retention information in the analysis task; a privacy protection strength evaluation unit, which quantifies the privacy protection strength of the preliminary feature retention information in the analysis task based on k-anonymity, l-diversity, and differential privacy theory; a performance-privacy tradeoff function construction unit, which establishes a multi-objective optimization function with the objectives of maximizing information availability and maximizing privacy protection strength as the performance-privacy tradeoff function; and a Pareto front search unit, which searches for the Pareto optimal solution set of the performance-privacy tradeoff function under constraints, including computational resource constraints, time constraints, and privacy budget constraints.

[0079] Furthermore, the trade-off coefficient between information availability and privacy protection strength in the performance-privacy trade-off function is determined through user preference settings and business requirement priorities: When the user's preference is privacy-first, the weight of privacy protection is greater than the weight of information availability; when the user's preference is utility-first, the weight of information availability is greater than the weight of privacy protection; when the user's preference is balanced, the two weights are set to be equal; the trade-off coefficient can be dynamically adjusted according to the feedback of actual application effects, and the adjustment range is determined based on the system learning algorithm and user satisfaction evaluation results.

[0080] By establishing a tradeoff coefficient determination mechanism based on user preference settings, adaptive weight allocation under three preferences—privacy priority, utility priority, and balance mode—is achieved. Combined with a dynamic adjustment mechanism of system learning algorithm and user satisfaction evaluation, the tradeoff strategy can be continuously optimized based on feedback from actual application effects. This enables the multi-objective optimization function to truly reflect the actual needs and preferences of users, significantly improving user satisfaction and practicality of the optimization results.

[0081] Furthermore, the specific process for determining the optimal feature preservation scheme is as follows: Combining the privacy risk weight vector and information utility weight vector in the dual evaluation result matrix, a comprehensive score is given to each scheme in the Pareto optimal solution set; a heuristic search algorithm is used to quickly locate the candidate region of the solution with a high comprehensive score in the solution space, narrowing the search range; a local optimization algorithm is used to perform a fine search within the candidate region to find an approximate optimal solution within a preset time, which is then used as the optimal feature retention scheme.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive feature-preserving analysis method for sensitive information, characterized in that, include: A privacy risk-information utility dual assessment model is constructed to evaluate sensitive information and generate a dual assessment result matrix; The sensitive information is medical data, including basic information, medical information, and examination data; the medical information includes medical history, diagnostic plans, and treatment plans; the examination data includes laboratory indicators and imaging data; the privacy risk-information utility dual assessment model includes: a privacy risk analysis unit, which performs privacy risk analysis in parallel at different granularities; and an information utility assessment unit, which assesses information utility through different combinations of various assessment modes. Based on the privacy risks and information utility of each sensitive information in the dual evaluation result matrix, a strategy compatibility analysis model is used to analyze the interaction relationship of different feature retention strategies when used in combination, resulting in a strategy compatibility matrix. The strategy compatibility analysis model includes: a privacy risk-information utility mapping unit, which establishes the adaptation relationship between sensitive information and feature retention strategies based on the privacy risks and information utility information of each sensitive information in the dual evaluation result matrix; a strategy interaction relationship identification unit, which identifies the positive synergistic effect, negative conflict effect, and neutral independent effect of different feature retention strategies when used in combination; a compatibility score calculation unit, which obtains the compatibility score between feature retention strategies based on the strategy function overlap, parameter conflict, and resource competition; and a strategy compatibility matrix generation unit, which generates a strategy compatibility matrix describing the compatibility relationship between each feature retention strategy based on the strategy interaction identification results and the compatibility score. Based on the dual evaluation result matrix and the strategy compatibility matrix, a hierarchical combination search model is used to generate multiple candidate feature retention strategy combination schemes and perform conflict detection to generate the optimal strategy combination without conflict. The hierarchical combination search model includes a compatibility filtering layer, which filters out a set of mutually compatible candidate strategies based on the strategy compatibility matrix; a complementarity selection layer, which selects strategy combinations with functional complementarity from the candidate strategy set, prioritizing strategies with positive synergistic effects; and an efficiency optimization layer, which optimizes the strategy execution order in the strategy combination based on computational complexity and resource consumption, generating multiple candidate feature retention strategy combination schemes. The impact of the optimal strategy combination on information availability and privacy protection is evaluated through a performance-privacy trade-off mechanism. By combining the dual evaluation result matrix, the optimal feature retention scheme is determined within a preset time using heuristic search and local optimization, and the features of sensitive information are retained to obtain feature-retained information.

2. The adaptive feature preservation analysis method for sensitive information according to claim 1, characterized in that: The privacy risk analysis unit includes: a privacy risk analysis layer, which performs privacy risk analysis at the record level, field level, value range level and pattern level to obtain privacy risk analysis results; and a granularity switching layer, which adaptively switches the privacy risk analysis granularity according to preset granularity switching conditions. The information utility evaluation unit includes: an information utility evaluation layer, which evaluates information utility using statistical importance evaluation mode, functional importance evaluation mode, reliability importance evaluation mode, dynamic importance evaluation mode, and dependency importance evaluation mode respectively; and a scenario feature matching layer, which dynamically selects a combination of evaluation modes based on the current analysis scenario type and user needs to generate information utility evaluation results. The privacy risk-information utility dual assessment model also includes a cross-validation unit to verify the consistency between the privacy risk analysis results and the information utility assessment results.

3. The adaptive feature preservation analysis method for sensitive information according to claim 2, characterized in that: The preset granularity switching conditions include: automatically switching from field level to pattern level when the data association density exceeds the association density threshold; automatically switching from value range level to field level when the data missing rate is higher than the data missing threshold; and triggering multi-granularity parallel analysis when the data distribution skewness is abnormal.

4. The adaptive feature preservation analysis method for sensitive information according to claim 1, characterized in that: The specific process for generating conflict-free optimal policy combinations is as follows: Conflict detection is performed on candidate feature retention strategy combinations to identify conflict types and locations, including parameter conflicts, logical conflicts, and resource conflicts. Detected conflicts are resolved based on the dual evaluation result matrix and the strategy compatibility matrix to generate conflict-free optimal strategy combinations. When parameter conflicts are detected, the value range and execution priority of conflict parameters are dynamically adjusted based on the privacy risk weight and information utility weight in the dual evaluation result matrix. When a logical conflict is detected, the policy execution order and combination method are reconstructed based on the policy compatibility matrix; when a resource conflict is detected, the allocation scheme of computing resources and storage resources is optimized based on the dual evaluation result matrix.

5. The adaptive feature preservation analysis method for sensitive information according to claim 1, characterized in that: The performance-privacy trade-off mechanism includes: The system comprises the following components: a preliminary feature retention unit, which retains features based on the optimal strategy combination to obtain preliminary feature retention information; an information availability evaluation unit, which calculates the availability retention degree of the preliminary feature retention information in the analysis task; a privacy protection strength evaluation unit, which quantifies the privacy protection strength of the preliminary feature retention information in the analysis task based on k-anonymity, l-diversity, and differential privacy theory; a performance-privacy tradeoff function construction unit, which establishes a multi-objective optimization function with the objectives of maximizing information availability and maximizing privacy protection strength as the performance-privacy tradeoff function; and a Pareto front search unit, which searches for the Pareto optimal solution set of the performance-privacy tradeoff function under constraints, including computational resource constraints, time constraints, and privacy budget constraints.

6. The adaptive feature preservation analysis method for sensitive information according to claim 5, characterized in that: The specific process for determining the optimal feature preservation scheme is as follows: combining the privacy risk weight vector and information utility weight vector in the dual evaluation result matrix, a comprehensive score is given to each scheme in the Pareto optimal solution set; a heuristic search algorithm is used to quickly locate the candidate region of the solution with a high comprehensive score in the solution space, thus narrowing the search range; a local optimization algorithm is used to perform a fine search within the candidate region to find an approximate optimal solution within a preset time, which is then used as the optimal feature preservation scheme.

Images