Statistical result release-oriented multi-mode optimization differential privacy mechanism recommendation method
By optimizing the differential privacy mechanism recommendation method in multiple modes, the problems of single evaluation indicators and insufficient data space adaptability in existing technologies are solved, and high-precision and stable recommendations are achieved in complex query scenarios.
Patent Information
- Application Number
- CN202510734725.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-09-26
AI Technical Summary
In the existing technology, the differential privacy mechanism recommendation method for publishing statistical results has problems such as single evaluation indicators and insufficient data space adaptability, making it difficult to maintain high accuracy and stability in complex and changeable data query scenarios.
A multi-mode optimized differential privacy mechanism recommendation method is adopted. By randomly generating training sets and dividing the feature space, setting multiple utility maintenance goals, and testing with multiple utility loss evaluation indicators, the differential privacy mechanism with the highest global near-optimal probability is selected as the recommendation.
Under the same privacy budget and diversified utility loss evaluation, the recommendation accuracy and stability of the differential privacy mechanism in complex query scenarios are significantly improved, and the versatility of the model is improved.
Smart Images
Figure CN120705393A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of information security and privacy protection, and in particular relates to a multi-mode optimized differential privacy mechanism recommendation method. Background Art
[0002] In data-driven decision-making scenarios, the publication of statistical results serves as a key means of revealing the inherent patterns of data.
[0003] The paper, "A Decision Tree-Based Statistical Differential Privacy Mechanism Recommendation Method," was published in the Proceedings of the 2017 ACM International Conference on Management of Data on May 9, 2017. It describes a two-stage training method for recommending decision tree models that implement differential privacy mechanisms. The first stage involves training set generation, where one or two attributes are selected from a public training set to construct a one-dimensional or two-dimensional dataset. For each dataset, a count query is performed to obtain its histogram data as the result dataset. For each result dataset, a feature vector is calculated for subsequent decision tree training. This feature vector consists of the domain of the selected attribute, the total number of data records, the number of non-zero counts in the result dataset, and the variance of the result dataset. The second stage is the decision tree training stage. For each result dataset generated in the first stage, different differential privacy mechanisms are used to add noise to the statistical results. The mean squared error between the original data and the noisy data is calculated as the utility loss. To account for the uncertainty of differential privacy noise, the noise addition and utility loss calculation are repeated 20 times and the average is used as the final evaluation result. The evaluation results of each result data set can be used to obtain the differential privacy mechanism with the best utility maintenance effect. Therefore, the result data set can be divided according to its feature vectors and the decision tree model can be trained so that the optimal differential privacy mechanism of all result data sets in the divided subspace is almost the same. This method has two defects: (1) The existing data utility evaluation index is too simplistic; the existing decision tree model usually uses a single error index (such as mean deviation) to evaluate the utility of the differential privacy mechanism, which is difficult to fully reflect the data quality of the statistical results in terms of distribution characteristics, variance preservation and multi-dimensional correlation. (2) The problem of insufficient adaptability to the real scene data space; the existing decision tree model is effective in the scenario of specific data, but when faced with complex and changing data queries, its mechanism recommendation accuracy drops significantly. First, the training data of this method comes from the public data set, which makes the decision tree model effective in the scenario of specific data, but when faced with complex and changing data queries, its mechanism recommendation accuracy drops significantly. Second, this method uses a single error index to evaluate the utility loss of the differential privacy mechanism, which is difficult to fully reflect the statistical results in maintaining data quality in a variety of data modes.
[0004] Differential privacy prevents the privacy leakage of database D through a random mechanism K. Its core goal is to make the information derived from D indistinguishable from any adjacent database D' that differs by only a single record. This indistinguishability can be described by the following formula:
[0005] Pr[K(D)∈S]≤exp(ε)×Pr[K(D')∈S]
[0006] Here, S represents the set of all possible outputs of mechanism K, and ε represents the privacy budget, which regulates the strength of privacy protection. In practice, smaller ε indicates stricter privacy protection. The same privacy protection standard can be achieved through multiple mechanisms, but with varying utility losses. The parameter combination consisting of the privacy budget ε and mechanism K determines a specific implementation of differential privacy. However, precisely specifying the appropriate implementation details in this way often requires a high degree of expertise and repeated trial and error. Summary of the Invention
[0007] Aiming to solve the problems of the existing technology of single evaluation indicators and insufficient adaptability to the complete data space, the present invention proposes a multi-modal optimized differential privacy mechanism recommendation method for statistical result publishing. Aiming at the user's multi-utility maintenance goal, it realizes the reliable recommendation of multi-modal optimal differential privacy mechanism in complex query scenarios.
[0008] To achieve the above-mentioned purpose, the present invention proposes a method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results, which includes the following steps:
[0009] S1. Randomly generate statistical result release data, filter some statistical distribution result release data as a training set, and perform feature space division on the training set according to a specific data pattern;
[0010] S2. Setting a utility maintenance target based on the data model, using the utility loss evaluation index to evaluate the test results, and obtaining an average utility maintenance target evaluation result;
[0011] S3. Execute a differentiated selection strategy on the average utility maintenance goal evaluation results, select the optimal differential privacy mechanism, and obtain the feature space output result;
[0012] S4. Perform feature space mining based on the training set and the optimal differential privacy mechanism, and select the differential privacy mechanism with the highest global near-optimal probability as the recommended differential privacy mechanism.
[0013] Furthermore, the utility loss assessment indicators in S2 also include a first type of utility loss assessment indicator and a second type of utility loss assessment indicator; wherein, the first type of utility loss assessment indicator is the probability of data pattern change; and the second type of utility loss assessment indicator is the difference value of data pattern change.
[0014] Furthermore, the average utility maintenance target in S2 also includes an average probability of data pattern changes and an average difference value of data pattern changes.
[0015] Furthermore, the differentiated selection strategy is to determine that any differential privacy mechanism whose average utility loss is 10% higher than the optimal differential privacy mechanism will be considered the optimal differential privacy mechanism.
[0016] Furthermore, the step S4 further includes:
[0017] Step 4.1: Feature space mining further includes partitioning the feature space into subspaces. The feature space is partitioned so that the optimal differential privacy mechanism for the training data in the subspaces is close. After the spatial partitioning, the optimal differential privacy mechanism corresponding to the subspace is used as the recommendation result, which is used as the recommendation result for the differential privacy mechanism of the feature space, and subspace statistics are performed;
[0018] Step 4.2: Based on the user's multiple utility maintenance goals, integrate the output results of the spatial partitioning statistics and weightedly calculate the global near-optimality probability of each optimal differential privacy mechanism for all feature space output results. This probability value reflects the ability of the differential privacy mechanism to meet the utility maintenance goals of all users. Finally, the differential privacy mechanism with the highest global near-optimality probability is selected as the recommended differential privacy mechanism.
[0019] Furthermore, the data pattern in S1 further includes six types of data patterns, namely 1) special identification class, which describes the special identification of a data ; 2) difference class, which describes the degree of difference within a set of statistical release results; 3) morphological class, which describes the distribution of a set of statistical release results; 4) special value class; 5) clustering, which describes the clustering results of a set of statistical release results.
[0020] The beneficial effects of the present invention are as follows: under the same privacy budget and diversified utility loss evaluation indicators, a differentiated selection strategy is implemented on the average utility maintenance target evaluation results to select the optimal differential privacy mechanism, so the recommendation effect of the differential privacy mechanism is better than that of using any one mechanism all the time; when there are multiple utility loss targets, based on the training set and the optimal differential privacy mechanism, the differential privacy mechanism recommendation based on the user's multiple utility maintenance targets is realized, and finally the differential privacy mechanism with the highest global near-optimal probability is selected as the recommended differential privacy mechanism, thereby improving the versatility of the model. For any statistical data as the training set, the recommendation accuracy and stability of the differential privacy mechanism in complex query scenarios can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1This is an overall flow chart of a method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to an embodiment of the present invention;
[0022] Figure 2 This is a technical roadmap for implementation examples. DETAILED DESCRIPTION
[0023] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0024] Figure 1 The overall process of the multi-mode optimized differential privacy mechanism recommendation method for publishing statistical results proposed in this invention is shown, which specifically includes the following steps:
[0025] Step 1: randomly generating statistical result release data, filtering part of the statistical distribution result release data to obtain a training set, and dividing the training set into feature spaces according to specific data patterns; further comprising randomly generating statistical result release data, filtering the generated statistical result release data features by feature filtering so that specific features of the statistical data are distributed as evenly as possible, and using the filtered statistical distribution result release data features as the training set;
[0026] Step 2: Set the utility maintenance target based on the data model, use the utility loss evaluation index to evaluate the test results, and obtain the average utility maintenance target evaluation result; the specific process includes:
[0027] Step 2.1. Design two utility maintenance target quantification methods: i. Utility maintenance target quantification method based on consistency judgment. The first utility maintenance target is designed to be the probability of failing to maintain the specified data insight. The first type of utility loss assessment indicator is the probability of data pattern change, that is, the frequency of data pattern changes in statistical tests that violate the preset consistency threshold. ii. A utility maintenance target quantification method based on similarity judgment is designed. The second type of utility maintenance target is designed to be the similarity of the utility evaluation index between the original statistical release result and the statistical release result after adding noise. The second type of utility loss evaluation index is the difference value of the data pattern change, that is, the average deviation between the perturbation value and the true value. Specifically, the true value is the original statistical release result, and the perturbation value is the data after the statistical release result is subjected to the differential privacy algorithm. Specifically, the test result is the result of multiple rounds of testing based on the statistical release results of the training set. The differential privacy algorithm is performed on the training set data multiple times with different privacy budgets and different differential privacy mechanisms to achieve multiple tests. Multiple test results are obtained, that is, each test will be evaluated using all the utility loss evaluation indicators provided by the user to generate a corresponding probability distribution of the data pattern change, and the probability distribution is summarized as the final test result.
[0028] Step 2.2: Use each utility loss evaluation metric to evaluate the final test results. For each statistical release result, two utility loss evaluation metrics can be used. Multiple rounds of testing are conducted using each evaluation metric to obtain test results, which are then used to build the recommendation model. In actual use, the user needs to specify the type of utility evaluation metric, which will serve as input to the recommendation model.
[0029] Specifically, the average utility maintenance objective evaluation result includes the average probability of data pattern change and the average difference value of data pattern change. For utility maintenance, i.e., the first type of utility loss evaluation indicator, the probability of data pattern change serves as the evaluation indicator of utility loss, and the average utility loss represents the average probability value of data pattern change. For similarity indicators, i.e., the second type of utility loss evaluation indicator, the change value of the data pattern serves as the evaluation indicator of utility loss, and the average utility loss represents the average difference value of data pattern change. The main concepts are average probability and average difference value. Given that the effect of differential privacy is random, the utility loss result is obtained through multiple averaging.
[0030] Step 3: Perform a differentiated selection strategy on the average utility maintenance target evaluation results, and select the optimal differential privacy mechanism as the feature space output result. This strategy adopts a "near-optimal" strategy to select the current optimal differential privacy mechanism due to the consideration of the randomness of differential privacy.
[0031] Step 3.1: Design a differentiated selection strategy, that is, any differential privacy mechanism with an average utility loss 10% higher than the optimal differential privacy mechanism will be judged as the optimal differential privacy mechanism;
[0032] Step 3.2: Evaluate the impact of the distribution characteristics of the test dataset on the utility loss of different differential privacy mechanisms based on the test results.
[0033] Step 4: Based on the training set obtained in step 1 and the optimal differential privacy mechanism obtained in step 3, feature space mining is performed, and the differential privacy mechanism with the highest global near-optimal probability is finally selected as the recommended differential privacy mechanism; thus, the differential privacy mechanism recommendation based on the user's multi-utility maintenance goal is achieved.
[0034] Step 4.1: Feature space mining further includes partitioning the feature space into subspaces. The feature space is partitioned so that the optimal differential privacy mechanism for the training data in the subspaces is close. After the spatial partitioning, the optimal differential privacy mechanism corresponding to the subspace is used as the recommendation result, which is used as the recommendation result for the differential privacy mechanism of the feature space, and subspace statistics are performed;
[0035] Step 4.2: Based on the user's multiple utility maintenance goals, integrate the output results of the spatial partitioning statistics and weightedly calculate the global near-optimality probability of each optimal differential privacy mechanism for all feature space output results. This probability value reflects the ability of the differential privacy mechanism to meet the utility maintenance goals of all users. Finally, the differential privacy mechanism with the highest global near-optimality probability is selected as the recommended differential privacy mechanism.
[0036] For example, for each utility maintenance goal, the feature space of the training dataset is divided into multiple ranges using data distribution characteristics and privacy budget, and all range statistics are performed. Specifically, under the same utility evaluation metric, data with continuous feature values or privacy budgets in the feature space may have the same optimal differential privacy algorithm. For example, the optimal algorithm for data with a maximum value of 20-30 is the Laplace algorithm, so the range division is performed according to this rule.
[0037] Specifically, in this embodiment, the statistical result release data is divided into six categories of data patterns: 1) Special identification class, which describes the special identification of a data , such as the age range of the group with the highest income; 2) Difference class, which describes the degree of difference within a set of statistical release results, such as variance; 3) Morphology class, which describes the distribution of a set of statistical release results, such as skewness and long tail; 4) Special value class, such as maximum value and outlier; 5) Clustering, which describes the clustering results of a set of statistical release results. The input is a set of statistical release results, and the output is "text" (first category) or "value" (other categories). (There are only 5 categories here, please confirm again by the inventor.)
[0038] Figure 2 For the embodiment of the present invention, Figure 2 To know, the feature space of data distribution is composed of each test data. Multiple tests are performed using each differential privacy mechanism. The dots in the figure represent the results of a single test. The recommended differential privacy mechanism is: Mechanism 2.
[0039] S1. Quantify the utility loss target based on the data pattern. Design differentiation. Generate quantitative evaluation indicators based on consistency judgment according to the specified data pattern to support utility loss measurement.
[0040] S2. Training data synthesis and statistics: The feature space of data distribution is composed of each test data. Each mechanism is tested multiple times and the results of a single test are obtained. The feature space is a 3×2×2 multidimensional vector space.
[0041] S3. Multi-objective mechanism recommendation: Table 1 shows the comparison results of multi-objective mechanisms. Taking Table 1 as an example, the final recommended differential privacy mechanism is: Mechanism 2.
[0042] Table 1
[0043]
[0044] In summary, the present invention first constructs a multi-dimensional evaluation index system for privacy budget and utility loss; then, based on feature space mining and approximate optimal strategies, spatial partitioning statistics are performed for each evaluation index; finally, multi-objective recommendation results are fused through global near-optimal probability to achieve a differential privacy mechanism that recommends the best possible results under diverse utility maintenance objectives. The present invention constructs a composite multi-dimensional quantitative evaluation index system that integrates statistical similarity, structural consistency, and other differentiations to achieve a precise balance between privacy protection strength and data practicality; and, for multiple utility maintenance objectives, a scalable recommendation algorithm is designed. Through feature space mining and approximate optimal techniques, the algorithm's generalization ability in the entire data space is improved, ensuring the reliability of recommendations in complex query scenarios.
[0045] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications made without departing from the principles of the present invention should also be considered to fall within the scope of protection of the present invention.
Claims
1. A multi-mode optimized differential privacy mechanism recommendation method for publishing statistical results, characterized by: The following steps are involved: S1. Randomly generate statistical result release data, filter some statistical distribution result release data as a training set, and perform feature space division on the training set according to a specific data pattern; S2. Setting a utility maintenance target based on the data model, using the utility loss evaluation index to evaluate the test results, and obtaining an average utility maintenance target evaluation result; S3. Execute a differentiated selection strategy on the average utility maintenance goal evaluation results, select the optimal differential privacy mechanism, and obtain the feature space output result; S4. Perform feature space mining based on the training set and the optimal differential privacy mechanism, and select the differential privacy mechanism with the highest global near-optimal probability as the recommended differential privacy mechanism.
2. The method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to claim 1 is characterized in that: The utility loss evaluation indicators in S2 also include a first type of utility loss evaluation indicator and a second type of utility loss evaluation indicator; wherein, the first type of utility loss evaluation indicator is the probability of data pattern change; and the second type of utility loss evaluation indicator is the difference value of data pattern change.
3. The method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to claim 1, characterized in that: The average utility maintenance target in S2 also includes the average probability of data pattern changes and the average difference value of data pattern changes.
4. The method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to claim 1, characterized in that: The differential selection strategy is to determine that any differential privacy mechanism whose average utility loss is 10% higher than the optimal differential privacy mechanism will be considered the optimal differential privacy mechanism.
5. The method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to claim 1, characterized in that: The step S4 further comprises: Step 4.1: Feature space mining further includes partitioning the feature space into subspaces. The feature space is partitioned so that the optimal differential privacy mechanism for the training data in the subspaces is close. After the spatial partitioning, the optimal differential privacy mechanism corresponding to the subspace is used as the recommendation result, which is used as the recommendation result for the differential privacy mechanism of the feature space, and subspace statistics are performed; Step 4.2: Based on the user's multiple utility maintenance goals, integrate the output results of the spatial partitioning statistics and weightedly calculate the global near-optimality probability of each optimal differential privacy mechanism for all feature space output results. This probability value reflects the ability of the differential privacy mechanism to meet the utility maintenance goals of all users. Finally, the differential privacy mechanism with the highest global near-optimality probability is selected as the recommended differential privacy mechanism.
6. The method for recommending a multi-mode optimized differential privacy mechanism for publishing statistical results according to claim 1, characterized in that: The data pattern in S1 further includes six types of data patterns, namely 1) a special identification class, which describes the special identification of a data ; 2) a difference class, which describes the degree of difference within a set of statistical release results; 3) a morphology class, which describes the distribution of a set of statistical release results; 4) a special value class; 5) Clustering, describing the clustering results of a set of statistical release results.