A privacy protection method and system for data mining

Abstract

This invention discloses a privacy protection method and system for data mining. The privacy protection method includes: calculating the influence weight values ​​of all features in the dataset based on the logistic algorithm; classifying the influence levels based on the influence weight values ​​of the features; performing privacy protection processing on multi-dimensional features based on the gradient response algorithm (K-RR); dynamically allocating a user-defined privacy budget based on the feature influence level and feature dimension; constructing a privacy budget loss calculation based on different influence levels; performing privacy protection processing on two-dimensional features based on the stochastic response algorithm; dynamically allocating perturbation probabilities based on the feature influence level; observing the spatial distribution of features before and after perturbation; and verifying the accuracy of the model analysis before and after perturbation. This invention achieves a balance between privacy protection and data analysis accuracy, protecting users' privacy data while improving the accuracy of data analysis results.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, and more specifically to a privacy protection method and system for data mining. Background Technology

[0002] With the rapid development of artificial intelligence and machine learning technologies, data mining has transformed all aspects of people's lives, manifesting in reforms in society, education, and the economy. Businesses of all sizes across industries collect user data and analyze it to extract useful information for their own commercial purposes or to better serve users. However, because user data contains a large amount of sensitive and private information, data mining also carries the risk of privacy leaks. If this private information falls into the hands of malicious individuals and is used for illegal purposes such as online fraud or information tampering, it will seriously harm users' interests.

[0003] In 2006, Dwork proposed the differential privacy model, which presupposes that an attacker can possess maximum background knowledge, meaning the attacker can have all data records except for the target's private information. The differential privacy model is built on solid mathematical proofs and can quantify the level of privacy protection through the parameter ε. Currently, differential privacy technology has become an effective method to circumvent the assumption of attacker knowledge and a hot topic in data privacy protection research.

[0004] While differential privacy technology has wide applications and significant research value in data mining and analysis, the degree of protection provided by differential privacy and the accuracy of data analysis are often contradictory. Introducing differential privacy typically involves adding noise or perturbations to the dataset to protect user privacy; however, introducing inappropriate noise and perturbations can negatively impact the accuracy of subsequent data mining and analysis results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a privacy protection method and system for data mining that can achieve a balance between privacy protection and data analysis accuracy, thereby protecting users' privacy data while improving the accuracy of data analysis results.

[0006] This invention discloses a privacy protection method for data mining, comprising:

[0007] The influence weight values ​​of all features in the dataset are calculated based on the logistic algorithm.

[0008] Influence levels are classified based on the influence weight values ​​of features;

[0009] Based on the gradient response algorithm K-RR algorithm, multidimensional feature privacy protection is processed; in the process of multidimensional feature privacy protection, the user-defined privacy budget is dynamically allocated based on the feature influence level and feature dimension;

[0010] Based on different impact levels, a privacy budget loss calculation is constructed;

[0011] Based on the random response algorithm, privacy protection processing of two-dimensional features is performed; during the two-dimensional feature privacy protection process, the perturbation probability is dynamically allocated based on the feature influence level;

[0012] Observe the spatial distribution patterns of features before and after the disturbance;

[0013] Verify the accuracy of the model analysis before and after.

[0014] As a further improvement of the present invention, the calculation of the influence weight values ​​of all features in the dataset based on the logistic algorithm includes:

[0015] Obtain the dataset;

[0016] The weight value of each feature's influence on the data analysis results is calculated based on the logistic regression algorithm.

[0017] The influence weight values ​​are normalized and mapped to the [0,1] interval.

[0018] As a further improvement of the present invention, the calculation of the influence weight value of each feature on the data analysis result based on the logistic regression algorithm includes:

[0019] Initialize the weighting coefficients w for each feature in the loss function cost; where,

[0020]

[0021] y * The original label representing the data features, z = w0 + w i x1 + w2x2 + w3x3 + ... + w n x n , where x i The feature represented by the subscript i, w i x represents the index i i Weighting coefficients of features;

[0022] Update the weighting coefficients w for each feature. In the formula, alpha represents the learning rate. This represents the derivative of cost with respect to the variable w;

[0023] Iterative updates are performed, and the weighted coefficient that minimizes the cost of the loss function is taken as the influence weight value of the feature.

[0024] As a further improvement to the present invention, the criteria for classifying the degree of influence are as follows:

[0025] Influence weight values ​​in the range [0, 0.3] are classified as having low influence.

[0026] Influence weight values ​​in the range (0.3, 0.6) are classified as having moderate influence.

[0027] Influence weight values ​​in the range (0.6, 1) are classified as having high influence.

[0028] As a further improvement of the present invention, the gradient response algorithm-based K-RR algorithm provides privacy protection processing for multidimensional features; including:

[0029] by The true result of the probabilistic response, with The probability response is one of the remaining k-1 outcomes, which preserves as much information as possible from the original data while protecting privacy; where k represents the data dimension size and ε represents the privacy budget.

[0030] As a further improvement of the present invention, the dynamic allocation of the user-defined privacy budget based on the feature influence level and feature dimension includes:

[0031] Each feature variable with low influence is counted individually, and the set K1 = {k} is recorded, which constitutes the total number s1 of the feature variables with low influence. 11, k 12, ....,k 1i The set K2 = {k} consists of the total number of feature variables s2 with moderate influence and the dimensions that record each feature variable. 21, k 22, ....,k 2j The set K3 = {k} is composed of the total number of highly influential feature variables s3 and the dimensions that record each feature variable. 31, k 32, ....,k 3k}, forming set S t ={s1,s2,s3};

[0032] The total privacy budget allocated to feature variables with low impact Total privacy budget allocated to high-impact feature variables The total privacy budget ε is allocated to feature variables with moderate impact. sum2 =ε sum -(ε sum1 +ε sum3 The Max() function finds the maximum value in the set, and the Min() function finds the minimum value in the set.

[0033] The privacy budget assigned to each feature variable with low impact is ε. i for Where k i This indicates that the privacy budget allocated to each feature variable, which corresponds to a certain dimension and has a moderate impact, is ε. j for The privacy budget assigned to each high-impact feature variable is ε. k for

[0034] As a further improvement to the present invention, the formula for calculating the privacy budget loss is as follows:

[0035]

[0036] ε ij =ε i(j-1) -P ij

[0037] Where, ε i Representing the characteristic variable t i The total privacy budget is dynamically allocated, w i Representing the characteristic variable t i The weighted normalized values, j∈(1,2,3,...k), ε j This indicates that for the characteristic variable t i The budget for the j-th single analysis query, P ij Representing the characteristic variable t i Total budget for the j-th single analysis query;

[0038] Based on the privacy budget loss calculation, the characteristic variable t is derived. i The remaining total privacy budget for the j-th single analysis query is ε. ij .

[0039] As a further improvement of the present invention, the perturbation probability is dynamically allocated based on the feature influence level; including:

[0040] For a feature set with high influence, first obtain the lowest normalized value l1 and the highest normalized value h1 of the feature weights in the set. The lower limit of the perturbation probability p1 of the high-influence feature is 0.62+Round(l1 / 10,2), and the upper limit of p1 is 0.72+Round(h1 / 10,2), which is expressed as 0.62+Round(l1 / 10,2)≦p1≦0.72+Round(h1 / 10,2); where the Round(number,2) function means to keep two decimal places for the number number.

[0041] For a feature set with moderate influence, first obtain the lowest value l2 and the highest value h2 of the feature weights after normalization. The lower limit of the feature perturbation probability p2 with moderate influence is 0.54+Round(l2 / 10,2), and the upper limit of p1 is 0.62+Round(h2 / 10,2), which can be expressed as 0.54+Round(l2 / 10,2)≦p2≦0.62+Round(h2 / 10,2).

[0042] For a feature set with low impact, first obtain the lowest value l3 and the highest value h3 of the feature weights after normalization. The lower limit of the perturbation probability p3 of the feature with low impact is 0.45+Round(l3 / 10,2), and the upper limit of p1 is 0.52+Round(h3 / 10,2), which can be expressed as 0.45+Round(l3 / 10,2)≦p3≦0.50+Round(h3 / 10,2).

[0043] This invention also discloses a privacy protection system for data mining, used to implement the above-mentioned privacy protection method; comprising:

[0044] The modeling and analysis module is used to calculate the influence weight values ​​of all features in the dataset based on the logistic algorithm.

[0045] The ranking module is used to classify the influence level based on the influence weight value of the feature.

[0046] The privacy protection module is used for privacy protection processing of multidimensional features based on the gradient response algorithm K-RR algorithm;

[0047] The dynamic allocation module is used to dynamically allocate the user-defined privacy budget based on the feature impact level and feature dimension.

[0048] The loss calculation module is used to construct privacy budget loss calculations based on different impact levels;

[0049] The two-dimensional feature protection module is used to perform privacy protection processing on two-dimensional features based on the random response algorithm; during the two-dimensional feature privacy protection process, the perturbation probability is dynamically allocated based on the feature influence level;

[0050] The morphological distribution module is used to observe the spatial distribution morphology of features before and after the disturbance.

[0051] The accuracy analysis module is used to verify the accuracy of the model analysis before and after the model analysis.

[0052] As a further improvement of the present invention

[0053] The privacy protection module includes a distance calculation module and a distance weighting module, which are connected together.

[0054] The dynamic allocation module includes a dimension statistics module, a total privacy budget module, and a single privacy budget module. The dimension statistics module and the total privacy budget module are connected, and the total privacy budget module and the single privacy budget module are connected.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. This invention introduces a feature influence weight value, and performs a series of personalized privacy protection processes based on the feature influence weight value;

[0057] 2. This invention combines the dimensions of features and the magnitude of their influence weights to achieve a dynamic and reasonable allocation of the privacy budget, ensuring data utility;

[0058] 3. This invention constructs different privacy budget loss calculations based on the magnitude of the feature's influence to ensure privacy.

[0059] 4. This invention performs personalized perturbations on features of different dimensions, taking into account the magnitude of the influence weights and different perturbation algorithms. Attached Figure Description

[0060] Figure 1 This is a flowchart of the privacy protection method for data mining disclosed in this invention;

[0061] Figure 2 This is a schematic diagram showing that the characteristic variables of the present invention are not disturbed.

[0062] Figure 3 This is a schematic diagram illustrating the interference of the characteristic variables in this invention;

[0063] Figure 4 This is a schematic diagram illustrating the training and testing of the model analysis accuracy without the introduction of differential privacy protection in this invention.

[0064] Figure 5 This is a schematic diagram illustrating the training and testing of the model analysis accuracy of the present invention with differential privacy protection.

[0065] Figure 6 This is a framework diagram of the privacy protection system for data mining according to the present invention;

[0066] Figure 7 for Figure 6 Framework diagram of the privacy protection module;

[0067] Figure 8 for Figure 6 The framework diagram of the dynamic allocation module.

[0068] In the picture:

[0069] 1. Modeling and Analysis Module; 2. Hierarchy Classification Module; 3. Privacy Protection Module; 4. Dynamic Allocation Module; 5. Loss Calculation Module; 6. Two-Dimensional Feature Protection Module; 7. Morphological Distribution Module; 8. Accuracy Analysis Module; 31. Distance Calculation Module; 32. Distance Weighting Module; 41. Dimension Statistics Module; 42. Total Privacy Budget Module; 43. Single Privacy Budget Module. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] The present invention will now be described in further detail with reference to the accompanying drawings:

[0072] like Figure 1 As shown, this invention provides a privacy protection method for data mining, which constructs different privacy budget loss calculations based on the magnitude of feature influence to ensure privacy; including:

[0073] S1. Calculate the influence weight values ​​of all features in the dataset based on the logistic algorithm;

[0074] Specifically, it includes:

[0075] S11. Obtain the dataset.

[0076] S12. Calculate the weight value of each feature's influence on the data analysis results based on the logistic regression algorithm; the logistic regression algorithm processing procedure is as follows:

[0077] The logistic regression algorithm introduces the sigmoid function. y∈(0,1); The sigmoid function outputs probabilistic values, using a probability of 0.5 as the dividing line to classify data into positive examples (represented by 1) and negative examples (represented by 0). That is, given an input, what is the probability that the input will be mapped to a certain category or task under the mapping of this sigmoid function? When the probability is greater than 0.5, it means that the input is classified as a positive example; when it is less than 0.5, it means that the input is classified as a negative example. The sigmoid function is based on linear regression, where the function expression... The expression z in the equation is a linear expression: z = w0 + w1x1 + w2x2 + w3x3 + ... + w n x n , where x iThe feature represented by the subscript i, w i x represents the index i i The weighting coefficients of the features, also known as feature weights;

[0078] During the classification process, we continuously fit and approximate a "boundary" to ensure that data is classified according to this boundary. Furthermore, we need to minimize the loss function cost, i.e., minimize the cost of classification, which is equivalent to finding the optimal decision boundary. The loss function cost is:

[0079]

[0080] In the formula, y * The original label (0 or 1) representing the data feature is z expressed as a linear expression: w0 + w1x1 + w2x2 + w3x3 + ... + w n x n Instead; in the process of finding the optimal decision boundary, we hope to minimize the loss function to obtain the optimal decision boundary; by observing the loss function cos, we can see that the only variable parameter is the feature x with index i. i weighting coefficient w i Therefore, iterative algorithms using gradient descent are needed to find the optimal w. i The iterative process is as follows:

[0081] 1) Initialize the weighting coefficients w for each feature;

[0082] 2) Update w, Where alpha represents the learning rate. This represents the derivative of cost with respect to the variable w;

[0083] 3) Iterate and update until the optimal decision boundary is found.

[0084] When the optimal decision boundary is found, the weighting coefficient of each feature is the final optimal and most reasonable weighting coefficient. This weighting coefficient is the influence weight value of the feature, which directly affects the subsequent decision classification and analysis prediction results.

[0085] S13. Normalize the influence weight values, map them to the [0,1] interval, and record the normalized values ​​in the set in ascending order.

[0086] S2. Classify the degree of influence based on the influence weight values ​​of features;

[0087] The specific classification criteria are as follows:

[0088] The influence weight values ​​in the range [0, 0.3] are divided into low influence values, and a feature set L with low influence values ​​is constructed.

[0089] The influence weight values ​​in the range (0.3, 0.6) are classified as having moderate influence, and a feature set M with moderate influence is constructed.

[0090] The influence weight values ​​in the (0.6,1] interval are classified as high influence values, and a high influence feature set H is constructed.

[0091] For features with different levels of impact, in order to ensure the accuracy of subsequent data analysis, data perturbation cannot be done simply and directly. Therefore, it is necessary to dynamically allocate a user-defined privacy budget ε based on the feature's impact level and feature dimension, and to perform personalized differential privacy protection based on the allocated privacy budget.

[0092] S3. Based on the gradient response algorithm K-RR algorithm, multi-dimensional feature privacy protection processing is performed;

[0093] Specifically, it includes:

[0094] For the processing of multidimensional features, this invention, based on the localized differential privacy gradient response algorithm K-RR, introduces considerations regarding the weight of feature influence, feature dimension, and privacy budget allocation to ensure the accuracy of subsequent data analysis as much as possible. Specifically, the K-RR algorithm can select the true result as the output with a certain probability, and when the true result cannot be selected, it selects other results that are close to the true result as the output with a certain probability. This way, privacy is protected while retaining as much information as possible from the original data, thus ensuring the utility of the data to a certain extent.

[0095] The algorithm process is as follows:

[0096] 1) Calculate the distance between each possible output and the original data using Hamiltonian or Euclidean distance measurement methods;

[0097] 2) These distances are weighted using an exponential function. The weighting method is as follows: for each result r, calculate the distance between it and the true result, and divide it by the privacy parameter ε to obtain an exponential value.

[0098]

[0099] Where dist(r,true) represents the distance between the result r and the true result; this index value can be regarded as a score for each result, and the higher the score, the more likely the result is to be selected;

[0100] by The probability response is the true result; if no true result is selected, then it is... Other results that are close to the true result are used as outputs, thus preserving as much information as possible from the original data while protecting privacy;

[0101] k-RR algorithm with The true result of the probabilistic response, with The probability response is one of the remaining k-1 outcomes, where k represents the data dimension and ε represents the privacy budget. Based on the probability formula, it can be found that the probability of sending a true value is directly proportional to the size of the privacy budget ε and inversely proportional to the dimension k, while the probability of sending a false value is inversely proportional to the size of the privacy budget ε. Therefore, the size of the privacy budget ε allocated to this feature variable directly affects the accuracy of subsequent data analysis. For features with high influence, if the allocated privacy budget is too small, resulting in excessive perturbation fluctuations, it will directly affect the accuracy of subsequent data analysis.

[0102] Currently, most differential privacy protection schemes use a pre-set total privacy budget for users, failing to consider that different data may present different privacy protection needs for each user. Therefore, to meet the personalized privacy protection needs of each user, this invention allows users to customize their own total privacy budget ε. sum Size, and based on the feature variables in S2 that are divided into different influence levels according to their weight values, and the feature dimensions, the privacy budget ε is affected. sum Dynamic and reasonable allocation is crucial. Most existing differential privacy dynamic data publishing methods employ simplistic approaches to allocate the privacy budget, such as distributing it evenly, incrementally, or decreasingly across the sampling points to be published. However, they fail to rationally allocate the limited privacy budget based on the characteristics of dynamic data, leading to premature depletion or waste of the privacy budget. Privacy budget ε sum The allocation of privacy is closely related to the utility of data. In particular, during the release of dynamic data, if the limited privacy budget is not allocated reasonably, the utility of dynamic data may be reduced.

[0103] S4. In the process of privacy protection, based on the feature impact level and feature dimension, the user-defined privacy budget is dynamically allocated to ensure data utility.

[0104] Privacy Budget ε sum The core principle of the dynamic allocation process is to allocate more privacy budget to high-impact, high-dimensional features and reduce the degree of disturbance to features with high impact, so as to ensure the accuracy of subsequent data analysis as much as possible.

[0105] Specifically, it includes:

[0106] 1) Calculate the total number s1 of feature variables with low influence and record the dimensions of each feature variable to form a set K1 = {k 11, k 12, ....,k 1iThe set K2 = {k} consists of the total number of feature variables s2 with moderate influence and the dimensions that record each feature variable. 21, k 22, ....,k 2j The set K3 = {k} is composed of the total number of highly influential feature variables s3 and the dimensions that record each feature variable. 31, k 32, ....,k 3k}, forming set S t ={s1,s2,s3};

[0107] 2) Total privacy budget allocated to feature variables with low impact Total privacy budget allocated to high-impact feature variables The total privacy budget ε is allocated to feature variables with moderate impact. sum2 =ε sum -(ε sum1 +ε su The Max() function finds the maximum value in the set, and the Min() function finds the minimum value in the set.

[0108] 3) The privacy budget allocated to each feature variable with low impact is ε. i for Where k i This indicates that the privacy budget allocated to each feature variable, which corresponds to a certain dimension and has a moderate impact, is ε. j for The privacy budget assigned to each high-impact feature variable is ε. k for

[0109] S5. Based on different impact levels, construct a privacy budget loss calculation to ensure privacy.

[0110] The core principle of privacy budget loss calculation is: features with higher impact suffer privacy loss faster, when feature variable t i When the dynamically allocated total privacy budget consumption is 0, this invention returns only a fixed value for each subsequent analysis query, ensuring that attackers cannot obtain any valuable results through the analysis of this user record.

[0111] Specifically, it includes:

[0112] Privacy budgeting based on characteristic variables within different impact levels ε sum Reasonable dynamic allocation ensures that features with high impact are allocated as much privacy budget ε as possible. iThis reduces the perturbation of features with high influence, but at the cost of sacrificing the privacy of those features. To prevent malicious attacks and privacy leaks, this invention constructs a privacy budget loss calculation based on different influence levels:

[0113]

[0114] ε ij =ε i(j-1) -P ij

[0115] Where, ε i Representing the characteristic variable t i The total privacy budget is dynamically allocated, w i Representing the characteristic variable t i The weighted normalized values, j∈(1,2,3,...k), ε j This indicates that for the characteristic variable t i The budget for the j-th single analysis query, P ij Representing the characteristic variable t i The total budget for the j-th single analytical query; based on the privacy budget loss calculation, the feature variable t is derived. i The remaining total privacy budget for the j-th single analysis query is ε. ij .

[0116] S6. Based on the random response algorithm, two-dimensional feature privacy protection processing is performed; during the two-dimensional feature privacy protection process, the perturbation probability is dynamically allocated based on the feature influence level;

[0117] Specifically, it includes:

[0118] Based on the idea of ​​a localized differential privacy random response algorithm, perturbation is performed according to probability p, sending true values ​​with probability p and false values ​​with probability 1-p. However, considering the accuracy of subsequent data analysis, we should not specify a uniform probability p for perturbation. Instead, we should consider the influence level of the feature variable. For example, for feature variables with high influence, if the p value is set low, the value of the feature variable will be perturbed more strongly, but this will have a relatively large impact on the accuracy of subsequent analysis. Conversely, if the p value is set high, privacy will be compromised. Therefore, this scheme dynamically specifies the upper and lower thresholds of p according to the influence level of the feature, l≦p≦h. The process of dynamically specifying the upper and lower thresholds of probability p for features with different influence levels is as follows:

[0119] 1) For a feature set with high influence, first obtain the lowest value l1 and the highest value h1 of the feature weights after normalization. The lower limit of the perturbation probability p1 of the feature with high influence is 0.62+Round(l1 / 10,2), and the upper limit of p1 is 0.72+Round(h1 / 10,2), which is expressed as 0.62+Round(l1 / 10,2)≦p1≦0.72+Round(h1 / 10,2); where the Round(number,2) function means to keep two decimal places for the number number.

[0120] 2) For a feature set with moderate influence, first obtain the lowest value l2 and the highest value h2 of the feature weights after normalization. The lower limit of the feature perturbation probability p2 with moderate influence is 0.54+Round(l2 / 10,2), and the upper limit of p1 is 0.62+Round(h2 / 10,2). Therefore, it can be expressed as 0.54+Round(l2 / 10,2)≦p2≦0.62+Round(h2 / 10,2).

[0121] 3) For a feature set with low influence, first obtain the lowest value l3 and the highest value h3 of the feature weights after normalization. The lower limit of the perturbation probability p3 of the feature with low influence is 0.45+Round(l3 / 10,2), and the upper limit of p1 is 0.52+Round(h3 / 10,2), which is expressed as 0.45+Round(l3 / 10,2)≦p3≦0.50+Round(h3 / 10,2).

[0122] S7. Observe the spatial distribution pattern of the features before and after the disturbance;

[0123] Specifically, it includes:

[0124] Simulation experiments were conducted on a medical image dataset containing 1702 feature labels and 492 original patient data. The spatial distribution of features before and after perturbation was observed. Figure 2 , 3 As shown, it is required that, while ensuring privacy, the spatial distribution pattern before and after the feature should not differ significantly, because it is necessary to ensure data availability and the accuracy of subsequent analysis results.

[0125] S8. Verify the accuracy of the model analysis before and after;

[0126] Specifically, it includes:

[0127] Experiments were conducted to train and test the data analysis accuracy of the model without differential privacy protection and the model with differential privacy protection, respectively, and ROC curves were constructed. Figure 4 , 5As shown, the area scores of the experimental groups remained relatively stable at around 0.02, indicating that the accuracy of data analysis was well guaranteed while ensuring user privacy.

[0128] like Figure 6-8 As shown, this invention provides a privacy protection system for data mining, comprising: a modeling and analysis module 1, a hierarchy classification module 2, a privacy protection module 3, a dynamic allocation module 4, a loss calculation module 5, a two-dimensional feature protection module 6, a morphological distribution module 7, and an accuracy analysis module 8; wherein,

[0129] The modeling and analysis module 1 is connected to the level division module 2, the level division module 2 is connected to the privacy protection module 3, the privacy protection module 3 is connected to the dynamic allocation module 4, the dynamic allocation module 4 is connected to the loss calculation module 5, the loss calculation module 5 is connected to the two-dimensional feature protection module 6, the two-dimensional feature protection module 6 is connected to the morphological distribution module 7, and the morphological distribution module 7 is connected to the accuracy analysis module 8.

[0130] The privacy protection module 3 includes a distance calculation module 31 and a distance weighting module 32, which are connected together.

[0131] The dynamic allocation module 4 includes a dimension statistics module 41, a total privacy budget module 42, and a single privacy budget module 43. The dimension statistics module 41 and the total privacy budget module 42 are connected, and the total privacy budget module 42 and the single privacy budget module 43 are connected.

[0132] Specifically:

[0133] The modeling and analysis module 1 of the present invention is used to implement the above-mentioned S1;

[0134] The grade division module 2 of the present invention is used to implement the above-mentioned S2;

[0135] The privacy protection module 3 of the present invention is used to implement the above-described S3;

[0136] The dynamic allocation module 4 of the present invention is used to implement the above-described S4;

[0137] The loss calculation module 5 of the present invention is used to implement the above-described S5;

[0138] The two-dimensional feature protection module 6 of the present invention is used to implement the above-described S6;

[0139] The morphology distribution module 7 of the present invention is used to implement the above-described S7;

[0140] The accuracy analysis module 8 of the present invention is used to implement the above-described S8.

[0141] The advantages of this invention are:

[0142] 1. This invention introduces a feature influence weight value, and performs a series of personalized privacy protection processes based on the feature influence weight value;

[0143] 2. This invention combines the dimensions of features and the magnitude of their influence weights to achieve a dynamic and reasonable allocation of the privacy budget, ensuring data utility;

[0144] 3. This invention constructs different privacy budget loss calculations based on the magnitude of the feature's influence to ensure privacy.

[0145] 4. This invention performs personalized perturbations on features of different dimensions, taking into account the magnitude of the influence weights and different perturbation algorithms.

[0146] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A privacy protection method for data mining, characterized by, The method comprises the following steps: calculating the influence weight value of all features in the data set based on the logistic algorithm; wherein, the calculation process of the influence weight value of each feature on the data analysis result comprises the following steps: initializing the weighting coefficient w of each feature in the loss function cost; wherein, y * represents the label of the original data feature, z = w0+ w1x1+ w2x2+ w3x3+... + w n x n , where x i represents the feature with subscript i, w i represents the weighting coefficient of x i with subscript i; updating the weighting coefficients w of the various features, where alpha represents the learning rate, represents the derivative of cost with respect to the variable w; iteratively updating the weighting coefficient to the minimum value of the loss function cost as the influence weight value of the feature; performing influence degree level division based on the influence weight value of the feature; performing multi-dimensional feature privacy protection based on the gradient response algorithm K-RR algorithm; in the multi-dimensional feature privacy protection process, the user-defined privacy budget is dynamically allocated based on the feature influence degree level and the feature dimension; specifically comprising: respectively, and record the dimensionality composition set of each feature variable K1 = {k 11 ,k 12 ,....,k 1i}, the total number of feature variables with moderate influence degree s2, and record the dimensionality composition set of each feature variable K2 = {k 21 ,k 22 ,....,k 2j}, the total number of feature variables with high influence degree s3, and record the dimensionality composition set of each feature variable K3 = {k 31 ,k 32 ,....,k 3k}, form a set S t = {s1, s2, s3}; total privacy budget allocated to feature variables with low influence degree total privacy budget allocated to feature variables with high influence degree total privacy budget allocated to feature variables with moderate influence degree sum2 = ε sum -(ε sum1 + ε sum3 ), where the Max() function represents the maximum value in a set, and the Min() function represents the minimum value in a set. Each feature variable with low influence degree is assigned a privacy budget of ε i is where k i represents the dimension size corresponding to the feature variable, each feature variable with moderate influence degree is assigned a privacy budget of ε j is Each feature variable with high influence degree is assigned a privacy budget of ε k is constructing a privacy budget loss calculation based on different influence degree levels; performing two-dimensional feature privacy protection based on the random response algorithm; in the two-dimensional feature privacy protection process, the perturbation probability is dynamically allocated based on the feature influence degree level; observing the spatial distribution form of the features before and after perturbation; verifying the model analysis accuracy before and after.

2. The privacy preserving method of claim 1, wherein, The method for calculating the influence weight value of all features in the data set based on the logistic algorithm comprises the following steps: obtaining the data set; calculating the influence weight value of each feature on the data analysis result based on the logistic regression algorithm; normalizing the influence weight value and mapping it to the [0, 1] interval.

3. The privacy preserving method of claim 2, wherein, The division standard of the influence degree level is: the influence weight value in the [0, 0.3] interval is divided into low influence degree; the influence weight value in the (0.3, 0.6] interval is divided into moderate influence degree; the influence weight value in the (0.6, 1] interval is divided into high influence degree.

4. The privacy preserving method of claim 3, wherein, The method for performing multi-dimensional feature privacy protection based on the gradient response algorithm K-RR algorithm comprises the following steps: To probabilistically respond to the true result, and probabilistically respond to one of the remaining k-1 results, while preserving as much information in the original data as possible while protecting privacy; where k represents the data dimension size and ε represents the privacy budget.

5. The privacy preserving method of claim 4, wherein, The formula for privacy budget loss calculation is: e ij = e i(j-1) - P ij where ε i denotes the characteristic variable t i total privacy budget dynamically allocated, w i denotes the characteristic variable t i weight normalized value of the characteristic variable t j denotes the characteristic variable t i budget of the jth single analysis query, P ij denotes the characteristic variable t i total budget of the jth single analysis query; According to the privacy budget loss calculation, the feature variable t i The total privacy budget remaining after the jth single analysis query is ε ij .

6. The privacy preserving method of claim 3, wherein, The perturbation probability is dynamically allocated based on the feature influence degree level; comprising: for the feature set with high influence degree, first obtain the lowest value l1 and the highest value h1 of the feature weight normalization in the set, the lower limit of the perturbation probability p1 of the feature with high influence degree is 0.62+Round(l1 / 10,2), and the upper limit of p1 is 0.72+Round(h1 / 10,2), that is, 0.62+Round(l1 / 10,2)≦p1≦0.72+Round(h1 / 10,2); wherein, the Round(number,2) function means that the number number is kept to two decimal places; for the feature set with moderate influence degree, first obtain the lowest value l2 and the highest value h2 of the feature weight normalization in the set, the lower limit of the perturbation probability p2 of the feature with moderate influence degree is 0.54+Round(l2 / 10,2), and the upper limit of p1 is 0.62+Round(h2 / 10,2), that is, 0.54+Round(l2 / 10,2)≦p2≦0.62+Round(h2 / 10,2); For the feature set with low influence degree, first, the lowest value l3 and the highest value h3 of the feature weight normalization in the set are obtained, the lower limit of the disturbance probability p3 of the feature with low influence degree is 0.45+Round(l3 / 10,2), and the upper limit of p1 is 0.52+Round(h3 / 10,2), that is, 0.45+Round(l3 / 10,2)≦p3≦0.50+Round(h3 / 10,2).

7. A privacy protection system for data mining, for implementing the privacy protection method according to any one of claims 1 to 6; characterized in that, Comprise: The modeling analysis module is used for calculating the influence weight value of all features in the data set based on the logistic algorithm; The grade division module is used for influence degree grade division based on the influence weight value of the feature; The privacy protection module is used for multi-dimensional feature privacy protection processing based on the gradient response algorithm K-RR algorithm; The dynamic allocation module is used for dynamic allocation of user-defined privacy budget based on the feature influence degree grade and feature dimension; The loss calculation module is used for constructing privacy budget loss calculation based on different influence degree grades; The two-dimensional feature protection module is used for two-dimensional feature privacy protection processing based on the random response algorithm; in the two-dimensional feature privacy protection process, the disturbance probability is dynamically allocated based on the feature influence degree grade; The morphological distribution module is used for observing the spatial distribution form of the feature before and after disturbance; The accuracy analysis module is used for verifying the model analysis accuracy before and after.

8. The privacy protection system of claim 7, wherein the privacy protection module comprises a distance calculation module and a distance weighting module, and the distance calculation module and the distance weighting module are connected; the dynamic allocation module comprises a dimension statistics module, a total privacy budget module, and a single privacy budget module, and the dimension statistics module and the total privacy budget module are connected, and the total privacy budget module and the single privacy budget module are connected. ​ ​

Images