Aerodynamic prediction method and data mining system based on privacy protection
By adding noise to aerodynamic prediction in a distributed environment and calculating nonlinear cores, sharing cores to establish an aerodynamic prediction model, the difficulties of data privacy protection and data file analysis in the aerodynamic field are solved, and efficient privacy protection and data mining are achieved. .
Patent Information
- Application Number
- CN202111272950.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-10-29
AI Technical Summary
The existing technology is difficult to effectively protect data privacy in the pneumatic field, especially when multi-party computing is collaborative, there is a problem of data leakage, and the existing privacy protection methods cannot meet the needs of the pneumatic field for data file analysis and processing.
By adding noise obeying normal distribution to each participant in a distributed environment, generating data sets with perturbations, and computing local nonlinear cores, sharing these cores to calculate global cores, and finally establishing an aerodynamic prediction model, realizing Privacy protection and data mining.
It effectively protects data privacy, prevents external nodes from obtaining original data, and improves computing speed. Especially when the number of classification samples is large and the number of dimensions is high, it is suitable for discrete and nonlinear data, and has better adaptability.
Smart Images

Figure CN114036862B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data mining, and more specifically, to an aerodynamic prediction method and a data mining system for performing regression prediction in a distributed environment under the premise of protecting privacy. Background Art
[0002] With the development of big data, all walks of life are generating and storing a large amount of data every moment. The high concentration and sharing of data has greatly facilitated the cooperation and research between organizations. At the same time, it has also increased the risk of privacy information leakage. Data mining is an extremely powerful data analysis tool. With the help of various data mining algorithms, it can discover potentially useful knowledge and rules in the data, and even dig out the huge economic and political interests hidden behind the data. Privacy-preserving data mining is a research branch in the field of data mining. It organically combines data mining technology with privacy-preserving technology, and integrates knowledge and technology from multiple fields such as databases, artificial intelligence, pattern recognition, and machine learning. The main purpose of privacy-preserving data mining is to obtain more accurate data mining results as much as possible under the premise of reasonable protection of privacy data or sensitive rules.
[0003] The main idea of protecting the privacy of raw data is to change the raw data set through predefined transformations and provide the transformed data to the miner for data mining. This idea is used in privacy-preserving data mining. Therefore, how to provide a data mining method and system to protect the raw data set to solve the defects of the prior art such as easy leakage of private data or information and low confidentiality of information has become a technical problem that practitioners in this field need to solve.
[0004] In the field of pneumatics, when collaborative calculations among multiple parties are required, some important data cannot be directly used by third parties, resulting in data leakage.
[0005] In the prior art, there are also methods for protecting data privacy. For example, the patent title is: "A privacy-preserving data mining system and method based on medical big data", which focuses on building a big data medical system to ensure that private data will not be obtained by malicious third parties. However, if the verification is correct, the provincial cloud platform can send medical data to the lower-level regional platform, that is, the data between the regional platforms are interactive, and the privacy protection effect cannot meet the requirements;
[0006] In addition, there is another patent named "A data mining method and system for protecting the association rules of original transaction data sets". This patent targets the mining of association rules in the data mining system, but cannot achieve prediction and cannot meet the needs of the pneumatic field for data file analysis and processing. Summary of the invention
[0007] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.
[0008] In order to achieve these purposes and other advantages according to the present invention, a method for aerodynamic force prediction based on privacy protection is provided, comprising:
[0009] Step 1: Each participant adds noise that follows a normal distribution to the local original data file set to obtain a disturbed data set;
[0010] Step 2: Each participant calculates the local nonlinear kernel based on the obtained local perturbation data set;
[0011] Step 3: All participating parties set their calculated local nonlinear kernels to be shared, so that the host computer calculates the corresponding global kernel based on the shared local nonlinear kernels;
[0012] Step 4: Solve the optimization problem based on the obtained global kernel function to obtain the optimal solution of the parameters;
[0013] Step 5: Use the optimal parameter solution obtained in step 4 to establish a prediction model and perform aerodynamic prediction.
[0014] Preferably, in step 1, the process of acquiring the data set is configured to include:
[0015] S10, in the linear centered support vector machine PSVM for discrete data, a kernel function is introduced to adjust the linear PSVM to a nonlinear PSVM, and the PSVM is adjusted to a nonlinear centered support vector machine PSVR that can be applicable to continuous data.
[0016] S11, random perturbation technology is introduced in PSVR, and a perturbed data set is obtained by adding noise that obeys the normal distribution to the local original data set.
[0017] Preferably, the method of introducing the kernel function in PSVM is configured to include:
[0018] S101, in PSVM, the class center of two types of samples is defined as the mean of the two types of samples, the positive class and the negative class are X+ and X- respectively, and the mean is recorded as The original optimization problem is as follows:
[0019]
[0020] st(ω·x i )+by i ≤ε, i=1,2,...l;
[0021] y i -(ω·xi )-b≤ε, i=1,2,...,l;
[0022]
[0023] By making the objective function of the original optimization problem strictly convex, the central hyperplane corresponding to each type of point is obtained;
[0024] The dual problem of the optimization problem obtained by introducing the Lagrangian function is:
[0025]
[0026] stα i ,β i ≥0,i=1,2,...,ly i ∈R
[0027] Among them, α = (α1, α2,..., α l ) T , β=(β1,β2,...,β l ) T is the Lagrange multiplier.
[0028] Assume α * =(α1,α2,...,α l ,β1,β2,...,β l ) is the optimal solution to the optimization problem, then f(x)=ω *T ·x+b *
[0029] in
[0030] S102, based on PSVM, we can get the optimization problem of PSVR. By introducing the kernel function in the linear regression method, the above original optimization problem is adjusted to:
[0031]
[0032] stα i ,β i ≥0,i=1,2,...,ly i ∈R;
[0033] Among them, K() is the kernel function.
[0034] Preferably, in step 2, the method for obtaining the local nonlinear kernel is configured to include:
[0035] S20, based on Formula A i =A i +N iAll participants in PSVR generate corresponding protection data with disturbances, where A i , represents the data set of each participant, N i represents the added noise that follows a normal distribution;
[0036] S21, calculating the local nonlinear kernel K for the data of each participant, and sharing the nonlinear kernel among all the participants.
[0037] Preferably, in step three, when the distributed database system uses SVR for regression, the global kernel can be obtained by first calculating the local kernel of each node and then summing the local kernels of each node.
[0038] Preferably, in step 5, the optimal solution obtained in step 4 is used to establish a corresponding prediction model;
[0039] Wherein, the prediction model is:
[0040]
[0041] in,
[0042] A data mining system for aerodynamic force prediction, comprising:
[0043] A privacy protection module, used for converting the original data set into an authorized data set according to a pre-made interference adding strategy;
[0044] In the joint calculation module, each participant calculates the local kernel based on the transformed data set, and the local sum is used to derive the global kernel and calculate the optimal solution;
[0045] The prediction module establishes an aerodynamic prediction model to analyze and predict the aerodynamic force based on the optimal solution jointly calculated by all participating parties.
[0046] The present invention includes at least the following beneficial effects: First, the present invention adds noise to each node in the distributed system, thereby generating a perturbed data set and a perturbed local kernel, so that external nodes cannot obtain the original data of the node, thereby playing a role in privacy protection.
[0047] The method in this paper is that each participant will not obtain the original data of other participants. In the process of cooperative computing, each participant only needs to provide the interference core of its own node to ensure that the data of its own node will not be obtained by other nodes.
[0048] Secondly, the method of the present invention is faster than the common regression algorithm in terms of operation speed, and when the number of classification samples is large and the dimension is high, the advantage of operation speed is more obvious.
[0049] Thirdly, the present invention introduces a nonlinear kernel, so that it is suitable for discrete data and nonlinear data and has better adaptability.
[0050] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 This is a schematic diagram of node data leakage when the privacy protection of the regression algorithm is tested on the basis of the original algorithm. DETAILED DESCRIPTION
[0052] The present invention is further described in detail below so that those skilled in the art can implement it according to the description.
[0053] The aerodynamic force prediction method based on privacy protection in the present invention comprises the following steps:
[0054] (1) All participants generate perturbed datasets, A i =A i +N i , where N i represents the added noise that follows a normal distribution, A i The original data of each participant.
[0055] (2) Each participant calculates the local nonlinear kernel K(A) according to formula 1 i , A i T ); where A i T Here we use the Gaussian radial basis kernel, as shown in Formula 1. Users can choose other kernel functions according to actual conditions.
[0056] K=exp(-||x-x'|| 2 / δ 2 ) (Formula 1)
[0057] Among them, x, x' represent the samples in the data set, δ represents the parameters of the Gaussian kernel function, and the parameters can be selected according to the actual situation.
[0058] (3) All participants share the nonlinear kernel obtained in step 2. The central node sums the kernels shared by all participants to obtain the global kernel.
[0059] In a distributed environment, each node is a data holder, so each node must ensure the privacy of its own data before aggregating data to the data center. At the same time, when performing collaborative computing among distributed nodes, each node must also prevent information leakage between each other. Therefore, from the perspective of data flow, each holder must use effective means to ensure data privacy before data flows out.
[0060] We introduce random perturbation technology into the model, adding noise that follows a normal distribution to the data set to protect the original information. Thus, we obtain a perturbed data set and a perturbed local kernel function to implement a privacy-preserving regression prediction method.
[0061] Furthermore, in a distributed system, we can calculate the global kernel from the local kernel.
[0062] K is the kernel function, then there exists a kernel function from χ={x1,x2,...,x m} to R m The mapping of
[0063]
[0064] Assume that there are two participants and their cores are:
[0065] Below we discuss two situations:
[0066] Case 1: For a vertically distributed distributed database, the global core can be expressed as:
[0067]
[0068] Therefore, for vertically distributed distributed databases, the global kernel can be obtained through the local kernel.
[0069] Case 2: For a horizontal distributed database system, we use the Gran matrix of the row vector of the matrix. Then the Gram matrix of each node can be expressed as follows.
[0070] We assume that there are two nodes, P1 and P2. P1 contains m1 records, and P2 protects m2 records. Each record contains all attributes. Then the data in P1 can be represented as a matrix of m1×n, and the data in P2 can be represented as a matrix of m2×n, m1+m2=m. Therefore, we cannot use the local SVM mode. We need the global SVM mode. The local kernels corresponding to P1 and P2 are expressed as follows.
[0071]
[0072]
[0073] Then the global kernel:
[0074]
[0075] From the above, we can see that the global kernel of distributed data is K = K1 + K2. When a distributed database system uses SVR for regression, the global kernel A can be obtained by first finding the local kernel of each node and then summing the local kernels of each node.
[0076] (5) Solve the optimization problem and obtain the optimal solution.
[0077] Specifically, firstly, the present invention derives a central support vector regression method based on a linear central support vector machine PSVM.
[0078] Generally speaking, the center support vector machine (PSVM) is for discrete data. We need to improve the algorithm to make it applicable to continuous data. PSVM follows the principle of structural risk minimization and finds an optimal differentiation hyperplane. The two types of points can be separated as much as possible, and based on PSVM we can get the optimization problem of PSVR.
[0079] By defining the class center of the two types of samples as the mean of the two types of samples, the positive and negative classes are X+ and X- respectively, and the mean is recorded as The original optimization problem is:
[0080]
[0081] The optimization problem satisfies the following constraints:
[0082] st(ω·x i )+by i ≤ε, i=1,2,...l;
[0083] y i -(ω·x i )-b≤ε, i=1,2,...,l;
[0084]
[0085] Where ω is the weight vector, b is the threshold, ε is the relaxation factor, and st means subject to, which means satisfying the constraints.
[0086] This change makes the objective function strictly convex and changes the nature of the optimization problem. The hyperplane is no longer a supporting hyperplane, but a central hyperplane passing through each type of point.
[0087] Furthermore, a nonlinear center vector support regression method is calculated based on the nonlinear center support vector machine, that is, the linear center support vector machine is extended to the nonlinear center vector regression method by introducing a kernel function to increase the adaptability of the algorithm.
[0088] The above linear regression method is extended to deal with nonlinear regression problems by introducing kernel functions. Then the above optimization problem becomes:
[0089]
[0090] The optimization problem needs to satisfy the following constraints:
[0091] α i ,β i ≥0,i=1,2,...,ly i ∈R
[0092] Among them, x i ,y i represents the sample points in the data set, where α = (α1, α2, ..., α l ) T , β=(β1,β2,...,β l ) T is the Lagrange multiplier.
[0093] (5) Using the optimal parameters obtained in step 4, a prediction model is established. The mathematical model is shown in formula (2). Aerodynamic force prediction is performed according to formula (2).
[0094] f(x)=ω *T ·x+b * (Formula 2)
[0095] in
[0096] A data mining system for aerodynamic force prediction, comprising:
[0097] A privacy protection module, used for converting the original data set into an authorized data set according to a pre-made interference adding strategy;
[0098] In the joint calculation module, each participant calculates the local kernel based on the transformed data set, and the local sum is used to derive the global kernel and calculate the optimal solution;
[0099] The prediction module establishes an aerodynamic prediction model to analyze and predict the aerodynamic force based on the optimal solution jointly calculated by all participating parties.
[0100] Furthermore, during the communication process among the participants, each participant only needs to share their own local core and does not need to share the original data. Even if other participants conduct a collusion attack, they cannot obtain the original data of the participant, which further protects data privacy. This method has good performance in predicting aerodynamics, while also protecting the privacy data of each participant.
[0101] Embodiment 1:
[0102] Based on the original algorithm, the privacy protection of the regression algorithm is tested. In the distributed database classification mining algorithm, we assume that there are 200 distributed nodes in each node. In the experiment, it is assumed that the data is split into m = 3, 5, 10, 20, 30. The data leakage of a certain node is as follows Figure 1 As shown. Figure 1 We can see that as the number of collusion nodes increases, the probability of data leakage increases. However, as m increases, the probability of data leakage becomes smaller and smaller. When there are fewer colluders, the probability of information being exposed is basically equal to zero. When there are more collusion nodes, the probability of being exposed decreases as m increases. Even if data leakage occurs, only the local perturbation core of each node is leaked, and the original data of each node will not be leaked. Therefore, we can conclude that this algorithm can meet the requirements of secure regression mining in distributed databases.
[0103] Numerical experiments were conducted using six data sets from the UCI database. The six data sets are Concrete CS, Boston housing, Machine CPU, Auto price, Wisconsin BC, and Auto-Mpg. These are commonly used data sets for testing the pros and cons of regression algorithms. The simulation results are shown in Table 1-2:
[0104] Table 1 RSEM root mean square error
[0105]
[0106] Table 2 Operation time
[0107]
[0108]
[0109] From the simulation results, it can be seen that the root mean square error (RMSE) of the PPSVR and PSVR algorithms is slightly different from the RSEM of the SVR algorithm, and is even higher than the SVR algorithm on some data sets. However, in terms of computing speed, the PPSVR and PSVR algorithms are faster than the SVR algorithm, and when the number of classification samples is large and the dimension is high, the advantage of computing speed is more obvious.
[0110] Embodiment 2:
[0111] The data mining method for aerodynamic force prediction based on privacy protection described in this embodiment is applied in the following embodiment, if the collaborators are A and B. A and B each have an original data set D1 and D2, and A and B both want to use D1 and D2 for aerodynamic force prediction. A and B do not want to directly provide the original transaction data set to each other for information privacy considerations. At this time, A and B must use some technical method to transform the original transaction data sets D1 and D2 into authorized data sets D1' and D2'. A and B can use the authorized data sets D1' and D2' for aerodynamic force prediction.
[0112] The privacy-preserving data mining method based on aerodynamic prediction includes the following steps:
[0113] Step 1: Party A and Party B convert the original transaction data sets D1 and D2 into authorized data sets D1' and D2' according to the pre-made interference strategy. This data set includes parameters such as angle of attack, sideslip angle, normal vector coefficient, axial force coefficient, etc. These participating values are all continuous values. We will add noise that obeys normal distribution to the parameter values to obtain a disturbed data set.
[0114] Step 2: Each participant calculates the local nonlinear kernel K(A) according to formula 1 i , A i T ); where A i T Represents the transpose of the data Ai. Here we use the Gaussian radial basis kernel, as shown in Formula 1.
[0115] Step 3: Participants A and B set the calculated local nonlinear kernel to be shared, so that the other party can calculate the corresponding global kernel based on the shared local nonlinear kernel.
[0116] Step 4: Based on the obtained global kernel function, solve the optimization problem and obtain the optimal solution of the parameters; the optimization problem model is shown below.
[0117]
[0118] stα i ,β i ≥0,i=1,2,...,ly i ∈R
[0119] Among them, α = (α1, α2,..., α l ) T , β=(β1,β2,...,β l ) T is the Lagrange multiplier.
[0120] Assume α * =(α1,α2,...,α l ,β1,β2,...,β l ) is the optimal solution to this optimization problem.
[0121] Step 5: Use the optimal parameters obtained in step 4 to establish a prediction model, and the mathematical model is shown in formula (2). Aerodynamic force prediction is performed according to formula (2).
[0122] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.
[0123] The number of devices and processing scales described here are used to simplify the description of the present invention. Applications, modifications and variations of the present invention will be obvious to those skilled in the art.
[0124] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.
Claims
1. A privacy-preserving aerodynamic prediction method, characterized in that: include: Step 1: Each participant adds noise that follows a normal distribution to the local original data file set to obtain a disturbed data set; Step 2: Each participant calculates a local nonlinear kernel based on the local perturbation data set obtained in step 1; Step 3: All participants set their calculated local nonlinear kernels to be shared, so that other participants can calculate the corresponding global kernel based on the shared local nonlinear kernels; Step 4: Solve the optimization problem based on the obtained global kernel function to obtain the optimal solution of the parameters; Step 5: Use the optimal solution of the parameters obtained in step 4 to establish a prediction model and perform aerodynamic prediction; In step 1, the data set acquisition process is configured to include: S10, in the linear center support vector machine PSVM for discrete data, a kernel function is introduced to adjust the linear PSVM to a nonlinear PSVM, and the PSVM is adjusted to a nonlinear center support vector machine PSVR that is applicable to continuous data; S11, random perturbation technology is introduced in PSVR, by adding noise that obeys normal distribution to the local original data set to obtain a perturbed data set; The method of introducing the kernel function in PSVM is configured to include: S101, in PSVM, the class center of two types of samples is defined as the mean of the two types of samples, the positive class and the negative class are X+ and X- respectively, and the mean is recorded as The original optimization problem is as follows: s.t.(ω·x i )+b-y i ≤ε,i=1,2,...l; y i -(ω·x i )-b≤ε,i=1,2,...,l; Where X represents the input sample data, x i represents the i-th component of X, ω represents the weight, b represents the threshold, ε represents the slack variable, y i Indicates the corresponding output value; By transforming the original optimization problem, the objective function is made strictly convex, and the central hyperplane corresponding to each type of point is obtained; The dual problem of the optimization problem obtained by introducing the Lagrangian function is: stα i ,b i ≥0,i=1,2,...,ly i ∈R Among them, α = (α1, α2,..., α l ) T , β=(β1,β2,...,β l ) T is the Lagrange multiplier, R represents a real number; Assume α * =(α1,α2,...,α l ,β1,β2,...,β l ) is the optimal solution to the optimization problem, then f(x)=ω *T ·x+b * in S102, based on PSVM, we can get the optimization problem of PSVR. By introducing the kernel function in the linear regression method, the above original optimization problem is adjusted to: stα i ,b i ≥0,i=1,2,...,ly i ∈R; Among them, K() is the kernel function.
2. The aerodynamic force prediction method based on privacy protection according to claim 1, characterized in that: In step 2, the local nonlinear kernel is obtained in a manner configured to include: S20, based on Formula A i =A i +N i All participants in PSVR generate corresponding protection data with disturbances, where A i Represents the data set of each participant, N i represents the added noise that follows a normal distribution; S21, calculating the local nonlinear kernel K for the data of each participant, and sharing the nonlinear kernel among all the participants.
3. The aerodynamic force prediction method based on privacy protection according to claim 1, characterized in that: In step three, when the distributed database system uses PSVR for regression, the global kernel can be obtained by first calculating the local kernel of each node and then summing the local kernels of each node.
4. The aerodynamic force prediction method based on privacy protection according to claim 1, characterized in that: In step 5, the optimal solution obtained in step 4 is used to establish a corresponding prediction model; Wherein, the prediction model is: in, 5. A data mining system using the aerodynamic force prediction method according to any one of claims 1 to 4, characterized in that: include: A privacy protection module, used for converting the original data set into an authorized data set according to a pre-made interference adding strategy; In the joint calculation module, each participant calculates the local kernel based on the transformed data set. Each participant can deduce and calculate the global kernel based on each local kernel and calculate the optimal solution. In the prediction module, the host computer establishes a prediction model based on the optimal solution jointly calculated by all participating parties, and uses the prediction model to analyze and predict the aerodynamic force.
Citation Information
Patent Citations
Multi-party collaborative model updating method, device and system for realizing privacy protection
CN113221183A
Model-based collaborative filtering method for collaborative web quality-of-service prediction for privacy protection
WO2019056572A1
