Data generation method and system for member privacy risk detection of credit investigation model

Abstract

The application discloses a kind of data generation method and system of member privacy risk detection of credit investigation model, comprising: respectively estimating the average prediction confidence of member data and non-member data on each credit investigation category as judgment threshold;Randomly sample a sample from the sample space of the target credit investigation model to be measured, and for each feature dimension of the sample, a new feature value is randomly sampled from the feature value space of the feature dimension to replace the current feature value, and a seed sample is generated;The seed sample is input into the target credit investigation model to be measured, and the posterior probability prediction output is obtained;When the prediction probability of the seed sample in target label c satisfies the judgment threshold condition of member data or non-member data, then it corresponds as member detection sample or non-member detection sample;Generation satisfies the sample quantity of detection data.The application can more accurately capture the performance difference of credit investigation model on two different data, and guarantee the effectiveness of the generated detection data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of privacy security in the application process of credit scoring model. More specifically, the present application relates to a data generation method and system for member privacy risk detection of credit scoring model. BACKGROUND

[0002] In the financial field, credit scoring system is a key tool for assessing the credit risk of individuals or enterprises. Traditional credit scoring models rely on historical data and statistical analysis to help financial institutions make decisions in the process of loaning, credit granting, etc. However, with the increase in data volume and data dimension, traditional methods gradually show limitations. The introduction of machine learning technology brings new development opportunities for credit scoring system. Machine learning models can handle large amounts of complex multi-dimensional data, discover hidden patterns and correlations, and improve the accuracy and efficiency of credit assessment. For example, a model based on classification algorithm can classify and predict the credit risk of borrowers according to their behavior data, financial records, and other external information. Although machine learning technology provides significant advantages for credit scoring system, it still faces many challenges in the application process, especially in the privacy risk detection of credit scoring system. In this application, we mainly focus on the problem of member privacy leakage risk detection of credit scoring model, that is, the risk level of credit scoring model leaking its training data sample information. It is worth noting that existing member privacy risk detection methods are usually carried out under ideal conditions, which assume that the risk detection personnel can obtain the real credit scoring model training data (i.e. member data) and non-training data (i.e. non-member data) as detection data, so as to evaluate the risk level of the model by using existing member privacy detection technology. However, in the real financial credit scoring scene, due to the consideration of privacy security, the detection personnel usually cannot obtain the real member and non-member data related to the credit scoring model. Therefore, the member privacy risk detection of credit scoring model needs a detection data generation device to generate detection data that can replace the real member and non-member data, and then complete the model risk detection task in the actual scene. SUMMARY

[0003] An object of the present application is to solve at least the above problems and to provide at least the advantages to be described later.

[0004] To achieve these objects and other advantages in accordance with the present application, a data generation method for member privacy risk detection of credit scoring model is provided, comprising the following steps:

[0005] S1, respectively estimating the average prediction confidence of the member data in each credit scoring category of the target credit scoring model M to be detected and respectively estimating the average prediction confidence of the non-member data in each credit scoring category of the target credit scoring model to be detected wherein n represents that the target credit scoring model M has n credit scoring categories, i represents the i-th credit scoring category, represents the average predicted confidence of the member data on the i-th credit scoring category of the target credit scoring model M, represents the average predicted confidence of the non-member data on the i-th credit scoring category of the target credit scoring model M;

[0006] S2, a sample is randomly sampled from the sample space of the target credit scoring model M, and for each feature dimension of the sample, a new feature value is randomly sampled from the feature value space of the feature dimension to replace the current feature value, thereby generating a seed sample, wherein the sample space comprises member data and non-member data;

[0007] S3, the seed sample is input into the target credit scoring model M to obtain a posterior probability prediction output, and a judgment threshold of the member data is set a judgment threshold of the non-member data wherein represents the judgment threshold of the member data on the i-th credit scoring category, represents the judgment threshold of the non-member data on the i-th credit scoring category;

[0008] S4, the target label of the credit scoring category is fixed as c, when the prediction probability y c of the target label c of the target credit scoring model M on the seed sample is higher than the judgment threshold of the member data and greater than the prediction probability of the label of any other credit scoring category, the seed sample is saved as a generated member detection sample, wherein c is one of the labels of the n credit scoring categories, represents the judgment threshold of the member data on the credit scoring category corresponding to the target label c;

[0009] when the prediction probability y c of the target label c of the target credit scoring model M on the seed sample is lower than the judgment threshold of the non-member data and greater than the prediction probability of the label of any other credit scoring category, the seed sample is saved as a generated non-member detection sample, wherein represents the judgment threshold of the non-member data on the credit scoring category corresponding to the target label c;

[0010] S5, the steps S2-S4 are repeated to generate member detection samples and non-member detection samples that meet the sample quantity.

[0011] Preferably, when the seed sample does not meet the condition of step S4, the following steps are continued to be executed:

[0012] S6.1, a parameter k max is set, kmin and k, where [k min , k max ] represent the value range of k, and k max Not greater than the characteristic dimension of the seed sample, the first k takes the maximum value k max , and then adjusted to 1 / 2 of the current value;

[0013] S6.2. Randomly select k feature dimensions from the feature dimensions of the sample, and randomly sample a new feature value from the feature value space of these k feature dimensions to replace the current feature value, thereby generating a sample, and repeating the judgment in step S4.

[0014] When the number of times the condition of step S4 is not met reaches the preset number of times rej max When , execute step S6.3;

[0015] S6.3. Adjust the value of k by 1 / 2 of the current value, and repeat step S6.2 until the condition of step S4 is met or k has completed all possible values.

[0016] Preferably, the step S1 of "respectively estimating the average prediction confidence of member data in each credit category of the target credit model M to be tested" And estimate the average prediction confidence of non-member data on each credit category of the target credit model M The specific methods include:

[0017] When the member data and non-member data of several target credit models M can be obtained,

[0018] Input the member data of each label belonging to each credit category into the target credit model M to be tested, and calculate the average prediction confidence of each credit category

[0019]

[0020] in, Represents the sub-dataset with the true label i in the member data;

[0021] Input the non-member data of the labels belonging to each credit category into the target credit model M to be tested, and calculate the average prediction confidence of each credit category

[0022]

[0023] in, Represents the sub-dataset whose true label is i in the non-member data.

[0024] Preferably, the method further comprises: when a plurality of data samples of the target credit risk model M can be obtained, but whether the data samples are member data or non-member data is unknown, then,

[0025] The obtained data samples are randomly divided into two parts, one part of the data samples is used as training samples, and the other part of the data samples is used as test samples;

[0026] The target credit risk model M is trained by using the training samples, and a “shadow” credit collection model M is obtained shadow ;

[0027] The training samples belonging to the labels of each credit risk category are respectively input into the “shadow” credit collection model M shadow , and the average prediction confidence on each credit risk category is calculated

[0028]

[0029] wherein, represents a sub-data set of the “shadow” credit collection model M shadow with the real label being i in the training samples;

[0030] The test samples belonging to the labels of each credit risk category are respectively input into the “shadow” credit collection model M shadow , and the average prediction confidence on each credit risk category is calculated

[0031]

[0032] wherein, represents a sub-data set of the “shadow” credit collection model M shadow with the real label being i in the test samples.

[0033] Preferably, the method further comprises: when the number of member data and non-member data of the target credit risk model M obtained is less than a certain number, then,

[0034] The average prediction confidence and the average prediction confidence are respectively assigned values, wherein,

[0035] The application also provides a system for generating data based on member privacy risk detection of the credit risk model, comprising:

[0036] A judgment threshold estimation module is configured to estimate the average prediction confidence of the member data on each credit risk category of the target credit risk model M and estimate the average prediction confidence of the non-member data on each credit risk category of the target credit risk model Set the judgment threshold of member data Judgment threshold for non-member data

[0037] Among them, n means that the target credit model M has n credit categories, i means the i-th credit category, It represents the average prediction confidence of the member data of the i-th credit category of the target credit model M to be tested, represents the average prediction confidence of non-member data on the i-th credit category of the target credit model M, represents the judgment threshold of member data in the i-th credit category, represents the judgment threshold of non-member data in the i-th credit category;

[0038] A sample generation module is configured to randomly sample a sample from the sample space of the target credit model M to be tested, and for each feature dimension of the sample, randomly sample a new feature value from the feature value space of the feature dimension to replace the current feature value, thereby generating a seed sample, wherein the sample space includes member data and non-member data;

[0039] The sample data judgment module is used to input the seed sample into the target credit model M to obtain the posterior probability prediction output;

[0040] And the target label for the fixed credit category is c, when the target credit model M predicts the probability y of the seed sample in the target label c c Higher than the judgment threshold of member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated member detection sample, where c is one of the labels of the n credit categories, Indicates the judgment threshold of member data on the credit category corresponding to the target label c;

[0041] When the target credit model M predicts the probability y of the seed sample in the target label c c Below the judgment threshold of non-member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated non-member detection sample, where Indicates the judgment threshold of non-member data in the credit category corresponding to the target label c;

[0042] Generate membership detection samples and non-membership detection samples that meet the sample quantity.

[0043] Preferably, the sample generation module is configured to continue executing the following steps when the seed sample does not meet the conditions of step S4:

[0044] S6.1. Setting parameter k max 、k min and k, where [k min , k max ] represent the value range of k, and k max Not greater than the characteristic dimension of the seed sample, the first k takes the maximum value k max , and then adjusted to 1 / 2 of the current value;

[0045] S6.2. Randomly select k feature dimensions from the feature dimensions of the sample, and randomly sample a new feature value from the feature value space of these k feature dimensions to replace the current feature value, thereby generating a sample, and repeating the judgment in step S4.

[0046] When the number of times the condition of step S4 is not met reaches the preset number of times rej max When , execute step S6.3;

[0047] S6.3. Adjust the value of k by 1 / 2 of the current value, and repeat step S6.2 until the condition of step S4 is met or k has completed all possible values.

[0048] An electronic device is provided, comprising:

[0049] one or more processors;

[0050] and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method.

[0051] A computer-readable storage medium is provided, wherein a computer program is stored on the storage medium, and when the program is executed by a processor, the method described above is implemented.

[0052] The present invention includes at least the following beneficial effects: The present invention constructs a method for generating privacy risk detection data for credit models. First, by analyzing the differences in the output behavior of the credit model for member data samples and non-member data samples, a threshold judgment method based on prediction confidence is proposed. Then, by designing a data generation technology that combines a sample feature space search method with a sampling method, the purpose of quickly and efficiently generating detection data is achieved. The present invention uses the above method to generate detection data for privacy risk detection of credit models, thereby supporting privacy risk detection tasks in actual scenarios and providing accurate privacy risk assessment of the credit model to be tested.

[0053] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of the process framework of one of the technical solutions of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0056] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified; in the description of the present invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0057] like Figure 1 As shown, the present invention provides a data generation method for member privacy risk detection of a credit investigation model, comprising the following steps:

[0058] S1. Estimate the average prediction confidence of member data in each credit category of the target credit model M And estimate the average prediction confidence of non-member data in each credit category of the target credit model to be tested Where n represents the target credit model M to be tested has n credit categories, i represents the i-th credit category, It represents the average prediction confidence of the member data of the i-th credit category of the target credit model M to be tested, It represents the average prediction confidence of non-member data on the i-th credit category of the target credit model M to be tested. We have designed different methods for different detection scenarios that may exist in reality, which will be described in detail in the technical solution below.

[0059] S2. A sample is randomly sampled from the sample space of the target credit model M to be tested, and for each feature dimension of the sample, a new feature value is randomly sampled from the feature value space of the feature dimension to replace the current feature value, thereby generating a seed sample, wherein the sample space includes member data and non-member data; this method can generate new samples quickly and efficiently.

[0060] S3. Input the seed sample into the target credit model M to obtain the posterior probability prediction output and set the judgment threshold of the member data. Judgment threshold for non-member data in, represents the judgment threshold of member data in the i-th credit category, represents the judgment threshold of non-member data in the i-th credit category;

[0061] S4, fix the target label of the credit category to be c, when the target credit model M predicts the probability y of the seed sample in the target label c c Higher than the judgment threshold of member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated member detection sample, where c is one of the labels of the n credit categories, Indicates the judgment threshold of member data on the credit category corresponding to the target label c;

[0062] When the target credit model M predicts the probability y of the seed sample in the target label c c Below the judgment threshold of non-member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated non-member detection sample, where Represents the judgment threshold of non-member data in the credit category corresponding to the target label c; combining steps S1, S3, and S4 can more accurately capture the performance differences of the credit model on two different types of data, thereby clarifying the generation standard of privacy risk detection data and ensuring the validity of the generated detection data.

[0063] S5. Repeat steps S2 to S4 to generate member detection samples and non-member detection samples that meet the sample quantity.

[0064] In the above technical solution, first, by analyzing the credit model's predictive output behavior for member data and non-member data, the present invention can more accurately capture the differences in the credit model's performance on these two different types of data, thereby clarifying the generation standards for privacy risk detection data and ensuring the validity of the generated detection data. Second, by utilizing feature space search technology, the efficiency of detection data generation can be effectively improved. Finally, the generation method of the present invention is highly versatile and flexible, and is applicable to credit classification prediction models constructed using different machine learning algorithms.

[0065] In another technical solution, when the seed sample does not meet the conditions of step S4, the following steps are continued:

[0066] S6.1. Setting parameter k max 、k minand k, where [k min , k max ] represent the value range of k, and k max Not greater than the characteristic dimension of the seed sample, the first k takes the maximum value k max , and then adjusted to 1 / 2 of the current value;

[0067] S6.2. Randomly select k feature dimensions from the feature dimensions of the sample, and randomly sample a new feature value from the feature value space of these k feature dimensions to replace the current feature value, thereby generating a sample, and repeating the judgment in step S4.

[0068] When the number of times the condition of step S4 is not met reaches the preset number of times rej max When , execute step S6.3;

[0069] S6.3. Adjust the value of k by 1 / 2 of the current value, and repeat step S6.2 until the condition of step S4 is met or k has completed all possible values.

[0070] In the above technical solution, the value of k represents the maximum number of feature dimensions randomly selected and reassigned in each perturbation operation. Initially, the maximum value k is taken. max , which means that at the beginning, the maximum number of feature dimensions will be selected for perturbation, and the randomly selected k feature dimensions are the targets of reassignment. For the selected features, different operations are performed according to their types:

[0071] Binary features: If the feature is binary (i.e., the feature value can only be 0 or 1), a flip operation is used. That is, if the current feature value is 0, it is changed to 1; if the current feature value is 1, it is changed to 0.

[0072] Other types of features: For non-binary features, a new feature value is randomly sampled from the feature value space to replace the current feature value. The feature value space refers to the set of all possible values ​​of the feature.

[0073] The purpose of this perturbation operation is to simulate the natural variation of the data by introducing randomness, thereby obtaining more diverse samples and more data that the model has not previously anticipated.

[0074] In another technical solution, the step S1 of "respectively estimating the average prediction confidence of member data in each credit category of the target credit model M to be tested" And estimate the average prediction confidence of non-member data on each credit category of the target credit model M The specific methods include:

[0075] When the member data and non-member data of several target credit models M can be obtained,

[0076] Input the member data of each label belonging to each credit category into the target credit model M to be tested, and calculate the average prediction confidence of each credit category

[0077]

[0078] in, Represents the sub-dataset with the true label i in the member data;

[0079] Input the non-member data of the labels belonging to each credit category into the target credit model M to be tested, and calculate the average prediction confidence of each credit category

[0080]

[0081] in, Represents the sub-dataset whose true label is i in the non-member data.

[0082] In the above technical solution, a detection scenario is provided. In this detection scenario, we can obtain a lot of member data and non-member data of the target credit model M to be tested. Therefore, the average prediction confidence of the two types of data is obtained by direct generation and calculation methods, which is used to evaluate the generated sample data.

[0083] In another technical solution, the step S1 of "respectively estimating the average prediction confidence of member data in each credit category of the target credit model M to be tested" And estimate the average prediction confidence of non-member data on each credit category of the target credit model M The specific methods include:

[0084] When several data samples of the target credit model M to be tested can be obtained, but whether the data samples are member data or non-member data is unknown, then,

[0085] The acquired data samples are randomly divided into two parts, one data sample is used as a training sample, and the other data sample is used as a test sample;

[0086] Use the training samples to train the target credit model M to obtain the "shadow" collection model M shadow ;

[0087] Input the training samples of labels belonging to each credit category into the “shadow” collection model M shadow , calculate the average prediction confidence of each credit category

[0088]

[0089] wherein, denotes the "shadow" enrollment model M shadow a sub-data set of the training sample of the "shadow" enrollment model M

[0090] respectively input the test sample of the label belonging to each enrollment category into the "shadow" enrollment model M shadow calculate the average prediction confidence on each enrollment category

[0091]

[0092] wherein, denotes the "shadow" enrollment model M shadow a sub-data set of the test sample of the "shadow" enrollment model M

[0093] In the above technical solution, a detection scenario is provided. Although the detection scenario can also obtain sample data of the to-be-tested target enrollment model M, it cannot distinguish whether the data belongs to member data or non-member data. Therefore, a "shadow" enrollment model is trained based on the obtained sample data and the to-be-tested target enrollment model M. The "shadow" enrollment model may change the internal parameters and weights of the to-be-tested target enrollment model. Then, the average prediction confidence of the two types of data is generated and calculated based on the "shadow" enrollment model, and is used for evaluation of the generated sample data.

[0094] In another technical solution, the method of "respectively estimating the average prediction confidence of the member data on each enrollment category of the to-be-tested target enrollment model M and respectively estimating the average prediction confidence of the non-member data on each enrollment category of the to-be-tested target enrollment model M " in step S1 specifically includes:

[0095] When the number of the member data and the non-member data of the to-be-tested target enrollment model M obtained is less than a certain number, then,

[0096] respectively assign values to the average prediction confidence and the average prediction confidence wherein,

[0097] In the above technical solution, a detection scenario is provided. The member data and the non-member data of the to-be-tested target enrollment model M can be rarely obtained, or even zero. Therefore, based on the principle and consensus of machine learning, the average prediction confidence of the two types of data is directly assigned a value, and is used for evaluation of the generated sample data.

[0098] Based on the same inventive concept, the present invention also provides a system for generating data for member privacy risk detection based on the credit reporting model, comprising:

[0099] The judgment threshold estimation module is used to estimate the average prediction confidence of member data in each credit category of the target credit model M to be tested. And estimate the average prediction confidence of non-member data in each credit category of the target credit model to be tested Set the judgment threshold of member data Judgment threshold for non-member data

[0100] Among them, n means that the target credit model M has n credit categories, i means the i-th credit category, It represents the average prediction confidence of the member data of the i-th credit category of the target credit model M to be tested, represents the average prediction confidence of non-member data on the i-th credit category of the target credit model M, represents the judgment threshold of member data in the i-th credit category, represents the judgment threshold of non-member data in the i-th credit category;

[0101] The sample generation module is used to randomly sample a sample from the sample space of the target credit model M to be tested, and for each feature dimension of the sample, randomly sample a new feature value from the feature value space of the feature dimension to replace the current feature value, thereby generating a seed sample, wherein the sample space includes member data and non-member data; new samples can be generated quickly and efficiently.

[0102] The sample data judgment module is used to input the seed sample into the target credit model M to obtain the posterior probability prediction output;

[0103] And the target label for the fixed credit category is c, when the target credit model M predicts the probability y of the seed sample in the target label c c Higher than the judgment threshold of member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated member detection sample, where c is one of the labels of the n credit categories, Indicates the judgment threshold of member data on the credit category corresponding to the target label c;

[0104] When the target credit model M predicts the probability y of the seed sample in the target label c c Below the judgment threshold of non-member data And it is greater than the predicted probability of any other credit category label, then save the seed sample as the generated non-member detection sample, where Indicates the judgment threshold of non-member data in the credit category corresponding to the target label c;

[0105] Generate membership detection samples and non-membership detection samples that meet the sample quantity.

[0106] In the above technical solution, first, by analyzing the credit model's predictive output behavior for member data and non-member data, the present invention can more accurately capture the differences in the credit model's performance on these two different types of data, thereby clarifying the generation standards for privacy risk detection data and ensuring the validity of the generated detection data. Second, by utilizing feature space search technology, the efficiency of detection data generation can be effectively improved. Finally, the system of the present invention is highly versatile and flexible, and is applicable to credit classification prediction models constructed using different machine learning algorithms.

[0107] In another technical solution, the sample generation module is configured to continue executing the following steps when the seed sample does not meet the conditions of step S4:

[0108] S6.1. Setting parameter k max 、k min and k, where [k min , k max ] represent the value range of k, and k max Not greater than the characteristic dimension of the seed sample, the first k takes the maximum value k max , and then adjusted to 1 / 2 of the current value;

[0109] S6.2. Randomly select k feature dimensions from the feature dimensions of the sample, and randomly sample a new feature value from the feature value space of these k feature dimensions to replace the current feature value, thereby generating a sample, and repeating the judgment in step S4.

[0110] When the number of times the condition of step S4 is not met reaches the preset number of times rej max When , execute step S6.3;

[0111] S6.3. Adjust the value of k by 1 / 2 of the current value, and repeat step S6.2 until the condition of step S4 is met or k has completed all possible values.

[0112] In the above technical solution, the sample generation module simulates the natural changes of data by introducing randomness, thereby obtaining more diverse samples and more data that the model has not previously anticipated.

[0113] The present invention further provides an electronic device, comprising:

[0114] one or more processors;

[0115] and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to perform the data generation method. The electronic device can be any terminal device including a mobile phone, a laptop computer, a desktop computer, a tablet computer, a PDA (Personal Digital Assistant), a POS (Point of Sales), an in-vehicle computer, or the like.

[0116] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the data generating method is implemented.

[0117] Through the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware, and of course can also be implemented by means of dedicated hardware including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. In general, all functions performed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the present invention, software program implementation is a better implementation method in most cases. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc., and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0118] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A data generation method for member privacy risk detection in a credit reporting model, characterized in that: The following steps are involved: S1. Estimate the member data in the target credit model to be tested M The average prediction confidence of each credit category , and respectively estimate the average prediction confidence of non-member data on each credit category of the target credit model to be tested ,in, n Represents the target credit model to be tested M have n credit categories, i Indicates the i credit categories, Represents the target credit model to be tested M No. i The average prediction confidence of member data on credit categories, Represents the target credit model to be tested M No. i The average prediction confidence of non-member data on the credit category; S2. From the credit model of the target to be tested M A sample is randomly sampled from the sample space of , and for each feature dimension of the sample, a new feature value is randomly sampled from the feature value space of the feature dimension to replace the current feature value, to generate a seed sample, wherein the sample space includes member data and non-member data; S3. Input the seed sample into the target credit model to be tested M , get the posterior probability prediction output, set the judgment threshold of member data , the judgment threshold of non-member data ,in, Indicates the i The judgment threshold of member data on each credit category, Indicates the i The judgment threshold for non-member data in each credit category; S4, the target label of the fixed credit category is c , when the target credit model to be tested M For seed samples in target label c The predicted probability y c Higher than the judgment threshold of member data , and is greater than the predicted probability of any other credit category label, then save the seed sample as the generated member detection sample, where, c For the n One of the labels in the credit category, Indicates the target label c The judgment threshold of member data in the corresponding credit category; When the target credit model to be tested M For seed samples in target label c The predicted probability y c Below the judgment threshold of non-member data , and is greater than the predicted probability of any other credit category label, then save the seed sample as the generated non-member detection sample, where, Indicates the target label c The judgment threshold of non-member data in the corresponding credit category; S5. Repeat steps S2 to S4 to generate member detection samples and non-member detection samples that meet the sample quantity.

2. The data generation method for member privacy risk detection of the credit reporting model according to claim 1 is characterized in that: When the seed sample does not meet the conditions of step S4, the following steps are continued: S6.

1. Setting parameters k max 、 k min and k ,in,[ k min , k max ] respectively represent k The value range of k max Not greater than the characteristic dimension of the seed sample, for the first time k Take the maximum value k max , and then adjusted to 1 / 2 of the current value; S6.

2. Randomly select from the feature dimension of the sample k feature dimensions, and k Randomly sample a new eigenvalue from the eigenvalue space of the feature dimension to replace the current eigenvalue, generate a sample, and repeat the judgment of step S4; When the number of times that the condition of step S4 is not met reaches the preset number rej max When , execute step S6.3; S6.3, adjust by 1 / 2 of the current value k Repeat step S6.2 until the condition of step S4 is met, or k Complete all values.

3. The data generation method for member privacy risk detection of the credit reporting model according to claim 1 is characterized in that: In step S1, the "respective estimation of member data in the target credit model to be tested" M The average prediction confidence of each credit category , and separately estimate the non-member data in the target credit model to be tested M The average prediction confidence of each credit category The specific methods include: When several target credit models to be tested are available M When the member data and non-member data of Input the member data of each credit category label into the target credit model to be tested M , calculate the average prediction confidence of each credit category ; in, Indicates that the true label in the member data is i sub-datasets of; Input the non-member data of each credit category label into the target credit model to be tested M , calculate the average prediction confidence of each credit category ; in, Indicates that the true label in the non-member data is i sub-datasets of .

4. The data generation method for member privacy risk detection of a credit reporting model according to claim 3, characterized in that: Also includes: When several target credit models to be tested are available M , but whether the data sample is member data or non-member data is unknown, then, The acquired data samples are randomly divided into two parts, one data sample is used as a training sample, and the other data sample is used as a test sample; Use training samples to train the target credit model M , get the "shadow" collection model ; Input the training samples of labels belonging to each credit category into the "shadow" collection model , calculate the average prediction confidence of each credit category ; in, Represents the "shadow" solicitation model The true label in the training sample is i sub-datasets of; Input the test samples of labels belonging to each credit category into the "shadow" collection model , calculate the average prediction confidence of each credit category ; in, Represents the "shadow" solicitation model The true label of the test sample is i sub-datasets of .

5. The data generation method for member privacy risk detection of the credit reporting model according to claim 3 is characterized in that: Also includes: When the target credit model to be tested is obtained M When the number of member data and non-member data is less than a certain number, then, The average prediction confidence and the average prediction confidence Assignment, where > .

6. A system for generating data for member privacy risk detection based on the credit reporting model according to any one of claims 1 to 5, characterized in that: include: The judgment threshold estimation module is used to estimate the member data in the target credit model to be tested M The average prediction confidence of each credit category , and respectively estimate the average prediction confidence of non-member data on each credit category of the target credit model to be tested , set the judgment threshold of member data , the judgment threshold of non-member data ; in, n Represents the target credit model to be tested M have n credit categories, i Indicates the i credit categories, Represents the target credit model to be tested M No. i The average prediction confidence of member data on credit categories, Represents the target credit model to be tested M No. i The average prediction confidence of non-member data on credit categories, Indicates the i The judgment threshold of member data on each credit category, Indicates the i The judgment threshold for non-member data in each credit category; The sample generation module is used to generate samples from the target credit model to be tested M A sample is randomly sampled from the sample space of , and for each feature dimension of the sample, a new feature value is randomly sampled from the feature value space of the feature dimension to replace the current feature value, to generate a seed sample, wherein the sample space includes member data and non-member data; The sample data judgment module is used to input the seed sample into the target credit model to be tested M , get the posterior probability prediction output; And the target label for fixed credit category is c , when the target credit model to be tested M For seed samples in target label c The predicted probability y c Higher than the judgment threshold of member data , and is greater than the predicted probability of any other credit category label, then save the seed sample as the generated member detection sample, where, c For the n One of the labels in the credit category, Indicates the target label c The judgment threshold of member data in the corresponding credit category; When the target credit model to be tested M The predicted probability of the seed sample in the target label c y c Below the judgment threshold of non-member data , and is greater than the predicted probability of any other credit category label, then save the seed sample as the generated non-member detection sample, where, Indicates the target label c The judgment threshold of non-member data in the corresponding credit category; Generate membership detection samples and non-membership detection samples that meet the sample quantity.

7. The system according to claim 6, wherein: The sample generation module is used to continue to perform the following steps when the seed sample does not meet the conditions of step S4: S6.

1. Setting parameters k max 、 k min and k ,in,[ k min , k max ] respectively represent k The value range of k max Not greater than the characteristic dimension of the seed sample, for the first time k Take the maximum value k max , and then adjusted to 1 / 2 of the current value; S6.

2. Randomly select from the feature dimension of the sample k feature dimensions, and k Randomly sample a new eigenvalue from the eigenvalue space of the feature dimension to replace the current eigenvalue, generate a sample, and repeat the judgment of step S4; When the number of times that the condition of step S4 is not met reaches the preset number rej max When , execute step S6.3; S6.3, adjust by 1 / 2 of the current value k Repeat step S6.2 until the condition of step S4 is met, or k Complete all values.

8. An electronic device, characterized in that include: one or more processors; and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor performs the method according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images