Method, device, equipment and storage medium for screening federated feature engineering data

Abstract

The present disclosure relates to the field of privacy computing technology, and in particular to federated learning technology, and specifically discloses a method, apparatus, device and storage medium for screening federated feature engineering data, the method comprising: sending a sample ID aligned with a sample on the data end and a designated label of homomorphic encryption to a data end; receiving a ciphertext summation result of each bin, a bin number and a total number of samples sent by the data end; decrypting the received ciphertext summation result of each bin to obtain a decryption result of each bin on the data end; determining an analysis result on the data end according to the bin number, the decryption result of each bin and the total number of samples, so as to determine whether to select data binned according to the features currently used for binning as data for federated modeling based on the analysis result. The data sent to the data end includes only one sample label. The analysis result of each bin is calculated by obtaining the total number of samples in each bin on the data end, which can greatly improve the efficiency of federated feature engineering.

Description

Technical Field

[0001] The present disclosure relates to the field of privacy computing technology, and in particular to federated learning technology, and specifically discloses a method and apparatus, device, and storage medium for screening federated feature engineering data. Background Art

[0002] With the rapid development of technologies like artificial intelligence, laws and regulations have mandated that user data be used only by the institution holding it, in order to protect user privacy. This has led to data being blocked from circulation and sharing between institutions. To address this issue, federated learning methods, based on multiple confidentiality technologies like differential privacy and homomorphic encryption, aim to achieve multi-party data joint modeling while protecting user data privacy, thereby leveraging the value of social data resources and strengthening orderly data sharing. Because the quality of external data sources is difficult to control during federated learning, it is necessary to perform feature engineering through federation before modeling to screen and evaluate the features of external data sources.

[0003] Taking feature engineering in financial risk control scenarios as an example, a common feature analysis method involves calculating the correlation between features and labels. This involves calculating metrics to evaluate the sample-discriminating effect and stability of features. However, these metrics often rely on sample labels. For example, in the federated feature engineering calculation process, assuming the total number of samples after alignment between the business and data entities is N, the business entity needs to perform additive homomorphic encryption on the N label values ​​y and the N 1-y values ​​to obtain ciphertext. This ciphertext, with a size of 2N, is then sent to the data entity. Subsequent binning and summing processes on the data entity also require a size of 2N. Furthermore, in real-world business scenarios, datasets are often very large, often in the hundreds of thousands or even millions. Therefore, operations such as encryption, summing, and decryption can significantly impact the efficiency of federated feature engineering and increase communication overhead and computing resource consumption for federated modeling participants. Summary of the Invention

[0004] In order to solve the above technical problems or at least partially solve the above technical problems, embodiments of the present disclosure provide a method and apparatus, device, and storage medium for screening federated feature engineering data.

[0005] In a first aspect, embodiments of the present disclosure provide a method for screening federated feature engineering data, comprising:

[0006] Sending a sample ID aligned with the sample of the data end and a designated label of homomorphic encryption to the data end, so that the data end can determine the ciphertext summation result of each bin according to the designated label of homomorphic encryption;

[0007] Receive the ciphertext summation result of each bin, bin number and total number of samples sent by the data end;

[0008] Decrypting the sum of the ciphertexts of each received sub-box to obtain the decryption results of each sub-box on the data end;

[0009] According to the bin number, the analysis result of the data end is determined according to the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling based on the analysis result, wherein the analysis result of each bin is used to represent the correlation between the characteristic parameters currently used for binning and the positive sample label of the business end.

[0010] In one possible implementation, the decryption result includes: the number of samples corresponding to the specified label, and determining the analysis result of the data end based on the decryption result of each bin and the total number of samples includes:

[0011] Calculating the total number of positive samples and the total number of negative samples in each bin of the data end, and the total number of positive samples and the total number of negative samples in all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples, wherein the designated label is a positive sample label or a negative sample label;

[0012] Calculate the evidence weight of each bin on the data side according to the total number of positive samples and the total number of negative samples in each bin on the data side and the total number of positive samples and the total number of negative samples in all bins;

[0013] The amount of information of each bin of the data end is calculated according to the evidence weight of each bin of the data end, and the total amount of information of the data end is obtained as the analysis result by summing up.

[0014] In a possible implementation, the designated label is a positive sample label, and the calculating of the total number of positive samples and the total number of negative samples for each bin of the data end, and the total number of positive samples and the total number of negative samples for all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples includes:

[0015] The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of positive samples of each bin of the data end;

[0016] The difference between the total number of samples in each bin at the data end and the total number of positive samples is used as the total number of negative samples in each bin at the data end;

[0017] The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins;

[0018] The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

[0019] In a possible implementation, the designated label is a negative sample label, and the calculating, based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples, the total number of positive samples and the total number of negative samples of each bin of the data end, and the total number of positive samples and the total number of negative samples of all bins, includes:

[0020] The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of negative samples of each bin of the data end;

[0021] The difference between the total number of samples in each bin at the data end and the total number of negative samples is used as the total number of positive samples in each bin at the data end;

[0022] The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins;

[0023] The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

[0024] In a possible implementation, determining the analysis result of the data end according to the decryption result of each sub-box and the total number of samples further includes:

[0025] According to the total number of positive samples and the total number of negative samples in each bin of the data end, the KS value of the data on the data end is calculated as the analysis result.

[0026] In one possible implementation, the KS value of the data on the data end is calculated according to the total number of positive samples and the total number of negative samples in each bin of the data end by the following expression:

[0027] KS=max{|cum i (Bad)-cum i (Good)|}

[0028] Among them, cum i (Bad) is the cumulative number of negative samples in each bin interval i, cum i (Good) is the cumulative number of positive samples in each bin interval i.

[0029] In a second aspect, embodiments of the present disclosure provide a method for screening federated feature engineering data, which is applied to a data end. The method includes:

[0030] Receive a sample ID and a homomorphically encrypted designated label sent by the service end, which are aligned with the sample on the data end, wherein the designated label is a positive sample label or a negative sample label;

[0031] Bin the sample IDs of the sample alignment according to the preset feature parameters;

[0032] Calculate the sum of the ciphertexts of each bin based on the homomorphically encrypted specified labels corresponding to the sample IDs in each bin;

[0033] The bin number, ciphertext sum result and total number of samples of each bin are sent to the business end.

[0034] In a possible implementation, the preset characteristic parameters are determined by the following steps:

[0035] Calculate the sample coverage and / or variance value of each selected feature parameter of all sample data;

[0036] Delete feature parameters whose sample coverage is less than a preset coverage threshold, and / or feature parameters whose variance is less than a preset variance threshold;

[0037] For any two of the remaining candidate feature parameters, the Pearson correlation coefficient between the two candidate feature parameters is calculated, and when the Pearson correlation coefficient is greater than or equal to the preset coefficient threshold, one of the two candidate feature parameters is deleted until the Pearson correlation coefficient is no longer greater than or equal to the preset coefficient threshold, and the last remaining candidate feature parameter is used as the preset feature parameter.

[0038] In a third aspect, an embodiment of the present disclosure provides a device for screening federated feature engineering data, which is applied to a business end. The device includes:

[0039] A sending module, configured to send a sample ID aligned with the sample of the data end and a homomorphically encrypted positive sample label to the data end, so that the data end can determine the ciphertext summation result of each bin according to the homomorphically encrypted positive sample label;

[0040] A receiving module is used to receive the ciphertext summation result of each bin, bin number and total number of samples sent by the data end;

[0041] A decryption module, configured to decrypt the received ciphertext summation results of each sub-box to obtain a decryption result of each sub-box on the data end;

[0042] A determination module is used to determine the analysis results of the data end according to the decryption results of each bin and the total number of samples according to the bin number, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federal modeling according to the analysis results, wherein the analysis results of each bin are used to represent the correlation between the characteristic parameters currently used for binning and the positive sample labels of the business end.

[0043] In a fourth aspect, embodiments of the present disclosure provide a device for screening federated feature engineering data, applied to a data end, the device comprising:

[0044] A receiving module, configured to receive a sample ID and a homomorphically encrypted designated label sent by a service end, the designated label being a positive sample label or a negative sample label;

[0045] A binning module, which is used to bin the sample IDs of the sample alignment according to preset feature parameters;

[0046] A calculation module is used to calculate the sum of the ciphertexts of each bin based on the homomorphically encrypted specified labels corresponding to the sample IDs in each bin;

[0047] The sending module is used to send the box number, ciphertext summation result and total number of samples of each box to the business end.

[0048] In a fifth aspect, an embodiment of the present disclosure provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0049] Memory for storing computer programs;

[0050] The processor is configured to implement the above-mentioned method for screening federated feature engineering data when executing the program stored in the memory.

[0051] In a sixth aspect, an embodiment of the present disclosure provides a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the above-mentioned method for screening federated feature engineering data.

[0052] The above technical solutions provided by the embodiments of the present disclosure have at least some or all of the following advantages compared to the prior art:

[0053] The method for screening federated feature engineering data described in the embodiment of the present disclosure includes sending a sample ID aligned with a sample on the data end and a designated homomorphically encrypted tag to a data end, so that the data end determines the ciphertext summation result of each bin according to the designated homomorphically encrypted tag; receiving the ciphertext summation result of each bin, the bin number, and the total number of samples sent by the data end; decrypting the received ciphertext summation result of each bin to obtain the decryption result of each bin on the data end; determining the analysis result of the data end according to the bin number, the decryption result of each bin, and the total number of samples, so as to determine whether to select data binned according to the features currently used for binning as data for federated modeling based on the analysis result, wherein the analysis result of each bin is used to represent the correlation between the feature parameters currently used for binning and the positive sample label of the business end. The data sent to the data end only includes one sample label, and the analysis result of each bin is calculated by obtaining the total number of samples in each bin on the data end, thereby reducing the amount of data sent, received, encrypted, and decrypted by at least half, and significantly improving the efficiency of federated feature engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0055] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0056] Figure 1 An exemplary system architecture of the method or apparatus for screening federated feature engineering data using an embodiment of the present disclosure is shown;

[0057] Figure 2 The following schematically illustrates a flow chart of a method for screening federated feature engineering data according to an embodiment of the present disclosure;

[0058] Figure 3 The following schematically illustrates a flow chart of a method for screening federated feature engineering data according to another embodiment of the present disclosure;

[0059] Figure 4 The following schematically shows a structural block diagram of a device for screening federated feature engineering data according to an embodiment of the present disclosure;

[0060] Figure 5 Schematically shows a structural block diagram of a device for screening federated feature engineering data according to another embodiment of the present disclosure;

[0061] Figure 6 Schematically illustrates an interactive process diagram of a system for screening federated feature engineering data according to an embodiment of the present disclosure;

[0062] Figure 7 A schematic diagram of a workflow of a system for screening federated feature engineering data according to an embodiment of the present disclosure is shown; and

[0063] Figure 8 The structural block diagram of an electronic device according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0065] Figure 1 An exemplary system architecture 100 is shown to which the method or apparatus for screening federated feature engineering data according to an embodiment of the present disclosure may be applied.

[0066] like Figure 1 As shown, system architecture 100 may include data terminals 101, 102, and 103, a network 104, and a service terminal 105. Network 104 is a medium for providing communication links between data terminals 101, 102, and 103 and service terminal 105. Network 104 may include various connection types, such as wired and wireless communication links.

[0067] Business side 105 is the business side of federated learning, while data sides 101, 102, and 103 are modeling participants. In theory, business side 105, as the business side, owns the label variables (dependent variables) and some independent variables, while the other participants own the other independent variables. The interaction process often involves encrypted dependent variables (or independent variables) from business side 105 interacting with the participants.

[0068] When multiple data owners (such as enterprises, governments, and other institutions) want to combine their respective data to train machine learning models, they must ensure that the original data owned by each party does not leave the local area while building a unified model.

[0069] The application scenario of this disclosure is federated learning. Federated learning requires that the sample IDs contained in each platform are basically the same, but the features are different. The business side has labels, while the data side does not.

[0070] In federated feature engineering, the data side only has features but no labels. Although the business side has both features and labels, the features on the business side also lack some features on the data side. Therefore, it needs to be implemented through data interaction under encrypted conditions.

[0071] It should be noted that the service end can be either hardware or software. When the service end is hardware, it can be implemented as a distributed service end cluster consisting of multiple service ends, or it can be implemented as a single service end. When the service end is software, it can be implemented as multiple software or software modules (for example, multiple software or software modules used to provide distributed services), or it can be implemented as a single software or software module. No specific limitations are given here.

[0072] It should be noted that the method for screening federated feature engineering data provided in the embodiments of the present application can be executed by the data terminals 101, 102, and 103, or by the service terminal 105. Accordingly, the device for screening federated feature engineering data can be provided in the data terminals 101, 102, and 103, or in the service terminal 105. This is not specifically limited here.

[0073] It should be understood that Figure 1 The numbers of data terminals, network terminals, and service terminals are merely illustrative. Any number of data terminals, network terminals, and service terminals may be provided as required.

[0074] See also Figure 2 , an embodiment of the present disclosure provides a method for screening federated feature engineering data applied to a business end, comprising the following steps:

[0075] S21, sending a sample ID aligned with the sample of the data end and a designated label of homomorphic encryption to the data end, so that the data end can determine the ciphertext summation result of each bin according to the designated label of homomorphic encryption;

[0076] In practical applications, sample alignment is achieved through the following process: Before conducting federated feature engineering, it is necessary to find the intersection of the user groups of the participants, including those on the business side and the data side. That is, based on mechanisms such as RSA encryption (RSA encryption is an asymmetric encryption that can complete decryption without directly transferring the key), the common user set of the participants in the federated modeling is found under the premise that the participants do not disclose their respective user lists, and at the same time, the private users of the participants, that is, non-overlapping users, are not exposed.

[0077] In actual applications, the business end will assign the specified label y to each sample after alignment. n After encryption based on additive homomorphic encryption, we get <y n >, then the sample primary key ID and the label ciphertext value corresponding to the ID <y n>Transmitted to the data end together. Assuming that the sample label is a binary classification label, the label will be mapped to 0 and 1 before encryption, and the label 1 is set to represent a good user and the label 0 is set to represent a bad user. Assuming that the total number of samples after alignment is N, N-level labels will be encrypted and 2N-level data will be transmitted.

[0078] S22, receiving the ciphertext summation result of each bin, the bin number and the total number of samples sent by the data end;

[0079] In actual applications, the data on the business side is as follows: Sample ID A, designated label: yes; Sample ID B, designated label: no; Sample ID C, designated label: yes; Sample ID D, designated label: no; the designated label is the positive sample label, with the designated label "yes" mapped to 1 and the homomorphically encrypted value being 255; the designated label "no" mapped to 0 and the homomorphically encrypted value being 256. Based on the target features, the data side divides samples with IDs A, B, C, and D into bins A, C and B, D. The ciphertext sum of bin 1, which includes A and B, is 255+255, and the total number of samples is 2; the ciphertext sum of bin 2, which includes B and D, is 256+256, and the total number of samples is 2.

[0080] S23, decrypting the sum of the received ciphertexts of each sub-box to obtain the decryption results of each sub-box on the data end;

[0081] In actual applications, the sum of the ciphertexts of bin 1 including A and B is 255+255, and the decryption result in bin 1 is 1+1, and the number of positive samples (the number of samples with the label "yes") is 2; the sum of the ciphertexts of bin 2 including B and D is 256+256, and the decryption result in bin 2 is 0+0, and the number of positive samples (the number of samples with the label "yes") is 0.

[0082] S24, according to the bin number, the analysis result of the data end is determined according to the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federal modeling based on the analysis result, wherein the analysis result of each bin is used to represent the correlation between the characteristic parameters currently used for binning and the positive sample label of the business end.

[0083] In step S24, the decryption result includes: the number of samples corresponding to the specified label, and the analysis result of the data end is determined based on the decryption result of each bin and the total number of samples, including:

[0084] Calculating the total number of positive samples and the total number of negative samples in each bin of the data end, and the total number of positive samples and the total number of negative samples in all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples, wherein the designated label is a positive sample label or a negative sample label;

[0085] Calculate the evidence weight of each bin on the data side according to the total number of positive samples and the total number of negative samples in each bin on the data side and the total number of positive samples and the total number of negative samples in all bins;

[0086] The amount of information of each bin of the data end is calculated according to the evidence weight of each bin of the data end, and the total amount of information of the data end is obtained as the analysis result by summing up.

[0087] When the total amount of information on the data end is used as an analysis result, determining whether to select data binned according to the features currently used for binning as data for federated modeling according to the analysis result includes:

[0088] In the preset binning method, the total amount of information on the data end is compared with a first preset threshold:

[0089] When the total amount of information on the data end is less than a first preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the total amount of information on the data end is greater than or equal to the preset threshold;

[0090] When the total amount of information on the data end is greater than or equal to a first preset threshold, the total amount of information on the data end is compared with a second preset threshold:

[0091] When the total amount of information on the data end is less than a second preset threshold, discarding the data that is binned according to the feature currently used for binning;

[0092] When the total amount of information on the data end is greater than or equal to a second preset threshold, the data that is binned according to the features currently used for binning is selected as data for federated modeling.

[0093] In practical applications, when the designated label is a positive sample label, the calculation of the total number of positive samples and the total number of negative samples for each bin of the data end and the total number of positive samples and the total number of negative samples for all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples includes:

[0094] The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of positive samples of each bin of the data end;

[0095] The difference between the total number of samples in each bin at the data end and the total number of positive samples is used as the total number of negative samples in each bin at the data end;

[0096] The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins;

[0097] The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

[0098] In practical applications, when the designated label is a negative sample label, the calculation of the total number of positive samples and the total number of negative samples for each bin of the data end and the total number of positive samples and the total number of negative samples for all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples includes:

[0099] The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of negative samples of each bin of the data end;

[0100] The difference between the total number of samples in each bin at the data end and the total number of negative samples is used as the total number of positive samples in each bin at the data end;

[0101] The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins;

[0102] The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

[0103] In step S24, the analysis result of the data end is determined according to the decryption result of each sub-box and the total number of samples, and further includes:

[0104] According to the total number of positive samples and the total number of negative samples in each bin of the data end, the KS value of the data on the data end is calculated as the analysis result, wherein the KS value of the data on the data end is calculated according to the total number of positive samples and the total number of negative samples in each bin of the data end by the following expression:

[0105] KS=max{|cum i (Bad)-cum i (Good)|}

[0106] Among them, cum i (Bad) is the cumulative number of negative samples in each bin interval i, cum i (Good) is the cumulative number of positive samples in each bin interval i.

[0107] When the KS value of the data on the data end is used as an analysis result, determining whether to select the data binned according to the features currently used for binning as the data for federated modeling according to the analysis result includes:

[0108] In the preset binning method, the KS value of the data end is compared with the third preset threshold:

[0109] When the total amount of information on the data end is less than a third preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the KS value of the data end is greater than or equal to the preset threshold;

[0110] When the total amount of information of the data end is greater than or equal to the third preset threshold, the KS value of the data end is compared with the fourth preset threshold:

[0111] When the KS value of the data end is less than a fourth preset threshold, discarding the data that is binned according to the feature currently used for binning;

[0112] When the KS value of the data end is greater than or equal to the fourth preset threshold, the data binned according to the feature currently used for binning is selected as the data for federal modeling, or the feature currently used for binning is compared with the KS values ​​corresponding to other features, and the binned data corresponding to the feature whose KS value is greater than or equal to the fifth preset threshold is selected.

[0113] See also Figure 3 , an embodiment of the present disclosure provides a method for screening federated feature engineering data applied to a data end, the method comprising:

[0114] S31, receiving a sample ID and a homomorphically encrypted designated label sent by a service end that are aligned with the sample of the data end, wherein the designated label is a positive sample label or a negative sample label;

[0115] S32, binning the sample IDs of the sample alignment according to preset feature parameters;

[0116] In practical applications, the preset characteristic parameters are determined by the following steps:

[0117] Calculate the sample coverage and / or variance value of each selected feature parameter of all sample data;

[0118] Delete feature parameters whose sample coverage is less than a preset coverage threshold, and / or feature parameters whose variance is less than a preset variance threshold;

[0119] For any two of the remaining candidate feature parameters, the Pearson correlation coefficient between the two candidate feature parameters is calculated, and when the Pearson correlation coefficient is greater than or equal to the preset coefficient threshold, one of the two candidate feature parameters is deleted until the Pearson correlation coefficient is no longer greater than or equal to the preset coefficient threshold, and the last remaining candidate feature parameter is used as the preset feature parameter.

[0120] S33, calculating the sum of the ciphertexts of each bin based on the homomorphically encrypted designated labels corresponding to the sample IDs in each bin;

[0121] In practical applications, the data end divides the samples into bins according to the features to be evaluated. That is, for the features to be analyzed, the data end divides the samples into bins according to preset binning methods such as equal frequency, equal distance, etc., so as to divide N samples into a few different grouping bins according to the binning threshold. i And count the total number of samples in each bin sum(bin i ). The data end obtains the ciphertext of each sample corresponding to the specified label from the business end <y n >After that, the ciphertext values ​​in each bin are summed up to get the sum in each bin. <y n >), and finally the bin number bin i 、The total number of samples in each bin sum(bin i ) and the sum of the secret labels of each bin ( <y n >) is synchronized to the business end.

[0122] S34, sending the box number, ciphertext summation result and total number of samples of each box to the business end.

[0123] See also Figure 4 , an embodiment of the present disclosure provides a screening device for federated feature engineering data applied to a business end, comprising:

[0124] A sending module 41 is configured to send a sample ID aligned with the sample of the data terminal and a homomorphically encrypted positive sample label to the data terminal, so that the data terminal can determine the ciphertext summation result of each bin according to the homomorphically encrypted positive sample label;

[0125] A receiving module 42 is configured to receive the ciphertext summation result of each bin, the bin number, and the total number of samples sent by the data terminal;

[0126] A decryption module 43, which is used to decrypt the received ciphertext summation results of each sub-box to obtain the decryption results of each sub-box on the data end;

[0127] A determination module 44 is used to determine the analysis results of the data end according to the decryption results of each bin and the total number of samples according to the bin number, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federal modeling according to the analysis results, wherein the analysis results of each bin are used to represent the correlation between the characteristic parameters currently used for binning and the positive sample labels of the business end.

[0128] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0129] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0130] In this embodiment, any multiple of the sending module 41, receiving module 42, decryption module 43, and determination module 44 can be combined into a single module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in a single module. At least one of the sending module 41, receiving module 42, decryption module 43, and determination module 44 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented in hardware or firmware through any other reasonable method of circuit integration or packaging, or can be implemented in any one of the three implementation methods of software, hardware, and firmware, or any appropriate combination of any of these. Alternatively, at least one of the sending module 41, receiving module 42, decryption module 43, and determination module 44 can be at least partially implemented as a computer program module that, when executed, can perform the corresponding function.

[0131] See also Figure 5 , an embodiment of the present disclosure provides a screening device for federated feature engineering data applied to a data end, comprising:

[0132] A receiving module 51 is configured to receive a sample ID and a homomorphically encrypted designated label sent by a service end, wherein the designated label is a positive sample label or a negative sample label;

[0133] A binning module 52, which is used to bin the sample IDs of the sample alignment according to preset characteristic parameters;

[0134] A calculation module 53 is used to calculate the sum of the ciphertexts of each bin based on the homomorphically encrypted designated labels corresponding to the sample IDs in each bin;

[0135] The sending module 54 is used to send the bin number, ciphertext summation result and total number of samples of each bin to the business end.

[0136] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0137] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0138] In this embodiment, any multiple of the receiving module 51, the binning module 52, the computing module 53, and the sending module 54 can be combined into a single module, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in a single module. At least one of the receiving module 51, the binning module 52, the computing module 53, and the sending module 54 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented in hardware or firmware by any other reasonable means of integrating or packaging circuits, or can be implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of these. Alternatively, at least one of the receiving module 51, the binning module 52, the computing module 53, and the sending module 54 can be at least partially implemented as a computer program module, which can perform the corresponding function when the computer program module is executed.

[0139] like Figure 1 In the federated feature engineering data screening system including the business side and the data side, the interaction process between the business side and the data side is as follows: Figure 6 shown.

[0140] The method provided by the embodiments of the present disclosure can achieve efficient feature evaluation by reducing the level of encryption and decryption samples. Compared with the prior art that requires encrypting the labels y and 1-y of each sample based on homomorphic encryption schemes such as Paillier and then sending them to the data party for secret calculation, only one of the labels y and 1-y needs to be encrypted and sent, thereby reducing the encryption, transmission and other overheads by nearly half, while not reducing the security of the calculation process and the accuracy of the results, and ultimately achieving more efficient and quick evaluation of the characteristics of each data party.

[0141] like Figure 1 The workflow of the federated feature engineering data screening system, including the business side and the data side, is shown as follows Figure 7 Shown, including:

[0142] (1) Filtering invalid features at the data end

[0143] In order to prevent the existence of some invalid features from affecting the evaluation efficiency and effectiveness of other features, it is necessary to filter and delete these invalid features before formally conducting feature evaluation. Invalidity specifically refers to the low degree of discrimination of features for samples. For example, all samples are null values ​​or the same value under a certain feature. The former can be judged by calculating the sample coverage of the feature, and the latter can be judged by calculating the variance value of each feature. If the variance of a feature is close to 0, that is, there is basically no difference between the eigenvalues ​​of the feature, then this feature is of no use in distinguishing samples. In addition, in addition to the effectiveness of the quantitative features themselves, the correlation between features also needs to be considered. For example, the Pearson correlation between features can be calculated to filter out highly correlated feature combinations, and ultimately only one feature is retained to participate in the subsequent modeling process.

[0144] (2) Aligning encrypted samples between the data end and the business end

[0145] After eliminating invalid features from the data provided by each participant, before conducting federated feature engineering, it is necessary to find the intersection of the user groups of the participants. That is, based on mechanisms such as RSA encryption, the common user set of the participants is found under the premise that the federated modeling participants do not disclose their respective user lists, while also not exposing the private users of the participants, that is, non-overlapping users.

[0146] (3) Business-side encryption tag

[0147] After all participants in the federated modeling align their samples, the business end will assign the label y of each aligned sample to the n After encryption based on additive homomorphic encryption, we get <y n >, then the sample primary key ID and the label ciphertext value corresponding to the ID <y n> are transmitted to the data end together. Assuming that the sample label is a binary label, the label will be mapped to 0 and 1 before encryption, and the label 1 is set to represent a good user and the label 0 is set to represent a bad user. Assuming that the total number of samples after alignment is N, N-level labels will be encrypted and 2N-level data will be transmitted.

[0148] (4) Data end bin statistics

[0149] The data side divides the samples into bins according to the features to be evaluated. That is, for the features to be analyzed, the data side divides the samples into bins according to the preset binning methods such as equal frequency, equal distance, etc., so as to divide N samples into a few different grouping bins according to the binning threshold. i And count the total number of samples in each bin sum(bin i ). Based on step (3), the data side obtains the ciphertext of the label corresponding to each sample from the business side. <y n >After that, the ciphertext values ​​in each bin are summed up to get the sum in each bin. <y n >), and finally the bin number bin i 、The total number of samples in each bin sum(bin i ) and the sum of the secret labels of each bin ( <y n >)Synchronize to the business side.

[0150] (5) Business-side decryption assessment

[0151] After receiving the information from the data party, the business party will first decrypt the ciphertext sum value to obtain the sum (y n ), which specifically means the total number of positive samples Good in the i-th bin i (Step (2) assumes that the labels are divided into two categories, 1 and 0, where 1 represents positive samples and 0 represents negative samples.) Since the labels in the binary classification are either positive or negative, the number of negative samples = the total number - the number of positive samples. The number of each bin can be calculated using the following formula: i The number of negative samples.

[0152] Bad i =sum(bin i )-Good i

[0153] In summary, the business side has obtained the number of good and bad people in each bin corresponding to the data side's features, and the business side itself has the label data. For the business side's own feature evaluation, it only needs to calculate the number of good and bad people in each bin locally. T and Bad T ; After getting the number of good and bad people in each bin, calculate the number of good and bad people in each bin in turn. And for each bin IV i The sum of the values Among them, IV is information value.

[0154] At the same time, after obtaining the number of good and bad people in each box, KS=max{|cum i (Bad)-cum i (Good)|}, where cum i (Bad) is the cumulative number of negative samples in each bin interval i, cum i (Good) is the cumulative number of positive samples in each bin interval i, and KS (the name (Kolmogorov-Smimov) comes from two Soviet mathematicians ANKolmogorov and NVSmirnov, and is constructed through the empirical cumulative distribution function) is the maximum of these absolute values.

[0155] The method provided by the embodiment of the present disclosure utilizes the fact that the number of negative samples in the binary classification can be calculated by the difference between the total number and the number of positive samples. By replacing the double homomorphic encryption in the prior art with a single homomorphic encryption of the label, calculations such as Woe (weight of evidence) can also be realized. Compared with the traditional federated modeling, which requires encryption and summation of the labels y and 1-y respectively, the computational efficiency of this solution can be improved by nearly 50%, so it is highly efficient. In terms of transmission efficiency, since the present invention only requires homomorphic encryption of the label once, only ciphertext of size N needs to be transmitted during communication. Compared with the existing solution, the transmission efficiency is greatly improved, which can reduce excessive communication consumption and reduce the communication requirements in actual business. In terms of security, the present disclosure uses homomorphic encryption to transmit and sum the labels throughout the process, and does not require additional information beyond the original solution, and does not reduce the security of the calculation process and the accuracy of the results.

[0156] Reference Figure 8 As shown, an embodiment of the present disclosure provides an electronic device, including a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140;

[0157] Memory 1130, for storing computer programs;

[0158] The processor 1110 is configured to implement the following method for screening federated feature engineering data when executing the program stored in the memory 1130:

[0159] Sending a sample ID aligned with the sample of the data end and a designated label of homomorphic encryption to the data end, so that the data end can determine the ciphertext summation result of each bin according to the designated label of homomorphic encryption;

[0160] Receive the ciphertext summation result of each bin, bin number and total number of samples sent by the data end;

[0161] Decrypting the sum of the ciphertexts of each received sub-box to obtain the decryption results of each sub-box on the data end;

[0162] According to the bin number, the analysis result of the data end is determined according to the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling based on the analysis result, wherein the analysis result of each bin is used to represent the correlation between the characteristic parameters currently used for binning and the positive sample label of the business end.

[0163] The communication bus 1140 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, the figure shows only one thick line, but this does not mean that there is only one bus or only one type of bus.

[0164] The communication interface 1120 is used for communication between the electronic device and other devices.

[0165] The memory 1130 may include a random access memory (RAM) or a non-volatile memory, such as at least one disk storage. Alternatively, the memory 1130 may be at least one storage device located away from the processor 1110.

[0166] The above-mentioned processor 1110 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0167] An embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for screening federated feature engineering data as described above.

[0168] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments, or may exist independently and not incorporated into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the method for screening federated feature engineering data according to the embodiments of the present disclosure.

[0169] According to an embodiment of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as, but not limited to, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0170] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0171] The foregoing description is intended only to provide specific embodiments of the present disclosure, intended to enable those skilled in the art to understand and implement the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure is not to be limited to the embodiments shown herein, but is to be construed in the broadest manner consistent with the principles and novel features claimed herein.

Claims

1. A method for screening federated feature engineering data, characterized in that: Applied to the business end, the method includes: Sending a sample ID aligned with the sample of the data end and a designated label of homomorphic encryption to the data end, so that the data end can determine the ciphertext summation result of each bin according to the designated label of homomorphic encryption; Receive the ciphertext summation result of each bin, bin number and total number of samples sent by the data end; Decrypting the sum of the ciphertexts of each received sub-box to obtain the decryption results of each sub-box on the data end; According to the bin number, the analysis result of the data end is determined according to the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling based on the analysis result, wherein the analysis result of each bin is used to represent the correlation between the characteristic parameters currently used for binning and the positive sample label of the business end, When the total amount of information on the data end is used as an analysis result, determining whether to select data binned according to the features currently used for binning as data for federated modeling according to the analysis result includes: In the preset binning method, the total amount of information on the data end is compared with a first preset threshold: When the total amount of information on the data end is less than a first preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the total amount of information on the data end is greater than or equal to the preset threshold; When the total amount of information on the data end is greater than or equal to a first preset threshold, the total amount of information on the data end is compared with a second preset threshold: When the total amount of information on the data end is less than a second preset threshold, discarding the data that is binned according to the feature currently used for binning; When the total amount of information on the data end is greater than or equal to a second preset threshold, the data that is binned according to the features currently used for binning is selected as data for federated modeling.

2. The method according to claim 1, characterized in that The decryption result includes: the number of samples corresponding to the specified label, and the analysis result of the data end is determined based on the decryption result of each bin and the total number of samples, including: Calculating the total number of positive samples and the total number of negative samples in each bin of the data end, and the total number of positive samples and the total number of negative samples in all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples, wherein the designated label is a positive sample label or a negative sample label; Calculate the evidence weight of each bin on the data side according to the total number of positive samples and the total number of negative samples in each bin on the data side and the total number of positive samples and the total number of negative samples in all bins; The amount of information of each bin of the data end is calculated according to the evidence weight of each bin of the data end, and the total amount of information of the data end is obtained as the analysis result by summing up.

3. The method according to claim 2, characterized in that The designated label is a positive sample label, and the calculation of the total number of positive samples and the total number of negative samples for each bin of the data end, and the total number of positive samples and the total number of negative samples for all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples includes: The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of positive samples of each bin of the data end; The difference between the total number of samples in each bin at the data end and the total number of positive samples is used as the total number of negative samples in each bin at the data end; The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins; The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

4. The method according to claim 2, characterized in that The designated label is a negative sample label, and the calculation of the total number of positive samples and the total number of negative samples for each bin of the data end, and the total number of positive samples and the total number of negative samples for all bins based on the number of samples corresponding to the designated label of each bin of the data end and the total number of samples includes: The number of samples corresponding to the specified labels of each bin of the data end is used as the total number of negative samples of each bin of the data end; The difference between the total number of samples in each bin at the data end and the total number of negative samples is used as the total number of positive samples in each bin at the data end; The sum of the total number of positive samples in each bin at the data end is taken as the total number of positive samples in all bins; The sum of the total number of negative samples in each bin at the data end is taken as the total number of negative samples in all bins.

5. The method according to claim 2, characterized in that The step of determining the analysis result of the data terminal according to the decryption result of each sub-box and the total number of samples further includes: According to the total number of positive samples and the total number of negative samples in each bin of the data end, the KS value of the data on the data end is calculated as the analysis result.

6. The method according to claim 5, characterized in that The KS value of the data on the data end is calculated according to the total number of positive samples and the total number of negative samples in each bin of the data end through the following expression: in, For each bin interval The cumulative number of negative samples, For each bin interval The cumulative number of positive samples.

7. A method for screening federated feature engineering data, characterized in that: Applied to the data end, the method includes: Receive a sample ID and a homomorphically encrypted designated label sent by the service end, which are aligned with the sample on the data end, wherein the designated label is a positive sample label or a negative sample label; Bin the sample IDs of the sample alignment according to the preset feature parameters; Calculate the sum of the ciphertexts of each bin based on the homomorphically encrypted specified labels corresponding to the sample IDs in each bin; The bin number, ciphertext sum result and total number of samples of each bin are sent to the business end, so that the ciphertext sum result of each bin received is decrypted at the business end to obtain the decryption result of each bin on the data end; and the analysis result of the data end is determined according to the bin number, the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling according to the analysis result. When the total amount of information on the data end is used as an analysis result, determining whether to select data binned according to the features currently used for binning as data for federated modeling according to the analysis result includes: In the preset binning method, the total amount of information on the data end is compared with a first preset threshold: When the total amount of information on the data end is less than a first preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the total amount of information on the data end is greater than or equal to the preset threshold; When the total amount of information on the data end is greater than or equal to a first preset threshold, the total amount of information on the data end is compared with a second preset threshold: When the total amount of information on the data end is less than a second preset threshold, discarding the data that is binned according to the feature currently used for binning; When the total amount of information on the data end is greater than or equal to a second preset threshold, the data that is binned according to the features currently used for binning is selected as data for federated modeling.

8. The method according to claim 7, characterized in that The preset characteristic parameters are determined by the following steps: Calculate the sample coverage and / or variance value of each selected feature parameter of all sample data; Delete feature parameters whose sample coverage is less than a preset coverage threshold, and / or feature parameters whose variance is less than a preset variance threshold; For any two of the remaining candidate feature parameters, the Pearson correlation coefficient between the two candidate feature parameters is calculated, and when the Pearson correlation coefficient is greater than or equal to the preset coefficient threshold, one of the two candidate feature parameters is deleted until the Pearson correlation coefficient is no longer greater than or equal to the preset coefficient threshold, and the last remaining candidate feature parameter is used as the preset feature parameter.

9. A device for screening federated feature engineering data, characterized in that: Applied to the business end, the device includes: A sending module, configured to send a sample ID aligned with the sample of the data end and a homomorphically encrypted positive sample label to the data end, so that the data end can determine the ciphertext summation result of each bin according to the homomorphically encrypted positive sample label; A receiving module is used to receive the ciphertext summation result of each bin, bin number and total number of samples sent by the data end; A decryption module, configured to decrypt the received ciphertext summation results of each sub-box to obtain a decryption result of each sub-box on the data end; A determination module, which is used to determine the analysis results of the data end according to the decryption results of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling according to the analysis results, wherein the analysis results of each bin are used to represent the correlation between the characteristic parameters currently used for binning and the positive sample labels of the business end; When the total amount of information on the data end is used as an analysis result, determining whether to select data binned according to the features currently used for binning as data for federated modeling according to the analysis result includes: In the preset binning method, the total amount of information on the data end is compared with a first preset threshold: When the total amount of information on the data end is less than a first preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the total amount of information on the data end is greater than or equal to the preset threshold; When the total amount of information on the data end is greater than or equal to a first preset threshold, the total amount of information on the data end is compared with a second preset threshold: When the total amount of information on the data end is less than a second preset threshold, discarding the data that is binned according to the feature currently used for binning; When the total amount of information on the data end is greater than or equal to a second preset threshold, the data that is binned according to the features currently used for binning is selected as data for federated modeling.

10. A device for screening federated feature engineering data, characterized in that: Applied to a data terminal, the device includes: A receiving module, configured to receive a sample ID and a homomorphically encrypted designated label sent by a service end, the designated label being a positive sample label or a negative sample label; A binning module, which is used to bin the sample IDs of the sample alignment according to preset feature parameters; A calculation module is used to calculate the sum of the ciphertexts of each bin based on the homomorphically encrypted specified labels corresponding to the sample IDs in each bin; The sending module is used to send the bin number, ciphertext sum result and total number of samples of each bin to the business end, so that the business end decrypts the received ciphertext sum result of each bin to obtain the decryption result of each bin on the data end; and determines the analysis result of the data end according to the bin number, the decryption result of each bin and the total number of samples, so as to determine whether to select the data binned according to the characteristics currently used for binning as the data for federated modeling based on the analysis result. When the total amount of information on the data end is used as an analysis result, determining whether to select data binned according to the features currently used for binning as data for federated modeling according to the analysis result includes: In the preset binning method, the total amount of information on the data end is compared with a first preset threshold: When the total amount of information on the data end is less than a first preset threshold, changing the binning method, and calculating the total amount of information corresponding to the changed binning method until the total amount of information on the data end is greater than or equal to the preset threshold; When the total amount of information on the data end is greater than or equal to a first preset threshold, the total amount of information on the data end is compared with a second preset threshold: When the total amount of information on the data end is less than a second preset threshold, discarding the data that is binned according to the feature currently used for binning; When the total amount of information on the data end is greater than or equal to a second preset threshold, the data that is binned according to the features currently used for binning is selected as data for federated modeling.

11. An electronic device, characterized in that: The processor, the communication interface, the memory and the communication bus are connected to each other via the communication bus. Memory for storing computer programs; The processor is configured to implement the method for screening federated feature engineering data according to any one of claims 1 to 8 when executing a program stored in the memory.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for screening federated feature engineering data according to any one of claims 1 to 8 is implemented.

Images