WO2022227644A1 - 数据处理方法、装置、设备、存储介质及程序产品 - Google Patents

数据处理方法、装置、设备、存储介质及程序产品 Download PDF

Info

Publication number
WO2022227644A1
WO2022227644A1 PCT/CN2021/140955 CN2021140955W WO2022227644A1 WO 2022227644 A1 WO2022227644 A1 WO 2022227644A1 CN 2021140955 W CN2021140955 W CN 2021140955W WO 2022227644 A1 WO2022227644 A1 WO 2022227644A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
data
matrix
sample
participant
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/140955
Other languages
English (en)
French (fr)
Inventor
魏文斌
范涛
陈天健
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
WeBank Co Ltd
Original Assignee
WeBank Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by WeBank Co Ltd filed Critical WeBank Co Ltd
Publication of WO2022227644A1 publication Critical patent/WO2022227644A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning

Definitions

  • the present application relates to the technical field of artificial intelligence, and relates to, but is not limited to, a data processing method, apparatus, device, storage medium and program product.
  • Machine learning is a science that studies how to use computers to simulate or realize human learning activities. It is one of the most intelligent and cutting-edge research fields in artificial intelligence.
  • the research of machine learning is mainly divided into two research directions: the first is the research of traditional machine learning, which mainly studies the learning mechanism, focusing on the exploration of the learning mechanism of simulated human; the second is the research of machine learning in the big data environment.
  • Research, this type of research mainly studies how to effectively use information, focusing on obtaining hidden, effective and understandable knowledge from huge amounts of data.
  • Federated learning technology is an emerging privacy protection technology that can effectively combine data from all parties for model training without the need for local data.
  • Many business problems in the field of big data can be solved by corresponding machine learning models.
  • Removing collinear data is the key to training a good model.
  • related technologies cannot quantify the collinearity of multi-party data in federated learning, and cannot efficiently screen and eliminate collinear training data, resulting in low accuracy and poor stability of the trained model. .
  • the embodiments of the present application provide a data processing method, device, device, computer-readable storage medium, and computer program product, which can eliminate collinear data in linear federated modeling, improve the accuracy and stability of the federated model, and improve the The modeling effect of the model.
  • the embodiment of the present application provides a data processing method, which is applied to the first participant of federated learning, and the method includes:
  • a virtual feature correlation matrix is constructed, and the security computing model is composed of the first participant and other participants in federated learning.
  • party is pre-trained based on secure multi-party computation;
  • the feature data of the target feature and the feature data of at least one feature among other features, and the other features include features other than the target feature held by the first participant and the other features. characteristics held by the party.
  • An embodiment of the present application provides a data processing apparatus, which is applied to a first participant of federated learning, and the apparatus includes:
  • the building module is configured to construct a virtual feature correlation matrix based on the first sample feature data held by the first participant and a pre-trained security computing model, and the security computing model is composed of the first participant and the pre-trained security computing model.
  • Other participants in federated learning are pre-trained based on secure multi-party computation;
  • a first determining module configured to determine, based on the feature correlation matrix, a collinear quantization factor of each feature corresponding to the first sample feature data
  • a second determining module configured to determine a target feature from each feature corresponding to the first sample feature data based on the collinearity quantization factor
  • a deletion module configured to delete the feature data of the target feature from the first sample feature data, to obtain the first training data for the joint training of the first participant and the other participants;
  • the feature data of the target feature and the feature data of at least one feature among other features, and the other features include features other than the target feature held by the first participant and the other features. characteristics held by the party.
  • An embodiment of the present application provides a data processing device, and the device includes:
  • a memory configured to store executable instructions
  • the processor when configured to execute the executable instructions stored in the memory, implements the method provided by the embodiment of the present application.
  • Embodiments of the present application provide a computer-readable storage medium, where executable instructions are stored on the computer-readable storage medium, and are configured to cause a processor to execute the method to implement the method provided by the embodiments of the present application.
  • the embodiments of the present application provide a computer program product, including a computer program, which implements the methods provided by the embodiments of the present application when the computer program is executed by a processor.
  • the embodiments of the present application have the following beneficial effects: in the data processing method provided by the embodiments of the present application, during data processing, first, the first participant and other participants of the federated learning are pre-trained to obtain a secure computing model based on secure multi-party computation; Then the first participant obtains the first sample feature data held by the first participant, and builds a virtual feature correlation matrix based on the first sample feature data and the pre-trained security computing model; and then determines the first sample based on the feature correlation matrix.
  • the feature data has a linear relationship, and other features here include not only the features held by the first participant except the target feature, but also all features held by other participants; after determining the target feature, from the first sample feature data
  • the feature data of the target feature is deleted to obtain the first training data for the joint training of the first participant and other participants. In this way, under the premise of protecting data privacy, the data with collinearity in the feature data held by each participant can be screened and eliminated, and the training data without linear relationship can be obtained.
  • Using training data without linear relationship for joint training can improve the accuracy and stability of the federated model and improve the modeling effect of the federated model.
  • FIG. 1 is a schematic diagram of a network architecture of a data processing method provided by an embodiment of the present application
  • FIG. 2 is a schematic diagram of the composition and structure of a data processing device provided by an embodiment of the present application.
  • FIG. 3 is a schematic diagram of a realization flow of the data processing method provided by the embodiment of the present application.
  • FIG. 4 is a schematic flowchart of another implementation of the data processing method provided by the embodiment of the present application.
  • FIG. 5 is a schematic flowchart of the calculation flow of the variance inflation factor under the vertical federation situation provided by the embodiment of the present application;
  • FIG. 6 is a schematic flowchart of a calculation flow of a determinant of a correlation matrix provided by an embodiment of the present application.
  • first ⁇ second ⁇ third is only used to distinguish similar objects, and does not represent a specific ordering of objects. It is understood that “first ⁇ second ⁇ third” Where permitted, the specific order or sequence may be interchanged to enable the embodiments of the application described herein to be practiced in sequences other than those illustrated or described herein.
  • VIF Variance Inflation Factor
  • Variance Inflation Factor also known as Variance Inflation Factor
  • VIF Variance Inflation Factor
  • Homomorphic Encryption is a cryptographic technique based on the computational complexity theory of mathematical problems. Processing the homomorphically encrypted data to obtain an output, decrypting this output yields the same output as processing the unencrypted raw data in the same way.
  • the following describes an exemplary application of the apparatus for implementing the embodiment of the present application, and the apparatus provided by the embodiment of the present application may be implemented as a terminal device.
  • exemplary applications covering terminal devices when the apparatus is implemented as a terminal device will be described.
  • FIG. 1 is a schematic diagram of a network architecture of a data processing method provided by an embodiment of the present application.
  • the network architecture at least includes a first participant 100 , a second participant 200 , and a network 300 .
  • the first participant 100 and the second participant 200 may be the participants in the vertical federated learning that jointly train the machine learning model.
  • the first participant 100 and the second participant 200 may be clients, for example, various banks or hospitals and other participant devices that store user characteristic data, and the clients may be laptops, tablet computers, desktop computers, special training equipment and other devices with model training capabilities.
  • the first participant 100 is connected to the second participant 200 through a network 300.
  • the network 300 may be a wide area network or a local area network, or a combination of the two, using wireless or wired links to realize data transmission.
  • the first participant 100 first obtains first sample feature data from its own data, and processes the first sample feature data to obtain processed first sample feature data.
  • the second participant 200 obtains second sample feature data from its own data, and processes the second sample feature data to obtain processed second sample feature data.
  • the first sample feature data and the second sample feature data have the same identifier, that is, the first sample feature data and the second sample feature data are the same identifiers held by the first participant 100 and the second participant 200 respectively. Data of different characteristics of batch samples.
  • the first participant 100 and the second participant 200 use the processed first sample feature data and the processed second sample feature data to determine the first matrix E based on secure multi-party computation, the first participant 100 and the second participant Each of the squares 200 holds the first matrix E.
  • the first participant 100 can only obtain the first matrix E, and cannot know the processed second sample feature data held by the second participant 200; similarly, the second participant 200 only has To obtain the first matrix E, it is impossible to obtain the processed first sample feature data held by the first participant 100 .
  • the first participant 100 constructs a virtual feature correlation matrix according to the first matrix E and the processed first sample feature data; then calculates the determinant of the feature correlation matrix, and the feature correlation matrix.
  • the determinant of each cosubformula corresponding to the matrix based on the determinant of the feature correlation matrix and the determinant of each cosubformula, determine the collinearity quantization factor of each feature corresponding to the first sample feature data, and the collinearity quantization factor can be the variance
  • the expansion coefficient is used to quantify the collinearity of each feature with all other features. All other features here include not only other features held by the first participant 100 except the quantified features this time, but also the second participant 200. have all the characteristics.
  • the first participant 100 determines which features have collinearity between the feature data according to the collinearity quantification factor of each feature, determines these features as target features, and finally deletes the feature data of the target feature from the first sample feature data , to obtain the first training data for the joint training of the first participant 100 and the second participant 200 .
  • the second participant 200 obtains second training data for joint training. Therefore, during joint training, the first participant 100 and the second participant 200 use the first training data and the second training data without collinearity to perform joint training, so that a federated model with high accuracy and good stability can be obtained.
  • the data with collinearity in the feature data held by each participant can be screened and eliminated, and the training data without linear relationship can be obtained.
  • each participant uses the training data without linear relationship for joint training, which can improve the accuracy and stability of the federated model and improve the modeling effect of the federated model.
  • the apparatuses provided in the embodiments of the present application may be implemented in a manner of hardware or a combination of software and hardware, and various exemplary implementations of the apparatuses provided in the embodiments of the present application are described below.
  • the data processing device 10 is shown by taking the device applied to the first participant of federated learning as an example, and other exemplary structures of the data processing device 10 can be foreseen, Therefore, the structure described here should not be regarded as a limitation. For example, some components described below may be omitted, or components not described below may be added to meet the special requirements of certain applications.
  • the data processing device 10 shown in FIG. 2 includes: at least one processor 110, a memory 140, at least one network interface 120 and a user interface 130. Each component in data processing device 10 is coupled together by bus system 150 . It can be understood that the bus system 150 is used to implement the connection communication between these components. In addition to the data bus, the bus system 150 also includes a power bus, a control bus, and a status signal bus. However, for clarity of illustration, the various buses are labeled as bus system 150 in FIG. 2 .
  • User interface 130 may include a display, keyboard, mouse, touch pad, touch screen, and the like.
  • Memory 140 may be volatile memory or nonvolatile memory, and may include both volatile and nonvolatile memory.
  • the non-volatile memory may be a read-only memory (ROM, Read Only Memory).
  • the volatile memory may be random access memory (RAM, Random Access Memory).
  • RAM Random Access Memory
  • the memory 140 in the embodiment of the present application can store data to support the operation of the data processing device 10 .
  • Examples of such data include: any computer programs used to operate on data processing device 10, such as operating systems and applications.
  • the operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks.
  • Applications can contain various applications.
  • the method provided by the embodiment of the present application may be directly embodied as a combination of software modules executed by the processor 110, and the software module may be located in a storage medium, and the storage medium is located in the memory 140,
  • the processor 110 reads the executable instructions included in the software module in the memory 140, and combines necessary hardware (for example, including the processor 110 and other components connected to the bus 150) to complete the method provided by the embodiments of the present application.
  • the processor 110 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices , discrete hardware components, etc., where a general-purpose processor may be a microprocessor or any conventional processor, or the like.
  • DSP Digital Signal Processor
  • FIG. 3 is a schematic diagram of an implementation flow of the data processing method provided by the embodiment of the present application, which is applied to the first participant of the network architecture shown in FIG. 1 , and will be described with reference to the steps shown in FIG. 3 .
  • Step S301 construct a virtual feature correlation matrix based on the first sample feature data held by the first participant and the pre-trained security computing model.
  • the feature data of different features may have a linear relationship. If a collinear data training model is used, the trained model has low accuracy and poor stability.
  • the first participant constructs a virtual feature correlation matrix, denoted as H, according to the first sample feature data and the pre-trained security computing model.
  • the secure computing model is pre-trained by the first participant and other participants of federated learning based on secure multi-party computation, and can perform matrix addition and subtraction and matrix multiplication under privacy.
  • the data processing method is described by taking two participants performing federated learning as an example. In practical applications, the data processing method can also be applied to three or more participants for federated learning.
  • the first participant and the second participant conduct training based on privacy protection technology to obtain a trained security computing model.
  • the trained security computing model is owned by each participant.
  • the training data needs to come from the same user to perform joint training, that is, the first sample feature used by the first participant for joint training
  • the data is the same as the second sample feature data used by the second participant for joint training, and their corresponding user identifiers (ID, Identity Document) are the same. Therefore, before step S301, the method further includes: the first participant acquires the first sample characteristic data.
  • acquiring the feature data of the first sample may be implemented as: acquiring a common training sample held by the first participant and other participants; screening out this training sample from the common training samples; The characteristic data of the sample is determined as the characteristic data of the first sample.
  • the common training samples are the common users held by each participant. Each participant obtains the identifier of the data held by each participant, and based on the encrypted calculation, the identifier of the data held by each participant is used to determine the common user of these participants. Then, the target users participating in the training are selected from these common users as the training samples, and the feature data of these target users is used as the first sample feature data.
  • Step S302 based on the feature correlation matrix, determine the collinearity quantization factor of each feature corresponding to the feature data of the first sample.
  • methods for determining the quantitative factor of collinearity among the feature data of multiple features include variance inflation factor method, eigenroot analysis method and condition number method.
  • an intuitive judgment method that can qualitatively analyze the degree of collinearity between feature data of multiple features. The intuitive judgment method is generally used for preliminary judgment and cannot be quantitatively analyzed.
  • the first sample feature data held by the first participant and the second sample feature data held by the second participant correspond to the data in different columns in the feature correlation matrix. Whether there is collinearity between the column data, it can be determined whether there is collinearity between the feature data of each feature in the first sample feature and the feature data of other features.
  • the existence of collinearity between the feature data of different features includes the following possibilities: the feature data of one feature is a multiple of the feature data of another feature; the feature data of one feature is equal to the feature data of another feature plus a constant term; There is a feature whose feature data is equal to the sum of the feature data of the other two features.
  • the collinearity quantization factor of each feature corresponding to the first sample feature data is determined, and which features can be determined according to the collinearity quantization factor of each feature corresponding to the first sample feature data There is feature data that is collinear with it.
  • Step S303 based on the collinearity quantization factor, determine the target feature from the features corresponding to the first sample feature data.
  • the target features with collinearity are determined.
  • the variance inflation factor in the absence of multicollinearity, the variance inflation factor is close to 1, and the stronger the multicollinearity, the larger the variance inflation factor. In fact, there is always more or less multicollinearity between the data, so it is unrealistic to use the variance expansion factor equal to 1 as the criterion for evaluating collinearity.
  • a boundary value can be preset according to the actual application scenario, In this embodiment of the present application, the preset boundary value may be 10. Determine whether the variance expansion factor of each feature is greater than 10. When the variance expansion factor of a feature is greater than 10, it is considered that the feature data of this feature and the feature data of other features have strong collinearity, and the feature is determined as the target. feature.
  • Step S304 delete the feature data of the target feature from the first sample feature data to obtain the first training data for the joint training of the first participant and other participants.
  • the other features here include features held by the first participant other than the target feature and features held by other participants.
  • the variance expansion factor of each feature in the first sample feature data of the first participant may be calculated based on the feature correlation matrix, and the target feature with the variance expansion factor greater than the preset boundary value is determined, and then the first sample The feature data of the target feature in the feature data is deleted, and the feature data without collinearity is obtained as the first training data. Therefore, when performing joint training, the first participant uses the first training data without collinearity to perform joint training, and can obtain a federated model with high accuracy and good stability.
  • the data processing method provided by the embodiment of the present application is applied to the first participant of federated learning, and the data processing method includes: based on the first sample feature data held by the first participant and the pre-trained security computing model, constructing A virtual feature correlation matrix, the secure computing model is pre-trained by the first participant and other participants in federated learning based on secure multi-party computation; based on the feature correlation matrix, determine the collinearity quantification of each feature corresponding to the feature data of the first sample factor; based on the collinearity quantification factor, the target feature is determined from the features corresponding to the first sample feature data; the feature data of the target feature is deleted from the first sample feature data, and the first participant and other participants are obtained to combine The first training data for training; wherein, the feature data of the target feature has a linear relationship with the feature data of at least one of the other features, and the other features include the features held by the first participant other than the target feature and the features held by other participants.
  • the data with collinearity in the feature data held by each participant can be screened and eliminated, and the training data without linear relationship can be obtained.
  • the participants use the training data without linear relationship for joint training, which can improve the accuracy and stability of the federated model and improve the modeling effect of the federated model.
  • the first participant before step S301 in the embodiment shown in FIG. 3 , the first participant first obtains the first sample feature data from the feature data held by itself.
  • the training data needs to come from the same user for joint training.
  • the second sample feature data used by the party for joint training, and their corresponding IDs are the same.
  • the first participant and the second participant each hold feature data of different features of a batch of users, for example, the first participant has feature data of a batch of users' birthday AA, age BB, weight CC, and deposit DD. , the second party has a batch of feature data of users' age BB, consumption ability EE, and hobby FF.
  • the database table of the first participant and the database table of the second participant are shown in Table 1 and Table 2 below:
  • the characteristic data held by the first participant is obtained Obtain characteristic data held by the second party according to the database table of the second party
  • the first participant determines the target users participating in this training, for example, randomly selects random users from the common users as the target users participating in this training, and then obtains these targets from the data stored by itself.
  • the feature data of the user is taken as the first sample feature data.
  • the first participant sends the IDs of the screened target users to the second participant, and the second participant obtains the characteristic data of these target users from the data stored by itself as the second sample characteristic data.
  • the second participant may also determine the target user participating in the training, and send the ID of the determined target user to the first participant.
  • the second participant selects the characteristic data of the target user S1 from the characteristic data held by the second participant to obtain the second sample characteristic data
  • the first participant obtains the first sample feature data
  • the second participant obtains the second sample feature data
  • step S301 based on the first sample feature data held by the first participant and the pre-trained security computing model, construct a virtual feature correlation matrix" can be performed through the following Steps to achieve:
  • Step S3011 based on the first sample feature data, determine the feature data of each feature corresponding to the first sample feature data and the number of samples corresponding to the first sample feature data.
  • the corresponding features of the first sample feature data are: AA, BB, CC, and DD
  • the feature data of feature AA is ⁇ a1, a3, a4 ⁇
  • the feature data of feature BB The feature data is ⁇ b1,b3,b4 ⁇
  • the feature data of feature CC is ⁇ c1,c3,c4 ⁇
  • the feature data of feature DD is ⁇ d1,d3,d4 ⁇
  • the samples corresponding to the first sample feature data are: ⁇ a1,b1,c1,d1 ⁇ , ⁇ a3,b3,c3,d3 ⁇ and ⁇ a4,b4,c4,d4 ⁇
  • Step S3012 Calculate the mean and standard deviation corresponding to the feature data of each feature respectively.
  • Step S3013 Determine the processed first sample feature data based on the feature data of each feature, the mean value corresponding to the feature data of each feature, the standard deviation corresponding to the feature data of each feature, and the number of samples;
  • the mean value corresponding to the feature data of each feature may be The standard deviation ⁇ x corresponding to the feature data of each feature and the number of samples n are used to update the feature data x of each feature to obtain each feature data x after processing.
  • Step S3014 Input the processed first sample characteristic data into the security calculation model to obtain a first matrix.
  • each participant performs training based on a secure multi-party computing protocol, such as the SPDZ protocol or MASCOT, to obtain a trained secure computing model.
  • the trained model is capable of matrix addition and subtraction and matrix multiplication under privacy.
  • Secure Multi-Party Computation (MPC, Secure Multi-Party Computation) research is mainly aimed at the problem of how to safely calculate a contract function in the absence of a trusted third party.
  • MPC Secure Multi-Party Computation
  • the m participants all hold the security computing model, and each participant inputs its own private input value to obtain the corresponding output value. For example, the i-th participant inputs x i and outputs y i , Among them, 1 ⁇ i ⁇ m.
  • the first matrix E is determined based on a secure multi-party computation protocol, each participant cannot know the private data of other participants, and the privacy and security of each participant's data can be ensured.
  • Step S3015 construct a virtual feature correlation matrix according to the processed first sample feature data and the first matrix.
  • Step S30151 Determine a first symmetric matrix according to the processed first sample feature data.
  • Symmetric Matrices refers to a square matrix with the main diagonal as the symmetry axis and each element corresponding to the same.
  • the first symmetric matrix F is constructed according to the processed first sample characteristic data C, and the transposed matrix of the processed first sample data C may be used. and the processed first sample data
  • Step S30152 generate an empty matrix whose number of rows and columns are equal to the number of columns of the obtained first matrix.
  • a virtual feature correlation matrix is constructed by the first participant, which is related to the feature data of the feature of the second participant.
  • the associated data is empty.
  • the first matrix E is a 4*3 matrix, and the number of columns is 3, so the generated empty matrix is a matrix with a dimension of 3*3, denoted as G'.
  • the second symmetric matrix G corresponding to the empty matrix G' is determined by the second participant according to the second sample feature data B.
  • Step S30153 construct a virtual feature correlation matrix according to the first symmetric matrix, the first matrix, the transposed matrix of the first matrix, and the empty matrix.
  • the first symmetric matrix F the first matrix E, the transposed matrix E T of the first matrix, and the empty matrix G', construct a virtual feature correlation matrix H, which is expressed as
  • the dimension of the first symmetric matrix F is 4*4
  • the dimension of the first matrix E is 4*3
  • the dimension of the transposed matrix E T of the first matrix is 3*4
  • the dimension of the empty matrix G' is 3* 3, so the dimension of the feature correlation matrix H is 7*7.
  • the first participant obtains a secure computing model based on the secure multi-party computing protocol and training with other participants, which cannot be quantified on the premise that the data held by each participant is not local and the data privacy is guaranteed.
  • the problem of collinearity of multi-party data in federated learning is converted into a determinant problem of block matrix that can be processed, so that the collinearity factor of multi-party data in federated learning can be quantified based on the feature correlation matrix, in order to eliminate the collinearity of training data,
  • the federated model provides the basis for obtaining accuracy and stability.
  • step S302 "based on the feature correlation matrix, determine the collinearity quantization factor of each feature corresponding to the first sample feature data" in the embodiment shown in FIG. 3 can be implemented by the following steps:
  • Step S3021 determine the determinant of the feature correlation matrix.
  • the determinant of the feature correlation matrix H is expressed as
  • determining the determinant of the feature correlation matrix can be implemented through steps S30211 to S30214:
  • Step S30211 generating a first random matrix whose determinant is a preset value and has the same dimension as the empty matrix.
  • the first random matrix generated by the first participant satisfies: a matrix whose determinant is a preset value and has the same dimension as the empty matrix G′.
  • the preset value can be set to 1.
  • the empty matrix G' is a matrix of 3*3, so the generated first random matrix is a matrix with a determinant of 1 and a dimension of 3*3, Denoted as R 1 .
  • Step S30212 the feature correlation matrix and the first random matrix are input into the security calculation model to obtain the second matrix.
  • the second matrix J is determined based on the secure multi-party computation protocol, and each participant uses the generated random matrix to confuse their data, so that each participant cannot know the private data of other participants, so it can Ensure the privacy and security of the data of all parties involved.
  • Step S30213 Calculate the determinants of the first symmetric matrix and the second matrix respectively.
  • step S30151 obtain the second matrix J according to step S30212, calculate the determinants of the first symmetric matrix F and the second matrix J respectively, and obtain the determinant
  • Step S30214 Multiply the determinant of the first symmetric matrix and the determinant of the second matrix to obtain the determinant of the feature correlation matrix.
  • each participant calculates the determinant of the virtual feature correlation matrix based on the security computing model, and the entire calculation process does not need to acquire the private data of other participants, which can ensure the data security of each participant.
  • Step S3022 delete the data in the ith row and the ith column of the feature correlation matrix, and obtain each cosubform corresponding to the feature correlation matrix.
  • i 1, 2, ..., m 1 , where m 1 is the number of features corresponding to the first sample feature data.
  • m 1 is the number of features corresponding to the first sample feature data.
  • Step S3023 determine the determinant of each cofactor.
  • the calculation method for determining the determinant of each cofactor may be the same as the method for determining the determinant of the feature correlation matrix in step S3021.
  • determining the determinant of the cofactor corresponding to the ith feature can be implemented as follows: generate a random matrix whose determinant is a preset value and has the same dimension as the empty matrix; input the ith cofactor and the random matrix into the security computing model , obtain the matrix corresponding to the ith cofactor; calculate the determinant of the first symmetric matrix and the matrix corresponding to the ith cofactor respectively; compare the determinant of the first symmetric matrix with the determinant of the matrix corresponding to the ith cofactor Multiply to get the determinant
  • Step S3024 based on the determinant of the feature correlation matrix and the determinant of each cosubformula, determine the collinear quantization factor of each feature corresponding to the feature data of the first sample.
  • the variance expansion factor is used as an example to determine the collinearity quantization factor
  • VIF i of the ith feature can be determined according to the determinant
  • the obtained VIF 1 is the collinear quantization factor of the feature AA
  • the obtained VIF 2 is the collinear quantization factor of the feature BB
  • the obtained VIF 3 is the feature
  • the obtained VIF 4 is the collinear quantization factor of the feature DD.
  • the first participant calculates the determinant of the feature correlation matrix and the determinant of the cofactor of each feature corresponding to the first sample feature data based on the secure multi-party computation protocol, so as to obtain the first sample
  • the variance expansion factor of each feature corresponding to this feature data provides a basis for eliminating collinear training data and obtaining a federated model with accuracy and stability.
  • step S303 in the above-mentioned embodiment shown in FIG. 3 "based on the collinearity quantization factor, determine the target feature from each feature corresponding to the first sample feature data", can be realized by the following steps:
  • Step S3031 judging whether the collinear quantization factor of each feature corresponding to the first sample feature data is greater than a preset boundary value.
  • the size of the collinear quantization factor of each feature corresponding to the first sample feature data and the preset boundary value is determined, that is, the sizes of VIF 1 , VIF 2 , VIF 3 and VIF 4 and the preset boundary value are determined.
  • step S3032 the feature with the collinear quantization factor greater than the preset boundary value is determined as the target feature.
  • the preset boundary value is taken as an example of 10. If it is determined that VIF 1 and VIF 2 are greater than 10, the feature AA corresponding to VIF 1 and the feature BB corresponding to VIF 2 are determined as target features.
  • the target feature AA of the first participant is birthday
  • the target feature BB of the first participant is age
  • the feature BB of the second participant is age
  • step S304 is executed to delete the feature data of the target feature in the first sample feature data, that is, delete the first sample feature data
  • the target feature in each feature corresponding to the second sample feature data held by the second participant is determined to be:
  • the feature BB of the second participant delete the feature data of the feature BB in the second sample feature data, that is, delete the second sample feature data In the first column of data (feature data of feature BB), the second training data obtained
  • the first participant and the second participant perform joint training according to the first training data A' and the second training data B', and the collinear relationship between the first training data A' and the second training data B' has been deleted. Therefore, during joint training, the first participant and the second participant use the first training data and the second training data that do not have a linear relationship to perform joint training, and the obtained federated model has a higher accuracy and higher stability. Yes, it can improve the modeling effect of the federated model.
  • FIG. 4 is a schematic flowchart of another implementation of the data processing method provided by the embodiment of the present application, which is applied to the network architecture shown in FIG. 1 , as shown in FIG. 4, the data processing method includes the following steps:
  • Step S401 the first participant and the second participant determine a common user held by the first participant and the second participant based on the secure multi-party computing protocol.
  • the training data needs to come from the same user for joint training.
  • the second sample feature data used by the party for joint training, and their corresponding IDs are the same.
  • the first participant and the second participant determine a common user based on privacy protection technology. Determining a common user based on the privacy protection technology can be implemented as follows: the first participant and the second participant obtain the identity of the user they hold respectively; then based on the privacy protection technology, the first participant holds the user's identity and the second participant. Hold the user's identity to calculate the intersection, and the result obtained is the common user.
  • Step S402 the first participant determines the target users participating in this training.
  • the first participant determines the target users participating in the training, for example, randomly selecting random users from the common users as the target users participating in the training.
  • Step S403 the first participant sends the identifier of the target user to the second participant.
  • the second participant can also determine the target user participating in this training, and in this case, steps S402 and S403 can be replaced with:
  • Step S402' the second participant determines the target users participating in this training.
  • Step S403' the second participant sends the identifier of the target user to the first participant.
  • Step S404 the first participant acquires the first sample feature data held by itself.
  • the first participant acquires the feature data of the target user from the data stored by itself as the first sample feature data.
  • Step S405 the first participant constructs a virtual first feature correlation matrix based on the first sample feature data and the pre-trained security computing model.
  • the construction of the first feature correlation matrix by the first participant may be implemented as: based on the first sample feature data, determining that feature data of each feature corresponding to the first sample feature data corresponds to the first sample feature data Calculate the mean value and standard deviation corresponding to the feature data of each feature; determine the corresponding mean value based on the feature data of each feature, the mean value corresponding to the feature data of each feature, the standard deviation corresponding to the feature data of each feature, and the number of samples.
  • the processed first sample feature data input the processed first sample feature data into the security calculation model to obtain a first matrix; determine a first symmetric matrix according to the processed first sample feature data; generate a row A first empty matrix whose number and number of columns are equal to the number of columns of the first matrix; a virtual first feature correlation matrix is constructed according to the first symmetric matrix, the first matrix, the transposed matrix of the first matrix, and the first empty matrix.
  • Step S406 the first participant determines, based on the first feature correlation matrix, the collinear quantization factor of each feature corresponding to the first sample feature data.
  • the collinearity quantification factor may be a variance expansion factor, a regression factor, or a correlation coefficient.
  • determining the collinearity quantization factor of each feature corresponding to the first sample feature data can be implemented as: according to the determinant of each cofactor and The determinant of the feature correlation matrix is used to calculate the variance expansion factor of each feature; the variance expansion factor of each feature is determined as the collinear quantization factor of each feature corresponding to the first sample feature data.
  • of the first feature correlation matrix can be implemented as: generating a first random matrix whose determinant is a preset value and has the same dimension as the first empty matrix; Input the matrix into the security calculation model to obtain the second matrix; calculate the determinant of the first symmetric matrix and the second matrix respectively; multiply the determinant of the first symmetric matrix and the determinant of the second matrix to obtain the first characteristic correlation matrix determinant of .
  • of the i-th cofactor corresponding to the i-th feature can be implemented as: generating a random matrix whose determinant is a preset value and has the same dimension as the first empty matrix; The formula and the random matrix are input into the security calculation model, and the matrix corresponding to the i-th cofactor is obtained; the determinant of the matrix corresponding to the first symmetric matrix and the i-th cofactor is calculated respectively; Multiply the determinants of the matrices corresponding to the cofactors to obtain the determinant
  • Step S407 the first participant determines the target feature from the features corresponding to the first sample feature data based on the collinearity quantization factor.
  • determining the target feature by the first participant according to the collinearity quantization factor may be implemented as: judging whether the collinearity quantization factor of each feature corresponding to the first sample feature data is greater than a preset boundary value; Features larger than the preset boundary value are determined as target features.
  • Step S408 the first participant deletes the feature data of the target feature from the first sample feature data to obtain the first training data for the joint training of the first participant and other participants.
  • Step S409 the second participant acquires the second sample feature data held by itself.
  • the second participant obtains the feature data of the target user from the data stored by itself as the second sample feature data.
  • Step S410 the second participant constructs a virtual second feature correlation matrix based on the second sample feature data and the pre-trained security computing model.
  • the construction of the second feature correlation matrix by the second participant may be implemented as: determining the feature data of each feature corresponding to the second sample feature data and the number of samples corresponding to the second sample feature data based on the second sample feature data ; Calculate the mean and standard deviation corresponding to the feature data of each feature respectively; based on the feature data of each feature, the mean value corresponding to the feature data of each feature, the standard deviation corresponding to the feature data of each feature and the number of samples, determine the processed first Two-sample feature data; input the processed second sample feature data into the security computing model to obtain a first matrix; determine a second symmetric matrix according to the processed second sample feature data; generate a number of rows and columns equal to the first matrix A second empty matrix of the number of columns of a matrix; according to the second symmetric matrix, the first matrix, the transposed matrix of the first matrix, and the second empty matrix, a virtual first feature correlation matrix is constructed.
  • the second empty matrix generated by the second participant is F'; the second feature correlation matrix constructed by the second participant is obtained
  • Step S411 the second participant determines the collinear quantization factor of each feature corresponding to the feature data of the second sample based on the feature correlation matrix.
  • determining the collinear quantization factor of each feature corresponding to the second sample feature data by the second participant may be implemented as: determining the determinant of the second feature correlation matrix; deleting the i+m 1th of the second feature correlation matrix row, i+m 1th column data (because the first m 1 column of the second feature correlation matrix H' is the feature of the first sample feature data, so the collinear quantization factor of each feature corresponding to the second sample feature data is determined.
  • of the second feature correlation matrix can be implemented as: generating a second random matrix whose determinant is a preset value and has the same dimension as the second empty matrix; Input the random matrix into the security calculation model to obtain the second matrix; calculate the determinant of the second symmetric matrix and the second matrix respectively; multiply the determinant of the second symmetric matrix and the determinant of the second matrix to obtain the second characteristic correlation The determinant of the matrix.
  • determine the determinant of the ith cofactor corresponding to the ith feature It can be implemented as: generating a random matrix whose determinant is a preset value and the dimension is the same as that of the second empty matrix; inputting the i-th cofactor and the random matrix into the security calculation model to obtain the matrix corresponding to the i-th cofactor; calculating separately The second symmetric matrix and the determinant of the matrix corresponding to the ith cofactor; multiply the determinant of the second symmetric matrix and the determinant of the matrix corresponding to the ith cofactor to obtain the determinant of the ith cofactor
  • Step S412 the second participant determines the target feature from the features corresponding to the second sample feature data based on the collinearity quantization factor.
  • determining the target feature by the second participant according to the collinearity quantization factor may be implemented as: judging whether the collinearity quantization factor of each feature corresponding to the second sample feature data is greater than a preset boundary value; The feature of the preset boundary value is determined as the target feature.
  • Step S413 the second participant deletes the feature data of the target feature from the second sample feature data to obtain second training data for the second participant to perform joint training with other participants.
  • the first participant and the second participant use the first training data and the second training data to perform a joint model based on the secure multi-party computation protocol, since the first training data and the second training data
  • the training data with no linear relationship in the data can improve the accuracy and stability of the federated model obtained by training, thereby improving the modeling effect of the federated model.
  • the first participant and the second participant obtain a secure computing model based on secure multi-party computation pre-training; then the first participant and the second participant obtain the first sample feature data and the third Two-sample feature data, the first participant constructs a virtual first feature correlation matrix based on the first sample feature data and the security computing model, and the second participant constructs a virtual second feature based on the second sample feature data and the security computing model.
  • feature correlation matrix then the first participant determines the collinear quantization factor of each feature corresponding to the first sample feature data based on the first feature correlation matrix, and the second participant determines the second sample feature data corresponding to the second sample feature data based on the second feature correlation matrix.
  • the collinearity quantization factor of each feature the first participant determines the target feature in each feature corresponding to the first sample feature data by the collinearity quantization factor of each feature corresponding to the first sample feature data, and the second participant determines the target feature of each feature corresponding to the first sample feature data.
  • the collinear quantification factor of each feature corresponding to the two-sample feature data determines the target feature in each feature corresponding to the second-sample feature data.
  • the feature data of these target features has a linear relationship with the feature data of at least one feature in the other features.
  • the Other features include all features held by the first participant and the second participant except the target feature; after determining the target feature, the first participant deletes the feature data of the target feature from the first sample feature data to obtain the first Training data; the second participant deletes the feature data of the target feature from the second sample feature data to obtain the second training data.
  • the collinearity data in the characteristic data held by the first participant and the second participant can be screened and eliminated, and the training data without linear relationship can be obtained.
  • the first participant and the second participant perform joint training using training data that does not have a linear relationship, which can improve the accuracy and stability of the federated model and improve the modeling effect of the federated model.
  • VIF Variance Inflation Factor
  • the calculation method usually needs to aggregate the data of all parties in one place.
  • the data of various parties may involve personal privacy or business secrets, and direct opening to other parties will result in information leakage.
  • the feature screening schemes for linear federation modeling in the related art are mainly for single-column data, lacking VIF, a method that can describe the collinearity of multi-column data.
  • Related technologies cannot combine data of multiple parties to calculate VIF while protecting data privacy.
  • the embodiment of the present application for the scenario of two independent participants, it may be assumed that data joint modeling is performed between two companies.
  • the raw data of the respective companies cannot leave the respective companies for reasons such as compliance, privacy and commercial confidentiality.
  • the intermediate data exchanged during the modeling process also cannot derive or reveal unnecessary original data information.
  • the two companies each hold different feature data of the same batch of ID users, and there may be collinearity between these features, which may affect the modeling effect of the subsequent linear model.
  • the embodiments of the present application use the technical solutions of privacy and security of the two parties to calculate the variance inflation factor VIF, and perform feature screening based on this.
  • FIG. 5 is a schematic flowchart of the calculation flow of the variance inflation factor in the vertical federation situation provided by the embodiment of the present application
  • FIG. 6 is a schematic flowchart of the calculation flowchart of the determinant of the correlation matrix provided by the embodiment of the present application. The calculation method of the variance inflation factor in the vertical federation situation provided by the application embodiments will be described in detail.
  • the embodiment of this application provides two independent participants (referred to as Alice and Bob respectively), each of which holds data with different characteristics of the same batch of IDs, which are respectively recorded as matrices A and B (where A and B have the same The number of rows is n, A has the number of columns m 1 , and B has the number of columns m 2 ).
  • Alice and Bob respectively normalize the features (columns of the matrix) of A and B locally to obtain matrices C and D.
  • the way to normalize is: x represents a column of features, represents the mean of the feature, and ⁇ x represents the standard deviation of the feature.
  • H ii is the cofactor of the matrix H (that is, the matrix left by deleting the i-th row and the i-th column of the matrix), and
  • the calculation problem of the determinant of the original matrix H which cannot be processed, is converted into the block matrix F ii , E i* , and determinant calculation problem.
  • Alice locally computes
  • Bob locally generates a random matrix R 2 whose determinant is 1 and has the same dimension as M 4 .
  • a special random matrix is constructed to confuse the original matrix, and the determinant problem of a matrix involving original data information is transformed into a determinant calculation problem of a random matrix with the same determinant.
  • the method provided by the embodiment of the present application converts the original determinant calculation problem that cannot be handled into a blockable matrix determinant problem through ingenious matrix transformation; and confuses the original matrix by constructing a special random matrix, which will involve the original data information.
  • the determinant problem of the matrix is transformed into the determinant calculation problem of a random matrix with the same determinant; through the ingenious cooperation of the SPDZ protocol, the whole calculation process does not leak any additional information other than the result, and has both high security and practical calculation efficiency.
  • the method provided by the embodiment of the present application makes it possible to safely calculate the variance expansion factor VIF, thereby enabling efficient feature screening and improving the overall effect of the subsequent linear model; Unnecessary data information is leaked, and the overall computing overhead is controlled within the practical efficiency range of production.
  • the data processing apparatus 70 stored in the memory 140 is applied to jointly train the model
  • the software module in the data processing device 70 may include:
  • the building module 71 is configured to construct a virtual feature correlation matrix based on the first sample feature data held by the first participant and a pre-trained security computing model, and the security computing model is determined by the first participant and other participants of federated learning are pre-trained based on secure multi-party computation;
  • a first determination module 72 configured to determine, based on the feature correlation matrix, a collinear quantization factor of each feature corresponding to the first sample feature data
  • the second determination module 73 is configured to determine a target feature from each feature corresponding to the first sample feature data based on the collinearity quantization factor;
  • a deletion module 74 configured to delete the feature data of the target feature from the first sample feature data, to obtain the first training data for the joint training of the first participant and the other participants;
  • the feature data of the target feature and the feature data of at least one feature among other features, and the other features include features other than the target feature held by the first participant and the other features. characteristics held by the party.
  • the building block 71 includes:
  • a first determination submodule configured to determine, based on the first sample feature data, feature data of each feature corresponding to the first sample feature data and the number of samples corresponding to the first sample feature data;
  • a calculation submodule configured to calculate the mean and standard deviation corresponding to the characteristic data of each characteristic
  • the second determination sub-module is configured to determine the processed sample based on the feature data of each feature, the mean value corresponding to the feature data of each feature, the standard deviation corresponding to the feature data of each feature, and the number of samples.
  • the first sample characteristic data is configured to determine the processed sample based on the feature data of each feature, the mean value corresponding to the feature data of each feature, the standard deviation corresponding to the feature data of each feature, and the number of samples.
  • an input submodule configured to input the processed first sample characteristic data into the security computing model to obtain a first matrix
  • a construction submodule is configured to construct a virtual feature correlation matrix according to the processed first sample feature data and the first matrix.
  • the building blocks include:
  • a first determining unit configured to determine a first symmetric matrix according to the processed first sample feature data
  • a first generating unit configured to generate an empty matrix whose number of rows and columns are equal to the number of columns of the first matrix
  • a construction unit configured to construct a virtual feature correlation matrix according to the first symmetric matrix, the first matrix, the transposed matrix of the first matrix, and the empty matrix.
  • the first determining module 72 includes:
  • a third determination submodule configured to determine the determinant of the feature correlation matrix
  • a fourth determination submodule configured to determine the determinant of each cosubformula
  • the fifth determination submodule is configured to determine the collinear quantization factor of each feature corresponding to the first sample feature data based on the determinant of the feature correlation matrix and the determinant of each cofactor.
  • the third determination submodule includes:
  • a second generating unit configured to generate a first random matrix whose determinant is a preset value and has the same dimension as the empty matrix
  • an input unit configured to input the feature correlation matrix and the first random matrix into the security computing model to obtain a second matrix
  • a first calculation unit configured to calculate the determinants of the first symmetric matrix and the second matrix respectively
  • the second calculation unit is configured to multiply the determinant of the first symmetric matrix and the determinant of the second matrix to obtain the determinant of the feature correlation matrix.
  • the fifth determination submodule includes:
  • a third calculation unit configured to calculate the variance expansion factor of each feature according to the determinant of each cofactor and the determinant of the feature correlation matrix
  • the second determining unit is configured to determine the variance expansion factor of each feature as a collinear quantization factor of each feature corresponding to the first sample feature data.
  • the second determining module 73 includes:
  • a judging submodule configured to judge whether the collinear quantization factor of each feature corresponding to the first sample feature data is greater than a preset boundary value
  • the sixth determination sub-module is configured to determine a feature whose collinearity quantization factor is greater than a preset boundary value as a target feature.
  • the data processing apparatus 70 further includes:
  • an acquisition module configured to acquire common training samples held by the first participant and the other participants
  • a screening module configured to screen out this training sample from the common training samples
  • the third determination module is configured to determine the characteristic data of the training sample this time as the characteristic data of the first sample.
  • the obtaining module includes:
  • an acquisition submodule configured to acquire the identifier of the data held by the first participant
  • the seventh determination sub-module is configured to, based on encrypted calculation, use the identifier of the data held by the first participant and the identifier of the data held by the other participants to determine the data of the first participant and the other participants. common training samples.
  • Embodiments of the present application provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data processing methods described above in the embodiments of the present application.
  • the embodiments of the present application provide a storage medium storing executable instructions, wherein the executable instructions are stored, and when the executable instructions are executed by a processor, the processor will cause the processor to execute the method provided by the embodiments of the present application, for example, as shown in FIG. 3 to the method shown in FIG. 6 .
  • the storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; it may also be various devices including one or any combination of the above-mentioned memories .
  • executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and which Deployment may be in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.
  • executable instructions may, but do not necessarily correspond to files in a file system, may be stored as part of a file that holds other programs or data, for example, in Hyper Text Markup Language (HTML)
  • HTML Hyper Text Markup Language
  • One or more scripts in a document stored in a single file dedicated to the program in question, or in multiple cooperating files (e.g., files that store one or more modules, subprograms, or code sections) .
  • executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or alternatively, distributed across multiple sites and interconnected by a communication network execute on.
  • the embodiments of the present application provide a data processing method, device, device, storage medium and program product, the method includes: constructing a virtual feature correlation matrix based on first sample feature data and a pre-trained security computing model; feature correlation matrix, to determine the collinearity quantization factor of each feature corresponding to the first sample feature data; based on the collinearity quantization factor, determine the target feature from each feature corresponding to the first sample feature data; from the first sample feature data Deleting the feature data of the target feature in the process to obtain the first training data for joint training; wherein, the feature data of the target feature has a linear relationship with the feature data of at least one of the other features.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Complex Calculations (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

提供一种数据处理方法、装置、设备、存储介质及程序产品，数据处理方法包括：基于第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵（S301）；基于特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子（S302）；基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征（S303）；从第一样本特征数据中删除目标特征的特征数据，得到用于进行联合训练的第一训练数据（S304）；其中，目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系。

Description

数据处理方法、装置、设备、存储介质及程序产品
相关申请的交叉引用
本申请基于申请号为202110454684.2、申请日为2021年04月26日、申请名称为“数据处理方法、装置、设备、存储介质及程序产品”的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请涉及人工智能技术领域，涉及但不限于一种数据处理方法、装置、设备、存储介质及程序产品。
背景技术
机器学习是研究怎样使用计算机模拟或实现人类学习活动的科学，是人工智能中最具智能特征，最前沿的研究领域之一。机器学习的研究主要分为两类研究方向：第一类是传统机器学习的研究，该类研究主要是研究学习机制，注重探索模拟人的学习机制；第二类是大数据环境下机器学习的研究，该类研究主要是研究如何有效利用信息，注重从巨量数据中获取隐藏的、有效的、可理解的知识。
联邦学习技术是新兴的一种隐私保护技术，能够保证数据在不出本地的前提下，有效联合各方数据进行模型训练。对于大数据领域的诸多业务问题，都可通过相应的机器学习模型来解决。而剔除共线数据是训练出好的模型的关键。在保护数据隐私的情况下，相关技术中无法量化联邦学习中多方数据的共线性，无法对存在共线性的训练数据进行高效的筛选与剔除，导致训练得到的模型准确率低、稳定性较差。
发明内容
本申请实施例提供一种数据处理方法、装置、设备、计算机可读存储介质及计算机程序产品，能够在线性联邦建模中剔除存在共线性的数据，提高联邦模型的准确率和稳定性，提高模型的建模效果。
本申请实施例提供一种数据处理方法，应用于联邦学习的第一参与方，所述方法包括：
基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，所述安全计算模型由所述第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；
基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子；
基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征；
从所述第一样本特征数据中删除所述目标特征的特征数据，得到所述第一参与方与所述其他参与方进行联合训练的第一训练数据；
其中，所述目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，所述其他特征包括所述第一参与方持有的除所述目标特征以外的特征和所述其他参与方持有的特征。
本申请实施例提供一种数据处理装置，应用于联邦学习的第一参与方，所述装置包括：
构建模块，配置为基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，所述安全计算模型由所述第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；
第一确定模块，配置为基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子；
第二确定模块，配置为基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征；
删除模块，配置为从所述第一样本特征数据中删除所述目标特征的特征数据，得到所述第一参与方与所述其他参与方进行联合训练的第一训练数据；
其中，所述目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，所述其他特征包括所述第一参与方持有的除所述目标特征以外的特征和所述其他参与方持有的特征。
本申请实施例提供一种数据处理设备，所述设备包括：
存储器，配置为存储可执行指令；
处理器，配置为执行所述存储器中存储的可执行指令时，实现本申请实施例提供的方法。
本申请实施例提供一种计算机可读存储介质，所述计算机可读存储介质上存储有可执行指令，配置为引起处理器执行时，实现本申请实施例提供的方法。
本申请实施例提供一种计算机程序产品，包括计算机程序，所述计算机程序被处理器执行时实现本申请实施例提供的方法。
本申请实施例具有以下有益效果：在本申请实施例提供的数据处理方法中，在进行数据处理时，首先第一参与方与联邦学习的其他参与方基于安全多方计算预先训练得到安全计算模型；然后第一参与方获取第一参与方持有的第一样本特征数据，基于第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵；然后基于特征相关矩阵确定第一样本特征数据对应的各特征的共线性量化因子；由共线性量化因子确定第一样本特征数据对应的各特征中的目标特征，这些目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，这里的其他特征不仅包括第一参与方持有的除目标特征以外的特征，还包括其他参与方持有的所有特征；确定目标特征后，从第一样本特征数据中删除目标特征的特征数据，得到第一参与方与其他参与方进行联合训练的第一训练数据。如此，能够在保护数据隐私的前提下，对各参与方持有的特征数据中存在共线性的数据进行筛选与剔除，得到不存在线性关系的训练数据，从而在进行联合训练时，各参与方利用不存在线性关系的训练数据进行联合训练，能够提高联邦模型的准确率和稳定性，提高联邦模型的建模效果。
附图说明
图1为本申请实施例提供的数据处理方法的网络架构示意图；
图2是本申请实施例提供的数据处理设备的组成结构示意图；
图3为本申请实施例提供的数据处理方法的一种实现流程示意图；
图4为本申请实施例提供的数据处理方法的另一种实现流程示意图；
图5为本申请实施例提供的纵向联邦情形下的方差膨胀因子的计算流程示意图；
图6为本申请实施例提供的相关矩阵的行列式的计算流程示意图。
具体实施方式
为了使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请作进一步地详细描述，所描述的实施例不应视为对本申请的限制，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
在以下的描述中，涉及到“一些实施例”，其描述了所有可能实施例的子集，但是可以理解，“一些实施例”可以是所有可能实施例的相同子集或不同子集，并且可以在不冲突的情况下相互结合。
在以下的描述中，所涉及的术语“第一\第二\第三”仅是区别类似的对象，不代表针对对象的特定排序，可以理解地，“第一\第二\第三”在允许的情况下可以互换特定的顺序或先后次序，以使这里描述的本申请实施例能够以除了在这里图示或描述的以外的顺序实施。
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请的技术领域的技术人员通常理解的含义相同。本文中所使用的术语只是为了描述本申请实施例的目的，不是旨在限制本申请。
对本申请实施例进行进一步详细说明之前，对本申请实施例中涉及的名词和术语进行说明，本申请实施例中涉及的名词和术语适用于如下的解释。
1)纵向联邦学习(Vertical Federated Learning)，在两个数据集的用户重叠较多、用户特征重叠较少的情况下，把数据集按照纵向(即特征维度)进行切分，并取出双方用户相同而用户特征不完全相同的部分数据进行训练的机器学习。
2)方差膨胀因子(VIF，Variance Inflation Factor)，又称方差扩大因子，是表征自变量观察值之间复共线性程度的数值，用于衡量多元线性回归模型中复(多重)共线性严重程度。
3)同态加密(Homomorphic Encryption)，同态加密是基于数学难题的计算复杂性理论的密码学技术。对经过同态加密的数据进行处理得到一个输出，将这一输出进行解密，其结果与用同一方法处理未加密的原始数据得到的输出结果是一样的。
下面说明实现本申请实施例的装置的示例性应用，本申请实施例提供的装置可以实施为终端设备。下面，将说明装置实施为终端设备时涵盖终端设备的示例性应用。
图1为本申请实施例提供的数据处理方法的网络架构示意图，如图1所示，在该网络架构中至少包括第一参与方100、第二参与方200和网络300。为实现支撑一个示例性应用，第一参与方100和第二参与方200可以为纵向联邦学习中联合训练机器学习模 型的各参与方。其中，第一参与方100和第二参与方200可以为客户端，例如各银行或医院等存储有用户特征数据的参与方设备，客户端可以是笔记本电脑，平板电脑，台式计算机，专用训练设备等具有模型训练功能的设备。第一参与方100通过网络300连接第二参与方200，网络300可以是广域网或者局域网，又或者是二者的组合，使用无线或有线链路实现数据传输。
第一参与方100首先从自身的数据中获取第一样本特征数据，对该第一样本特征数据进行处理，得到处理后的第一样本特征数据。第二参与方200从自身的数据中获取第二样本特征数据，对该第二样本特征数据进行处理得到处理后的第二样本特征数据。这里的第一样本特征数据和第二样本特征数据具有相同的标识，即第一样本特征数据和第二样本特征数据分别为第一参与方100和第二参与方200各自持有的同一批样本的不同特征的数据。第一参与方100和第二参与方200基于安全多方计算，利用处理后的第一样本特征数据和处理后的第二样本特征数据确定第一矩阵E，第一参与方100和第二参与方200均持有该第一矩阵E。基于安全多方计算，第一参与方100只能得到第一矩阵E，是无法获知第二参与方200所持有的处理后的第二样本特征数据的；同样的，第二参与方200也只能得到第一矩阵E，是无法获知第一参与方100所持有的处理后的第一样本特征数据的。
第一参与方100得到第一矩阵E后，根据第一矩阵E和处理后的第一样本特征数据，构建一虚拟的特征相关矩阵；然后计算该特征相关矩阵的行列式，以及该特征相关矩阵对应的各余子式的行列式，基于特征相关矩阵的行列式和各余子式的行列式，确定第一样本特征数据对应的各特征的共线性量化因子，该共线性量化因子可以为方差膨胀系数，用于量化每一特征与其他所有特征的共线性，这里的其他所有特征不仅包括第一参与方100持有的除本次量化特征以外的其他特征，还包括第二参与方200持有的所有特征。第一参与方100根据各特征的共线性量化因子，确定哪些特征的特征数据之间存在共线性，将该些特征确定为目标特征，最后从第一样本特征数据中删除目标特征的特征数据，得到第一参与方100与第二参与方200进行联合训练的第一训练数据。基于同样的方法，第二参与方200得到进行联合训练的第二训练数据。从而在进行联合训练时，第一参与方100和第二参与方200利用不存在共线性的第一训练数据和第二训练数据进行联合训练，能够得到准确率高、稳定性好的联邦模型。
通过本申请实施例提供的方法，能够在保护数据隐私的前提下，对各参与方持有的特征数据中存在共线性的数据进行筛选与剔除，得到不存在线性关系的训练数据，从而在进行联合训练时，各参与方利用不存在线性关系的训练数据进行联合训练，能够提高联邦模型的准确率和稳定性，提高联邦模型的建模效果。
本申请实施例提供的装置可以实施为硬件或者软硬件结合的方式，下面说明本申请实施例提供的装置的各种示例性实施。
根据图2示出的数据处理设备10的示例性结构，这里，数据处理设备10以应用于联邦学习的第一参与方的设备为例示出，可以预见数据处理设备10的其他的示例性结构，因此这里所描述的结构不应视为限制，例如可以省略下文所描述的部分组件，或者，增设下文所未记载的组件以适应某些应用的特殊需求。
图2所示的数据处理设备10包括：至少一个处理器110、存储器140、至少一个网 络接口120和用户接口130。数据处理设备10中的每个组件通过总线系统150耦合在一起。可理解，总线系统150用于实现这些组件之间的连接通信。总线系统150除包括数据总线之外，还包括电源总线、控制总线和状态信号总线。但是为了清楚说明起见，在图2中将各种总线都标为总线系统150。
用户接口130可以包括显示器、键盘、鼠标、触感板和触摸屏等。
存储器140可以是易失性存储器或非易失性存储器，也可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(ROM，Read Only Memory)。易失性存储器可以是随机存取存储器(RAM，Random Access Memory)。本申请实施例描述的存储器140旨在包括任意适合类型的存储器。
本申请实施例中的存储器140能够存储数据以支持数据处理设备10的操作。这些数据的示例包括：用于在数据处理设备10上操作的任何计算机程序，如操作系统和应用程序。其中，操作系统包含各种系统程序，例如框架层、核心库层、驱动层等，用于实现各种基础业务以及处理基于硬件的任务。应用程序可以包含各种应用程序。
作为本申请实施例提供的方法采用软件实施的示例，本申请实施例所提供的方法可以直接体现为由处理器110执行的软件模块组合，软件模块可以位于存储介质中，存储介质位于存储器140，处理器110读取存储器140中软件模块包括的可执行指令，结合必要的硬件(例如，包括处理器110以及连接到总线150的其他组件)完成本申请实施例提供的方法。
作为示例，处理器110可以是一种集成电路芯片，具有信号的处理能力，例如通用处理器、数字信号处理器(DSP，Digital Signal Processor)，或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等，其中，通用处理器可以是微处理器或者任何常规的处理器等。
将结合本申请实施例提供的设备的示例性应用和实施，说明本申请实施例提供的数据处理方法。
图3为本申请实施例提供的数据处理方法的一种实现流程示意图，应用于图1所示网络架构的第一参与方，将结合图3示出的步骤进行说明。
步骤S301，基于第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵。
本申请实施例中，第一样本特征数据对应的各特征中，不同特征的特征数据可能存在线性关系，若采用存在共线性的数据训练模型，训练好的模型准确度低、稳定性差。为了剔除存在线性关系的数据，第一参与方根据第一样本特征数据和预先训练好的安全计算模型，构建一虚拟的特征相关矩阵，记为H。
该安全计算模型由第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到，能够进行隐私下的矩阵加减法和矩阵乘法。示例性的，本申请实施例中，以两个参与方进行联邦学习为例对该数据处理方法进行说明。在实际应用中，该数据处理方法也可以应用于三个及三个以上参与方进行联邦学习。在各参与方持有的数据不出本地、保证数据隐私的前提下，第一参与方和第二参与方基于隐私保护技术进行训练，得到训练好的安全计算模型，训练好的安全计算模型为各参与方所持有。
本申请实施例中，由于第一参与方和第二参与方进行联合训练，进行训练的数据需 要来自相同的用户才能进行联合训练，即第一参与方用于进行联合训练的第一样本特征数据，与第二参与方用于进行联合训练的第二样本特征数据，其各自对应的用户标识(ID，Identity Document)是相同的。因此在步骤S301之前，该方法还包括：第一参与方获取第一样本特征数据。
在一种实现方式中，获取第一样本特征数据可以实现为：获取第一参与方和其他参与方持有的共同训练样本；从共同训练样本中筛选出本次训练样本；将本次训练样本的特征数据确定为第一样本特征数据。共同训练样本即各参与方持有的共同用户，各参与方获取自身持有数据的标识，基于加密计算，利用各参与方持有数据的标识，确定该些参与方的共同用户。然后从这些共同用户中筛选出参与本次训练的目标用户作为本次训练样本，将这些目标用户的特征数据作为第一样本特征数据。
步骤S302，基于特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子。
在实际应用中，确定多个特征的特征数据之间的共线性定量因子的方法包括方差膨胀因子法、特征根分析法和条件数法。另外，还存在一可定性的分析多个特征的特征数据之间共线程度的直观判定法，直观判断法一般用于初步判断，无法进行定量的分析。
第一参与方持有的第一样本特征数据，以及第二参与方持有的第二样本特征数据，与特征相关矩阵中不同列的数据相对应，确定特征相关矩阵中各列数据与其他列数据之间是否存在共线性，即可确定第一样本特征中各特征的特征数据与其他特征的特征数据之间是否存在共线性。这里不同特征的特征数据之间存在共线性包括以下可能：存在一个特征的特征数据是另一个特征的特征数据的倍数；存在一个特征的特征数据等于另一个特征的特征数据加上一个常数项；存在一个特征的特征数据等于另两个特征的特征数据相加。
本申请实施例中，根据特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子，根据第一样本特征数据对应的各特征的共线性量化因子，即可确定哪些特征存在与之共线的特征数据。
步骤S303，基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征。
根据共线性量化因子，确定出存在共线性的目标特征。以方差膨胀因子为例，在不存在多重共线性的情况下，方差膨胀因子接近于1，多重共线性越强，方差膨胀因子越大。实际上数据之间总是或多或少的存在多重共线性，因此将方差膨胀因子等于1作为评价共线性的标准是不现实的，一般的，可根据实际应用场景，预设一边界值，本申请实施例中，预设边界值可以取值为10。判断各特征的方差膨胀因子是否大于10，当存在某特征的方差膨胀因子大于10时，认为该特征的特征数据与其他特征的特征数据之间具有较强的共线性，将该特征确定为目标特征。
步骤S304，从第一样本特征数据中删除目标特征的特征数据，得到第一参与方与其他参与方进行联合训练的第一训练数据。
其中，目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，这里的其他特征包括第一参与方持有的除目标特征以外的特征和其他参与方持有的特征。本申请实施例中，可基于特征相关矩阵计算第一参与方的第一样本特征数据中各特 征的方差膨胀因子，确定出方差膨胀因子大于预设边界值的目标特征，然后对第一样本特征数据中目标特征的特征数据进行删除，得到不存在共线性的特征数据作为第一训练数据。从而在进行联合训练时，第一参与方利用不存在共线性的第一训练数据进行联合训练，能够得到准确率高、稳定性好的联邦模型。
本申请实施例提供的数据处理方法，应用于联邦学习的第一参与方，该数据处理方法包括：基于第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，该安全计算模型由第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；基于特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子；基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征；从第一样本特征数据中删除目标特征的特征数据，得到第一参与方与其他参与方进行联合训练的第一训练数据；其中，目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，其他特征包括第一参与方持有的除目标特征以外的特征和其他参与方持有的特征。通过该方法，能够在保护数据隐私的前提下，对各参与方持有的特征数据中存在共线性的数据进行筛选与剔除，得到不存在线性关系的训练数据，从而在进行联合训练时，各参与方利用不存在线性关系的训练数据进行联合训练，能够提高联邦模型的准确率和稳定性，提高联邦模型的建模效果。
在一些实施例中，上述图3所示实施例中步骤S301之前，第一参与方先从自身持有的特征数据中获取第一样本特征数据。
由于第一参与方和第二参与方进行联合训练，进行训练的数据需要来自相同的用户才能进行联合训练，即第一参与方用于进行联合训练的第一样本特征数据，与第二参与方用于进行联合训练的第二样本特征数据，其各自对应的ID是相同的。获取第一样本特征数据时，首先需要确定各参与方持有的共同用户，然后从这些共同用户中筛选出参与本次训练的目标用户，将这些目标用户的特征数据作为第一样本特征数据。
下面举例说明第一参与方获取第一样本特征数据的过程。
第一参与方和第二参与方各自持有一批用户的不同特征的特征数据，如第一参与方拥有一批用户的生日AA、年龄BB、体重CC、存款DD这几个特征的特征数据，第二参与方拥有一批用户的年龄BB、消费能力EE、爱好FF这几个特征的特征数据。第一参与方的数据库表和第二参与方的数据库表分别以下表1、表2示出：
表1第一参与方的数据库表
ID AA BB CC DD
01 a1 b1 c1 d1
02 a2 b2 c2 d2
03 a3 b3 c3 d3
04 a4 b4 c4 d4
06 a6 b6 c6 d6
表2第二参与方的数据库表
ID BB EE FF
01 b1 e1 f1
03 b3 e3 f3
04 b4 e4 f4
05 b5 e5 f5
06 b6 e6 f6
根据第一参与方的数据库表，得到第一参与方持有的特征数据
Figure PCTCN2021140955-appb-000001
根据第二参与方的数据库表，得到第二参与方持有的特征数据
Figure PCTCN2021140955-appb-000002
在隐私保护的前提下，由于各参与方无法获知其他参与方存储有哪些用户的数据，因此本申请实施例中第一参与方和第二参与方基于隐私保护技术确定共同用户。基于隐私保护技术确定共同用户可以实现为：第一参与方和第二参与方分别获取自身持有用户的标识；然后基于隐私保护技术，对第一参与方持有用户的标识和第二参与方持有用户的标识计算交集，得到的结果即为共同用户。利用该方法确定得到第一参与方和第二参与方持有的共同用户为ID为01、03、04和06的用户，记为S＝{01,03,04,06}。
得到共同用户S后，第一参与方确定参与本次训练的目标用户，如随机地从共同用户中筛选出随机个用户作为参与本次训练的目标用户，然后从自身存储的数据中获取这些目标用户的特征数据作为第一样本特征数据。并且，第一参与方将筛选出的目标用户的ID发送给第二参与方，第二参与方从自身存储的数据中获取这些目标用户的特征数据，作为第二样本特征数据。或者，也可以由第二参与方确定参与本次训练的目标用户，将确定的目标用户的ID发送给第一参与方。例如，第一参与方随机地从S中筛选出ID为01、03和04这3个用户作为目标用户，记为S1＝{01,03,04}，第一参与方从自身持有的特征数据中筛选出目标用户S1的特征数据，得到第一样本特征数据
Figure PCTCN2021140955-appb-000003
在第二参与方侧，第二参与方从自身持有的特征数据中筛选出目标用户S1的特征数据，得到第二样本特征数据
Figure PCTCN2021140955-appb-000004
本申请实施例中通过上述步骤，在各参与方数据不出本地、实现隐私保护的前提下， 第一参与方获得第一样本特征数据，第二参与方获得第二样本特征数据。
在一些实施例中，上述图3所示实施例中步骤S301“基于第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵”可以通过以下步骤来实现：
步骤S3011，基于第一样本特征数据，确定第一样本特征数据对应的各特征的特征数据和第一样本特征数据对应的样本个数。
仍以上述举例数据进行说明，如表3所示，第一样本特征数据对应的特征有：AA、BB、CC和DD，特征AA的特征数据为{a1,a3,a4}，特征BB的特征数据为{b1,b3,b4}，特征CC的特征数据为{c1,c3,c4}，特征DD的特征数据为{d1,d3,d4}，可见，各特征的特征数据对应A的一列数据。第一样本特征数据对应的样本有：{a1,b1,c1,d1}、{a3,b3,c3,d3}和{a4,b4,c4,d4}，确定样本个数n＝3。
步骤S3012，分别计算各特征的特征数据对应的均值和标准差。
根据
Figure PCTCN2021140955-appb-000005
计算特征AA、BB、CC和DD的特征数据对应的均值，根据
Figure PCTCN2021140955-appb-000006
计算特征AA、BB、CC和DD的特征数据对应的标准差，其中x＝a,b,c,d，n＝3。
步骤S3013，基于各特征的特征数据、各特征的特征数据对应的均值、各特征的特征数据对应的标准差和样本个数，确定处理后的第一样本特征数据；
本申请实施例中，可根据各特征的特征数据对应的均值
Figure PCTCN2021140955-appb-000007
各特征的特征数据对应的标准差σ x和样本个数n对各特征的特征数据x进行更新处理，得到处理后的各个特征数据x。这里处理后的各个特征数据x的处理公式可以为
Figure PCTCN2021140955-appb-000008
从而得到处理后的第一样本特征数据
Figure PCTCN2021140955-appb-000009
其中，
Figure PCTCN2021140955-appb-000010
Figure PCTCN2021140955-appb-000011
x＝a,b,c,d，n＝3。
步骤S3014，将处理后的第一样本特征数据输入至安全计算模型，得到第一矩阵。
本申请实施例中各参与方基于安全多方计算协议如SPDZ协议或MASCOT等进行训练，得到训练好的安全计算模型。该训练好的模型能够进行隐私下的矩阵加减法和矩阵乘法。安全多方计算(MPC，Secure Multi-Party Computation)研究的主要是针对无可信第三方的情况下，如何安全地计算一个约定函数的问题。若参与联邦学习训练的参与方有m个，训练好的安全计算模型可以表示为y 1,…,m＝f(x 1,…,m)，其中m≥2。训练完成 后，该m个参与方均持有该安全计算模型，各参与方输入自身私有的输入值，得到对应的输出值，例如，第i个参与方输入x i，输出的是y i，其中，1≤i≤m。
本申请实施例中以两个参与方为例，训练好的安全计算模型表示为y 1,2＝f(x 1,2)，第一参与方和第二参与方均持有该模型，第一参与方将处理后的第一样本特征数据C输入y 1,2＝f(x 1,2)，得到第一矩阵E，该第一矩阵E＝C TD，其中D为处理后的第二样本特征数据。需要说明的是，这里第一矩阵E＝C TD仅是为了说明E的值等于C TD的值，并非表明E是根据C TD计算得到的，第一参与方是根据y 1,2＝f(x 1,2)计算得到的E，第一参与方无需从第二参与方获取处理后的第二样本特征数据D。由C TD可知，第一矩阵E为4*3的矩阵。
对应的，在第二参与方侧，第二参与方将处理后的第二样本特征数据D输入y 1,2＝f(x 1,2)，得到第一矩阵E，该第一矩阵E＝C TD，其中C T为处理后的第一样本特征数据的转置矩阵。需要说明的是，这里第一矩阵E＝C TD仅是为了说明E的值等于C TD的值，并非表明E是根据C TD计算得到的，第二参与方是根据y 1,2＝f(x 1,2)计算得到的E，第二参与方无需从第一参与方获取处理后的第一样本特征数据的转置矩阵C T。
本申请实施例中，第一矩阵E是基于安全多方计算协议确定的，各参与方无法获知其他参与方的私有数据，能够确保各参与方数据的隐私安全。
步骤S3015，根据处理后的第一样本特征数据和第一矩阵，构建虚拟的特征相关矩阵。
本申请实施例中，“构建虚拟的特征相关矩阵”，可以通过以下步骤来实现：
步骤S30151，根据处理后的第一样本特征数据确定第一对称矩阵。
对称矩阵(Symmetric Matrices)是指以主对角线为对称轴，各元素对应相等的方形矩阵。本申请实施例中，根据处理后的第一样本特征数据C构造第一对称矩阵F，可以根据处理后的第一样本数据C的转置矩阵
Figure PCTCN2021140955-appb-000012
和处理后的第一样本数据
Figure PCTCN2021140955-appb-000013
确定的第一对称矩阵F，F的 计算公式F＝C TC，F的维度为4*4。
步骤S30152，生成一个行数和列数均等于得到第一矩阵的列数的空矩阵。
在根据第一参与方持有的特征和第二参与持有的特征构建特征相关矩阵时，在各参与方数据不出本地的前提下，第一参与方只能获知第二参与方持有的特征，无法获得第二参与方持有的各特征的特征数据，因此，本申请实施例中，在第一参与方构建的是一个虚拟的特征相关矩阵，其中与第二参与方特征的特征数据相关的数据为空。
根据上述步骤S3014可知第一矩阵E为4*3的矩阵，其列数为3，因此生成的空矩阵为维度为3*3的矩阵，记为G′。与该空矩阵G′相对应的第二对称矩阵G，由第二参与方根据第二样本特征数据B确定的。
步骤S30153，根据第一对称矩阵、第一矩阵、第一矩阵的转置矩阵和空矩阵，构建虚拟的特征相关矩阵。
根据第一对称矩阵F、第一矩阵E、第一矩阵的转置矩阵E T，以及空矩阵G′，构建虚拟的特征相关矩阵H，表示为
Figure PCTCN2021140955-appb-000014
其中，第一对称矩阵F的维度为4*4，第一矩阵E的维度为4*3，第一矩阵的转置矩阵E T的维度为3*4，空矩阵G′的维度为3*3，因此特征相关矩阵H的维度为7*7。
本申请实施例提供的方法，第一参与方基于安全多方计算协议与其他参与方训练获得安全计算模型，能够在各参与方持有的数据不出本地、保证数据隐私的前提下，将无法量化联邦学习中多方数据的共线性的问题，转换为可以处理地分块矩阵行列式问题，从而能够基于该特征相关矩阵量化联邦学习中多方数据的共线性因子，为剔除存在共线性的训练数据、得到准确率和稳定性的联邦模型提供基础。
在一些实施例中，上述图3所示实施例中步骤S302“基于特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子”，可以通过以下步骤来实现：
步骤S3021，确定特征相关矩阵的行列式。
特征相关矩阵H的行列式表示为|H|，本申请实施例中，确定该特征相关矩阵的行列式可以通过步骤S30211至步骤S30214来实现：
步骤S30211，生成一个行列式为预设值、维度与空矩阵相同的第一随机矩阵。
在计算特征相关矩阵的行列式时，由于各参与方的数据不能出本地，为了确保私有数据不泄露，本申请实施例中生成一个用于混淆原始数据的第一随机矩阵，如此实现隐私保护。
本申请实施例中，第一参与方生成的第一随机矩阵满足：行列式为预设值、维度与空矩阵G′相同的矩阵。为方便计算，可以将预设值设置为1，根据上述步骤S30152可知空矩阵G′为3*3的矩阵，因此生成的第一随机矩阵为行列式为1，维度为3*3的矩阵，记为R 1。
步骤S30212，将特征相关矩阵和第一随机矩阵输入至安全计算模型，得到第二矩阵。
第一参与方和第二参与方训练好的安全计算模型表示为y 1,2＝f(x 1,2)，第一参与方将特征相关矩阵H和第一随机矩阵R 1输入该模型，得到第二矩阵J，该第二矩阵J＝R 1 (D-CA -1B)R 2，其中A -1为第一样本特征数据A的逆，D为处理后的第二样本特征数据，R 2为第二参与方生成的行列式为1、维度为4*4(维度与第一对称矩阵F相同)的第二随机矩阵。需要说明的是，这里第二矩阵J＝R 1(D-CA -1B)R 2仅是为了说明第二矩阵J的值等于R 1(D-CA -1B)R 2的值，并非表明第二矩阵J是根据R 1(D-CA -1B)R 2计算得到的，第一参与方是根据y 1,2＝f(x 1,2)计算得到的J，第一参与方无需从第二参与方获取处理后的第二样本特征数据D以及第二随机矩阵R 2。
本申请实施例中，第二矩阵J是基于安全多方计算协议确定的，并且各参与方利用生成的随机矩阵对各自的数据进行混淆，使得各参与方无法获知其他参与方的私有数据，因此能够确保各参与方数据的隐私安全。
步骤S30213，分别计算第一对称矩阵和第二矩阵的行列式。
根据步骤S30151得到第一对称矩阵F，根据步骤S30212获得第二矩阵J，分别计算第一对称矩阵F和第二矩阵J的行列式，得到第一对称矩阵的行列式|F|和第二矩阵的行列式|J|。
这里，当生成的第一随机矩阵的行列式|R 1|或第二随机矩阵的行列式|R 1|不为1时，计算第二矩阵的行列式|J|时，需要乘以
Figure PCTCN2021140955-appb-000015
即
Figure PCTCN2021140955-appb-000016
步骤S30214，将第一对称矩阵的行列式和第二矩阵的行列式相乘，得到特征相关矩阵的行列式。
将第一对称矩阵的行列式|F|与第二矩阵的行列式|J|相乘，得到特征相关矩阵的行列式|H|＝|F|*|J|。本申请实施例中，各参与方基于安全计算模型计算得到虚拟的特征相关矩阵的行列式，并且整个计算过程无需获取其他参与方的私有数据，能够确保各参与方的数据安全。
步骤S3022，删除特征相关矩阵第i行、第i列数据，得到特征相关矩阵对应的各余子式。
其中，i＝1，2，…，m 1，其中，m 1为第一样本特征数据对应的特征个数。如上举例，第一样本特征数据A的特征个数为4，m 1＝4。由于第一参与方只能确定自身持有的第一样本特征数据对应的各特征的共线性量化因子，因此这里只需得到特征相关矩阵对应的第1余子式、第2余子式、第3余子式和第4余子式。删除特征相关矩阵H 7*7的第i行、第i列的数据，剩下的数据不改变原来的顺序，构成维度为6*6的矩阵，该矩阵即为第一样本特征数据的第i个特征对应的第i余子式H ii，例如，当i为2时，得到的即为特征BB对应的第2余子式H 22。
步骤S3023，确定各余子式的行列式。
本申请实施例中，确定各余子式的行列式的计算方式，可以与步骤S3021中确定特征相关矩阵的行列式的方式相同。例如，确定第i个特征对应余子式的行列式，可以实现为：生成一个行列式为预设值、维度与空矩阵相同的随机矩阵；将第i余子式和该随机矩阵输入至安全计算模型，得到第i余子式对应的矩阵；分别计算第一对称矩阵和该第i余子式对应的矩阵的行列式；将第一对称矩阵的行列式和第i余子式对应的矩阵矩阵的行列式相乘，得到第i余子式的行列式|H ii|。
步骤S3024，基于特征相关矩阵的行列式和各余子式的行列式，确定第一样本特征数据对应的各特征的共线性量化因子。
本申请实施例中，以方差膨胀因子为例确定共线性量化因子，第i个特征的方差膨胀因子VIF i，可以根据特征相关矩阵的行列式|H|和第i余子式的行列式|H ii|确定，计算公式如式(1)所示：
Figure PCTCN2021140955-appb-000017
其中，当i＝1时，得到的VIF 1为特征AA的共线性量化因子，i＝2时，得到的VIF 2为特征BB的共线性量化因子，i＝3时，得到的VIF 3为特征CC的共线性量化因子，i＝4时，得到的VIF 4为特征DD的共线性量化因子。
本申请实施例提供的方法，第一参与方基于安全多方计算协议，计算得到特征相关矩阵的行列式，以及第一样本特征数据对应的各特征的余子式的行列式，从而得到第一样本特征数据对应的各特征的方差膨胀因子，为剔除存在共线性的训练数据、得到准确率和稳定性的联邦模型提供基础。
在一些实施例中，上述图3所示实施例中步骤S303“基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征”，可以通过以下步骤来实现：
步骤S3031，判断第一样本特征数据对应的各特征的共线性量化因子是否大于预设边界值。
这里判断第一样本特征数据对应的各特征的共线性量化因子与预设边界值的大小，即判断VIF 1、VIF 2、VIF 3和VIF 4与预设边界值的大小。
步骤S3032，将共线性量化因子大于预设边界值的特征确定为目标特征。
预设边界值以10为例，判断得到VIF 1和VIF 2大于10，则将VIF 1对应的特征AA和VIF 2对应的特征BB确定为目标特征。
结合上述举例，第一参与方的目标特征AA为生日，第一参与方的目标特征BB为年龄，第二参与方的特征BB为年龄，确定生日与年龄之间存在线性关系(年龄等于今年减去出生年)、年龄与年龄之间存在线性关系(相等，即倍数为1)。即第一参与方的目标特征AA的特征数据与第一参与方的特征BB的特征数据、以及第二参与方的特征BB的特征数据存在线性关系，第一参与方的目标特征BB的特征数据与第一参与方的特征AA的特征数据、以及第二参与方的特征BB的特征数据存在线性关系。
得到目标特征后，执行步骤S304，在第一样本特征数据中删除目标特征的特征数据，即删除第一样本特征数据
Figure PCTCN2021140955-appb-000018
中第1列数据(特征AA的特征数据)和第2列数据(特征BB的特征数据)，得到的第一训练数据
Figure PCTCN2021140955-appb-000019
与上述确定第一参与方持有的第一样本特征数据对应的各特征中的目标特征步骤相同，确定得到第二参与方持有的第二样本特征数据对应的各特征中的目标特征为第二 参与方的特征BB，在第二样本特征数据中删除特征BB的特征数据，即删除第二样本特征数据
Figure PCTCN2021140955-appb-000020
中第1列数据(特征BB的特征数据)，得到的第二训练数据
Figure PCTCN2021140955-appb-000021
然后，第一参与方和第二参与方根据第一训练数据A′和第二训练数据B′进行联合训练，第一训练数据A′和第二训练数据B′中已经删除存在共线关系的特征数据，因此在进行联合训练时，第一参与方和第二参与方利用不存在线性关系的第一训练数据和第二训练数据进行联合训练，得到的联邦模型准确率较高、稳定性较好，能够提高联邦模型的建模效果。
基于前述的实施例，本申请实施例再提供一种数据处理方法，图4为本申请实施例提供的数据处理方法的再一种实现流程示意图，应用于图1所示的网络架构，如图4所示，该数据处理方法包括以下步骤：
步骤S401，第一参与方和第二参与方基于安全多方计算协议，确定第一参与方和第二参与方持有的共同用户。
由于第一参与方和第二参与方进行联合训练，进行训练的数据需要来自相同的用户才能进行联合训练，即第一参与方用于进行联合训练的第一样本特征数据，与第二参与方用于进行联合训练的第二样本特征数据，其各自对应的ID是相同的。获取第一样本特征数据时，首先需要确定各参与方持有的共同用户，然后从这些共同用户中筛选出参与本次训练的目标用户，将这些目标用户的特征数据作为第一样本特征数据。
在隐私保护的前提下，由于各参与方无法获知其他参与方存储有哪些用户的数据，因此本申请实施例中第一参与方和第二参与方基于隐私保护技术确定共同用户。基于隐私保护技术确定共同用户可以实现为：第一参与方和第二参与方分别获取自身持有用户的标识；然后基于隐私保护技术，对第一参与方持有用户的标识和第二参与方持有用户的标识计算交集，得到的结果即为共同用户。
步骤S402，第一参与方确定参与本次训练的目标用户。
得到共同用户后，第一参与方确定参与本次训练的目标用户，如随机地从共同用户中筛选出随机个用户作为参与本次训练的目标用户。
步骤S403，第一参与方将目标用户的标识发送给第二参与方。
在另一些实施例中，也可以由第二参与方确定参与本次训练的目标用户，此时步骤S402和步骤S403可以替换为：
步骤S402’，第二参与方确定参与本次训练的目标用户。
步骤S403’，第二参与方将目标用户的标识发送给第一参与方。
步骤S404，第一参与方获取自身持有的第一样本特征数据。
第一参与方从自身存储的数据中获取目标用户的特征数据作为第一样本特征数据。
步骤S405，第一参与方基于第一样本特征数据和预先训练好的安全计算模型，构 建虚拟的第一特征相关矩阵。
在一些实施例中，第一参与方构建第一特征相关矩阵可以实现为：基于第一样本特征数据，确定第一样本特征数据对应的各特征的特征数据和第一样本特征数据对应的样本个数；分别计算各特征的特征数据对应的均值和标准差；基于各特征的特征数据、各特征的特征数据对应的均值、各特征的特征数据对应的标准差和样本个数，确定处理后的第一样本特征数据；将处理后的第一样本特征数据输入至安全计算模型，得到第一矩阵；根据处理后的第一样本特征数据确定第一对称矩阵；生成一个行数和列数均等于第一矩阵的列数的第一空矩阵；根据第一对称矩阵、第一矩阵、第一矩阵的转置矩阵和第一空矩阵，构建虚拟的第一特征相关矩阵。
例如，第一样本特征数据为A；根据
Figure PCTCN2021140955-appb-000022
对A中各个特征数据x进行处理，得到处理后的第一样本特征数据为C；将处理后的第一样本特征数据C输入至预先训练好的安全计算模型y 1,…,m＝f(x 1,…,m)，得到第一矩阵E；根据处理后的第一样本特征数据的转置矩阵C T和处理后的第一样本特征数据C，确定第一对称矩阵F＝C TC；第一参与方生成的第一空矩阵为G′；得到第一参与方构建的第一特征相关矩阵
Figure PCTCN2021140955-appb-000023
步骤S406，第一参与方基于第一特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子。
在一些实施例中，第一参与方确定第一样本特征数据对应的各特征的共线性量化因子可以实现为：确定第一特征相关矩阵的行列式；删除第一特征相关矩阵第i行、第i列数据，得到第一特征相关矩阵对应的各余子式，其中，i＝1，2，…，m 1，m 1为第一样本特征数据对应的特征个数；确定各余子式的行列式；基于第一特征相关矩阵的行列式和各余子式的行列式，确定第一样本特征数据对应的各特征的共线性量化因子。
本申请实施例中，共线性量化因子可以为方差膨胀因子、回归因子或相关系数。以方差膨胀因子为例，基于特征相关矩阵的行列式和各余子式的行列式，确定第一样本特征数据对应的各特征的共线性量化因子可以实现为：根据各余子式的行列式与特征相关矩阵的行列式，计算各特征的方差膨胀因子；将各特征的方差膨胀因子确定为第一样本特征数据对应的各特征的共线性量化因子。删除第一特征相关矩阵H的第i行、第i列的数据，得到第一样本特征数据的第i个特征对应的第i余子式H ii，计算第一特征相关矩阵的行列式|H|和第i余子式的行列式|H ii|，将第i余子式的行列式与第一特征相关矩阵的行列式的商
Figure PCTCN2021140955-appb-000024
确定为第i特征的方差膨胀因子，其中i＝1，2，…，m 1，m 1为第一样本特征数据对应的特征个数，如此得到第一样本特征数据对应的各特征的共线性量化因子。
其中，确定第一特征相关矩阵的行列式|H|可以实现为：生成一个行列式为预设值、维度与第一空矩阵相同的第一随机矩阵；将第一特征相关矩阵和第一随机矩阵输入至安全计算模型，得到第二矩阵；分别计算第一对称矩阵和第二矩阵的行列式；将第一对称矩阵的行列式和第二矩阵的行列式相乘，得到第一特征相关矩阵的行列式。
同理，确定第i个特征对应的第i余子式的行列式|H ii|，可以实现为：生成一个行列式为预设值、维度与第一空矩阵相同的随机矩阵；将第i余子式和该随机矩阵输入至安全计算模型，得到第i余子式对应的矩阵；分别计算第一对称矩阵和该第i余子式对应的矩阵的行列式；将第一对称矩阵的行列式和第i余子式对应的矩阵矩阵的行列式相乘，得到第i余子式的行列式|H ii|。
步骤S407，第一参与方基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征。
在一些实施例中，第一参与方根据共线性量化因子确定目标特征可以实现为：判断第一样本特征数据对应的各特征的共线性量化因子是否大于预设边界值；将共线性量化因子大于预设边界值的特征确定为目标特征。
步骤S408，第一参与方从第一样本特征数据中删除目标特征的特征数据，得到第一参与方与其他参与方进行联合训练的第一训练数据。
其中，第一参与方中目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，其他特征包括第一参与方持有的除目标特征以外的特征和第二参与方持有的所有特征。
步骤S409，第二参与方获取自身持有的第二样本特征数据。
第二参与方从自身存储的数据中获取目标用户的特征数据作为第二样本特征数据。
步骤S410，第二参与方基于第二样本特征数据和预先训练好的安全计算模型，构建虚拟的第二特征相关矩阵。
在一些实施例中，第二参与方构建第二特征相关矩阵可以实现为：基于第二样本特征数据确定第二样本特征数据对应的各特征的特征数据和第二样本特征数据对应的样本个数；分别计算各特征的特征数据对应的均值和标准差；基于各特征的特征数据、各特征的特征数据对应的均值、各特征的特征数据对应的标准差和样本个数，确定处理后的第二样本特征数据；将处理后的第二样本特征数据输入至安全计算模型，得到第一矩阵；根据处理后的第二样本特征数据确定第二对称矩阵；生成一个行数和列数均等于第一矩阵的列数的第二空矩阵；根据第二对称矩阵、第一矩阵、第一矩阵的转置矩阵和第二空矩阵，构建虚拟的第一特征相关矩阵。
例如，第二样本特征数据为B；根据
Figure PCTCN2021140955-appb-000025
对B中各个特征数据x进行处理，得到处理后的第二样本特征数据为D；将处理后的第二样本特征数据D输入至预先训练好的安全计算模型y 1,…,m＝f(x 1,…,m)，得到第一矩阵E；根据处理后的第二样本特征数据的转置矩阵D T和处理后的第二样本特征数据D，确定第二对称矩阵G＝D TD；第二参与方生成的第二空矩阵为F′；得到第二参与方构建的第二特征相关矩阵
Figure PCTCN2021140955-appb-000026
步骤S411，第二参与方基于特征相关矩阵，确定第二样本特征数据对应的各特征的共线性量化因子。
在一些实施例中，第二参与方确定第二样本特征数据对应的各特征的共线性量化因子可以实现为：确定第二特征相关矩阵的行列式；删除第二特征相关矩阵第i+m 1行、 第i+m 1列数据(因为第二特征相关矩阵H′的前m 1列为第一样本特征数据的特征，因此在确定第二样本特征数据对应的各特征的共线性量化因子时，需要加上m 1)，得到第二特征相关矩阵对应的各余子式，其中，i＝1，2，…，m 2，m 2为第二样本特征数据对应的特征个数；确定各余子式的行列式；基于第二特征相关矩阵的行列式和各余子式的行列式，确定第二样本特征数据对应的各特征的共线性量化因子。
删除第二特征相关矩阵H′的第i+m 1行、第i+m 1列的数据，得到第二样本特征数据的第i个特征对应的第i余子式
Figure PCTCN2021140955-appb-000027
计算第二特征相关矩阵的行列式|H′|和第i余子式的行列式
Figure PCTCN2021140955-appb-000028
确定第一样本特征数据对应的各特征的共线性量化因子(以方差膨胀因子为例)
Figure PCTCN2021140955-appb-000029
其中，确定第二特征相关矩阵的行列式|H′|可以实现为：生成一个行列式为预设值、维度与第二空矩阵相同的第二随机矩阵；将第二特征相关矩阵和第二随机矩阵输入至安全计算模型，得到第二矩阵；分别计算第二对称矩阵和第二矩阵的行列式；将第二对称矩阵的行列式和第二矩阵的行列式相乘，得到第二特征相关矩阵的行列式。
同理，确定第i个特征对应的第i余子式的行列式
Figure PCTCN2021140955-appb-000030
可以实现为：生成一个行列式为预设值、维度与第二空矩阵相同的随机矩阵；将第i余子式和该随机矩阵输入至安全计算模型，得到第i余子式对应的矩阵；分别计算第二对称矩阵和该第i余子式对应的矩阵的行列式；将第二对称矩阵的行列式和第i余子式对应的矩阵矩阵的行列式相乘，得到第i余子式的行列式
Figure PCTCN2021140955-appb-000031
步骤S412，第二参与方基于共线性量化因子，从第二样本特征数据对应的各特征中确定目标特征。
在一些实施例中，第二参与方根据共线性量化因子确定目标特征可以实现为：判断第二样本特征数据对应的各特征的共线性量化因子是否大于预设边界值；将共线性量化因子大于预设边界值的特征确定为目标特征。
步骤S413，第二参与方从第二样本特征数据中删除目标特征的特征数据，得到第二参与方与其他参与方进行联合训练的第二训练数据。
其中，第二参与方中目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，其他特征包括第二参与方持有的除目标特征以外的特征和第一参与方持有的所有特征。
在另一些实施例中，在步骤S413之后，第一参与方和第二参与方基于安全多方计算协议，利用第一训练数据和第二训练数据进行联合模型，由于第一训练数据和第二训练数据中不存在线性关系的训练数据，能够提高训练得到的联邦模型的准确率和稳定性，从而提高联邦模型的建模效果。
本申请实施例提供的数据处理方法，第一参与方与第二参与方基于安全多方计算预先训练得到安全计算模型；然后第一参与方和第二参与方分别获取第一样本特征数据和第二样本特征数据，第一参与方基于第一样本特征数据和安全计算模型，构建虚拟的第一特征相关矩阵，第二参与方基于第二样本特征数据和安全计算模型，构建虚拟的第二 特征相关矩阵；然后第一参与方基于第一特征相关矩阵确定第一样本特征数据对应的各特征的共线性量化因子，第二参与方基于第二特征相关矩阵确定第二样本特征数据对应的各特征的共线性量化因子；第一参与方由第一样本特征数据对应的各特征的共线性量化因子确定第一样本特征数据对应的各特征中的目标特征，第二参与方由第二样本特征数据对应的各特征的共线性量化因子确定第二样本特征数据对应的各特征中的目标特征，这些目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，这里的其他特征包括第一参与方和第二参与方持有的除目标特征以外的所有特征；确定目标特征后，第一参与方从第一样本特征数据中删除目标特征的特征数据，得到第一训练数据；第二参与方从第二样本特征数据中删除目标特征的特征数据，得到第二训练数据。如此，能够在保护数据隐私的前提下，对第一参与方和第二参与方持有的特征数据中存在共线性的数据进行筛选与剔除，得到不存在线性关系的训练数据，从而在进行联合训练时，第一参与方和第二参与方利用不存在线性关系的训练数据进行联合训练，能够提高联邦模型的准确率和稳定性，提高联邦模型的建模效果。
下面，将说明本申请实施例在一个实际的应用场景中的示例性应用。
纵向联邦学习通常由不同的参与方联合训练机器学习模型。在线性模型的建模过程中，共线性的存在会显著影响线性模型的稳定和效果，因此在建模过程中需要剔除。方差膨胀因子(VIF，Variance Inflation Factor)可以很好的量化某一个特征与其他所有特征的共线性，在实际建模中非常常用。相关技术中，计算VIF时，通常计算方式需要将各方数据汇总到一个地方进行。但是在各方(如银行企业)的数据可能涉及个人隐私或者商业机密，直接开放给其他参与方会造成信息泄露。相关技术中针对线性联邦建模的特征筛选方案主要都是针对单列数据，缺少VIF这种能够刻画多列数据的共线性的方法。相关技术无法在保护数据隐私的情况下联合多方的数据计算VIF。
本申请实施例针对两个独立参与方的场景，可以假设为两个公司之间进行数据联合建模。由于合规、隐私和商业机密等原因，各自公司的原始数据不能离开各自的公司。建模过程中交换的中间数据也不能推导或者泄露不必要的原始数据信息。两方公司各自持有多个同一批ID用户的不同特征数据，这些特征之间可能存在共线性，可能影响后续线性模型的建模效果。本申请实施例利用两方的隐私安全的技术方案来计算方差膨胀因子VIF，基于此来进行特征筛选。
图5为本申请实施例提供的纵向联邦情形下的方差膨胀因子的计算流程示意图，图6为本申请实施例提供的相关矩阵的行列式的计算流程示意图，下面结合图5和图6，对本申请实施例提供的纵向联邦情形下的方差膨胀因子的计算方法进行详细说明。
1)本申请实施例提供有两个独立的参与方(分别记为Alice跟Bob)，各自持有同一批ID的不同特征的数据，分别记为矩阵A，B(这里A，B具有相同的行数n，A具有列数m 1，B具有列数m 2)。
2)Alice，Bob在本地分别对A，B的特征(矩阵的列)进行归一化，得到矩阵C，D。归一化的方式为：
Figure PCTCN2021140955-appb-000032
x代表一列特征，
Figure PCTCN2021140955-appb-000033
代表特征的均值，σ x代表特征的标准差。
3)通过扩展的SPDZ协议，进行隐私下的矩阵加减法和矩阵乘法，计算矩阵乘积 C TD＝E，计算的结果Alice，Bob均持有。本申请实施例中，通过安全多方计算协议SPDZ，实现隐私保护下矩阵的加减法和矩阵乘法计算，各参与方输入自身的私有数据，得到计算结果，无法获得其他参与方的私有数据，使得整个计算过程不泄漏结果之外的任何额外信息，兼具高安全性与实用计算效率。
4)Alice本地计算矩阵乘法C TC＝F，Bob本地计算矩阵乘法D TD＝G。
5)上一步得到了皮尔逊Pearson相关性矩
Figure PCTCN2021140955-appb-000034
注1：这里Alice持有F、E、E T，Bob持有E、E T、G。
根据VIF定义，第i个特征的
Figure PCTCN2021140955-appb-000035
这里H ii为矩阵H的余子式(即删除矩阵的第i行和第i列剩下的矩阵)，|H|表示矩阵H的行列式。
6)假设第i个特征在Alice方(在Bob方需要指标i加上F ii的列数m 1)，那么Alice可以删除矩阵F的第i行跟第i列得到矩阵F ii。Alice跟Bob都可以删除矩阵E的第i行得到E i*，删除矩阵E T的第i列得到
Figure PCTCN2021140955-appb-000036
那么余子式
Figure PCTCN2021140955-appb-000037
注2：这里Alice持有F ii、E i*、
Figure PCTCN2021140955-appb-000038
Bob持有E i*、
Figure PCTCN2021140955-appb-000039
本申请实施例中，通过巧妙的矩阵变换将无法处理的原始矩阵H的行列式的计算问题，转换为可以处理的分块矩阵F ii、E i*、
Figure PCTCN2021140955-appb-000040
和
Figure PCTCN2021140955-appb-000041
的行列式计算问题。
下面解决根据分块矩阵F ii、E i*、
Figure PCTCN2021140955-appb-000042
和
Figure PCTCN2021140955-appb-000043
的行列式计算问题。
7)根据注1跟注2，计算|H ii|与|H|所需要解决的是同一个问题。在计算|H|时，下面的M 1、M 2、M 3、M 4分别对应F、E、E T、G，在计算|H ii|时，下面的M 1、M 2、M 3、M 4分别对应F ii、E i*、
Figure PCTCN2021140955-appb-000044
Alice持有矩阵M 1、M 2、M 3，Bob持有矩阵M 2、M 3、M 4，计算行列式
Figure PCTCN2021140955-appb-000045
8)从这里开始解决7)的问题(由于Alice无法得到M 4，Bob无法得到M 1，所以7)中行列式无法直接计算得到)。当M 1可逆时，
Figure PCTCN2021140955-appb-000046
9)Alice本地计算|M 1|和
Figure PCTCN2021140955-appb-000047
并本地生成行列式为1、维度与M 4一致的随机矩阵R 1。Bob本地生成行列式为1与M 4维度一致的随机矩阵R 2。
10)再次利用扩展的SPDZ协议计算矩阵减法跟乘法并恢复结果矩阵J＝R 1
Figure PCTCN2021140955-appb-000048
R 2。
由于随机矩阵的存在，通过J，Alice跟Bob均无法恢复对方的原始矩阵。
本申请实施例中，通过构造特殊的随机矩阵来混淆原始矩阵，将涉及原始数据信息的矩阵的行列式问题转化具有相同行列式的随机矩阵的行列式计算问题。通过随机矩阵对原始矩阵进行混淆处理，各参与方无法恢复其他参与方的数据，能够对各参与方的数据进行隐私保护，避免造成数据泄露。
11)Alice本地计算
Figure PCTCN2021140955-appb-000049
从而得到行列式
Figure PCTCN2021140955-appb-000050
解决了7)的问题。
12)根据7)至11)描述的方法，依次计算|H 11|、|H 22|、…、
Figure PCTCN2021140955-appb-000051
和|H|，从而得到VIF 1，VIF 2，…，VIF m1+m2。
本申请实施例提供的方法，通过巧妙的矩阵变换将无法处理的原始行列式计算问题转换为可以处理的分块矩阵行列式问题；通过构造特殊的随机矩阵来混淆原始矩阵，将涉及原始数据信息的矩阵的行列式问题转化具有相同行列式的随机矩阵的行列式计算问题；通过SPDZ协议的巧妙配合，整个计算过程不泄漏结果之外的任何额外信息，兼具高安全性与实用计算效率。通过本申请实施例提供的方法，使得方差膨胀因子VIF的安全计算成为可能，进而能进行高效的特征筛选，提高后续线性模型的整体效果；并且兼顾高安全性与实用性，除结果外，不泄漏不必要的数据信息，且整体计算开销都控制在生产实用的效率范围。
下面继续说明本申请实施例提供的数据处理装置的实施为软件模块的示例性结构，在一些实施例中，如图2所示，存储在存储器140的数据处理装置70应用于对模型进行联合训练的第二参与方，所述数据处理装置70中的软件模块可以包括：
构建模块71，配置为基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，所述安全计算模型由所述第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；
第一确定模块72，配置为基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子；
第二确定模块73，配置为基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征；
删除模块74，配置为从所述第一样本特征数据中删除所述目标特征的特征数据，得到所述第一参与方与所述其他参与方进行联合训练的第一训练数据；
其中，所述目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，所述其他特征包括所述第一参与方持有的除所述目标特征以外的特征和所述其他参与方持有的特征。
在一些实施例中，所述构建模块71，包括：
第一确定子模块，配置为基于所述第一样本特征数据，确定第一样本特征数据对应的各特征的特征数据和所述第一样本特征数据对应的样本个数；
计算子模块，配置为分别计算所述各特征的特征数据对应的均值和标准差；
第二确定子模块，配置为基于所述各特征的特征数据、所述各特征的特征数据对应的均值、所述各特征的特征数据对应的标准差和所述样本个数，确定处理后的第一样本特征数据；
输入子模块，配置为将所述处理后的第一样本特征数据输入至所述安全计算模型，得到第一矩阵；
构建子模块，配置为根据所述处理后的第一样本特征数据和所述第一矩阵，构建虚 拟的特征相关矩阵。
在一些实施例中，所述构建子模块，包括：
第一确定单元，配置为根据所述处理后的第一样本特征数据确定第一对称矩阵；
第一生成单元，配置为生成一个行数和列数均等于所述第一矩阵的列数的空矩阵；
构建单元，配置为根据所述第一对称矩阵、所述第一矩阵、所述第一矩阵的转置矩阵和所述空矩阵，构建虚拟的特征相关矩阵。
在一些实施例中，所述第一确定模块72，包括：
第三确定子模块，配置为确定所述特征相关矩阵的行列式；
删除子模块，配置为删除所述特征相关矩阵第i行、第i列数据，得到所述特征相关矩阵对应的各余子式，其中，i＝1，2，…，m 1，m 1为所述第一样本特征数据对应的特征个数；
第四确定子模块，配置为确定所述各余子式的行列式；
第五确定子模块，配置为基于所述特征相关矩阵的行列式和所述各余子式的行列式，确定所述第一样本特征数据对应的各特征的共线性量化因子。
在一些实施例中，所述第三确定子模块，包括：
第二生成单元，配置为生成一个行列式为预设值、维度与所述空矩阵相同的第一随机矩阵；
输入单元，配置为将所述特征相关矩阵和所述第一随机矩阵输入至所述安全计算模型，得到第二矩阵；
第一计算单元，配置为分别计算所述第一对称矩阵和所述第二矩阵的行列式；
第二计算单元，配置为将所述第一对称矩阵的行列式和所述第二矩阵的行列式相乘，得到所述特征相关矩阵的行列式。
在一些实施例中，所述第五确定子模块，包括：
第三计算单元，配置为根据所述各余子式的行列式与所述特征相关矩阵的行列式，计算各特征的方差膨胀因子；
第二确定单元，配置为将所述各特征的方差膨胀因子确定为所述第一样本特征数据对应的各特征的共线性量化因子。
在一些实施例中，所述第二确定模块73，包括：
判断子模块，配置为判断所述第一样本特征数据对应的各特征的共线性量化因子是否大于预设边界值；
第六确定子模块，配置为将共线性量化因子大于预设边界值的特征确定为目标特征。
在一些实施例中，所述数据处理装置70，还包括：
获取模块，配置为获取所述第一参与方和所述其他参与方持有的共同训练样本；
筛选模块，配置为从所述共同训练样本中筛选出本次训练样本；
第三确定模块，配置为将所述本次训练样本的特征数据确定为所述第一样本特征数据。
在一些实施例中，所述获取模块，包括：
获取子模块，配置为获取所述第一参与方持有数据的标识；
第七确定子模块，配置为基于加密计算，利用所述第一参与方持有数据的标识和所述其他参与方持有数据的标识，确定所述第一参与方和所述其他参与方的共同训练样本。
这里需要指出的是：以上数据处理装置实施例项的描述，与上述方法描述是类似的，具有同方法实施例相同的有益效果。对于本申请数据处理装置实施例中未披露的技术细节，本领域的技术人员请参照本申请方法实施例的描述而理解。
本申请实施例提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行本申请实施例上述的数据处理方法。
本申请实施例提供一种存储有可执行指令的存储介质，其中存储有可执行指令，当可执行指令被处理器执行时，将引起处理器执行本申请实施例提供的方法，例如，如图3至图6示出的方法。
在一些实施例中，存储介质可以是FRAM、ROM、PROM、EPROM、EEPROM、闪存、磁表面存储器、光盘、或CD-ROM等存储器；也可以是包括上述存储器之一或任意组合的各种设备。
在一些实施例中，可执行指令可以采用程序、软件、软件模块、脚本或代码的形式，按任意形式的编程语言(包括编译或解释语言，或者声明性或过程性语言)来编写，并且其可按任意形式部署，包括被部署为独立的程序或者被部署为模块、组件、子例程或者适合在计算环境中使用的其它单元。
作为示例，可执行指令可以但不一定对应于文件系统中的文件，可以可被存储在保存其它程序或数据的文件的一部分，例如，存储在超文本标记语言(HTML，Hyper Te xt Markup Language)文档中的一个或多个脚本中，存储在专用于所讨论的程序的单个文件中，或者，存储在多个协同文件(例如，存储一个或多个模块、子程序或代码部分的文件)中。
作为示例，可执行指令可被部署为在一个计算设备上执行，或者在位于一个地点的多个计算设备上执行，又或者，在分布在多个地点且通过通信网络互连的多个计算设备上执行。
以上所述，仅为本申请的实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。
工业实用性
本申请实施例提供一种数据处理方法、装置、设备、存储介质及程序产品，所述方法包括：基于第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵；基于特征相关矩阵，确定第一样本特征数据对应的各特征的共线性量化因子；基于共线性量化因子，从第一样本特征数据对应的各特征中确定目标特征；从第一样本特征数据中删除目标特征的特征数据，得到用于进行联合训练的第一训练数据；其中，目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系。如此，在保护数 据隐私的前提下，实现对各参与方持有的特征数据中存在共线性的数据进行筛选与剔除，能够提高联合训练的到的联邦模型的准确率和稳定性，提高联邦模型的建模效果。

Claims (20)

  1. 一种数据处理方法，应用于联邦学习的第一参与方，所述方法包括：
    基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，所述安全计算模型由所述第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；
    基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子；
    基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征；
    从所述第一样本特征数据中删除所述目标特征的特征数据，得到所述第一参与方与所述其他参与方进行联合训练的第一训练数据；
    其中，所述目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，所述其他特征包括所述第一参与方持有的除所述目标特征以外的特征和所述其他参与方持有的特征。
  2. 根据权利要求1所述的方法，其中，所述基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，包括：
    基于所述第一样本特征数据，确定第一样本特征数据对应的各特征的特征数据和所述第一样本特征数据对应的样本个数；
    分别计算所述各特征的特征数据对应的均值和标准差；
    基于所述各特征的特征数据、所述各特征的特征数据对应的均值、所述各特征的特征数据对应的标准差和所述样本个数，确定处理后的第一样本特征数据；
    将所述处理后的第一样本特征数据输入至所述安全计算模型，得到第一矩阵；
    根据所述处理后的第一样本特征数据和所述第一矩阵，构建虚拟的特征相关矩阵。
  3. 根据权利要求2所述的方法，其中，所述根据所述处理后的第一样本特征数据和所述第一矩阵，构建虚拟的特征相关矩阵，包括：
    根据所述处理后的第一样本特征数据确定第一对称矩阵；
    生成一个行数和列数均等于所述第一矩阵的列数的空矩阵；
    根据所述第一对称矩阵、所述第一矩阵、所述第一矩阵的转置矩阵和所述空矩阵，构建虚拟的特征相关矩阵。
  4. 根据权利要求1至3任一项所述的方法，其中，所述基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子，包括：
    确定所述特征相关矩阵的行列式；
    删除所述特征相关矩阵第i行、第i列数据，得到所述特征相关矩阵对应的各余子式，其中，i＝1，2，…，m 1，m 1为所述第一样本特征数据对应的特征个数；
    确定所述各余子式的行列式；
    基于所述特征相关矩阵的行列式和所述各余子式的行列式，确定所述第一样本特征数据对应的各特征的共线性量化因子。
  5. 根据权利要求4所述的方法，其中，所述确定所述特征相关矩阵的行列式，包 括：
    生成一个行列式为预设值、维度与所述空矩阵相同的第一随机矩阵；
    将所述特征相关矩阵和所述第一随机矩阵输入至所述安全计算模型，得到第二矩阵；
    分别计算所述第一对称矩阵和所述第二矩阵的行列式；
    将所述第一对称矩阵的行列式和所述第二矩阵的行列式相乘，得到所述特征相关矩阵的行列式。
  6. 根据权利要求4所述的方法，其中，所述基于所述特征相关矩阵的行列式和所述各余子式的行列式，确定所述第一样本特征数据对应的各特征的共线性量化因子，包括：
    根据所述各余子式的行列式与所述特征相关矩阵的行列式，计算各特征的方差膨胀因子；
    将所述各特征的方差膨胀因子确定为所述第一样本特征数据对应的各特征的共线性量化因子。
  7. 根据权利要求1所述的方法，其中，所述基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征，包括：
    判断所述第一样本特征数据对应的各特征的共线性量化因子是否大于预设边界值；
    将共线性量化因子大于预设边界值的特征确定为目标特征。
  8. 根据权利要求1所述的方法，其中，所述方法还包括：
    获取所述第一参与方和所述其他参与方持有的共同训练样本；
    从所述共同训练样本中筛选出本次训练样本；
    将所述本次训练样本的特征数据确定为所述第一样本特征数据。
  9. 根据权利要求8所述的方法，其中，所述获取所述第一参与方和所述其他参与方持有的共同训练样本，包括：
    获取所述第一参与方持有数据的标识；
    基于加密计算，利用所述第一参与方持有数据的标识和所述其他参与方持有数据的标识，确定所述第一参与方和所述其他参与方的共同训练样本。
  10. 一种数据处理装置，应用于联邦学习的第一参与方，所述装置包括：
    构建模块，配置为基于所述第一参与方持有的第一样本特征数据和预先训练好的安全计算模型，构建虚拟的特征相关矩阵，所述安全计算模型由所述第一参与方和联邦学习的其他参与方基于安全多方计算预先训练得到；
    第一确定模块，配置为基于所述特征相关矩阵，确定所述第一样本特征数据对应的各特征的共线性量化因子；
    第二确定模块，配置为基于所述共线性量化因子，从所述第一样本特征数据对应的各特征中确定目标特征；
    删除模块，配置为从所述第一样本特征数据中删除所述目标特征的特征数据，得到所述第一参与方与所述其他参与方进行联合训练的第一训练数据；
    其中，所述目标特征的特征数据与其他特征中至少一个特征的特征数据存在线性关系，所述其他特征包括所述第一参与方持有的除所述目标特征以外的特征和所述其他参 与方持有的特征。
  11. 根据权利要求10所述的装置，其中，所述构建模块，包括：
    第一确定子模块，配置为基于所述第一样本特征数据，确定第一样本特征数据对应的各特征的特征数据和所述第一样本特征数据对应的样本个数；
    计算子模块，配置为分别计算所述各特征的特征数据对应的均值和标准差；
    第二确定子模块，配置为基于所述各特征的特征数据、所述各特征的特征数据对应的均值、所述各特征的特征数据对应的标准差和所述样本个数，确定处理后的第一样本特征数据；
    输入子模块，配置为将所述处理后的第一样本特征数据输入至所述安全计算模型，得到第一矩阵；
    构建子模块，配置为根据所述处理后的第一样本特征数据和所述第一矩阵，构建虚拟的特征相关矩阵。
  12. 根据权利要求11所述的装置，其中，所述构建子模块，包括：
    第一确定单元，配置为根据所述处理后的第一样本特征数据确定第一对称矩阵；
    第一生成单元，配置为生成一个行数和列数均等于所述第一矩阵的列数的空矩阵；
    构建单元，配置为根据所述第一对称矩阵、所述第一矩阵、所述第一矩阵的转置矩阵和所述空矩阵，构建虚拟的特征相关矩阵。
  13. 根据权利要求10至12任一项所述的装置，其中，所述第一确定模块，包括：
    第三确定子模块，配置为确定所述特征相关矩阵的行列式；
    删除子模块，配置为删除所述特征相关矩阵第i行、第i列数据，得到所述特征相关矩阵对应的各余子式，其中，i＝1，2，…，m 1，m 1为所述第一样本特征数据对应的特征个数；
    第四确定子模块，配置为确定所述各余子式的行列式；
    第五确定子模块，配置为基于所述特征相关矩阵的行列式和所述各余子式的行列式，确定所述第一样本特征数据对应的各特征的共线性量化因子。
  14. 根据权利要求13所述的装置，其中，所述第三确定子模块，包括：
    第二生成单元，配置为生成一个行列式为预设值、维度与所述空矩阵相同的第一随机矩阵；
    输入单元，配置为将所述特征相关矩阵和所述第一随机矩阵输入至所述安全计算模型，得到第二矩阵；
    第一计算单元，配置为分别计算所述第一对称矩阵和所述第二矩阵的行列式；
    第二计算单元，配置为将所述第一对称矩阵的行列式和所述第二矩阵的行列式相乘，得到所述特征相关矩阵的行列式。
  15. 根据权利要求13所述的装置，其中，所述第五确定子模块，包括：
    第三计算单元，配置为根据所述各余子式的行列式与所述特征相关矩阵的行列式，计算各特征的方差膨胀因子；
    第二确定单元，配置为将所述各特征的方差膨胀因子确定为所述第一样本特征数据对应的各特征的共线性量化因子。
  16. 根据权利要求10所述的装置，其中，所述第二确定模块，包括：
    判断子模块，配置为判断所述第一样本特征数据对应的各特征的共线性量化因子是否大于预设边界值；
    第六确定子模块，配置为将共线性量化因子大于预设边界值的特征确定为目标特征。
  17. 根据权利要求10所述的装置，其中，所述装置还包括：
    获取模块，配置为获取所述第一参与方和所述其他参与方持有的共同训练样本；
    筛选模块，配置为从所述共同训练样本中筛选出本次训练样本；
    第三确定模块，配置为将所述本次训练样本的特征数据确定为所述第一样本特征数据。
  18. 一种数据处理设备，所述设备包括：
    存储器，配置为存储可执行指令；
    处理器，配置为执行所述存储器中存储的可执行指令时，实现权利要求1至9任一项所述的数据处理方法。
  19. 一种计算机可读存储介质，所述计算机可读存储介质上存储有可执行指令，配置为引起处理器执行时，实现权利要求1至9任一项所述的数据处理方法。
  20. 一种计算机程序产品，包括计算机程序，所述计算机程序被处理器执行时实现权利要求1至9任一项所述的数据处理方法。
PCT/CN2021/140955 2021-04-26 2021-12-23 数据处理方法、装置、设备、存储介质及程序产品 Ceased WO2022227644A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110454684.2 2021-04-26
CN202110454684.2A CN113095514B (zh) 2021-04-26 2021-04-26 数据处理方法、装置、设备、存储介质及程序产品

Publications (1)

Publication Number Publication Date
WO2022227644A1 true WO2022227644A1 (zh) 2022-11-03

Family

ID=76679959

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/140955 Ceased WO2022227644A1 (zh) 2021-04-26 2021-12-23 数据处理方法、装置、设备、存储介质及程序产品

Country Status (2)

Country Link
CN (1) CN113095514B (zh)
WO (1) WO2022227644A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113095514B (zh) * 2021-04-26 2025-09-02 深圳前海微众银行股份有限公司 数据处理方法、装置、设备、存储介质及程序产品
CN113345597B (zh) * 2021-07-15 2021-11-16 中国平安人寿保险股份有限公司 传染病概率预测模型的联邦学习方法、装置及相关设备
CN113935050B (zh) * 2021-09-26 2024-09-27 平安科技（深圳）有限公司 基于联邦学习的特征提取方法和装置、电子设备、介质
CN114692201B (zh) * 2022-03-31 2023-03-31 北京九章云极科技有限公司 一种多方安全计算方法及系统
CN115293252A (zh) * 2022-07-29 2022-11-04 脸萌有限公司 信息分类的方法、装置、设备和介质
CN114996749B (zh) * 2022-08-05 2022-11-25 蓝象智联(杭州)科技有限公司 一种用于联邦学习的特征过滤方法
CN115545216B (zh) * 2022-10-19 2023-06-30 上海零数众合信息科技有限公司 一种业务指标预测方法、装置、设备和存储介质
CN117252287B (zh) * 2023-08-04 2024-07-05 上海零数众合信息科技有限公司 一种基于联邦皮尔逊相关性分析的指标预测方法及系统
CN118246524B (zh) * 2024-05-24 2024-07-26 蓝象智联(杭州)科技有限公司 一种用于联邦学习的多重共线性分析方法及电子设备

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160283735A1 (en) * 2015-03-24 2016-09-29 International Business Machines Corporation Privacy and modeling preserved data sharing
CN110909216A (zh) * 2019-12-04 2020-03-24 支付宝(杭州)信息技术有限公司 检测用户属性之间的关联性的方法及装置
CN111062487A (zh) * 2019-11-28 2020-04-24 支付宝(杭州)信息技术有限公司 基于数据隐私保护的机器学习模型特征筛选方法及装置
CN111160573A (zh) * 2020-04-01 2020-05-15 支付宝(杭州)信息技术有限公司 保护数据隐私的双方联合训练业务预测模型的方法和装置
US20200285984A1 (en) * 2019-03-06 2020-09-10 Hcl Technologies Limited System and method for generating a predictive model
CN111654853A (zh) * 2020-08-04 2020-09-11 索信达(北京)数据技术有限公司 一种基于用户信息的数据分析方法
CN111966473A (zh) * 2020-07-24 2020-11-20 支付宝(杭州)信息技术有限公司 一种线性回归任务的运行方法及装置、电子设备
CN112597540A (zh) * 2021-01-28 2021-04-02 支付宝(杭州)信息技术有限公司 基于隐私保护的多重共线性检测方法、装置及系统
CN113095514A (zh) * 2021-04-26 2021-07-09 深圳前海微众银行股份有限公司 数据处理方法、装置、设备、存储介质及程序产品

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110163378B (zh) * 2019-03-04 2024-12-06 腾讯科技（深圳）有限公司 特征处理方法、装置、计算机可读存储介质和计算机设备

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160283735A1 (en) * 2015-03-24 2016-09-29 International Business Machines Corporation Privacy and modeling preserved data sharing
US20200285984A1 (en) * 2019-03-06 2020-09-10 Hcl Technologies Limited System and method for generating a predictive model
CN111062487A (zh) * 2019-11-28 2020-04-24 支付宝(杭州)信息技术有限公司 基于数据隐私保护的机器学习模型特征筛选方法及装置
CN110909216A (zh) * 2019-12-04 2020-03-24 支付宝(杭州)信息技术有限公司 检测用户属性之间的关联性的方法及装置
CN111160573A (zh) * 2020-04-01 2020-05-15 支付宝(杭州)信息技术有限公司 保护数据隐私的双方联合训练业务预测模型的方法和装置
CN111966473A (zh) * 2020-07-24 2020-11-20 支付宝(杭州)信息技术有限公司 一种线性回归任务的运行方法及装置、电子设备
CN111654853A (zh) * 2020-08-04 2020-09-11 索信达(北京)数据技术有限公司 一种基于用户信息的数据分析方法
CN112597540A (zh) * 2021-01-28 2021-04-02 支付宝(杭州)信息技术有限公司 基于隐私保护的多重共线性检测方法、装置及系统
CN113095514A (zh) * 2021-04-26 2021-07-09 深圳前海微众银行股份有限公司 数据处理方法、装置、设备、存储介质及程序产品

Also Published As

Publication number Publication date
CN113095514A (zh) 2021-07-09
CN113095514B (zh) 2025-09-02

Similar Documents

Publication Publication Date Title
WO2022227644A1 (zh) 数据处理方法、装置、设备、存储介质及程序产品
Wang et al. An intelligent blockchain-based access control framework with federated learning for genome-wide association studies
Gascón et al. Secure linear regression on vertically partitioned datasets.
Shi et al. Secure multi-pArty computation grid LOgistic REgression (SMAC-GLORE)
WO2021120676A1 (zh) 联邦学习网络下的模型训练方法及其相关设备
CN110851869A (zh) 敏感信息处理方法、设备及可读存储介质
JP2020525814A (ja) 秘密分散を使用したロジスティック回帰モデリング方式
AlSereidi et al. Novel federated decision making for distribution of anti-SARS-CoV-2 monoclonal antibody to eligible high-risk patients
WO2021204268A1 (zh) 基于隐私数据进行模型训练
Koley et al. To use, or not to use the spatial Durbin model?–that is the question
Li et al. Supporting regularized logistic regression privately and efficiently
CN117675196A (zh) 保护输入数据和参数的量子神经网络隐私保护方法及系统
CN114611008B (zh) 基于联邦学习的用户服务策略确定方法、装置及电子设备
Zhang et al. Joint intelligence ranking by federated multiplicative update
CN114547658A (zh) 数据处理方法、装置、设备及计算机可读存储介质
CN114742239A (zh) 基于联邦学习的金融保险理赔风险模型训练方法及装置
Sun et al. A systematic review on privacy-preserving distributed data mining
CN112000988A (zh) 因子分解机回归模型构建方法、设备及可读存储介质
Omer et al. Privacy-preserving of SVM over vertically partitioned with imputing missing data
Chen et al. Fed-eini: an efficient and interpretable inference framework for decision tree ensembles in federated learning
Shah et al. Maintaining privacy in medical imaging with federated learning, deep learning, differential privacy, and encrypted computation
CN116776376A (zh) 一种基于联邦学习的元宇宙隐私保护方法
WO2021166242A1 (ja) 秘匿計算方法、秘匿計算システム及びプログラム
CN112949866B (zh) 泊松回归模型的训练方法、装置、电子设备及存储介质
Chowdhury et al. Integrating Federated Machine Learning in Blockchain Crowdfunding Ecosystem for Increased Security and Transparency

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21939103

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22.03.2024)

122 Ep: pct application non-entry in european phase

Ref document number: 21939103

Country of ref document: EP

Kind code of ref document: A1