A data processing method and device, electronic equipment and storage medium

Abstract

The present disclosure provides a data processing method and device, electronic equipment and storage medium, wherein the method comprises: obtaining label residual values determined for each data sample; each label residual value is used to represent the deviation degree between the true label of the corresponding data sample and the predicted label of the target model; based on each label residual value, residual decomposition is performed on each data sample to obtain the model parameter correction amount corresponding to each data sample; the model parameter correction amount is sent to the data receiver, so that the data receiver reconstructs the model parameter information of the target model according to the model parameter correction amount. Through residual decomposition, the present disclosure can effectively perturb the residual, which enables the data sender to protect the true label it owns, while the data sender can deliver the necessary correction information (i.e. model parameter correction amount) to the data receiver to reconstruct accurate model parameter information, further ensuring the training performance of the model.

Description

Technical Field

[0001] This disclosure relates to the field of data security technology, and more specifically, to a data processing method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of applications such as artificial intelligence and big data mining and analysis, the demand for data volume is increasing. To extract greater value, the integration of data from multiple sources is often necessary. This multi-source data can originate from different organizations; for example, transaction data can be distributed across financial institutions, and medical diagnostic records can come from medical facilities. Furthermore, it can come from different industries; for instance, transaction data can originate from both e-commerce and financial institutions.

[0003] The compliance constraints imposed by regulatory authorities on the circulation of multiple data elements make it impractical to centrally collect detailed data and train models for business operations, thus creating a data barrier. To break down this barrier, federated learning can be employed, allowing participating parties to obtain a global model without sharing their private data. This private data could include tagging information about target users, such as whether they are high-net-worth individuals. While reasonably exploring the application value of data, it is also necessary to prevent data misuse and protect privacy. In practical applications, encrypted exchange of model-related information can be used to achieve collaborative optimization of the federated model.

[0004] One related technology provides a residual encryption method. Taking labeled data as privacy data as an example, the labeled party can perform differential privacy in the form of addition before sending model-related information. After satisfying the differential privacy condition, the labeled party then sends the data to the unlabeled party in a homomorphic encrypted form. Due to the presence of noise, the unlabeled party cannot reconstruct the true residual based on the model-related parameters it possesses, thereby achieving the purpose of protecting privacy data.

[0005] However, due to the inherent noise in differential privacy, the unlabeled side suffers some performance loss in model accuracy. Summary of the Invention

[0006] This disclosure provides at least one data processing method, apparatus, electronic device, and storage medium. The data sender encrypts the private data, such as tags, through residual decomposition. The data receiver reconstructs the model parameters based on the model parameter correction amount obtained from the encryption, resulting in a model with high accuracy.

[0007] In a first aspect, embodiments of this disclosure provide a data processing method, including:

[0008] Obtain the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0009] Based on the residual values ​​of each label, residual decomposition is performed on each data sample to obtain the model parameter correction amount corresponding to each data sample;

[0010] The model parameter correction amount is sent to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0011] In one possible implementation, the step of performing residual decomposition on each data sample based on each of the label residual values ​​to obtain the model parameter correction amount corresponding to each data sample includes:

[0012] The residual values ​​of each label are sorted according to the order of data size to obtain the sorted residual values ​​of each label.

[0013] When a target data sample that needs to be decomposed is selected from the various data samples, the label residual value of the sorted target data sample is changed to obtain the residual change information corresponding to each data sample.

[0014] Based on the residual change information corresponding to each data sample and the true label of each data sample, the model parameter correction amount corresponding to each data sample is determined.

[0015] In one possible implementation, the target data sample to be residual decomposition is selected from the various data samples according to the following steps:

[0016] The sorted label residual values ​​are grouped according to a preset number of groups to obtain the grouped label residual values.

[0017] Select a preset number of target label residual values ​​from the label residual values ​​after each grouping;

[0018] The data sample corresponding to the target label residual value is determined as the target data sample.

[0019] In one possible implementation, when the residual change information corresponding to each data sample corresponds to a residual change vector, the step of changing the label residual value of the sorted target data sample to obtain the residual change information corresponding to each data sample includes:

[0020] For each data sample, if it is determined that the data sample is not the target data sample, the first residual change value corresponding to the data sample is determined; or,

[0021] If it is determined that the data sample is the target data sample, and the label residual value of the data sample is greater than zero, then the second residual change value corresponding to the data sample is determined; or,

[0022] If it is determined that the data sample is the target data sample and the label residual value of the data sample is less than zero, the third residual change value corresponding to the data sample is determined.

[0023] The residual change values ​​corresponding to each of the data samples are collected to determine the residual change vector; wherein, the residual change vector is used to characterize whether the label residual value of each data sample has changed.

[0024] In one possible implementation, determining the model parameter correction amount for each data sample based on the residual change information corresponding to each data sample and the true label of each data sample includes:

[0025] The first operator is determined by performing a dot product operation on the residual change vector and the transpose of the residual change vector; and, given the label value corresponding to the true label of each data sample, the second operator corresponding to the label value is determined.

[0026] Based on the first operator and the second operator, the model parameter correction amount corresponding to each data sample is determined.

[0027] In one possible implementation, after obtaining the residual change information corresponding to each data sample, the method further includes:

[0028] The changed label residual value is determined by multiplying the residual change information corresponding to each data sample with the sorted label residual values.

[0029] Based on the changed label residual value, determine the model parameter information to be sent to the data receiver;

[0030] The model parameter information is sent to the data recipient.

[0031] In one possible implementation, sending the model parameter information to the data receiver includes:

[0032] If the model convergence condition is not met, the model parameter information is sent to the data receiver.

[0033] Sending the model parameter correction amount to the data receiver includes:

[0034] Once the model convergence condition is met, the model parameter correction amount is sent to the data receiver.

[0035] Secondly, this disclosure also provides a data processing method, including:

[0036] The system receives model parameter correction amounts corresponding to each data sample sent by the data sender; wherein the model parameter correction amounts are determined based on residual decomposition of the label residual values ​​determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0037] The model parameter information of the target model is reconstructed based on the model parameter correction amount.

[0038] In one possible implementation, reconstructing the model parameter information of the target model based on the model parameter correction amount includes:

[0039] Obtain the data feature information input to the target model;

[0040] Based on the data feature information and the model parameter correction amount, the model parameter information of the target model is determined.

[0041] In one possible implementation, when the data feature information includes data feature vectors, determining the model parameter information of the target model based on the data feature information and the model parameter correction amount includes:

[0042] Perform a dot product operation on the data feature vector and the transpose of the data feature vector to determine the third operator; and perform a dot product operation on the transpose of the data feature vector and the data feature vector to determine the fourth operator.

[0043] Based on the third operator, the fourth operator, and the model parameter correction amount, the model parameter information of the target model is determined.

[0044] Thirdly, this disclosure also provides a data processing apparatus, comprising:

[0045] The acquisition module is used to acquire the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model.

[0046] The decomposition module is used to perform residual decomposition on each data sample based on each label residual value to obtain the model parameter correction amount corresponding to each data sample.

[0047] The sending module is used to send the model parameter correction amount to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0048] Fourthly, this disclosure also provides a data processing apparatus, comprising:

[0049] The receiving module is used to receive the model parameter correction amount corresponding to each data sample sent by the data sender; wherein, the model parameter correction amount is determined by residual decomposition based on the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0050] The reconstruction module is used to reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0051] Fifthly, this disclosure also provides an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, they perform a data processing method as described in any one of the first aspect and its various embodiments, and the second aspect and its various embodiments.

[0052] In a sixth aspect, this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the data processing method as described in any one of the first aspect and its various embodiments, and the second aspect and its various embodiments.

[0053] Using the aforementioned data processing method, apparatus, electronic device, and storage medium, after obtaining the label residual values ​​determined for each data sample, residual decomposition can be performed on each data sample based on the label residual values. Then, the model parameter correction amounts corresponding to each data sample obtained from the residual decomposition can be sent to the data receiver. The data receiver can then reconstruct the model parameter information based on the model parameter correction amounts. In this disclosure, the data sender (i.e., the labeled party) can effectively perturb the residuals during the residual decomposition process, ensuring that the data receiver (i.e., the unlabeled party) can only construct a model with incorrect label information. This allows the data sender to protect its true labels. Simultaneously, the labeled party can transmit necessary correction information (i.e., model parameter correction amounts) to the unlabeled party, enabling the unlabeled party to reconstruct accurate model parameter information according to the correction information, further ensuring the model's training performance.

[0054] Other advantages of this disclosure will be explained in more detail below with reference to the accompanying drawings.

[0055] It should be understood that the above description is merely an overview of the technical solution of this disclosure, so as to enable a clearer understanding of the technical means of this disclosure, and thus enable its implementation in accordance with the contents of the specification. To make the above and other objects, features, and advantages of this disclosure more apparent and understandable, specific embodiments of this disclosure are illustrated below. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0057] Furthermore, the same reference numerals are used to denote the same parts throughout the accompanying drawings. In the accompanying drawings:

[0058] Figure 1 A flowchart of a data processing method provided by an embodiment of this disclosure is shown;

[0059] Figure 2 A flowchart is shown illustrating a specific method for determining residual change information in the data processing method provided in this embodiment of the present disclosure;

[0060] Figure 3A timing flowchart of a data processing method provided in an embodiment of this disclosure is shown;

[0061] Figure 4 This illustration shows an application diagram of a data processing method provided in an embodiment of the present disclosure;

[0062] Figure 5 A flowchart of another data processing method provided by an embodiment of this disclosure is shown;

[0063] Figure 6 A schematic diagram of a data processing apparatus provided in an embodiment of this disclosure is shown;

[0064] Figure 7 A schematic diagram of another data processing apparatus provided in an embodiment of this disclosure is shown;

[0065] Figure 8 A schematic diagram of an electronic device provided in an embodiment of the present disclosure is shown. Detailed Implementation

[0066] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0067] In the description of embodiments disclosed herein, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, numbers, steps, behaviors, components, portions or combinations thereof disclosed herein, and are not intended to exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, portions or combinations thereof.

[0068] Unless otherwise stated, " / " means "or". For example, A / B can mean A or B. In this article, "and / or" is merely a way of describing the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone.

[0069] The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more.

[0070] Research has found that related technologies can achieve collaborative optimization of federated models by exchanging model-related information in an encrypted manner.

[0071] One related technology provides a residual encryption method. Taking labeled data as privacy data as an example, the labeled party can perform differential privacy in the form of addition before sending model-related information. After satisfying the differential privacy condition, the labeled party sends the data to the unlabeled party in the form of homomorphic encryption.

[0072] In the process of differential privacy, that is, giving the residual r... i The process of adding noise compromises differential privacy. The added noise satisfies the condition noise~Lap(2ε -1 (That is, the noise comes from a standard deviation of 2ε) -1 (The generated Laplace random variable). Due to the presence of noise, the unlabeled side cannot reconstruct the true residual r based on the available model-related parameters. i This is to achieve the goal of protecting privacy data.

[0073] Another encryption method is localized differential privacy random response. The idea is to randomly perturb the composition of the batch so that the residuals of the interaction remain unchanged in form, but the actual rank is less than the number of samples in the batch, thus the resulting residuals are not unique.

[0074] Specifically, the labeled party selects a subset C from its samples based on the agreed batch size, generates a random response vector m for the label y∈{0,1} according to the category distribution of the labels, and then performs differential privacy processing to obtain m. rr Send to the unlabeled party. When the unlabeled party determines based on feature x... b After constructing the local model, the labeled party will obtain m from the random response. i =1 residual gives the true residual value r i For samples not in the batch, the value is set to 0. While ensuring the residual dimension of the interaction remains the same as the batch size (|B|), the constructed residual matrix is ​​homomorphically encrypted and sent to the unlabeled party. The formal representation of the sent content D is as follows:

[0075]

[0076] in σ(·) is the sigmoid function, and <·> represents homomorphic encryption. This method is called a hybrid label protection strategy combining differential privacy and homomorphic encryption, and it can achieve ε-differential privacy under the following conditions:

[0077]

[0078] q is the proportion of m marked as 1. Through the above operations on the labeled side, it becomes impossible for the unlabeled side to construct a system of linear equations about the residuals.

[0079] However, although the above methods achieve label protection and the classification accuracy loss is reduced compared to the first encryption method, the presence of differential privacy and random zeroing does not fully utilize the sample information in the batch. Therefore, there is still a performance loss compared to plaintext computation. Moreover, the reduction in classification accuracy loss comes at the cost of increased training costs, and it is not feasible to put it into production use in the short term.

[0080] To at least partially address one or more of the aforementioned problems and other potential issues, this disclosure provides at least one data processing scheme in which the labeled party encrypts the label, a private data, through residual decomposition, and the unlabeled party reconstructs the model parameters based on the encrypted model parameter corrections, resulting in a model with higher accuracy.

[0081] It should be noted that in this embodiment, both the labeled and unlabeled parties need to perform local computation and collaboratively interact to correct information, without significantly increasing computational overhead. Here, the labeled party corresponds to the data sender, and the unlabeled party corresponds to the data receiver. Residual decomposition is performed on the labeled party, while parameter reconstruction is completed on the unlabeled party.

[0082] In practical applications, the data sender here refers to the initiator with real labels and its processing equipment used to perform operations including residual decomposition. The data receiver here refers to the responder with model input features but without real labels and its processing equipment used to perform operations including parameter reconstruction.

[0083] To facilitate understanding of this embodiment, a data processing method disclosed in this disclosure will first be described in detail. The data processing method provided in this disclosure is generally executed by an electronic device with a certain computing capability. This electronic device may include, for example, a terminal device or other processing device. The terminal device may be a user equipment (UE), mobile device, user terminal, cellular phone, personal digital assistant (PDA), handheld device, etc. In some possible implementations, this data processing method can be implemented by a processor calling computer-readable instructions stored in memory.

[0084] See Figure 1The diagram shows a flowchart of a data processing method provided in an embodiment of this disclosure. This method is primarily executed at the data sender and specifically includes steps S101 to S103, wherein:

[0085] S101: Obtain the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0086] S102: Perform residual decomposition on each data sample based on the residual values ​​of each label to obtain the model parameter correction amount corresponding to each data sample;

[0087] S103: Send the model parameter correction amount to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0088] To facilitate understanding of the data processing method provided in this disclosure, the application scenarios of the method will be described in detail below. The data processing method in this disclosure can be mainly applied to the field of federated learning. Here, federated learning can refer to a multi-source data fusion learning method, such as fused data formed from transaction data from financial institutions and medical diagnostic records from medical institutions.

[0089] In related technologies, longitudinal logistic regression is often used to construct joint models. However, due to the inherent limitations of longitudinal logistic regression, the privacy data of one party can potentially be inferred from the other. For example, data generated within a financial institution regarding whether a user is a high-net-worth individual could potentially be leaked to medical institutions through the sharing of model parameter information, posing a significant security risk to user privacy.

[0090] In order to better promote federated learning while enhancing the protection of privacy data, this disclosure provides a collaborative data processing scheme, which mainly corresponds to two stages: a model training stage and a model correction stage.

[0091] During the model training phase, the residual decomposition process is primarily performed on the labeled party. After decomposition, only a portion of the residual-related obfuscation information is exchanged with the unlabeled party under homomorphic encryption protection. The remaining residual information is held solely by the labeled party and does not participate in the model training process. This ensures that even if the unlabeled party solves the residual using a label leakage attack method (satisfying the full rank condition), the accuracy of its label inference about the labeled party will be comparable to that of random binary classification because the correspondence between the residual and the true label is no longer one-to-one. Furthermore, it cannot infer the true label of the labeled party from the sign of the residual obtained using a known label leakage attack method, thus ensuring the security of privacy data.

[0092] The model correction phase primarily involves parameter reconstruction performed by the unlabeled party. After the labeled party's model training converges, the process is initiated by the labeled party and responded to by the unlabeled party. Only one additional round of interaction between the labeled and unlabeled parties regarding the unlabeled party's weight information is required to complete the correction of each feature weight for the unlabeled party. Since the additional information provided by the labeled party cannot be obtained during the model training phase, and this information represents the entirety of the residual preservation implemented by the real label and the protective label, the unlabeled party, even after reconstructing a high-precision model, cannot obtain the labeled party's label information during the model correction phase, thus further ensuring the security of privacy data.

[0093] In this embodiment of the disclosure, the data sender, having obtained the label residual values ​​determined for each data sample, can perform residual decomposition on each data sample based on the label residual values ​​to obtain the model parameter correction amount corresponding to each data sample, and finally send the model parameter correction amount to the data receiver. Here, the model parameter correction amount is additional information provided by the labeled party to the unlabeled party, and this additional information is determined after decomposing the residuals.

[0094] The labeled party randomly performs residual decomposition locally based on how well the unlabeled party fits its own labels, thereby achieving the effect of randomizing and replacing label information, so that the unlabeled party cannot infer the true label from the residual information processed by the labeled party.

[0095] As can be seen, in this embodiment of the present disclosure, the perturbation residuals enable the unlabeled party to construct only a model with erroneous label information, and the erroneous information contained therein is generated by the labeled party decomposing the residual information locally, which the unlabeled party cannot distinguish; then, the labeled party transmits the necessary correction information to the unlabeled party, so that the unlabeled party can correct the weights of the features, but cannot obtain the labels of the labeled party.

[0096] With the cooperation of the labeled party, the unlabeled party corrects the biased feature weights and obtains the correct model parameter information. Since this information consists of the labels from the labeled party and perturbation information of the labels, and the perturbation information is unrelated to the information known to the unlabeled party during the model training phase, the unlabeled party cannot obtain the label information from the labeled party.

[0097] Considering the crucial role of determining the model parameter correction amounts in reconstructing the model parameters for unlabeled squares, the following section will focus on explaining the process of determining these correction amounts, which can be achieved through the following steps:

[0098] Step 1: Sort the residual values ​​of each label according to the order of data size to obtain the sorted residual values ​​of each label.

[0099] Step 2: After selecting the target data sample that needs to be decomposed from each data sample, change the label residual value of the target data sample in the sorting to obtain the residual change information corresponding to each data sample.

[0100] Step 3: Based on the residual change information corresponding to each data sample and the true label of each data sample, determine the model parameter correction amount corresponding to each data sample.

[0101] Here, the residual values ​​of each label can be sorted first to obtain residual change information, and then the model parameter correction amount can be determined based on the residual change information and the true label.

[0102] The aforementioned residual change information is used to indicate whether the label residual value has changed. For the label residual value that has changed, a first label can be applied, and for the label residual value that has not changed, a second label can be applied, thereby achieving the purpose of randomly replacing the real label. Considering that the relevant model parameter correction amount is obtained by combining the residual change information and the real label, that is, the unlabeled party does not obtain the direct real label but the correction amount after label mapping. Based on this correction amount and the corresponding input features, the model parameter information can be constructed.

[0103] It should be noted that the labeled party only needs to perform local calculations to synthesize the actual labels and the labels it has mapped after the model converges before sending it to the unlabeled party.

[0104] To achieve the goal of randomly replacing real labels, target samples need to be selected through a series of operations such as grouping and selection before residual transformation. Specifically, firstly, the sorted label residual values ​​can be grouped according to a preset number of groups to obtain grouped label residual values. Then, a preset number of target label residual values ​​are selected from the label residual values ​​of each group. Finally, the data samples corresponding to the target label residual values ​​are determined as the target data samples. The grouping here can be preset, for example, divided into two groups, four groups, etc., without specific restrictions. The purpose is to prevent the unlabeled party from obtaining additional information about the labeled party's label information based on the similarity between features. At the same time, in the process of selecting the data samples to be transformed from each group, random quantitative sampling can be used to ensure the randomness of label replacement and further improve data security.

[0105] In this embodiment, the residual change information can point to a residual change vector, where each vector corresponds to the residual change of a data sample. Specifically, if a data sample is determined to be a non-target data sample, a first residual change value is identified; if a data sample is determined to be a target data sample and its label residual value is greater than zero, a second residual change value is identified; and if a data sample is determined to be a target data sample and its label residual value is less than zero, a third residual change value is identified. Then, the residual change values ​​corresponding to each data sample are aggregated to determine the residual change vector.

[0106] Based on the above residual change vector, the model parameter correction amount can be determined, which is achieved through the following steps:

[0107] Step 1: Perform a dot product operation on the residual change vector and the transpose of the residual change vector to determine the first operator; and, given the label values ​​corresponding to the true labels of each data sample, determine the second operator corresponding to the label values.

[0108] Step 2: Based on the first and second operators, determine the model parameter correction amount corresponding to each data sample.

[0109] Here, the model parameter correction amount can be determined based on the first and second operators. This model parameter correction amount can also be a multi-dimensional vector, with each dimension pointing to the correction amount of a data sample, thereby achieving overall correction for the target model.

[0110] In the process of determining residual change information, the changed label residual value can also be determined by multiplying the residual change information corresponding to each data sample with the sorted label residual values. Then, the model parameter information to be sent to the data receiver can be determined based on the changed label residual value, and the model parameter information can be sent to the data receiver.

[0111] The calculation process of the modified label residual value can be a step that only needs to be performed once before the model converges. For example, the labeling party can perform the above steps in the first round to change the residual value of the randomly selected sample. This is equivalent to changing the true label of the sample, which reduces the computational cost while ensuring the safety of the label.

[0112] It should be noted that the model parameter information here can be sent from the labeled party to the unlabeled party during the intermediate training process, i.e., before the model convergence condition is met. Because residual changes occur here, even if the unlabeled party receives the model parameter information, it cannot deduce the true label. After the model training is completed, i.e., the model convergence condition is met, the labeled party can send correction values ​​to the unlabeled party so that the unlabeled party can reconstruct accurate model parameter information.

[0113] In practical applications, model convergence can be associated with training batches. Once all data samples in a batch have completed a round of parameter updates, it can be determined whether the model meets the convergence condition. If it converges, the correction amount is sent in the manner described above; otherwise, model training continues.

[0114] To facilitate understanding of the process for determining the model parameter correction amounts described above, the following examples can be used for further explanation.

[0115] First, the prerequisites are as follows:

[0116] 1) Modeling needs to be completed between two parties. Party A acts as the data user (Guest), and Party B acts as the data provider (Host) and collaborator (Arbiter). Labels are on Party A (the labeled party), and features are on Party B (the unlabeled party).

[0117] 2) Party A's label is binary (0 / 1);

[0118] 3) Party B only knows that Party A's labels are in binary classification 0 / 1 form. Party B does not know other information about Party A's labels, but Party B is interested in the labels of any category (i.e., Party B wants to know the label information of any sample from Party A).

[0119] 4) Both parties agree to complete the joint modeling according to the basic steps of vertical logistic regression, and Party B may only launch a semi-honest attack against Party A.

[0120] Here, the labeled and unlabeled parties agree in a secure manner that joint modeling will use longitudinal logistic regression to determine the size |B| and composition of the batch B in each round of modeling, the model parameters, and the learning rate η. Alignment samples are determined through secure intersection. Party B knows the features {(x... i A knows the labels {y} of the aligned samples. i Each round is initiated by Party B, with Party A responding, and Party B updates the local model accordingly. The parameters are now complete. Meanwhile, parties A and B agree that once the model converges (flag = 1), they will collaboratively reconstruct the true weights W of party B.

[0121] During the model training phase, if the model fails to reach convergence (flag = 0 at this point), the following steps are performed:

[0122] In a batch B, B party uses its own model parameters. and feature set Constructing a local model (Calculation method is) ), and then Send it to party A using homomorphic encryption (e.g., the Paillier algorithm).

[0123] When the model reaches the convergence condition (flag = 0), both parties have completed the training of the model with the incorrect labels and will perform the following operations:

[0124] Party A will adjust the model parameters. The data is sent to Party B, who then reconstructs the actual model parameters w as follows:

[0125]

[0126] Where x corresponds to the data feature vector, y corresponds to the true label value, and S is S B The label mapping vector obtained by batch splicing corresponds to the residual change information.

[0127] It should be noted that the above residual change information may correspond to the retained information of the residuals of this batch of samples. Furthermore, this retained information can be determined in the first round of training. Specifically, this can be achieved through the following steps:

[0128] 1) Party A, based on the residual r i B(t) Sort all samples within a batch (from smallest to largest or from largest to smallest).

[0129] 2) Party A performs randomized residual decomposition on the residuals of this batch of samples as follows: Randomly select a number m∈{2,4} as the group size of the samples in this batch.

[0130] For the sorted sample sequence, starting from the first sample, every m samples form a group. Among the symbols Let G represent the smallest integer not less than |B| / m. j Randomly select from within the group The samples form a candidate set C. B ,and And for all in C B For the samples within, the corresponding retention information is determined as follows:

[0131] Where, if r i B(t) >0, then If r i B(t) <0, then Regardless of whether it's 1 or -1, the residual will be changed subsequently. For all values ​​not in C... B Sample x within i ,have No further residual changes will be performed. The specific determination can be made using the following formula:

[0132]

[0133] Where sort(i) is sorted according to the residual Δr i B The sorted order of the i-th sample within the batch. Here, to prevent B from reverse-engineering, the data samples can be rearranged according to the original order of B.

[0134] To further explain the residual decomposition corresponding to the rearrangement operation, we can refer to... Figure 2 The residual decomposition diagram shown below will be further explained.

[0135] like Figure 2 As shown, the batch size |B| = 8. After sorting by residual values ​​in ascending order, the bold text in the second and fourth columns are respectively... For samples randomly selected by group at m=2 and m=4, the non-zero representation is shown, and the corresponding randomized residual decomposition can be seen in the 1 / -1 part of the third and fifth columns.

[0136] Here, information is retained based on the tag. The residual change vector S can be determined. BS B Each element in the table reflects whether the residual information of the sample changed during the first round of residual decomposition. When the residual changes, otherwise Right now:

[0137]

[0138] Determine the residual change vector S B In this case, the first operator SS can be determined by performing a dot product operation on the residual change vector and the transpose of the residual change vector. T Furthermore, given the label values ​​corresponding to the true labels of each data sample, determine the second operator corresponding to the label values. Then determine the model parameter correction amount.

[0139] It should be noted that the aforementioned model parameter corrections can be sent from A to B after model training has ended. During model training, the pseudo-residuals sent from A to B can be calculated using the following formula:

[0140]

[0141] In this way, B updates its own model parameters in the dense state. The calculation method is as follows:

[0142]

[0143] in, The parameters obtained by Party B based on the identity information The gradient information is equivalent to the pseudo-gradient information obtained by multiplying the features and corresponding pseudo-residues on each sample by A.

[0144] Repeat the above steps for each batch B until all samples have completed one round of parameter updates. Then determine whether the model meets the convergence condition. If it converges, set flag = 1; otherwise, continue. The timing diagram of the main operations is as follows: Figure 3 As shown.

[0145] As can be seen, the embodiments disclosed herein perform two steps on the labeled side: residual decomposition (only in the first round of joint modeling) and transmission of model parameter reconstruction information. Meanwhile, on the unlabeled side, weight construction is required. For details, please refer to... Figure 4 Please provide an explanation.

[0146] like Figure 4As shown, after batch initialization, the unlabeled party can build a local model based on the model parameters provided by the labeled party. Then, the labeled party determines whether it is in the first round of training. If so, the labeled party can construct the retained residuals and label transformation according to the above steps, while the unlabeled party can update based on the pseudo-residuals. Afterward, if it is determined that the model has converged, the labeled party can send additional information to the unlabeled party so that the unlabeled party can reconstruct the weights.

[0147] It should be noted that the embodiments disclosed herein do not limit the specific implementation methods of the batch initialization and convergence condition modules, nor does the alignment of the two samples before batch initialization limit the implementation method. The focus of the embodiments disclosed herein is that the labeled side performs residual decomposition on the unlabeled side and collaboratively implements weight reconstruction with the unlabeled side. Other steps in the model training phase can be solved using the classic gradient descent method.

[0148] Furthermore, data security can be analyzed from two aspects: model training and model modification. During model training, residual information is held solely by the labeled party and does not participate in the model training process. The unlabeled party cannot determine, based on known information, whether the labeled party has modified the labels of a specific example.

[0149] For the labeled party, this security is at the information theory level. Its effect prevents the unlabeled party from obtaining its own label by increasing computing power, and only requires one round of local operations. The additional computational overhead is negligible compared to the overall joint modeling overhead. Therefore, the unlabeled party cannot obtain the labeled party's true label y using a semi-honest attack.

[0150] The theoretical basis for the weight reconstruction is the equation XWS = YS. This is because the true label and its transformed information are integrated... The data is sent to the unlabeled party. This process does not involve gradient information exchange and is unrelated to known label leakage methods. This prevents the unlabeled party from calculating the label mapping S and the actual label Y made by the labeled party. Therefore, if the attacker uses the known label leakage attack method, the random correspondence between the residual and the label cannot be eliminated from the beginning. In other words, regardless of whether the result obtained by the unlabeled party is a positive or negative residual value, its corresponding real label may be either positive (1) or negative (0).

[0151] Furthermore, model usability can be analyzed from the perspective of weight reconstruction. Given that the problem can be limited to binary classification, if the labeled side changes the residual values ​​of randomly selected samples in the first round, the effect is equivalent to changing the true label of the sample, from positive (negative) to negative (positive). The significance of ranking lies in counteracting the leakage of more label information from the labeled side due to differences in randomly initialized parameters by the unlabeled side. Therefore, applying S or... There are no obvious differences in information, and the final result was not obfuscated, including differential privacy, before being sent by the labeled party. In theory, the weights can achieve better distinguishing ability for sample categories.

[0152] See Figure 5 The diagram shows a flowchart of a data processing method provided in an embodiment of this disclosure. This method is primarily executed at the data receiving end and specifically includes steps S501 to S502, wherein:

[0153] S501: Receive the model parameter correction amount corresponding to each data sample sent by the data sender; wherein, the model parameter correction amount is determined by residual decomposition based on the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0154] S502: Reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0155] Here, the parameters of the target model can be reconstructed based on the received model parameter correction amount. Since the model parameter correction amount is determined based on the real label of the data sender and the corresponding residual change information, the corresponding label replacement situation can be marked. Furthermore, the real label cannot be parsed based on the correction amount. Therefore, the accurate reconstruction of the model parameters can be achieved under the premise of protecting the real label of the data sender.

[0156] The process for determining the model parameter correction amount can be referred to the relevant description in the above embodiments, and will not be repeated here.

[0157] In the process of reconstructing model parameter information, in addition to relying on the aforementioned model parameter correction amounts, it is also necessary to consider the data feature information input to the target model. This can be achieved through the following steps:

[0158] Step 1: Perform a dot product operation on the data feature vector and its transpose to determine the third operator; and perform a dot product operation on the transpose of the data feature vector and the data feature vector to determine the fourth operator.

[0159] Step 2: Based on the third operator, the fourth operator, and the model parameter correction amount, determine the model parameter information of the target model.

[0160] The process for determining the model parameter information in this embodiment can refer to the formula for the model parameter w described above:

[0161]

[0162] XX here T This corresponds to the third operator, X. T X corresponds to the fourth operator. This corresponds to the correction amount, which in turn allows us to determine the model parameters.

[0163] In the description of this specification, the references to terms such as "some possible implementations," "some implementations," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that implementation or example is included in at least one implementation or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same implementation or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more implementations or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different implementations or examples described in this specification, as well as the features of different implementations or examples.

[0164] Regarding the method flowcharts of embodiments of this disclosure, certain operations are described as different steps performed in a certain order. Such flowcharts are illustrative and not restrictive. Some steps described herein may be grouped together and performed in a single operation, some steps may be divided into multiple sub-steps, and some steps may be performed in an order different from that shown herein. The various steps shown in the flowcharts may be implemented in any way by any circuit structure and / or tangible mechanism (e.g., by software running on a computer device, hardware (e.g., logic functions implemented by a processor or chip), and / or any combination thereof).

[0165] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0166] Based on the same inventive concept, this disclosure also provides a data processing device corresponding to the data processing method. Since the principle of the device in this disclosure for solving the problem is similar to that of the data processing method described above, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0167] Reference Figure 6 The diagram shown is a schematic representation of a data processing apparatus provided in an embodiment of this disclosure. The apparatus includes: an acquisition module 601, a decomposition module 602, and a sending module 603; wherein,

[0168] The acquisition module 601 is used to acquire the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model.

[0169] The decomposition module 602 is used to perform residual decomposition on each data sample based on the residual values ​​of each label, and to obtain the model parameter correction amount corresponding to each data sample.

[0170] The sending module 603 is used to send model parameter correction amounts to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amounts.

[0171] Using the aforementioned data processing device, after obtaining the label residual values ​​for each data sample, residual decomposition can be performed on each data sample based on these residual values. Then, the model parameter correction values ​​corresponding to each data sample obtained from the residual decomposition can be sent to the data receiver. The data receiver can then reconstruct the model parameter information based on these correction values. In this disclosure, the data sender (i.e., the labeled party) can effectively perturb the residuals during residual decomposition, ensuring that the data receiver (i.e., the unlabeled party) can only construct a model with incorrect label information. This allows the data sender to protect its true labels. Simultaneously, the labeled party can transmit necessary correction information (i.e., model parameter correction values) to the unlabeled party, enabling the unlabeled party to reconstruct accurate model parameter information according to the correction information, further ensuring the model's training performance.

[0172] In one possible implementation, the decomposition module 602 is configured to perform residual decomposition on each data sample based on each label residual value according to the following steps, to obtain the model parameter correction amount corresponding to each data sample, including...

[0173] Sort the residual values ​​of each label to obtain the sorted residual values ​​of each label;

[0174] When the target data sample to be decomposed is selected from the data samples, the label residual value of the target data sample is changed to obtain the residual change information corresponding to each data sample.

[0175] Based on the residual change information corresponding to each data sample and the true label of each data sample, the correction amount of the model parameters corresponding to each data sample is determined.

[0176] In one possible implementation, the decomposition module 602 is configured to select target data samples for residual decomposition from the various data samples according to the following steps:

[0177] The sorted label residual values ​​are grouped according to the preset grouping number to obtain the grouped label residual values;

[0178] Select a preset number of target label residual values ​​from the label residual values ​​after each grouping;

[0179] The data sample corresponding to the target label residual value is determined as the target data sample.

[0180] In one possible implementation, given that the residual change information for each data sample corresponds to a residual change vector, the decomposition module 602 is used to change the label residual value of the sorted target data sample according to the following steps to obtain the residual change information for each data sample:

[0181] For each data sample, if it is determined that the data sample is not the target data sample, determine the change value of the first residual corresponding to the data sample; or,

[0182] If it is determined that the data sample is the target data sample and the label residual value of the data sample is greater than zero, then determine the corresponding second residual change value of the data sample; or,

[0183] If it is determined that the data sample is the target data sample and the label residual value of the data sample is less than zero, determine the third residual change value corresponding to the data sample.

[0184] The residual change values ​​corresponding to each data sample are collected to determine the residual change vector; where the residual change vector is used to characterize whether the label residual value of each data sample has changed.

[0185] In one possible implementation, the decomposition module 602 is used to determine the model parameter correction amount for each data sample based on the residual change information corresponding to each data sample and the true label of each data sample, according to the following steps:

[0186] The first operator is determined by performing a dot product operation on the residual change vector and the transpose of the residual change vector; and the second operator corresponding to the label value is determined after determining the true label value of each data sample.

[0187] Based on the first and second operators, determine the model parameter correction amount corresponding to each data sample.

[0188] In one possible implementation, the sending module 603 is further configured to:

[0189] After obtaining the residual change information corresponding to each data sample, the changed label residual value is determined by multiplying the residual change information corresponding to each data sample with the sorted label residual values.

[0190] The model parameter information to be sent to the data receiver is determined based on the changed label residual value;

[0191] Send model parameter information to the data recipient.

[0192] In one possible implementation, the sending module 603 is configured to send model parameter information to the data receiver according to the following steps:

[0193] If the model convergence condition is not met, send model parameter information to the data receiver;

[0194] The sending module 603 is used to send model parameter corrections to the data receiver according to the following steps:

[0195] Once the model convergence condition is met, the model parameter correction amount is sent to the data receiver.

[0196] Reference Figure 7 The diagram shown is a schematic of another data processing apparatus provided in an embodiment of this disclosure. The apparatus includes: a receiving module 701 and a reconstruction module 702; wherein,

[0197] The receiving module 701 is used to receive the model parameter correction amount corresponding to each data sample sent by the data sender; wherein, the model parameter correction amount is determined by residual decomposition based on the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model;

[0198] The reconstruction module 702 is used to reconstruct the model parameter information of the target model based on the model parameter correction amount.

[0199] In one possible implementation, the reconstruction module 702 is used to reconstruct the model parameter information of the target model according to the model parameter correction amount according to the following steps:

[0200] Obtain the data feature information input to the target model;

[0201] Based on data feature information and model parameter correction amounts, the model parameter information of the target model is determined.

[0202] In one possible implementation, when the data feature information includes data feature vectors, the reconstruction module 702 is used to determine the model parameter information of the target model based on the data feature information and the model parameter correction amount according to the following steps:

[0203] Perform a dot product operation on the data feature vector and the transpose of the data feature vector to determine the third operator; and perform a dot product operation on the transpose of the data feature vector and the data feature vector to determine the fourth operator.

[0204] Based on the third operator, the fourth operator, and the model parameter correction, the model parameter information of the target model is determined.

[0205] It should be noted that the apparatus in the embodiments of this application can implement each process of the aforementioned method and achieve the same effect and function, which will not be elaborated here.

[0206] This disclosure also provides an electronic device, such as... Figure 8 The diagram shown is a schematic representation of an electronic device structure provided in this embodiment of the present disclosure, including: a processor 801, a memory 802, and a bus 803. The memory 802 stores machine-readable instructions executable by the processor 801 (e.g., ...). Figure 6 The device includes the acquisition module 601, the decomposition module 602, and the sending module 603, which correspond to the execution instructions, etc.; for example... Figure 7 In the device, the receiving module 701 and the reconfiguration module 702 (and the corresponding execution instructions, etc.) are used. When the electronic device is running, the processor 801 and the memory 802 communicate via the bus 803. When machine-readable instructions are executed by the processor 801, they are executed as follows: Figure 1 Or such as Figure 5 The steps of the data processing method shown are as follows.

[0207] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the data processing method described in the above method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0208] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the data processing method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0209] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0210] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the descriptions of the apparatus, device, and computer-readable storage medium embodiments are simplified because they are substantially similar to the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0211] The apparatus, device, and computer-readable storage medium provided in this application correspond one-to-one with the method. Therefore, the apparatus, device, and computer-readable storage medium also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, device, and computer-readable storage medium will not be repeated here.

[0212] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, apparatus (devices or systems), or computer-readable storage media. Therefore, this disclosure can take the form of a completely hardware implementation, a completely software implementation, or an implementation combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer-readable storage medium implemented on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-readable program code.

[0213] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices or systems), and computer-readable storage media according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0214] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0215] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0216] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0217] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0218] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. Furthermore, although the operations of the methods of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0219] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A data processing method, characterized in that, include: Obtain the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model; Based on the residual values ​​of each label, residual decomposition is performed on each data sample to obtain the model parameter correction amount corresponding to each data sample; The residual decomposition includes: sorting the label residual values ​​according to their numerical values; selecting target data samples for residual decomposition from the data samples based on the sorting results; performing change processing on the label residual values ​​of the selected target data samples to obtain residual change information; and determining the model parameter correction amount based on the residual change information and the true labels of each data sample. The model parameter correction amount is sent to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amount.

2. The method according to claim 1, characterized in that, Select the target data samples for residual decomposition from the various data samples according to the following steps: The sorted label residual values ​​are grouped according to a preset number of groups to obtain the grouped label residual values. Select a preset number of target label residual values ​​from the label residual values ​​after each grouping; The data sample corresponding to the target label residual value is determined as the target data sample.

3. The method according to claim 1, characterized in that, In the case where the residual change information corresponds to the residual change vector for each data sample, the step of changing the label residual value of the sorted target data sample to obtain the residual change information corresponding to each data sample includes: For each data sample, if it is determined that the data sample is not the target data sample, the first residual change value corresponding to the data sample is determined; or, If it is determined that the data sample is the target data sample, and the label residual value of the data sample is greater than zero, then the second residual change value corresponding to the data sample is determined; or, If it is determined that the data sample is the target data sample and the label residual value of the data sample is less than zero, the third residual change value corresponding to the data sample is determined. The residual change values ​​corresponding to each of the data samples are collected to determine the residual change vector; wherein, the residual change vector is used to characterize whether the label residual value of each data sample has changed.

4. The method according to claim 3, characterized in that, The step of determining the model parameter correction amount for each data sample based on the residual change information corresponding to each data sample and the true label of each data sample includes: The first operator is determined by performing a dot product operation on the residual change vector and the transpose of the residual change vector; and, given the label value corresponding to the true label of each data sample, the second operator corresponding to the label value is determined. Based on the first operator and the second operator, the model parameter correction amount corresponding to each data sample is determined.

5. The method according to any one of claims 1 to 4, characterized in that, After obtaining the residual change information corresponding to each data sample, the method further includes: The changed label residual value is determined by multiplying the residual change information corresponding to each data sample with the sorted label residual values. Based on the changed label residual value, determine the model parameter information to be sent to the data receiver; The model parameter information is sent to the data recipient.

6. The method according to claim 5, characterized in that, Sending the model parameter information to the data receiver includes: If the model convergence condition is not met, the model parameter information is sent to the data receiver. Sending the model parameter correction amount to the data receiver includes: Once the model convergence condition is met, the model parameter correction amount is sent to the data receiver.

7. A data processing method, characterized in that, include: The system receives model parameter correction amounts corresponding to each data sample sent by the data sender. These correction amounts are determined based on residual decomposition using the label residual values ​​of each data sample. The residual decomposition includes: sorting the label residual values ​​according to their numerical values; selecting target data samples from the data samples for residual decomposition based on the sorting results; performing change processing on the label residual values ​​of the selected target data samples to obtain residual change information; and determining the model parameter correction amounts based on the residual change information and the true labels of each data sample. Each label residual value characterizes the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model. The model parameter information of the target model is reconstructed based on the model parameter correction amount.

8. The method according to claim 7, characterized in that, The step of reconstructing the model parameter information of the target model based on the model parameter correction amount includes: Obtain the data feature information input to the target model; Based on the data feature information and the model parameter correction amount, the model parameter information of the target model is determined.

9. The method according to claim 8, characterized in that, When the data feature information includes data feature vectors, determining the model parameter information of the target model based on the data feature information and the model parameter correction amount includes: Perform a dot product operation on the data feature vector and the transpose of the data feature vector to determine the third operator; and perform a dot product operation on the transpose of the data feature vector and the data feature vector to determine the fourth operator. Based on the third operator, the fourth operator, and the model parameter correction amount, the model parameter information of the target model is determined.

10. A data processing apparatus, characterized in that, include: The acquisition module is used to acquire the label residual value determined for each data sample; each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model. The decomposition module is used to perform residual decomposition on each data sample based on each label residual value to obtain the model parameter correction amount corresponding to each data sample. The residual decomposition includes: sorting the label residual values ​​according to their numerical values; selecting target data samples for residual decomposition from the data samples based on the sorting results; performing change processing on the label residual values ​​of the selected target data samples to obtain residual change information; and determining the model parameter correction amount based on the residual change information and the true labels of each data sample. The sending module is used to send the model parameter correction amount to the data receiver so that the data receiver can reconstruct the model parameter information of the target model based on the model parameter correction amount.

11. A data processing apparatus, characterized in that, include: A receiving module is used to receive the model parameter correction amounts corresponding to each data sample sent by the data sender. The model parameter correction amounts are determined based on residual decomposition of the label residual values ​​of each data sample. The residual decomposition includes: sorting the label residual values ​​according to their numerical values; selecting target data samples from the data samples that require residual decomposition based on the sorting results; performing change processing on the label residual values ​​of the selected target data samples to obtain residual change information; and determining the model parameter correction amounts based on the residual change information and the true labels of each data sample. Each label residual value is used to characterize the degree of deviation between the true label of the corresponding data sample and the label predicted by the target model. The reconstruction module is used to reconstruct the model parameter information of the target model based on the model parameter correction amount.

12. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the memory via the bus, and the machine-readable instructions, when executed by the processor, perform the data processing method as described in any one of claims 1 to 9.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the data processing method as described in any one of claims 1 to 9.

Images