Multi-party training sample selection method and device, electronic equipment and medium

Abstract

The specification discloses a privacy-protected data transmission method and device, electronic equipment and medium. The method comprises: an enabled participant of federated learning obtains interference feature data of a sample of a local target annotation result provided by a non-enabled participant based on a differential privacy algorithm, and obtains an interference feature dataset by summarizing. The enabled participant constructs a decision tree of the interference feature dataset based on a node splitting manner, and any target node in the decision tree is split according to an optimal splitting condition of the node. The splitting condition comprises a feature data filtering rule under at least one target annotation result. The enabled participant determines a sample selection rule of the target annotation result based on a splitting condition corresponding to at least one node in the decision tree when actually splitting, and provides the sample selection rule to the non-enabled participant. The non-enabled participant selects a training sample to be used as a target annotation result from local unlabeled samples based on the received sample selection rule.

Description

Technical Field

[0001] This document relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device and medium for selecting training samples involving multiple parties. Background Technology

[0002] Federated learning is a machine learning training framework that enables multiple participants to jointly train a model while meeting data privacy requirements. While federated learning solves the data privacy issue during the training phase, in the earlier training sample preparation phase, participants need to negotiate and summarize the selection rules for training samples based on the model's application requirements and the characteristics of their respective samples. These rules are then used to select training samples from their local private data that align with the training objectives. However, to summarize these selection rules, participants inevitably need to share some of their private data as support.

[0003] Therefore, in scenarios involving multiple parties in joint modeling, ensuring data privacy during the training sample preparation stage is a pressing technical issue that needs to be addressed. Summary of the Invention

[0004] This specification provides a method, apparatus, electronic device, and medium for selecting training samples involving multiple parties. While ensuring data privacy, it combines the data characteristics of samples from each participant in federated learning to analyze a unified selection rule for training samples. This helps each participant select training samples with consistent training objectives from their private data based on the selection rule.

[0005] To solve the above-mentioned technical problems, the embodiments in this specification are implemented as follows:

[0006] Firstly, a multi-party participation method for selecting training samples is proposed, including:

[0007] The non-enabling participants in the federated learning process generate interference feature data of the local target annotation results based on the differential privacy algorithm, and send the interference feature data to the enabling participants in the federated learning process.

[0008] The enabling participant integrates the interference feature data received from the non-enabling participant into an interference feature dataset based on the differential privacy algorithm. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding noise-reduced feature data.

[0009] The empowering participants construct a decision tree for the interference feature dataset using a node-based splitting method. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0010] The enabling participant determines the sample selection rule for the target labeling result based on the target candidate splitting condition corresponding to at least one node in the decision tree, and feeds back the sample selection rule to the non-enabling participant;

[0011] The non-empowered participant selects training samples from the local unlabeled samples based on the received sample selection rules, which will be used as the target labeling result.

[0012] Secondly, a federated learning device is proposed, including:

[0013] The first sending module, based on a differential privacy algorithm, generates interference feature data of samples from the local target annotation results and sends the interference feature data to the enabling participants in the federated learning process. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0014] The first receiving module receives the sample selection rule of the target labeling result sent by the empowering participant, wherein the sample selection rule is determined by the empowering participant based on the target candidate splitting condition corresponding to at least one node in the decision tree;

[0015] The filtering module selects training samples from the local unlabeled samples, based on the received sample selection rules, to be used as the target labeling result.

[0016] Thirdly, an electronic device is proposed, comprising: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the following operations:

[0017] Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of the federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0018] The sample selection rule for receiving target annotation results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree.

[0019] Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples.

[0020] Fourthly, a computer-readable storage medium is provided that stores one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform the following operations:

[0021] Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of the federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0022] The sample selection rule for receiving target annotation results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree.

[0023] Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples.

[0024] Fifthly, a federated learning device is proposed, comprising:

[0025] The second receiving module receives interference feature data of the target annotation result sample sent by the non-empowered participant in federated learning. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0026] The integration module, based on the differential privacy algorithm, integrates the received interference feature data provided by the non-empowered participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0027] The decision tree construction module constructs a decision tree for the interference feature dataset based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0028] The second sending module determines the sample selection rule for the target labeling result based on the target candidate splitting condition corresponding to at least one node in the decision tree, and feeds back the sample selection rule to the non-empowered participant. The non-empowered participant selects training samples to be used as the target labeling result from the local unlabeled samples based on the received sample selection rule.

[0029] In a sixth aspect, an electronic device is proposed, comprising: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the following operations:

[0030] The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0031] Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset, and the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0032] A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node, where each candidate splitting condition contains at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0033] Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules.

[0034] In a seventh aspect, a computer-readable storage medium is provided that stores one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform the following operations:

[0035] The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0036] Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset, and the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0037] A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node, where each candidate splitting condition contains at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0038] Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules.

[0039] Based on the scheme of the embodiments in this specification, each non-empowered participant in federated learning can de-identify the feature data of the samples of the local target labeling results using a differential privacy algorithm, and then upload the de-identified interference feature data to the empowered participant. Afterwards, the empowered participant summarizes the interference feature data and constructs a decision tree for the interference feature dataset based on node splitting. The splitting of any target node in the decision tree is constrained by statistically significant changes in splitting purity and constraint fitness, selecting the best feature data selection rule, without relying on sensitive information in the original data. Finally, the empowered participant constructs a sample selection rule for determining the target labeling results based on the feature data selection rule used in the actual splitting of at least one node in the decision tree, and provides it to the non-empowered participants. It can be seen that the scheme of the embodiments in this specification, while ensuring data privacy, can combine the data characteristics of the samples from each participant in federated learning to analyze a unified selection rule for training samples, thereby helping each participant to select training samples with consistent training objectives from their own private data based on the selection rule. Attached Figure Description

[0040] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings:

[0041] Figure 1 A diagram illustrating federated learning.

[0042] Figure 2 This is a flowchart illustrating the training sample selection method in an embodiment of this specification.

[0043] Figure 3 This is a schematic diagram of the decision tree structure in the training sample selection method of this specification embodiment.

[0044] Figure 4 This is a schematic diagram of a first structure of a multi-party training sample selection device according to an embodiment of this specification.

[0045] Figure 5 This is a schematic diagram of a second structure of a multi-party training sample selection device according to an embodiment of this specification.

[0046] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this specification. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this document clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments in this document, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this document.

[0048] As mentioned earlier, federated learning is a machine learning training framework that helps multiple stakeholders jointly train models while meeting data privacy protection requirements.

[0049] To better understand the scheme in this document, a brief introduction to federated learning is provided below.

[0050] refer to Figure 1 Suppose that institutions A and B, as participants, jointly train a machine learning model. Their respective business systems possess data related to their users. In addition, institution B also possesses the label data that the model needs to predict. For data privacy and security reasons, A and B cannot directly exchange data; therefore, they can use a federated learning model.

[0051] Federated learning includes:

[0052] Part 1: (Encrypted) Sample Alignment. Since the user groups of the two companies do not completely overlap, the system uses encrypted user sample alignment technology to identify the common users of both companies without A and B disclosing their respective data, and without exposing non-overlapping users, so as to combine the characteristics of these users for modeling.

[0053] Part Two: Encrypted Model Training. After identifying the shared user group, this data can be used to train a machine learning model. To ensure data confidentiality during training, encrypted training is required using a third-party server (C). Taking a linear regression model as an example, the training process can be divided into the following four steps:

[0054] Step 1: Server C distributes the public key to A and B to encrypt the data that needs to be exchanged during the training process.

[0055] Step 2: A and B exchange intermediate results for gradient calculation in an encrypted manner.

[0056] Step 3: A and B calculate the gradient value based on the encrypted gradient value, while B calculates the loss based on its label data and summarizes the results to C. C calculates the total gradient value based on the summarized results and decrypts it.

[0057] Step 4: C sends the decrypted total gradient value back to A and B respectively. A and B update the parameters of their respective models based on the total gradient value. Here, A and B's models have the same risk feature dimension, but the methods for updating their model parameters are not necessarily the same. That is to say, the weight values ​​of the risk feature in model A and model B may differ.

[0058] By iterating through the above steps until the loss function converges, the entire federated learning process is completed. During sample alignment and model training, the private data of A and B remain locally, eliminating the risk of exposure.

[0059] The above explains the principle of federated learning. It can be seen that federated learning solves the data privacy problem during the training phase. However, in the sample preparation phase before training, the participating parties need to negotiate the selection rules for training samples based on the application requirements of the model and the data characteristics of their respective samples. Then, according to the negotiated selection rules, training samples with consistent training objectives are selected from their local private data. In order to summarize the selection rules for training samples, the participating parties inevitably need to share some of their own private data as support, thus still raising the issue of data privacy protection.

[0060] In view of this, this specification aims to propose a multi-party training sample selection scheme that, while ensuring data privacy, combines the data characteristics of samples from each participant in federated learning to analyze a unified training sample selection rule. This helps each participant select training samples with consistent training objectives from their own private data based on the selection rule.

[0061] Specifically, one embodiment of this specification provides a multi-party participation method for selecting training samples, which may be executed by the corresponding apparatus described below. Figure 2 This is a flowchart illustrating the training sample selection method, which includes the following steps:

[0062] In S202, the non-enabling participants in federated learning generate interference feature data of the samples of local target annotation results based on the differential privacy algorithm, and send the interference feature data to the enabling participants in federated learning.

[0063] In the embodiments described in this specification, the samples of the target annotation results refer to samples that have been labeled with training labels. It should be understood that the training labels depend on the specific application scenario of the model, and this document does not impose specific limitations. For ease of understanding, a simple example is given here: Suppose the federated learning model is used to predict risky users, then "risky users" is the training label of the model, and correspondingly, the samples of the target annotation results are user samples labeled as risky users; similarly, suppose the federated learning model is used to predict blacklisted users, then the samples of the target annotation results are user samples labeled as blacklisted users.

[0064] Typically, sample labeling is done manually, while federated learning involves institutions or platforms that have accumulated large amounts of data. Manually labeling training data would obviously be very costly. Therefore, it is necessary to summarize the sample patterns under the target labeling results based on the data characteristics of a small subset of the labeled samples—that is, the selection rules described in this paper. Clearly, once the selection rules are determined, machines can then select suitable samples for training the model from a massive dataset according to these rules; these are the training samples described in this paper.

[0065] In the embodiments described in this specification, the enabling participant refers to a special participant responsible for centralized computation of the data provided by various non-enabling participants. The non-enabling participants use a differential privacy algorithm to add noise to the feature data of the samples in the local target annotation results, and then provide the noise-added interference feature data to the enabling participant. While the enabling participant cannot directly obtain the sensitive information from the interference feature data, it can still complete data-related calculations through the interference feature data to subsequently summarize and analyze the selection rules for training samples.

[0066] Differential privacy is a technique in cryptography that maximizes the accuracy of data retrieval when querying statistical databases, while minimizing the chance of identifying the records.

[0067] Specifically, in the embodiments of this specification, the non-empowered participant sums the feature data of the samples of the local target annotation result to obtain a summation result; then, the non-empowered participant splits the locally obtained summation result into multiple data blocks conforming to the prior distribution of the differential privacy algorithm and sends them to other non-empowered participants, and receives the summation results and corresponding multiple data blocks provided by other non-empowered participants; when the non-empowered participant collects the summation results and corresponding multiple data blocks provided by other non-empowered participants, it further calculates the interference feature data of the samples of the local target annotation result based on the locally obtained summation data and corresponding multiple data blocks, as well as the summation data and corresponding multiple data blocks provided by other non-empowered participants.

[0068] As an example, the steps of differential privacy described above are as follows:

[0069] 1) Each non-empowering participant {P1, P2, P3, ...} sums the feature data of the samples of the local target annotation results to obtain a summation result x. i And each of them chooses a relatively small positive number ε. i and a real number u;

[0070] 2) All parties P iGenerate n-1 data blocks x i1 ,x i2 ,…x in-1 It satisfies the Laplace distribution.

[0071] 3) Any non-empowered participant P i Send x ij To other non-empowering participants P j (i≠j);

[0072] 4) Any non-empowered participant P i Calculate Y i =x i -(x i1 +x i2 +…+x in-1 )+(x 1i +x 2i +…+x ni ), and publish Y i As a result, Y i P represents i Interference feature data, x 1i x represents P1 i , ..., x ni P represents n x i .

[0073] S204, the empowered participants, based on the differential privacy algorithm, integrate the interference feature data received from the non-empowered participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0074] For the embodiments in this specification, the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data; following the example of differential privacy described above, the enabling party can obtain Y after receiving the published Y from all non-enabling parties.

[0075] S206, empowering participants to construct a decision tree for the interference feature dataset using a node-based splitting method, wherein the process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node, each candidate splitting condition containing at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0076] It should be understood that Y obtained based on the differential privacy algorithm i Based on the computational characteristics of data statistics, for the embodiments described herein, the participants can initially conceive some feature data filtering rules for the target annotation results, such as whether the sample data features should be greater than a certain value, less than a certain value, or equal to a certain value. It should be noted that these initially conceived feature data filtering rules are not necessarily accurate, and the embodiments in this specification also require the use of node splitting to construct a decision tree for intelligent filtering.

[0077] This example uses a binary tree decision tree structure. (Reference) Figure 3 As shown, the root node of the decision tree is the initial original interference feature dataset S.

[0078] First, the splitting begins from the root node. The process involves using the feature data filtering rule corresponding to the interference feature dataset S as the splitting condition to divide the interference feature dataset S into child node 1 and child node 2. For example, if the feature data filtering rule is that the interference feature data value is less than 20, then the part of the interference feature dataset S with a value less than 50 is assigned to node 1, and the part with a value not less than 50 is assigned to node 2.

[0079] Then, following the same logic, child node 1 is further split. The process includes dividing the interference feature dataset s1 into sub-nodes 11 and 12 according to the feature data filtering rules corresponding to the interference feature dataset s1 of child node 1. It should be understood that the feature data filtering rules corresponding to the interference feature dataset s1 are a further refinement of the feature data filtering rules corresponding to the original interference feature dataset S. For example, if the feature data filtering rule for the original interference feature dataset S is that the interference feature data value should be less than 50, then the feature data filtering rule for the interference feature dataset s1 is that the interference feature data value should be less than a threshold of 50.

[0080] As can be seen, each node in the decision tree is essentially the result of splitting the interference feature data according to different levels of interference feature data filtering rules. If the split is reasonable, the interference feature data root of that part is suitable for the model to predict the target labeling result, and the corresponding interference feature data filtering rule can be used as the sample selection rule for the target labeling result.

[0081] Based on the above decision tree design, this embodiment can use purity and constraint fitness as the splitting criteria of the decision tree to ensure that the purity is maximized so that the left and right branch samples have high consistency after splitting, while maximizing the constraint fitness so that the left and right branch samples meet the constraint conditions. Thus, in each node, the best feature data screening rules are selected for splitting.

[0082] Here, we define s as a feature, t as the feature split point, and D as... L and D r Let φ(D) and φ(D) represent the left and right child nodes after the split, respectively. L ) and φ(D R ) represent the purity functions of the parent node and the left and right child nodes, respectively. In the embodiments of this specification, the change in purity of the simulated splitting result corresponding to each alternative splitting condition is determined based on the purity function of the target node and the purity functions of the child nodes obtained after the simulated splitting based on the alternative splitting condition. That is, the change in purity of the node before and after can be expressed as:

[0083]

[0084] In addition, the constraint fitness index also needs to be considered; constraint fitness can be understood as the degree to which the splitting result conforms to the preset data constraints; the preset data constraints can be flexibly selected according to actual needs. As an example, the preset data constraints in the embodiments of this specification include: the minimum proportion of the interference feature data of each splitting result's child node whose difference between the prediction result corresponding to the preset model and the target annotation result is less than a preset threshold, that is, using an existing preset model in this application scenario to evaluate the interference feature data of the child node corresponding to the splitting result; if the prediction result given by the preset model is closer to the target annotation result, it indicates that the splitting of the node at that position is more reasonable, that is, this paper uses the accuracy index of the preset model as the preset data constraint, i.e., ψ(D)=min(0,Precision(D)-), where θ represents the accuracy constraint threshold. The constraint fitness brought about by the splitting behavior is defined as:

[0085] Ψ(s,t)=max(ψ(D L ),(D R ))

[0086] The splitting criteria for a decision tree need to consider both purity variation and constraint fitness. Here, for each alternative splitting condition, a weighted sum of the purity variation and constraint fitness of the simulated splitting results can be performed to obtain the comprehensive score ρ(s,t) for each alternative splitting condition:

[0087] ρ(s,t)=Φ(s,t)+Ψ(s,t)

[0088] Based on the above principles, we assume that each feature data filtering rule corresponding to a node is a candidate splitting condition. When splitting the target node, we can try to use each candidate splitting condition corresponding to the target node to perform a simulated split and determine the comprehensive score ρ(s,t) of the simulated splitting result corresponding to each candidate splitting condition. Then, we select the one with the highest comprehensive score ρ(s,t) as the target candidate splitting condition for actual classification until the decision tree is constructed.

[0089] S208, the empowering participant determines the sample selection rules for the target labeling results based on the target candidate splitting conditions corresponding to at least one node in the decision tree, and feeds back the sample selection rules to the non-empowering participant.

[0090] It should be noted that the implementation of this manual requires the use of feature data filtering rules selected from one or more layers of nodes in the decision tree during the actual split to construct the sample selection rules for the target annotation results.

[0091] The following describes a solution where the components are assembled by machine.

[0092] Specifically, the embodiments of this specification can optimize the sample coverage of the combination results of the target candidate splitting conditions corresponding to each section in the decision tree, and thus determine the sample selection rules of the target labeling results based on the optimized solution of the combination results of the target candidate splitting conditions.

[0093] Here, to simplify the computation, the fitness constraint mentioned above can be introduced as a constraint in the optimization process. Under this constraint, a greedy algorithm is further used to maximize the coverage of the combination of target candidate splitting conditions for each section of the decision tree, thereby selecting the optimal combination of target candidate splitting conditions. It should be understood that each selected target candidate splitting condition is a feature data selection rule based on the target labeled results. These feature data selection rules combined together constitute the sample selection rules needed to train the model in federated learning.

[0094] In the process of maximizing coverage, the following definition is made:

[0095] Formula for maximizing the number of covered samples

[0096] Constraint formula based on constraint fitness

[0097] Where ij represents the rule index and the sample index, respectively, and p ij ∈{0,1} represents rule r i For sample s j The result of the label rule, l j ∈{0,1} is the sample s j Training labels.

[0098] Based on a greedy algorithm, we can first determine a matching... The target candidate splitting condition; after determining a target candidate splitting condition, based on this target candidate splitting condition, it is combined with other target candidate splitting conditions to determine the conditions that meet the requirements. The other, and so on until Once the maximum value is reached, there is no longer a need to select alternative splitting conditions.

[0099] It should be understood that the final determined target candidate splitting conditions are the optimal solution for maximizing coverage.

[0100] For the purposes of this specification, the feature data selection rules in the optimal solution of the target candidate splitting conditions can be used as the sample selection rules for the target annotation results.

[0101] S210, the non-empowered participant selects training samples from the local unlabeled samples to be used as target labeling results based on the received sample selection rules.

[0102] Based on the method described in this specification, each non-empowered participant in federated learning can de-identify the feature data of the samples of the local target labeling results using a differential privacy algorithm, and then upload the de-identified interference feature data to the empowered participant. Subsequently, the empowered participant summarizes the interference feature data and constructs a decision tree for the interference feature dataset based on node splitting. The splitting of any target node in the decision tree is constrained by statistically significant changes in splitting purity and constraint fitness, selecting the best feature data selection rule, without relying on sensitive information in the original data. Finally, the empowered participant constructs a sample selection rule for determining the target labeling results based on the feature data selection rule used in the actual splitting of at least one node in the decision tree, and provides it to the non-empowered participants. It can be seen that the scheme in this specification, while ensuring data privacy, can combine the data characteristics of the samples from each participant in federated learning to analyze a unified selection rule for training samples, thereby helping each participant to select training samples with consistent training objectives from their own private data based on the selection rule.

[0103] The above is an exemplary description of the methods described in the embodiments of this specification. In practical applications, the enabling parties may also provide their own interference feature data, which will not be specifically limited here.

[0104] In addition, for the Figure 2 The training sample selection method is shown in this specification. An embodiment of this specification also provides a training sample selection device. Figure 4 This is a schematic diagram of a multi-party training sample device as described in this specification, including:

[0105] The first sending module 410, based on a differential privacy algorithm, generates interference feature data of samples of local target annotation results and sends the interference feature data to the enabling participants of the federated learning; wherein, the enabling participants, based on the differential privacy algorithm, integrate the interference feature data received from non-enabling participants into an interference feature dataset, and the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data; and constructs a decision tree for the interference feature dataset based on node splitting, the process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node to simulate splitting the target node, each candidate splitting condition containing at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the best comprehensive score.

[0106] The first receiving module 420 receives the sample selection rule of the target labeling result sent by the empowering participant, wherein the sample selection rule is determined by the empowering participant based on the target candidate splitting condition corresponding to at least one node in the decision tree.

[0107] The filtering module 430 selects training samples from the local unlabeled samples to be used as the target labeling result, based on the received sample selection rules.

[0108] Optionally, the first sending module 410 is specifically used for: summing the feature data of the samples of the local target annotation result to obtain a summation result; splitting the locally obtained summation result into multiple data blocks conforming to the prior distribution of the differential privacy algorithm; sending the multiple data blocks obtained locally to other non-empowering participants in the federated learning, and receiving the summation result and corresponding multiple data blocks provided by other non-empowering participants; and calculating the interference feature data of the samples of the local target annotation result based on the locally obtained summation data and corresponding multiple data blocks, as well as the summation data and corresponding multiple data blocks provided by other non-empowering participants.

[0109] It should be understood that the processing device of the federated learning device in the embodiments of this specification is as... Figure 2 The method shown is executed by a non-empowering participant, thus enabling [the implementation of the method]. Figure 2 The steps and functions of the method shown are as follows.

[0110] also, Figure 5 This is a schematic diagram of another multi-party training sample device in this specification, including:

[0111] The second receiving module 510 receives interference feature data of the target annotation result sample sent by the non-empowered participant in the federated learning, wherein the interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0112] The integration module 520, based on the differential privacy algorithm, integrates the received interference feature data provided by the non-empowered participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0113] The decision tree construction module 530 constructs a decision tree for the interference feature dataset based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0114] The second sending module 540 determines the sample selection rule for the target labeling result based on the target candidate splitting condition corresponding to at least one node in the decision tree, and feeds back the sample selection rule to the non-empowered participant. The non-empowered participant selects training samples to be used as the target labeling result from the local unlabeled samples based on the received sample selection rule.

[0115] Optionally, the second sending module 540 determines the sample selection rules for the target labeling results based on the target candidate splitting conditions used for the actual splitting of at least one node in the decision tree. This includes: optimizing the sample coverage of the combination results of the target candidate splitting conditions corresponding to each section of the decision tree; and then determining the sample selection rules for the target labeling results based on the combination results of the optimized solution corresponding to the target candidate splitting conditions. Specifically, optimizing the sample coverage of the combination results of the target candidate splitting conditions corresponding to each section of the decision tree can be achieved by: maximizing the coverage of the combination results of the target candidate splitting conditions corresponding to each section of the decision tree under the constraint of the fitness constraint.

[0116] Optionally, the preset data constraints include:

[0117] The minimum percentage of interference feature data of child nodes in each simulated splitting result where the difference between the prediction result corresponding to the preset model and the target annotation result is less than a preset threshold.

[0118] It should be understood that the processing device of the federated learning device in the embodiments of this specification is as... Figure 2 The method shown empowers the execution entity of the participants, thus enabling [the implementation of the method]. Figure 2 The steps and functions of the method shown are as follows.

[0119] Figure 6 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this specification. Please refer to it. Figure 6 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0120] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0121] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0122] The processor reads the corresponding computer program from non-volatile memory into main memory and then executes it, forming the aforementioned federated learning device at the logical level. The processor executes the program stored in memory and specifically performs the following operations:

[0123] Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of the federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0124] The sample selection rule for receiving target annotation results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree.

[0125] Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples.

[0126] Alternatively, the processor reads the corresponding computer program from non-volatile memory into main memory and then runs it, forming the aforementioned federated learning device at the logical level. The processor executes the program stored in memory and specifically performs the following operations:

[0127] The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0128] Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0129] A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node, where each candidate splitting condition contains at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0130] Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules.

[0131] The above is as described in this instruction manual. Figure 1The methods disclosed in the illustrated embodiments can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in one or more embodiments of this specification. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in one or more embodiments of this specification can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0132] The electronic device can also perform Figure 1 The methods are not described in detail here.

[0133] Of course, in addition to software implementation, the electronic device described in this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0134] This specification also provides an embodiment of a computer-readable storage medium that stores one or more programs.

[0135] The program includes instructions that, when executed by a portable electronic device comprising multiple applications, enable the portable electronic device to perform... Figure 2 The method of the illustrated embodiment is specifically used to perform the following operations:

[0136] Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of the federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0137] The sample selection rule for receiving target annotation results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree.

[0138] Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples.

[0139] Alternatively, the one or more programs include instructions that, when executed by a portable electronic device including multiple applications, enable the portable electronic device to perform... Figure 2 The method of the illustrated embodiment is specifically used to perform the following operations:

[0140] The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm.

[0141] Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data.

[0142] A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node, where each candidate splitting condition contains at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to the preset data constraint conditions, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting result corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score.

[0143] Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules.

[0144] In summary, the above description is merely a preferred embodiment of this specification and is not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of one or more embodiments of this specification.

[0145] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0146] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0147] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0148] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

Claims

1. A training sample selection method involving multiple parties, comprising: The non-enabling participants in the federated learning process generate interference feature data of the local target annotation results based on the differential privacy algorithm, and send the interference feature data to the enabling participants in the federated learning process. The enabling participant integrates the interference feature data received from the non-enabling participant into an interference feature dataset based on the differential privacy algorithm. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding noise-reduced feature data. The empowering participants construct the decision tree for the interference feature dataset using a node splitting method. The process of splitting any target node includes: determining each alternative splitting condition corresponding to the target node and simulating splitting the target node. Each alternative splitting condition contains at least one feature data filtering rule under the target annotation result. The constraint fitness of each simulated splitting result relative to preset data constraints and the splitting purity variation of each simulated splitting result relative to the interference feature dataset are determined. The constraint fitness is determined based on the difference between the prediction results of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation results. The splitting purity variation is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting. Based on the splitting purity variation and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition, a comprehensive score corresponding to each candidate splitting condition is determined. The target node is actually split based on the target candidate splitting condition with the optimal comprehensive score. The enabling participant determines the sample selection rule for the target labeling result based on the target candidate splitting condition corresponding to at least one node in the decision tree, and feeds back the sample selection rule to the non-enabling participant; The non-empowered participant selects training samples from the local unlabeled samples based on the received sample selection rules, and performs federated learning training based on the training samples.

2. The method according to claim 1, The enabling participant determines the sample selection rules for the target labeling results based on the target candidate splitting conditions adopted for the actual splitting corresponding to at least one node in the decision tree, including: The enabling participants optimize the sample coverage of the combination results of the target candidate splitting conditions corresponding to each section in the decision tree. The enabling participants will determine the sample selection rules for the target labeling results based on the combination results of the target candidate splitting conditions corresponding to the optimization solution.

3. The method according to claim 2, The enabling participants optimize the sample coverage of the combination results of the target candidate splitting conditions corresponding to each section in the decision tree, including: The empowering participants, under the constraint of the fitness constraint, solve for the maximum coverage of the combination results of the target candidate splitting conditions corresponding to each section in the decision tree.

4. The method according to any one of claims 1 to 3, The preset data constraints include: The minimum percentage of interference feature data of child nodes in each simulated splitting result where the difference between the prediction result corresponding to the preset model and the target annotation result is less than a preset threshold.

5. The method according to any one of claims 1 to 3, The non-empowered participants in federated learning generate perturbation feature data for samples of local target annotation results based on differential privacy algorithms, including: The non-empowered participants in federated learning sum the feature data of the samples of the local target annotation results to obtain the summation result; The non-enabled participant splits the locally obtained summation result into multiple data blocks that conform to the prior distribution of the differential privacy algorithm; The non-empowered participants send the multiple data blocks obtained from the local split to other non-empowered participants in the federated learning, and receive the summation results and corresponding multiple data blocks provided by the other non-empowered participants; The non-empowered participant calculates the interference feature data of the local target annotation result sample based on the summation data and corresponding multiple data blocks obtained locally, as well as the summation data and corresponding multiple data blocks provided by other non-empowered participants.

6. The method according to any one of claims 1 to 3, Based on the degree of change in splitting purity and the constraint fitness of the simulated splitting results corresponding to each alternative splitting condition, a comprehensive score is determined for each alternative splitting condition, including: For each alternative splitting condition, the weighted sum of the degree of change in splitting purity and the constraint fitness of the simulated splitting results is obtained to obtain the comprehensive score corresponding to each alternative splitting condition.

7. The method according to any one of claims 1 to 3, The target labeling results include user samples locally labeled as blacklisted users or user samples locally labeled as risky users.

8. A federated learning device, comprising: The first sending module, based on a differential privacy algorithm, generates interference feature data of samples from the local target annotation results and sends the interference feature data to the enabling participants in the federated learning process. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data from non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each alternative splitting condition corresponding to the target node and simulating splitting the target node; each alternative splitting condition contains at least one feature data from the target annotation results. The selection rules include: determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset. The constraint fitness is determined based on the difference between the prediction results of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation results. The degree of change in splitting purity is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting. Based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition, a comprehensive score is determined for each candidate splitting condition. The target node is actually split based on the target candidate splitting condition with the best comprehensive score. The first receiving module receives the sample selection rule of the target labeling result sent by the empowering participant, wherein the sample selection rule is determined by the empowering participant based on the target candidate splitting condition corresponding to at least one node in the decision tree; The filtering module selects training samples from the local unlabeled samples based on the received sample selection rules, so as to perform federated learning training based on the training samples.

9. An electronic device, comprising: processor; and memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the following operations: Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data provided by non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each alternative splitting condition corresponding to the target node and simulating splitting the target node; each alternative splitting condition contains at least one feature data filtering rule under the target annotation result. The constraint fitness of each simulated splitting result relative to preset data constraints and the splitting purity variation of each simulated splitting result relative to the interference feature dataset are determined. The constraint fitness is determined based on the difference between the prediction results of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation results. The splitting purity variation is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting. Based on the splitting purity variation and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition, a comprehensive score corresponding to each candidate splitting condition is determined. The target node is actually split based on the target candidate splitting condition with the optimal comprehensive score. The sample selection rule for receiving target annotation results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree. Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples, so as to perform federated learning training based on the training samples.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store computer-executable instructions, which, when executed by a processor, perform the following operations: Based on the differential privacy algorithm, interference feature data of samples of local target annotation results is generated and sent to the enabling participants of federated learning. The enabling participants, based on the differential privacy algorithm, integrate the received interference feature data provided by non-enabling participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each alternative splitting condition corresponding to the target node and simulating splitting the target node; each alternative splitting condition contains at least one feature data filtering rule under the target annotation result. The constraint fitness of each simulated splitting result relative to preset data constraints and the splitting purity variation of each simulated splitting result relative to the interference feature dataset are determined. The constraint fitness is determined based on the difference between the prediction results of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation results. The splitting purity variation is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting. Based on the splitting purity variation and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition, a comprehensive score corresponding to each candidate splitting condition is determined. The target node is actually split based on the target candidate splitting condition with the optimal comprehensive score. The sample selection rule for receiving the target labeling results sent by the enabling participant, wherein the sample selection rule is determined by the enabling participant based on the target candidate splitting conditions corresponding to at least one node in the decision tree; Based on the received sample selection rules, training samples to be used as the target annotation results are selected from the local unlabeled samples, so as to perform federated learning training based on the training samples.

11. A federated learning device, comprising: The second receiving module receives interference feature data of the target annotation result sample sent by the non-empowered participant in federated learning. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm. The integration module, based on the differential privacy algorithm, integrates the interference feature data provided by the non-empowered participants into an interference feature dataset. The summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. The decision tree construction module constructs a decision tree for the interference feature dataset based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; the constraint fitness is determined based on the difference between the prediction result of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation result; the degree of change in splitting purity is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score. The second sending module determines the sample selection rule for the target labeling result based on the target candidate splitting condition corresponding to at least one node in the decision tree, and feeds back the sample selection rule to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling result based on the received sample selection rule, so as to perform federated learning training based on the training samples.

12. An electronic device, comprising: processor; and memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the following operations: The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm. Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset, and the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; the constraint fitness is determined based on the difference between the prediction result of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation result; the degree of change in splitting purity is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score. Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules, so as to perform federated learning training based on the training samples.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store computer-executable instructions, which, when executed by a processor, perform the following operations: The interference feature data of the sample of the target annotation result sent by the non-empowered participant in the federated learning is received. The interference feature data sent by the non-empowered participant is locally generated by the non-empowered participant based on the differential privacy algorithm. Based on the differential privacy algorithm, the interference feature data received from the non-empowered participants are integrated into an interference feature dataset, and the summation result of the interference feature data in the interference feature dataset is consistent with the summation result of the corresponding denoised feature data. A decision tree for the interference feature dataset is constructed based on node splitting. The process of splitting any target node includes: determining each candidate splitting condition corresponding to the target node and simulating splitting the target node; each candidate splitting condition includes at least one feature data filtering rule under the target annotation result; determining the constraint fitness of each simulated splitting result relative to preset data constraints, and the degree of change in splitting purity of each simulated splitting result relative to the interference feature dataset; the constraint fitness is determined based on the difference between the prediction result of the interference feature data of the child nodes obtained after simulated splitting on the preset model and the target annotation result; the degree of change in splitting purity is determined based on the difference between the purity function of the target node before splitting and the purity function of the child nodes obtained after simulated splitting; determining the comprehensive score corresponding to each candidate splitting condition based on the degree of change in splitting purity and constraint fitness of the simulated splitting results corresponding to each candidate splitting condition; and performing actual splitting of the target node based on the target candidate splitting condition with the optimal comprehensive score. Based on the target candidate splitting conditions corresponding to at least one node in the decision tree, the sample selection rules for the target labeling results are determined, and the sample selection rules are fed back to the non-empowered participant. The non-empowered participant selects training samples from the local unlabeled samples to be used as the target labeling results based on the received sample selection rules, so as to perform federated learning training based on the training samples.

Images