Participant selection method based on model precision estimation in longitudinal federated learning

By using encrypted mutual information and sample intersection number in vertical federated learning to estimate the model performance, combined with PCA dimensionality reduction and feature discretization, fit the model accuracy prediction function, and select participants through greedy strategies, participation in vertical federated learning is solved. The issue of user selection and privacy protection improves model performance and privacy security.

CN119026671AActive Publication Date: 2024-11-26NANJING UNIV OF AERONAUTICS & ASTRONAUTICS +1

Patent Information

Application Number
CN202411520613.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-29
Publication Date
2024-11-26
Estimated Expiration
2044-10-29

AI Technical Summary

Technical Problem

In vertical federated learning, how to efficiently select the right participants for machine learning model training and assign prediction tasks, ensure model performance and protect data privacy, especially if existing contribution measurement methods cannot estimate participants before vertical federated learning begins Contribute and there is a risk of privacy leakage.

Method used

The participant selection method based on model accuracy estimate is adopted to estimate the model performance by encrypting mutual information and the number of sample intersections, and to realize privacy calculations using combined feature value coding, combining PCA dimensionality reduction and feature discretization to fit the model accuracy estimate function. and select participants through greedy strategies to maximize the performance benefits of test samples.

Benefits of technology

This enables quantification of participants' contribution to model performance before vertical federated learning begins, ensuring privacy protection, improving model inference performance on new samples, and reducing training costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026671A_ABST
    Figure CN119026671A_ABST
Patent Text Reader

Abstract

The invention discloses a participant selection method based on model precision estimation in longitudinal federated learning, and the method comprises the steps: estimating the performance of a longitudinal federated learning model based on a longitudinal federated learning framework according to encrypted mutual information and the number of sample intersections, and obtaining an optimal estimation function of model precision through fitting; realizing a privacy calculation mutual information algorithm by using combined characteristic value coding, and embedding the privacy calculation mutual information algorithm into an encrypted sample alignment process of longitudinal federated learning through PCA dimensionality reduction and characteristic discretization means to realize a privacy protection task; the method also comprises a step of realizing a participant selection algorithm based on a greedy strategy so as to maximize the performance benefit of a test sample, thereby completing participant distribution of a model prediction task. Besides, a participant selection strategy based on a greedy algorithm is realized through a fitted estimation precision function, a model prediction task is efficiently allocated to each participant in a reasoning stage, and the performance benefit of joint training is maximized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to federated learning technology, and specifically relates to a participant selection technology in vertical federated learning, and in particular to a participant selection method based on model accuracy estimation in vertical federated learning. Background Art

[0002] Traditional AI model training usually requires a large amount of data. However, these data are often distributed in different institutions and regions. Due to the restrictions of privacy protection laws and other regulations, they cannot be directly concentrated on a centralized server for processing. Federated learning, a privacy-preserving distributed machine learning method, can collaborate with multiple participants to jointly train models without leaving the local device. It can not only use the data of each participant for model training, but also protect data privacy well. Different from the horizontal federated learning scenario where the data sets of each participant have the same feature space and different sample spaces, vertical federated learning is committed to allowing each participant who shares the same sample space and different feature spaces to collaborate in training by utilizing the dispersed features of their aligned samples. It has important research value in the fields of financial credit, medical health, etc. In the vertical federated learning model, due to the different distribution of data features, it is crucial to efficiently select appropriate participants for machine learning model training and assign prediction tasks, because this will directly affect the performance of the final trained joint model. In order to maximize the reasoning performance of the federated model on new samples, the participant selection problem of vertical federated learning faces the following challenges:

[0003] 1) In vertical federated learning, each participant contributes to the federated machine learning model by sharing its features with the federated task publisher. In practical applications, there are many feature dimensions held by participants that are not highly relevant to the training tasks or even harmful to the model performance. How to quantify the contribution of each potential participant in vertical federated learning to the final model performance to help the task publisher screen participants and thus improve the efficiency of training is a major challenge. This is also of great significance for saving task publishers and training costs and promoting healthy and sustainable federated learning.

[0004] 2) Currently, most of the calculation methods for model contribution and profit distribution in the context of federated machine learning are based on the assumption that the participants are certain. They study how to evaluate the contribution of each participant after the vertical federated learning training is completed. However, in real application scenarios, the task publisher needs to pay a reward each time it cooperates with the participant. It is obviously unreasonable to evaluate the contribution of the participant after the cooperation is over and then make a choice. A method is needed to estimate the contribution of the participants before the vertical federated training begins.

[0005] 3) Developing a highly adaptive participant selection mechanism for vertical federated learning and finding a balance between contribution measurement and data privacy is not easy for the following reasons:

[0006] a. Model performance is highly random and is affected by many factors, such as the number of public samples, feature importance, task difficulty, and the order in which training samples are selected. It is difficult to use a single indicator to directly estimate to what extent each participant can improve a specific model performance, which makes problem modeling more complicated.

[0007] b. Since federated learning requires that data must remain locally owned by the data owner, task publishers cannot access data, and important factors such as the importance of features to training tasks are difficult to obtain. In addition, in the vertical federation scenario, it is usually assumed that the participants are honest but curious, and each participant will try to infer as much specific content as possible from the encrypted information obtained by other participants. Existing contribution measurement methods will inevitably expose the feature values ​​of each participant, leading to privacy leakage. How to obtain the interactive information of each participant while ensuring privacy security is also a huge challenge.

[0008] c. After the training of the vertical federated learning model is completed, in order to obtain the best reasoning performance on more new samples, the task publisher needs to select appropriate participants according to its own budget to cooperate in the corresponding reasoning tasks, which in turn requires further solving new constrained optimization problems based on the original participant selection results. Summary of the invention

[0009] Purpose of the invention: In response to the problem of participant selection in vertical federated learning existing in the above-mentioned existing work, the present invention provides a participant selection method based on model accuracy estimation in vertical federated learning, which is a participant selection method based on model accuracy estimation with privacy protection. This method can take effect before the start of vertical federated training.

[0010] Technical solution: A participant selection method based on model accuracy estimation in vertical federated learning. The method is based on the vertical federated learning framework, estimates the performance of the vertical federated learning model according to the encrypted mutual information and the number of sample intersections, and fits the best estimation function of the model accuracy. The privacy calculation mutual information algorithm is implemented using combined eigenvalue coding, and is embedded into the encrypted sample alignment process of vertical federated learning through PCA dimensionality reduction and feature discretization to achieve privacy protection tasks.

[0011] The method also includes implementing a participant selection algorithm based on a greedy strategy to maximize the test sample performance benefit, thereby completing the participant allocation of the model prediction task;

[0012] The method comprises the following steps:

[0013] (1) Conduct correlation analysis on the number of training samples and the mutual information between features and labels, including training the two on public datasets to verify the positive correlation between the accuracy of the longitudinal federated learning model and the number of training samples and mutual information;

[0014] (2) Based on the analysis results in step (1), a model accuracy estimation function is built, and the data of the task publisher is split and combined to simulate the longitudinal federated learning process, and the corresponding data points representing the relationship between mutual information and model accuracy and the relationship between the number of training samples and the estimation error are obtained; then a linear function is used to fit the relationship between model accuracy and mutual information. , and then use the quadratic function to further fit the relationship between the number of training samples and the accuracy estimation error , correct and fine-tune the precision estimation error based on mutual information, and then fit the joint effect of mutual information and the number of training samples on the accuracy of the longitudinal federated learning model;

[0015] The model accuracy estimation function is fitted using the following formula:

[0016]

[0017] Represents the model accuracy estimated based on two factors: mutual information and the number of training samples;

[0018] (3) Reduce the actual performance A and estimated performance The residual between is used to correct the fitting error of the model performance prediction function, and the steps include:

[0019] (31) Calculate the absolute error of each sample and get The error is divided into intervals based on the calculated value, and the number of errors in each interval is counted. Finally, the number of errors in each interval is divided by the total number to calculate the probability distribution.

[0020] (32) Repeat the model accuracy fitting operation in step (2) based on the data points obtained from the real federated training, and also obtain a probability distribution of the fitting error;

[0021] (33) After each round of the real longitudinal federation training process, the estimation error is calculated to obtain an increasingly accurate error mean offset err, and then the estimation performance is corrected according to the following formula so that the average estimation error is close to 0:

[0022]

[0023] In the formula, This is the final model accuracy estimation function, which is fitted by the task publisher based on its own data and part of the historical training data before the start of the vertical federated training;

[0024] (4) Privacy calculation involves the mutual information between the features involved in training and the task labels. The calculation process includes:

[0025] (41) The task publisher and federated participant k each discretize the features of each dimension, divide the data range of the feature value of each feature dimension into b equally spaced intervals, numbered 1-b, and for each floating-point feature, calculate the interval number in which it falls according to its feature value, and then use the corresponding number value to replace the continuous floating-point value as the discretized feature value;

[0026] (42) Based on the privacy intersection technology of the Diffie-Hellman encryption algorithm, the task publisher and the federation participant k can align their encrypted samples;

[0027] In step (42), the task publisher and participant k first encrypt their own sample IDs with their own private keys and send them to each other. Then participant k encrypts the sample ID encrypted by the task publisher with its own private key again and sends it to the task publisher. Based on the Diffie-Hellman encryption algorithm, a symmetric encryption algorithm is used for encryption and decryption. The task publisher only needs to compare the sample ID of participant k encrypted with the private keys of e and d and the sample ID of the task publisher encrypted with the private keys of d and e to see if they are consistent.

[0028] (43) During the encrypted sample alignment process, a combined eigenvalue encoding is transmitted to calculate the mutual information of features and labels shared by the privacy computing participants and the task publisher. The combined eigenvalue encoding calculation includes:

[0029] Suppose the dataset of federated participants is reduced by PCA. dimensional features, the eigenvalues ​​are discretized and scattered to In the interval, the mapping function relationship between the discretized value and the original eigenvalue is: , assuming the sample record format is:

[0030] [ id , f 1 , f 2 ,..., f d k ]

[0031] The combined encoding of the feature values ​​of the i-th aligned sample of participant k is expressed as follows:

[0032]

[0033] It can be concluded that for any combination of feature values, the corresponding combined feature value code is uniquely determined;

[0034] (5) The task publisher combines the feature value code sent by the participant with its own feature value code and label column to calculate the following formula: The value of Represents the number of data samples under a given variable value, so as to further calculate the mutual information estimate of the joint feature and label:

[0035] ;

[0036] (6) Based on the task publisher’s training set D and test set T, the goal is converted to maximizing the expected model performance of all samples in the test set T, including:

[0037] According to the model accuracy prediction function and the distribution of estimation errors obtained in step (3), the expected model maximization problem is re-modeled as follows:

[0038] max ∑ i ∈ T E [ U i ] s . t . U i = max { k | i ∈ T k & k ∈ S ∪ { O } } Q k Q k = f ( I k , n k ) + e k , k ∈ S e k ~ N ( μ , σ 2 ) , k ∈ S n k ≤ | D ∩ D k | T k = T ∩ D k ∑ k n k ≤ B

[0039] Among them, the task publisher is O, its training budget is B, the set of participants is S, and for each participant k in S, its sample set is The intersection with the test set of the task publisher is , The samples in have both k and the feature dimensions of the task publisher. For these samples, the model trained by the task publisher and participant k can be used. Perform inference predictions, The estimated performance is ,variable is the decision variable, corresponding to the number of training samples brought by participant k. The larger the value, the better the estimated model performance. The larger the value, the more quota the participant k's sample will occupy in the budget B, and the best inference performance obtained by the test sample i in the test set T is the best among all The maximum value in ;

[0040] Follow steps (1) to (5) to calculate the model that each participant k trains in collaboration with the task publisher The estimated accuracy , all participants are divided into groups according to the estimated accuracy Sort from high to low, if any The accuracy is lower than the model trained by the task publisher using only its own feature dimension data. If the accuracy of the participant is lower than B, it will be removed from the selection range. Then, starting from the participant with the highest estimated accuracy, we will try to purchase training data from them one by one until the cumulative amount of purchased data exceeds the budget limit B or there are no more participants to choose from. Finally, based on the purchase situation, we will inform the selected participants of the amount of training data they need to provide and carry out vertical federated learning training.

[0041] (7) Consider the problem that after the participants are selected using the greedy strategy in step (6) before training, there are several participants and several longitudinal federated learning models will be obtained. In the model prediction and inference stage, it is necessary to perform the best model task allocation. The prediction task of the corresponding test set samples is assigned to each participant according to the priority of the participants. Then, the prediction results of the test data are analyzed and compared with the true labels. The proportion of samples correctly classified by the model to the total number of samples is calculated to obtain the prediction accuracy of the longitudinal federated learning model.

[0042] Specifically, the prediction accuracy of the longitudinal federated learning model described in this method is calculated according to the following formula:

[0043]

[0044] Among them, TP represents the number of samples correctly predicted by the model as positive, TN represents the number of samples correctly predicted by the model as negative, FP represents the number of samples incorrectly predicted by the model as negative, and FN represents the number of samples incorrectly predicted by the model as positive.

[0045] Furthermore, the correlation analysis process of the number of training samples in step (1) includes training them using three machine learning models, including random forest, logistic regression and support vector machine, and then recording the relationship between the number of training samples and the model prediction accuracy, and verifying that under the respective model and data set configurations, the model accuracy and the number of training samples are significantly positively correlated;

[0046] The correlation analysis process of mutual information in this step includes:

[0047] The mutual information MI of two random variables is a measure of the interdependence between the variables. Mutual information measures the information shared by two random variables X and Y. The standard definition of mutual information is as follows:

[0048]

[0049] Assume that the feature set held by the task publisher of vertical federated learning is X, and the feature set held by federated participant k is , then the mutual information between all features and labels for federated learning can be expressed as , the mutual information can be measured and The mutual dependence between them, according to the definition of mutual information, The calculation formula is as follows:

[0050]

[0051] in and y represents and possible values ​​of Y, yes The probability of satisfying the It is the KL divergence between the two distributions. Since the true distribution of the data is difficult to obtain, we consider using the maximum likelihood estimation method to approximate them based on the feature distribution and label distribution in the training set, as shown in the following formula:

[0052]

[0053] In the formula, It is estimated based on the training set The distribution of is the number of data samples under a given variable value, is the total number of public data samples between participant k and the task publisher;

[0054] For continuous features, they are converted into discrete ones, and then the probability distribution is calculated by maximum likelihood estimation. or The dimension of includes converting it into a low-dimensional vector using PCA dimensionality reduction;

[0055] The higher the mutual information between features and labels, the higher the contribution of the features to predicting labels, which helps improve the reasoning performance of the model, that is, the accuracy of label prediction.

[0056] Furthermore, in step (2), the relationship between model performance and mutual information is first fitted, and then fine-tuned using the number of training samples, expressed as:

[0057] Assume that task publisher T has a total of z dimensions of features, and task publisher T chooses to retain a feature combination containing z' features , the rest The characteristics are used as the characteristics of the federated participants, and then the task publisher Randomly select an integer in the range , and then from the remaining Randomly select features including The feature combination of ; then use As a feature set of the task publisher, As a feature set of federated participants, by dividing the feature dimension data of the task publisher, the task publisher can simulate and perform longitudinal federated learning alone and obtain the performance of the model before cooperating with other participants.

[0058] Furthermore, in step (4), the model performance estimation function is applied to the actual longitudinal federated training process, which requires obtaining the mutual information and the number of aligned training samples. The number of samples is naturally obtained through data interaction, and the calculation of the mutual information needs to be privacy-processed.

[0059] In the process of mutual information between the features and task labels involved in privacy computing training, it includes dimensionality reduction processing of the data feature dimensions of the participants or task publishers;

[0060] The dimensionality reduction processing refers to PCA dimensionality reduction, and processing is performed on data with feature dimensions exceeding three dimensions to reduce data transmission overhead and calculation overhead.

[0061] Furthermore, in step (7), assuming that there are several federated participants who have different sample intersections with the task publisher T, in the model inference phase, the common samples between participant k and the task publisher can use the model trained by both of them. ; For samples shared by some participants and task publishers, the estimated performance is selected A relatively better model is used to make predictions; for some samples that are only held by the task publisher, the model M0 trained by the task publisher itself is used to make predictions.

[0062] Step (7) includes using the model accuracy estimation function to substitute the encrypted mutual information and the number of sample intersections to calculate the accuracy estimation value of the longitudinal federated learning model obtained by each participant in cooperation with the task publisher, and assigning corresponding priorities to each participant by sorting the estimated values ​​of the model accuracy, wherein the participant with a greater model estimation accuracy has a higher priority;

[0063] For each sample in the task publisher's test set, the system checks which participants have records related to the sample and assigns prediction tasks based on priority; the specific steps are as follows:

[0064] For each sample in the task publisher's test set, the system checks which participants have records related to the sample and assigns prediction tasks based on priority;

[0065] For each sample ID in the test set, the system will check whether each participant has a record corresponding to the sample ID in order of priority. If a participant has a record corresponding to the sample ID, the system will mark the participant as "1", otherwise it will be marked as "0";

[0066] If only one participant has a record of the sample, the task publisher will choose to use the model trained in collaboration with this participant to predict the sample;

[0067] If multiple participants have records corresponding to the sample ID, the model trained in collaboration with the participant with the highest priority is selected to be responsible for the prediction task of the sample.

[0068] Beneficial effects: Compared with the prior art, the substantial progress and significant effects of the present invention are as follows:

[0069] 1) This paper designs a simulation experiment of vertical federated learning and explores the factors that affect the accuracy of the vertical federated learning model, including feature importance and the number of samples that can be aligned. Inspired by the mutual information selection method in feature engineering, the mutual information value is used to reflect the feature importance, and the joint impact of the above two factors on the performance of the vertical federated model is modeled. The model accuracy estimation function is fitted and the estimation function is corrected based on the error probability distribution.

[0070] 2) The contribution measurement methods that have been proposed are not suitable for use before the start of longitudinal federated training, and most of the measurement criteria are irrelevant to the training tasks. The present invention innovatively proposes to split the feature data of the task publisher itself and use the results of historical training data to fit the low-error model accuracy estimation function, thereby achieving a preliminary quantitative estimate of the contribution of participants to the improvement of model performance in longitudinal federated learning for specific training tasks.

[0071] 3) The method described in the present invention discretizes the continuous eigenvalues ​​after PCA dimensionality reduction and encodes them according to the designed rules, completing the privacy calculation of mutual information under the data encryption state. This method can not only be efficiently embedded in the sample alignment process of vertical federated training, but also take into account the data transmission efficiency and computational complexity during the training process.

[0072] 4) The present invention proposes a participant selection strategy that is applicable when multiple participants can improve model performance. By allocating model prediction tasks based on a greedy algorithm, the model performance expected by the task publisher can be maximized. Experimental results show that the test set accuracy obtained by performing longitudinal federation prediction according to the participant selection strategy proposed by the present invention is generally higher than the result of single federation model prediction without selection. In addition, the change of participant priority will also affect the effect of the model, which further illustrates the effectiveness of the strategy of prioritizing participants according to the estimated accuracy calculated by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a schematic diagram of the vertical federated learning model training transaction service framework in the present invention;

[0074] Figure 2 The present invention shows a fitting example diagram of the relationship between mutual information and model accuracy;

[0075] Figure 3 It is a schematic diagram of the probability distribution of the error of the fitting function;

[0076] Figure 4 This is an image of the final model prediction function obtained based on the breast cancer dataset;

[0077] Figure 5 It is an example diagram of the model prediction task allocation strategy described in the present invention. DETAILED DESCRIPTION

[0078] In order to explain the technical solution disclosed in the present invention in detail, further elaboration is made below in conjunction with the accompanying drawings and specific embodiments of the specification.

[0079] For vertical federated learning in the scenario of collaborative training models between multiple organizations, assuming that there are K federated participants who can collaborate with the federated task publisher to train a machine learning model through the vertical federated learning mechanism, they first need to align encrypted entities with the task publisher, find common data samples, and then jointly train a vertical federated model based on richer feature information while ensuring that the original data is not leaked. The vertical federated learning setting of the present invention includes a task publisher and a federated participant.

[0080] The task publisher is an organization that hopes to build a better-performing model by combining more relevant features of other participants through vertical federated learning. Its goal is to improve the prediction accuracy of the trained model for all new customer samples. The task publisher itself holds a certain amount of feature dimension information and sample label information, and its sample ID overlaps with that of the federated participants.

[0081] The federated participants do not have label information, but they have feature dimension information that the task publisher does not have. They overlap with the sample ID of the task publisher and can provide model training services for the task publisher. The federated participants cooperate with the task publisher to perform vertical federated training, which not only provides the task publisher with their own data value, but also consumes their own computing resources and communication bandwidth, so the task publisher is often required to pay them equal compensation.

[0082] In fact, not all participants can bring significant performance improvements to the model. For task publishers, since the profits brought by model performance improvement are limited, it is very important to screen out suitable federated participants before the start of vertical federated training. To address this problem, the present invention provides a low-error participant performance evaluation method, which preferentially selects participants with excellent performance from K participants to conduct two-way vertical federated training with the task publisher, ensuring that the prediction accuracy of the obtained model on new samples is higher than that of the model trained by the task publisher itself, and designs a suitable participant selection strategy to assign prediction tasks to suitable participants to cover more test cases and further improve the accuracy of the model on new samples.

[0083] In the method described in the present invention, the method is a participant selection algorithm based on model accuracy estimation suitable for vertical federated learning. Unlike related technologies such as the existing participant contribution measurement method in federated learning, this method can quantify the potential contribution of each participant to the performance of the joint model before the start of vertical federated training, and at the same time take into account the privacy protection task of vertical federated learning. In order to maximize the expected model performance obtained by all samples in the test set of the task publisher, the method described in the present invention includes constructing a joint prediction function of the encrypted mutual information and the number of intersections of training samples for the model performance, and implementing the participant selection algorithm based on the greedy strategy to complete the participant allocation of the model prediction task.

[0084] Furthermore, combined with Figure 1 As shown, the specific implementation steps of the technical solution provided by the present invention can be described as follows:

[0085] Step 1: Conduct correlation analysis on factors that affect the performance of the vertical federated learning model.

[0086] This step focuses on analyzing the number of training samples and the mutual information between the features and labels involved in the training as factors affecting model performance.

[0087] (11) Correlation analysis of the number of training samples.

[0088] In vertical federated learning training, only aligned samples can be used as training sets for subsequent model training. When the number of sample intersections between a federated participant and the task publisher is small, the model is likely to produce a higher variance, lack of data diversity, and be more prone to overfitting, resulting in poor generalization performance of the model on test data.

[0089] The present invention conducts experiments on three machine learning models, namely Random Forest Classification (RFC), Logistic Regression (LR) and Support Vector Classification (SVC), on two classic public data, the breast cancer dataset and the adult income dataset, and records the relationship between the number of training samples and the model prediction accuracy. It verifies that under the respective model and dataset configurations, the model accuracy is significantly positively correlated with the number of training samples. Therefore, the number of public samples of the participants and the task publisher is an important reference indicator for estimating the accuracy of the longitudinal federated model.

[0090] (12) Correlation analysis of mutual information.

[0091] The mutual information (MI) of two random variables is a measure of the interdependence between the variables. Mutual information measures the information shared by two random variables X and Y, expressed as The standard definition of mutual information is as follows:

[0092]

[0093] Assume that the feature set held by the task publisher of vertical federated learning is X, and the feature set held by federated participant k is , then the mutual information between all features and labels for federated learning can be expressed as , the mutual information can be measured and The mutual dependence between them, according to the definition of mutual information, The calculation formula is as follows:

[0094]

[0095] in and y represents and possible values ​​of Y, yes The probability of satisfying the It is the KL divergence between the two distributions. Since the true distribution of the data is difficult to obtain, we consider using the maximum likelihood estimation method to approximate them based on the feature distribution and label distribution in the training set, as shown in the following formula:

[0096]

[0097] In the formula It is estimated based on the training set The distribution of is the number of data samples under a given variable value. is the total number of public data samples between participant k and the task publisher. When some features are continuous, they are converted into discrete ones, and then the maximum likelihood estimation is applied to calculate the probability distribution. If or The dimension is high and the amount of calculation is large, so PCA dimensionality reduction can be performed to convert it into a low-dimensional vector.

[0098] The higher the mutual information between features and labels, the higher the contribution of the features to predicting labels, which helps improve the reasoning performance of the model (i.e., label prediction accuracy).

[0099] According to our experimental results, under the above three machine learning models, as the mutual information value MI increases, the model accuracy increases rapidly. Therefore, the mutual information value between the features and labels shared by the participants and the task publisher is also an important reference indicator for estimating the accuracy of the vertical federation model.

[0100] Step 2: Fitting of model accuracy estimation function.

[0101] Fitting the model accuracy estimation function, it can be concluded from the correlation analysis results of step 1 that the impact of mutual information on model performance is more significant than the impact of the number of training samples on model performance. Therefore, this step first fits the relationship between model performance and mutual information, and then uses the number of training samples as a factor for fine-tuning.

[0102] Assume that the task publisher T has a total of z dimensions of features, and the task publisher T can retain z' features ( ) , the rest As the characteristics of the federation participants, the task publisher Randomly select an integer in the range , and then from the remaining Randomly select features including The feature combination of .use As a feature set of the task publisher, As a feature set of federated participants, by dividing the feature dimension data of the task publisher, the task publisher can simulate and perform longitudinal federated learning alone and obtain the performance of the model before cooperating with other participants.

[0103] This embodiment uses the breast cancer dataset and the adult income dataset to simulate a real longitudinal federated learning scenario. Taking the adult income dataset as an example, the dataset has a total of 14 dimensions of features. The features of 8 dimensions are randomly selected as the features of the task publisher, and the remaining 6 features are optional features for each federated participant. The task publisher randomly extracts m features from his 8 dimensions of features. For the same m (m<8), there will be multiple feature combinations. Then, n combinations are sampled from these feature combinations to calculate their mutual information with the labels and the model accuracy obtained by training only with these feature dimensions. Note that these are the results that the task publisher can obtain by training only with his own data. Figure 2 The circular dots are used to mark the data points. The present invention uses these simulated data points to preliminarily fit the relationship between mutual information and model performance. In order to verify the effect of fitting these data points using linear and quadratic functions, feature combinations with different numbers of features are randomly selected from the remaining 6 features, and they are combined with the 8-dimensional features of the task publisher to calculate the mutual information between the labels and test the accuracy of the jointly trained model. Using these data, we can draw Figure 2 The triangle points in the figure are fitted with quadratic functions to obtain the corresponding line "quadratic function fitting true training accuracy". One of the goals of this step is to hope that the line fitted by the circle points can be as close as possible to the trend of the line fitted by the triangle points. After repeated experiments, it was found that the relationship between the mutual information and model performance of the linear function fitting is better than that of the quadratic function fitting in most cases. For the breast cancer dataset, the fitting function is 0.1263 x + 0.8664, and for the adult income dataset, the fitting function is: 0.1808 x + 0.7977.

[0104] Then, the relationship between the number of training samples and the estimation error of model performance based on mutual information is modeled. In order to further use the number of training set samples as a factor to fine-tune the fitting effect, the experiment sets the ratio of training samples involved in calculating mutual information to all samples: {0.2, 0.4, 0.6, 0.8, 1}. The mutual information and the corresponding model accuracy are calculated once at each ratio. With the sampling rate as the independent variable and the difference (i.e., error) between the estimated accuracy predicted value of the fitted line of the dot and the actual accuracy corresponding to the dot as the dependent variable, a quadratic function is fitted. For the breast cancer data set, the fitting function is y=-0.04061 x2 + 0.07673 x - 0.02817, and for the adult income data set, the fitting function is: y=-0.01432 x2+ 0.02825 x - 0.01064. Adding the estimated error fitting amount based on the sampling rate to the function based on mutual information fitting can correct a more accurate model performance, namely: .

[0105] Step 3: Correction of the fitting error of the model performance prediction function.

[0106] There will still be a certain error between the model performance estimated by the model accuracy prediction function obtained in step 2 and the actual performance. The reason for this error is that on the one hand, there may be other influencing factors besides mutual information and the number of training samples. On the other hand, the method of calculating mutual information using the feature discretization method itself has a certain error. In order to adjust the average estimation error to a state close to 0, the present invention proposes to calculate the absolute error of each sample, that is, to substitute the mutual information of the dots fitted in step 2 and the number of training samples into the fitting function, calculate the error between the fitting accuracy and the accuracy of the dot marking, divide this error into a certain interval, and calculate the probability distribution of the error, corresponding to Figure 3 In the same way, the error between the estimated result of the triangle point fitting function in step 2 and the actual accuracy is calculated to obtain the probability distribution, corresponding to Figure 3 The solid line in the figure is “Err of F^r for VFL”. Figure 3 It can be seen that the means of the above two error probability distributions are very close to 0, indicating that the fitting method proposed in the present invention is effective and reasonable. Next, the estimation function based on the task publisher's data fitting (i.e., the estimation function fitting the circle point data) is used in the real federated training, that is, the mutual information and sample number corresponding to the triangle points are substituted into the model accuracy estimation function, and the error is calculated by comparing the estimated accuracy with the actual accuracy of the triangle point data. Figure 3 In the error probability distribution image "Err of F^s for VFL", it should be emphasized that this is the error distribution when the estimation function is used for real longitudinal federated training. For the breast cancer dataset, the mean of this error distribution is about 0.0128, and for the adult income dataset, the mean of this error distribution is about 0.008. Note that the shapes of these error probability distributions are very similar, and can be regarded as normal distributions with similar variances but different means. Therefore, we only need to correct "Err of F^s forSVFL" to make the mean of its error distribution as close as possible to "Err of F^s for VFL" to achieve the correction of the estimation function in step 2.

[0107] The present invention obtains an increasingly accurate mean offset by calculating and recording the estimated error after each new longitudinal federated training process. In practical applications, there may not be enough historical longitudinal federated training results for further error correction. At this time, based on the idea of ​​transfer learning, the task publisher can use his own data to simulate longitudinal federated training and obtain the performance results of the model. Since it is much easier to obtain the offset of the estimated error than to directly fit the model performance prediction function, the number of pre-trained longitudinal federated training results required is much smaller. Figure 3 The "Corrected Err of F^s" in the figure reflects the effect of the error correction. It can be seen that according to the method proposed by the present invention, the corrected error distribution image is closer to the real error distribution image. In addition, from the perspective of the participant selection task, it is only necessary to estimate the relative advantages and disadvantages between the participants. The slight deviation of the overall estimated error value will not greatly affect the final result of the participant's selection.

[0108] Step 4: Implement private computation of mutual information between features and labels of federated participants.

[0109] Applying the model performance estimation function to the actual longitudinal federated training process requires obtaining the mutual information and the number of aligned training samples. The number of samples can be naturally obtained through data interaction, while the calculation of mutual information requires certain privacy processing.

[0110] It is further pointed out that according to step 1 From the calculation formula, we can see that the calculation of mutual information depends on , How to obtain the original data of federated participants without access to them? The value of is an important problem to be solved. This paper proposes a mutual information calculation method with high computational efficiency and privacy protection. The algorithm can be naturally embedded in the encrypted sample alignment process in vertical federated learning. The specific process of the algorithm is introduced below.

[0111] In this embodiment, first, in order to reduce the data transmission overhead and calculation overhead, the federation participants and task publishers need to perform PCA dimensionality reduction on their own data sets in advance to convert high-dimensional features into low-dimensional features. Next, the task publisher and the federation participants each discretize the features of each dimension after the dimensionality reduction. The present invention uses the discretization method of equal-width binning to divide the features of each dimension into b boxes. Specifically, the minimum value "min_val" and the maximum value "max_val" of each feature dimension data are first calculated, and then the data range of the feature dimension is divided into b equidistant intervals, then the width of each interval is "dis=(max_val-min_val) / 5". For each floating-point type feature, the minimum value min_val is subtracted from the feature value, and then divided by the interval dis of the dimension feature, and the result is rounded down to obtain the interval number (that is, it falls in which interval), and this number value is used instead of the continuous floating-point value as the discretized feature value.

[0112] Then the task publisher and federation participant k can start the normal encrypted sample alignment process. The difference is that in addition to encrypting and transmitting their own sample IDs, they also need to transmit a combined feature value encoding, which is used to calculate the mutual information between the features and labels shared by the two. After the values ​​of each feature dimension are mapped to their respective discretization intervals, the discretized feature combination of each aligned sample is encoded using the following calculation formula:

[0113]

[0114] in, is the mapping function relationship between the discretized value and the original eigenvalue, is the number of eigenvalue discretization intervals, and the eigenvector of a sample is recorded as , where each feature The code in its corresponding discrete interval is . The encoding process can be regarded as a weighted summation. The eigenvalues ​​of features of different dimensions have different corresponding weights, and the weights are based on the total number of features and the range of values. Through this encoding method, any combination of eigenvalues ​​can be mapped to a unique encoding value, thereby avoiding confusion during the processing process and facilitating further mutual information calculations. After the task publisher obtains this code, it cannot accurately know the actual eigenvalues ​​because it does not know how many discrete intervals the federal participants have divided the eigenvalues ​​into when discretizing the eigenvalues. Even if it knows, it can only deduce the discretized interval number. Therefore, it can achieve private calculation of mutual information. In order to further improve privacy, this encoded value can be mapped to a function again and sent to the task publisher, so that the task publisher cannot infer the original data. The task publisher combines the eigenvalue encoding sent by the participants with its own eigenvalue encoding and label column to calculate The value of is used to further calculate the mutual information between the joint features and the labels.

[0115] Step 5: Embed the mutual information privacy calculation into the encrypted sample alignment process of vertical federated learning.

[0116] a) Assuming that at this time, federated participant k wants to align samples with the task publisher, participant k needs to first hash the ID of each sample he holds locally to obtain the hash value The task publisher also hashes the ID of each sample he holds to obtain the hash value .

[0117] b) The task publisher uses the encryption key to encrypt the hashed sample ID and sends the encrypted ID to participant k. After receiving the encrypted ID from the task publisher, participant k further encrypts it and encodes the sample’s feature value combination It is sent back to the task publisher together with the encrypted ID.

[0118] c) The task publisher finds matching sample pairs by comparing the received encrypted IDs and calculates the number of shared samples. , thus realizing the encrypted sample alignment process of vertical federated learning.

[0119] d) For each pair of matching samples, the task publisher encodes the combined feature value of the matching samples and the combined feature value of its own data And the corresponding labels are counted, the frequency of occurrence is recorded, and the required mutual information is obtained as well as The value of The feature encoding is and The number of samples with label y, .

[0120] e) Finally, the task publisher calculates the mutual information based on the statistical results .

[0121] Through this algorithm, the task publisher can calculate the mutual information between feature combinations and labels without leaking the participants' private data. This method ensures data security while obtaining valuable statistical information.

[0122] Step 6: Implementation of the participant selection algorithm based on model accuracy estimation.

[0123] Before starting longitudinal federated learning training, the task publisher needs to select cooperative participants based on the model accuracy estimation results obtained in the previous five steps and determine the amount of sample data that each participant needs to provide to maximize the performance of the expected model within the budget limit.

[0124] The task publisher first needs to train a model on his own local data And test the accuracy on the test set Assuming that the set of participants is S and the total budget of training samples is B, the task publisher needs to first calculate the model obtained by cooperating with each participant k The estimated accuracy of , then follow Sort the priorities of the participants from high to low to get the sorted participant set .

[0125] The participant selection algorithm initializes two empty sets to store the selected participants and the training samples they provide. Then, the algorithm traverses the sorted participant sets in turn. , for each participant k, if its accuracy Greater than or equal to , the algorithm will calculate the number of samples needed from the participant based on the greedy strategy. As long as the number of samples held by the participant is within the remaining budget of the total number of training samples of the task publisher, the training data will be purchased first, and the participant and the corresponding number of training samples will be added to the corresponding set. If the total number of training samples exceeds the budget limit B after obtaining the samples, the algorithm will stop selecting more participants and terminate the loop. If the accuracy of the current participant is Less than threshold , the algorithm will skip the participant and continue to check the next participant. When the algorithm has traversed all participants or reached the target number of samples B, the algorithm returns the final selected participant set and the corresponding training sample set.

[0126] Combination Figure 4 The image of the final model prediction function based on the breast cancer dataset is shown. Through this step, the algorithm ensures that only those participants with high accuracy and significant sample contribution are selected, while avoiding exceeding the budget, thus achieving a balance between the optimal sample quantity and quality.

[0127] Step 7: Participants in the model inference phase predict task allocation.

[0128] After selecting participants using the greedy strategy described in step 6 before training, since selecting multiple participants will result in multiple longitudinal federated models, optimal model task allocation is required in the model prediction and reasoning stage. Figure 5 For example, we assume that there are two federated participants P1 and P2, which have different sample intersections with the task publisher T. In the model reasoning stage, the common samples between participants P1 and T can use the model M1 trained by P1 and T, and the common samples between participants P2 and T can use the model M2 trained by P2 and T. For samples shared by P1, P2 and T, the one with better estimated performance in M1 and M2 is selected for prediction. For samples that are not shared by P1 and P2 but only by T, the model M0 trained by T itself is used for prediction. Note that the reasoning performance of M1 and M2 should be better than that of M0, otherwise the task publisher will not select them to jointly train the model.

[0129] Specifically, after using the model accuracy estimation function to substitute the encrypted mutual information and the number of sample intersections to calculate the accuracy estimate of the longitudinal federated learning model obtained by each participant in cooperation with the task publisher, the corresponding priority can be assigned to each participant by sorting the estimated values ​​of the model accuracy, where the participant with the greater model estimation accuracy has a higher priority. For each sample in the task publisher's test set, the system will check which participants have records related to the sample and assign prediction tasks based on priority. The specific steps are as follows:

[0130] a) For each sample ID in the test set, the system checks each participant (in order of priority) to see if they have a record corresponding to that sample ID. If a participant has a record corresponding to that sample ID, the system marks that participant as "1", otherwise it marks it as "0".

[0131] b) If only one participant has a record for the sample, then this participant will be responsible for predicting the sample. Even if this participant is not the highest priority, because other participants do not have relevant records, this participant automatically takes on this task. If multiple participants have records corresponding to the sample ID, the system will select the participant with the highest priority to be responsible for the prediction task of the sample according to the priority order.

[0132] Finally, the system assigns a specific participant to perform the prediction task for each sample based on the above steps. This ensures that the task is assigned to the participant with the highest priority and relevant records as much as possible.

[0133] After the task publisher cooperates with each participant to carry out model prediction according to the assigned task, the task publisher records the label prediction results of all its test samples, and counts the predicted labels and true labels to obtain the confusion matrix of the classification model, which is a tool for evaluating the performance of the classification model. In this matrix, there are four key concepts:

[0134] TP (True Positives): This is the number of samples that the model correctly predicts as positive, that is, the samples where both the model prediction result and the actual label are positive.

[0135] TN (True Negatives): This is the number of samples that the model correctly predicts as negative, that is, samples where both the model prediction result and the actual label are negative.

[0136] FP (False Positives): This is the number of examples where the model incorrectly predicted the negative class as positive.

[0137] FN (False Negatives): This is the number of examples where the model incorrectly predicted the positive class as negative.

[0138] The prediction accuracy of the final longitudinal federated learning model is calculated according to the following formula:

[0139]

[0140] Based on the accuracy result, the effectiveness of the participant selection algorithm based on model accuracy estimation and the prediction task allocation strategy proposed in the present invention can be verified.

Claims

1. A participant selection method based on model accuracy estimation in vertical federated learning, characterized in that: The method is based on the framework of vertical federated learning. It estimates the performance of the vertical federated learning model according to the encrypted mutual information and the number of sample intersections, and fits the best estimation function for the model accuracy. It uses combined eigenvalue coding to implement the privacy calculation mutual information algorithm, and embeds it into the encrypted sample alignment process of vertical federated learning through PCA dimensionality reduction and feature discretization to achieve privacy protection tasks. The method also includes implementing a participant selection algorithm based on a greedy strategy to maximize the test sample performance benefit, thereby completing the participant allocation of the model prediction task; The method specifically comprises the following steps: (1) Conduct correlation analysis on the number of training samples and the mutual information between features and labels, including training the two on public datasets to verify the positive correlation between the accuracy of the longitudinal federated learning model and the number of training samples and mutual information; (2) Based on the analysis results in step (1), a model accuracy estimation function is constructed, and the data of the task publisher is split and combined to simulate the longitudinal federated learning process, and the corresponding data points representing the relationship between mutual information and model accuracy and the relationship between the number of training samples and the estimated error are obtained; then a linear function is used to fit the relationship between model accuracy and mutual information I , and then use the quadratic function to further fit the relationship between the number of training samples n and the accuracy estimation error , correct and fine-tune the precision estimation error based on mutual information, and then fit the joint effect of mutual information and the number of training samples on the accuracy of the longitudinal federated learning model; The model accuracy estimation function is fitted using the following formula: ; Represents the model accuracy estimated by the mutual information I between all training features and task labels and the number of training samples n; (3) Reduce the actual performance A and the estimated performance calculated by the model accuracy estimation function in step (2) The residual between is used to correct the fitting error of the model performance prediction function. The steps include: (31) Calculate the absolute error of each sample and get The error is divided into intervals based on the calculated value, and the number of errors in each interval is counted. Finally, the number of errors in each interval is divided by the total number to calculate the probability distribution. (32) Repeat the model accuracy fitting operation in step (2) based on the data points obtained from the real federated training, and also obtain a probability distribution of the fitting error; (33) After each round of the real longitudinal federation training process, the estimation error is calculated to obtain an increasingly accurate error mean offset err, and then the estimation performance is corrected according to the following formula so that the average estimation error is close to 0: ; In the formula, This is the final model accuracy estimation function, which is fitted by the task publisher based on its own data and part of the historical training data before the start of the vertical federated training; (4) Privacy calculation involves the mutual information between the features involved in training and the task labels. The calculation process includes: (41) The task publisher and federated participant k each discretize the features of each dimension, divide the data range of the feature value of each feature dimension into b equally spaced intervals, numbered 1-b, and for each floating-point feature, calculate the interval number in which it falls according to its feature value, and then use the corresponding number value to replace the continuous floating-point value as the discretized feature value; (42) Based on the privacy intersection technology of the Diffie-Hellman encryption algorithm, the task publisher and the federation participant k can align their encrypted samples; (43) During the encrypted sample alignment process, a combined eigenvalue encoding is transmitted to calculate the mutual information of features and labels shared by the privacy computing participants and the task publisher. The combined eigenvalue encoding calculation includes: Suppose the dataset of federated participants is reduced by PCA. dimensional features, the eigenvalues ​​are discretized and scattered to In the interval, the mapping function relationship between the discretized value and the original eigenvalue is: , assuming the sample record format is: ; In the formula, - Indicates samples 1 to The eigenvalues ​​of the dimensional features, the combined encoding of the eigenvalues ​​of the i-th aligned sample of participant k is expressed as follows: ; It can be concluded that for any combination of feature values, the corresponding combined feature value code is uniquely determined; (5) The task publisher combines the feature value code sent by the participant with its own feature value code and label column to calculate the following formula: The value of Represents the number of data samples under a given variable value, so as to further calculate the mutual information estimate of the joint feature and label: ; For each pair of matching samples, the task publisher encodes the combined feature values ​​of the matching samples. Combined feature value encoding with its own data And the corresponding label y is counted, the frequency of occurrence is recorded, and the required calculation of mutual information is obtained as well as The value of The feature encoding is and The number of samples with label y, ; (6) Based on the task publisher’s training set D and test set T, the goal is converted to maximizing the expected model performance of all samples in the test set T , including: According to the model accuracy prediction function and the distribution of estimation errors obtained in step (3), the expected model maximization problem is re-modeled as follows: , Among them, the task publisher is O, its training budget is B, the set of participants is S, and for each participant k in S, its sample set is The intersection with the test set of the task publisher is , The samples in have both k and the feature dimensions of the task publisher. For these samples, the model trained by the task publisher and participant k can be used. Perform inference predictions, The estimated performance is ,variable is the decision variable, corresponding to the number of training samples brought by participant k. The larger the value, the better the estimated model performance. The larger the value, the more quota the participant k's sample will occupy in the budget B, and the best inference performance obtained by the test sample i in the test set T is the best among all The maximum value in ; Follow steps (1) to (5) to calculate the model that each participant k trains in collaboration with the task publisher The estimated accuracy , all participants are divided into groups according to the estimated accuracy Sort from high to low, if there is a model The accuracy is lower than the model trained by the task publisher using only its own feature dimension data. If the accuracy of the participant is lower than B, it will be removed from the selection range. Then, starting from the participant with the highest estimated accuracy, we will try to purchase training data from them one by one until the cumulative amount of data purchased exceeds budget B or there are no more participants to choose from. Finally, based on the purchase situation, we will inform the selected participants of the amount of training data they need to provide and carry out vertical federated learning training. (7) Consider the problem that after the participants are selected using the greedy strategy in step (6) before training, there are several participants and several longitudinal federated learning models will be obtained. In the model prediction and inference stage, it is necessary to perform the best model task allocation. The prediction task of the corresponding test set samples is assigned to each participant according to the priority of the participants. Then, the prediction results of the test data are analyzed and compared with the true labels. The proportion of samples correctly classified by the model to the total number of samples is calculated to obtain the prediction accuracy of the longitudinal federated learning model.

2. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: Step (1) The correlation analysis process for the number of training samples includes training them using three machine learning models, including random forest, logistic regression and support vector machine, and then recording the relationship between the number of training samples and the model prediction accuracy, and verifying that under the respective model and data set configurations, the model accuracy and the number of training samples are significantly positively correlated; The correlation analysis process of mutual information in this step includes: The mutual information MI of two random variables is a measure of the interdependence between the variables. Mutual information measures the information shared by two random variables X and Y. The standard definition of mutual information is as follows: ; Assume that the feature set held by the task publisher of vertical federated learning is , federation participants The feature set held is , then the mutual information between all features and labels for federated learning can be expressed as , the mutual information can be measured and The mutual dependence between them, according to the definition of mutual information, The calculation formula is as follows: , in and y represents and possible values ​​of Y, yes The probability of satisfying the It is the KL divergence between the two distributions. Since the true distribution of the data is difficult to obtain, we consider using the maximum likelihood estimation method to approximate them based on the feature distribution and label distribution in the training set, as shown in the following formula: , In the formula, It is estimated based on the training set The distribution of is the number of data samples under a given variable value, is the total number of public data samples between participant k and the task publisher; For continuous features, they are converted into discrete ones, and then the probability distribution is calculated by maximum likelihood estimation. Or the dimension of X includes converting it into a low-dimensional vector using PCA dimensionality reduction; The higher the mutual information between features and labels, the higher the contribution of the features to predicting labels, which helps improve the reasoning performance of the model, that is, the accuracy of label prediction.

3. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: In step (2), the relationship between model performance and mutual information is first fitted, and then fine-tuned using the number of training samples, expressed as: Assume that task publisher T has a total of z dimensions of features, and task publisher T chooses to retain a feature combination containing z' features , the rest The characteristics are used as the characteristics of the federated participants, and then the task publisher Randomly select an integer in the range , and then from the remaining Randomly select features including The feature combination of ; then use As a feature set of the task publisher, As a feature set of federated participants, by dividing the feature dimension data of the task publisher, the task publisher can simulate and perform longitudinal federated learning alone and obtain the performance of the model before cooperating with other participants.

4. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: In step (4), the model performance estimation function is applied to the actual longitudinal federated training process, which requires obtaining the mutual information and the number of aligned training samples. The number of samples is naturally obtained through data interaction, and the calculation of mutual information needs to be privacy-processed.

5. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 4, characterized in that: In the process of mutual information between the features and task labels involved in privacy computing training, it includes dimensionality reduction processing of the data feature dimensions of the participants or task publishers; The dimensionality reduction processing refers to PCA dimensionality reduction, and processing is performed on data with feature dimensions exceeding three dimensions to reduce data transmission overhead and calculation overhead.

6. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: In step (42), the task publisher and participant k first encrypt their own sample IDs with their own private keys e and d and send them to each other. Then participant k encrypts the sample ID encrypted by the task publisher again with its own private key e and sends it to the task publisher. Based on the Diffie-Hellman encryption algorithm, a symmetric encryption algorithm is used for encryption and decryption. The task publisher only needs to compare the sample ID of participant k encrypted with the private keys of e and d and the sample ID of the task publisher encrypted with the private keys of d and e to see if they are consistent to obtain the sample intersection.

7. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: In step (7), it is assumed that there are several federated participants who have different sample intersections with the task publisher T. In the model inference phase, the common samples between participant k and the task publisher can use the model trained by both of them. ; For samples shared by some participants and task publishers, the estimated performance is selected A relatively better model is used to make predictions; for some samples that are only held by the task publisher, the model M0 trained by the task publisher itself is used to make predictions.

8. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1 or 7, characterized in that: Step (7) includes using the model accuracy estimation function to substitute the encrypted mutual information and the number of sample intersections to calculate the accuracy estimation value of the longitudinal federated learning model obtained by each participant in cooperation with the task publisher, and assigning corresponding priorities to each participant by sorting the estimated values ​​of the model accuracy, wherein the participant with a greater model estimation accuracy has a higher priority; For each sample in the task publisher's test set, the system checks which participants have records related to the sample and assigns prediction tasks based on priority; For each sample ID in the test set, the system will check whether each participant has a record corresponding to the sample ID in order of priority. If a participant has a record corresponding to the sample ID, the system will mark the participant as "1", otherwise it will be marked as "0"; If only one participant has a record of the sample, the task publisher will choose to use the model trained in collaboration with this participant to predict the sample; If multiple participants have records corresponding to the sample ID, the model trained in collaboration with the participant with the highest priority is selected to be responsible for the prediction task of the sample.

9. The method for selecting participants based on model accuracy estimation in vertical federated learning according to claim 1, characterized in that: The prediction accuracy of the longitudinal federated learning model described in this method is calculated according to the following formula: , Among them, TP represents the number of samples correctly predicted by the model as positive, TN represents the number of samples correctly predicted by the model as negative, FP represents the number of samples incorrectly predicted by the model as negative, and FN represents the number of samples incorrectly predicted by the model as positive.

Citation Information

Patent Citations

  • Training method and system of longitudinal federal learning model

    CN114330759A

  • Model training method and related equipment

    CN115630297A

  • Data enhancement method based on semi-supervised federated learning under privacy protection

    CN115829009A

  • Federal learning client contribution evaluation method based on significant score

    CN115905859A

  • Method for predicting indoor personal thermal comfort based on transfer learning

    CN118364949A

Cited By

  • Method for estimating contribution degree of participant model in longitudinal federated learning

    CN118734939A

  • A method for estimating the contribution of participant models in vertical federated learning

    CN118734939B

  • Electronic device and method for entity alignment in vertical federated learning

    WO2026138741A1