Federal learning optimization method for improving security collaborative modeling efficiency between banks
By introducing entropy weighting and feature dimensionality reduction into interbank federated learning, and optimizing feature fusion, the problems of data quality differences and security were solved, enabling efficient and secure model training.
Patent Information
- Application Number
- CN202511208693.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-27
- Publication Date
- 2026-01-09
AI Technical Summary
Existing technologies do not consider differences in data quality and feature correlations in interbank security collaborative modeling. Direct data fusion leads to slow model convergence, low accuracy, poor security, and low computational efficiency.
We employ entropy weighting to weight and fuse feature data, and optimize the interaction layer of federated learning through feature dimensionality reduction and homomorphic encryption. We then use the encrypted feature matrix for model training.
This improved the training efficiency and accuracy of the model, ensured data security, prevented data leakage, and enhanced the overall efficiency and security of the model.
Smart Images
Figure CN121302222A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and more specifically to a federated learning optimization method for improving the efficiency of interbank security collaborative modeling. Background Technology
[0002] Secure Multi-Party Computation (MPC): Allows multiple participants to perform computations collaboratively without disclosing their individual privacy data. Cryptographic protocols enable encrypted computation and data exchange, ensuring correct results while protecting data privacy. For example, in joint risk control scenarios within financial institutions, multiple banks can use secure multi-party computation to share data features for risk assessment without exposing their clients' sensitive information. Federated Learning: A distributed machine learning framework where participants retain data locally and collaborate on modeling by exchanging model parameters or intermediate results, achieving "data remains still, model moves, model privacy is protected." For example, different medical institutions can jointly train disease diagnostic models without sharing patients' original medical records. Trusted Execution Environment (TEE): Utilizes a secure area provided by hardware to execute computational tasks, ensuring data confidentiality and integrity. For example, Intel's SGX technology isolates sensitive computations within a trusted execution environment, preventing external access to private data, and can be used in cloud computing and other scenarios to protect user data privacy. Homomorphic Encryption: A special encryption technique that allows direct computation on encrypted data, with the decrypted result identical to the result of performing the same computation on the plaintext. This allows data owners to entrust encrypted data to others for computation without worrying about data leakage. Differential privacy: By adding noise or other methods to the data, the privacy of individual data is protected while maintaining its statistical properties. This can be used for privacy protection in scenarios such as data publishing and querying. Zero-knowledge proof: The prover can convince the verifier that a statement is true without providing any useful information. For example, in identity verification scenarios, a user can prove they know a password without actually revealing it.
[0003] Existing applications of secure and privacy-preserving computation in banking have the following problems: 1. When combining data from various parties at the interaction layer, the issue of different data quality from different data sources was not considered. Instead, weights were simply assigned based on quantity. When fusing data features from multiple institutions, the current approach is mostly direct fusion. For example, if institution A has 100,000 data features and institution B has only 100 data features, direct fusion after dimensionality balancing will result in slow model convergence and low model accuracy.
[0004] 2. When combining data, the direct addition method ignores the correlation between features. For example, if the data features of organization A and the features of organization B are internally related, direct addition will ignore the features.
[0005] 3. Directly using raw data for joint training has extremely poor security and is very likely to cause data leakage and other security and privacy issues.
[0006] 4. Encrypting the data before integration and training protects the data security, but the computational load increases exponentially, resulting in extremely low efficiency and making it unsuitable for industrial production. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a federated learning optimization method to improve the efficiency of interbank security collaborative modeling.
[0008] To achieve the above objectives, this invention provides a federated learning optimization method for improving the efficiency of interbank security collaborative modeling, comprising: Receive data input from each participant and extract feature data from the input data; The extracted feature data is encrypted to obtain the encrypted feature matrix; The encrypted feature matrices are concatenated to obtain the concatenated feature matrix; Perform feature dimensionality reduction on the concatenated feature matrix to obtain the dimensionality-reduced feature matrix; The entropy weighting method is used to determine the dimensionality-reduced feature matrix and then fused to obtain a data feature matrix with reasonable weights. The model is trained using the data feature matrix with appropriate weights, and the training is completed by calculating whether the loss function meets the preset threshold.
[0009] Furthermore, the specific method for concatenating the encrypted feature matrix is as follows: ; in, This represents the concatenated feature matrix, and [] denotes the encryption operation. This represents the j-th extracted feature matrix, where j = 1, 2, ..., n. This indicates matrix concatenation.
[0010] Furthermore, the specific method for performing feature dimensionality reduction on the concatenated feature matrix is as follows: ; Where z is the dimensionality-reduced feature matrix, and ReLU(.) is the linear rectified activation function. This represents the weight matrix used for dimensionality reduction. , For real numbers, Here, k represents the size of the concatenated features, and k represents the size of the dimension after dimensionality reduction. For bias.
[0011] Furthermore, the method for determining the dimensionality-reduced feature matrix using the entropy weight method and then fusing it is as follows: Standardize the attribute data in the dimensionality-reduced feature matrix; Calculate the information entropy of each attribute; Based on the information entropy of each attribute Calculate the weights of each feature; The data feature matrix with reasonable weights is obtained by calculating the feature matrix z after dimensionality reduction and the weights of each feature.
[0012] Furthermore, the standardization process for the attribute data of the dimensionality-reduced feature matrix is as follows: ; in, For standardized attribute data, Let i be the i-th attribute in the reduced-dimensional feature matrix, i = 1, 2, ..., k. This represents the minimum value of the attribute in the dimensionality-reduced feature matrix. This represents the maximum value of the attribute in the reduced-dimensional feature matrix.
[0013] Furthermore, the specific method for calculating the information entropy of each attribute is as follows: ; in, Let be the information entropy of the i-th attribute.
[0014] Furthermore, based on the information entropy of each attribute The specific method for calculating the weights of each feature is as follows: ; in, Let be the weight of the i-th feature.
[0015] Furthermore, the calculation method for the data feature matrix with reasonable weights is as follows: ; in, This is to calculate the data feature matrix with reasonable weights.
[0016] Beneficial effects: This invention introduces an entropy weighting method to allocate feature data according to different weights, and improves the additive method used in feature fusion. The above methods have the following advantages: 1. Using data federation and entropy weighting, the interaction layer of vertical federated learning is redesigned to combine the data features of all parties. 2. The entropy weight method is an objective method of assigning values. In practical application, the entropy weight of each data feature is calculated using information entropy, and then the weight of each data feature is adjusted using the entropy weight method. 3. An innovative approach of feature fusion followed by dimensionality reduction is adopted to improve the accuracy and efficiency of training models under secure and privacy-preserving computation. 4. Instead of using the original data for joint training, a homomorphic encryption algorithm is used to encrypt the data features, which ensures data security while also meeting the needs of model training. 5. The parties share encrypted feature data, rather than directly using the original data for training, which improves the overall efficiency of the model under secure and privacy-preserving computation. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating a federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to an embodiment of the present invention. Figure 2 This is a comparison chart of the training accuracy of this application and existing technologies; Figure 3 This is a comparison chart of the F1 scores of this application and the prior art; Figure 4 This is a comparison chart of the time consumed in each round between this application and existing technologies. Detailed Implementation
[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.
[0019] like Figure 1 As shown, this embodiment of the invention provides a federated learning optimization method to improve the efficiency of interbank security collaborative modeling, including: It receives data input from various participants and extracts feature data from the input data. Specifically, this can be represented as: ; in, Indicates the first Data input for each parameter, For feature data extracted from the data input, Let j represent the extracted j-th feature matrix, where j = 1, 2, ..., n.
[0020] The extracted feature data is encrypted to obtain an encrypted feature matrix. The encrypted feature matrix can be represented as follows: The brackets [] represent encryption operations. Paillier semi-homomorphic encryption can be used. The core content of this application is to optimize the splicing method when splicing features and to introduce the entropy weight method to optimize the fusion of data features from different institutions. The relevant encryption parts will not be elaborated.
[0021] The encrypted feature matrices are concatenated to obtain the concatenated feature matrix. The specific method for concatenating the encrypted feature matrices is as follows: ; in, This represents the concatenated feature matrix. This indicates matrix concatenation.
[0022] To address the issue of excessive feature dimensions in concatenated matrix features, this application performs dimensionality reduction on the concatenated feature matrices to obtain a dimensionality-reduced feature matrix. The specific method for dimensionality reduction of the concatenated feature matrix is as follows: ; Where z is the dimensionality-reduced feature matrix, and ReLU(.) is the linear rectified activation function. This represents the weight matrix used for dimensionality reduction. , For real numbers, Here, k represents the size of the concatenated features, and k represents the size of the dimension after dimensionality reduction. For bias.
[0023] The entropy weighting method is used to determine the dimensionality-reduced feature matrices for fusion, resulting in a data feature matrix with appropriate weights. The method for fusing the dimensionality-reduced feature matrices using the entropy weighting method is as follows: The attribute data in the dimensionality-reduced feature matrix is then standardized. The specific method for standardizing the attribute data of the dimensionality-reduced feature matrix is as follows: ; in, For standardized attribute data, Let i be the i-th attribute in the reduced-dimensional feature matrix, i = 1, 2, ..., k. This represents the minimum value of the attribute in the dimensionality-reduced feature matrix. This represents the maximum value of the attribute in the reduced-dimensional feature matrix.
[0024] Calculate the information entropy of each attribute. The specific method for calculating the information entropy of each attribute is as follows: ; in, Let be the information entropy of the i-th attribute.
[0025] Based on the information entropy of each attribute Calculate the weights of each feature. Based on the information entropy of each attribute. The specific method for calculating the weights of each feature is as follows: ; in, Let be the weight of the i-th feature.
[0026] Based on the dimensionality-reduced feature matrix z and the weights of each feature, a data feature matrix with appropriate weights is calculated. The specific calculation method for the data feature matrix with appropriate weights is as follows: ; in, This is to calculate the data feature matrix with reasonable weights.
[0027] Model training is performed using a data feature matrix with appropriate weights (the model is a specific deep learning model). Each institution should adopt a unified network structure and privacy protocol when training the model, and determine whether the model has been trained by calculating whether the loss function meets the preset threshold.
[0028] To verify the effectiveness of this invention, a traditional federated learning model and an optimized federated learning model were compared: (1) Dataset This method uses the Kaggle Portuguese Bank Marketing Dataset, which contains data related to a Portuguese bank's telemarketing campaign. The total sample size is 11,162, with 8,929 in the training set and 2,233 in the test set. The data is pre-split, with one side holding 4-dimensional basic attribute features and the other holding 12-dimensional banking transaction features; the corresponding labels are held only by one side. Variable types include various categories such as basic customer information (age, occupation, marital status, education level, etc.), financial information (credit default, mortgage status, personal loan status, etc.), and marketing-related information (contact type, month of contact, last contact date, last marketing campaign result, etc.). The target variable is whether the customer has subscribed to a fixed deposit, a binary categorical variable: "yes" indicates subscription, and "no" indicates no subscription.
[0029] (2) Experimental model design This method uses a heterogeneous logistic regression (Hetero-LR) model for training. Hetero-LR can utilize homomorphic encryption or secret sharing techniques to jointly train linear models without sharing the original features.
[0030] (3) To comprehensively evaluate model performance, accuracy and F1 score were used as evaluation metrics in the experiment. The calculation formulas are as follows: Accuracy = (TP + TN) / (TP + FN + FP + TN) F1 -score = (2 × TP) / (2 × TP + FP + FN) Among them, TP is the true positive sample, that is, the number of samples that are actually positive and are predicted as positive; FP is the false positive sample, that is, the number of samples that are actually negative and are predicted as positive; TN is the true negative sample, that is, the number of samples that are actually negative and are predicted as negative; FN is the false negative sample, that is, the number of samples that are actually positive and are predicted as negative. In order to verify the effectiveness of the federated learning method for improving the efficiency of interbank security collaborative modeling proposed in this invention, the following three sets of comparative experiments were conducted: (1) Training was performed using the original FATE of WeBank, with Acc_we representing its accuracy, F1_we representing the F1 value, and Time_we representing the average time consumed per round of training from model to convergence. The entropy weight method was used in the same way, but the splicing-dimensionality reduction fusion method proposed in this invention was not used when fusion. Similarly, Acc_s, F1_s, and Time_s represented the accuracy, F1 value, and average time consumed per round of training; finally, Acc_s+, F1_s+, The experiment, with a time limit of 10 seconds, was conducted under the following basic parameters: two participants, 14-dimensional features, batch size of 128, and each participant using an NVIDIA Tesla V100 graphics card and 32 GB of RAM. See details for further information. Figures 2 to 4 By comparing the average training time per round, it can be found that the optimized splicing method has the shortest time. The reason is that the splicing-dimensionality reduction method can significantly reduce the amount of related calculations. The entropy weight method slightly increases the amount of calculations, but it still improves the average time per round by 8 seconds compared to the original FATE method.
[0031] For example, the method and implementation steps for applying this invention in a financial scenario are as follows: (1) Joint modeling of micro and small enterprise risk control models between credit reporting agencies and banks; The specific implementation steps are as follows: 1. Data preparation stage Banks collect data such as loan applications, repayment records, and financial statements from micro and small enterprises; credit reporting agencies compile data such as credit reports and default records from micro and small enterprises. Each institution uses data cleaning algorithms locally to remove duplicate and invalid data, processes missing values using methods such as mean imputation and multiple imputation, and identifies and corrects outliers.
[0032] Banks extract loan-related characteristics such as loan amount, loan term, and repayment method; credit reporting agencies extract credit characteristics such as corporate credit score and number of overdue payments.
[0033] 2. Model Training Phase Before training, algorithm engineers need to design relevant risk control models, extract features from the feature data obtained in the previous step, encrypt them to obtain feature matrices, use the splicing-entropy weight fusion-dimensionality reduction proposed in this invention to finally obtain fused feature data, use the feature data to train the model, and at the same time, backpropagate the training results of the fundamental rounds to update the underlying model feature data extraction device. 3. Verification Phase Offline verification: Banks and credit reporting agencies use locally stored test datasets to calculate the model's accuracy, recall, AUC, and other metrics, and compare the data with the model finally obtained by training the model using only local features. 4. Online deployment The jointly trained risk control model is deployed into the bank's loan approval system to conduct risk assessments on newly applying micro and small enterprises. Continuous monitoring is conducted for 3-6 months, and the deviation between the actual default rate and the predicted default rate is statistically analyzed. If the deviation is within an acceptable range (e.g., no more than 5%), the effectiveness of the model in actual business operations is verified.
[0034] (2) Implementation steps of the marketing conversion rate prediction model between third-party e-commerce platforms and banks 1. Data and Feature Integration Stage Banks collect customer account information, transaction records, and other data; e-commerce platforms compile customer shopping preferences, spending amounts, and other data; social media platforms acquire customer social relationships, interests, and other data. Each institution extracts features from the data locally; for example, banks extract features such as customer deposit balances and transaction frequency, while e-commerce platforms extract features such as customer purchase categories and purchase frequency.
[0035] 2. Model Training Phase The model was jointly determined by the marketing team and algorithm engineers, and the training steps were the same as those for the micro-risk control model mentioned above.
[0036] 3. Verification Phase Each institution uses historical data to divide the training set and the test set, and calculates the mean squared error, accuracy and other indicators of the model.
[0037] 4. Online deployment In the actual marketing campaign, customers were randomly divided into Group A and Group B. Group A received marketing strategies based on recommended products and marketing copy generated by a jointly trained marketing conversion rate prediction model; Group B used traditional marketing methods. After the campaign, the conversion rates and purchase amounts of the two groups were compared.
[0038] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A federated learning optimization method to improve the efficiency of interbank security collaborative modeling, characterized in that, include: Receive data input from each participant and extract feature data from the input data; The extracted feature data is encrypted to obtain the encrypted feature matrix; The encrypted feature matrices are concatenated to obtain the concatenated feature matrix; Perform feature dimensionality reduction on the concatenated feature matrix to obtain the dimensionality-reduced feature matrix; The entropy weighting method is used to determine the dimensionality-reduced feature matrix and then fused to obtain a data feature matrix with reasonable weights. The model is trained using the data feature matrix with appropriate weights, and the training is completed by calculating whether the loss function meets the preset threshold.
2. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 1, characterized in that, The specific method for concatenating the encrypted feature matrix is as follows: ； in, This represents the concatenated feature matrix, and [] denotes the encryption operation. This represents the j-th extracted feature matrix, where j = 1, 2, ..., n. This indicates matrix concatenation.
3. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 2, characterized in that, The specific method for performing feature dimensionality reduction on the concatenated feature matrix is as follows: ； Where z is the dimensionality-reduced feature matrix, and ReLU(.) is the linear rectified activation function. This represents the weight matrix used for dimensionality reduction. , For real numbers, Here, k represents the size of the concatenated features, and k represents the size of the dimension after dimensionality reduction. For bias.
4. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 3, characterized in that, The method for determining the dimensionality-reduced feature matrix using the entropy weighting method and then fusing it is as follows: Standardize the attribute data in the dimensionality-reduced feature matrix; Calculate the information entropy of each attribute; Based on the information entropy of each attribute Calculate the weights of each feature; The data feature matrix with reasonable weights is obtained by calculating the feature matrix z after dimensionality reduction and the weights of each feature.
5. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 4, characterized in that, The specific method for standardizing the attribute data of the dimensionality-reduced feature matrix is as follows: ； in, For standardized attribute data, Let i be the i-th attribute in the reduced-dimensional feature matrix, i = 1, 2, ..., k. This represents the minimum value of the attribute in the dimensionality-reduced feature matrix. This represents the maximum value of the attribute in the reduced-dimensional feature matrix.
6. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 5, characterized in that, The specific method for calculating the information entropy of each attribute is as follows: ； in, Let be the information entropy of the i-th attribute.
7. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 6, characterized in that, Based on the information entropy of each attribute The specific method for calculating the weights of each feature is as follows: ； in, Let be the weight of the i-th feature.
8. The federated learning optimization method for improving the efficiency of interbank security collaborative modeling according to claim 7, characterized in that, The specific method for calculating the data feature matrix with appropriate weights is as follows: ； in, This is to calculate the data feature matrix with reasonable weights.