Business approval method, device, equipment, medium and program product

Abstract

The application provides a business approval method and device, equipment, medium and program product, which can be applied to the field of artificial intelligence and the field of financial technology. The method comprises the following steps: in response to a user submitting a business application, obtaining business behavior data associated with the business application; calling a local approval model corresponding to the business application to process the business behavior data to obtain a business approval result; wherein the local approval model is obtained by federated learning training of a global model and a local model based on a dynamic adjustment strategy of a privacy perturbation parameter according to historical business data through m client terminals and a server; wherein m is an integer greater than 1, the m client terminals include a target client, and the privacy perturbation parameter includes differential privacy noise and a perturbation radius; and returning the business approval result to the user.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and fintech, specifically to a business approval method, apparatus, equipment, medium, and program product, and more specifically to a business approval method, apparatus, equipment, medium, and program product based on federated learning. Background Technology

[0002] Business approval is a core and crucial step for financial institutions to control credit risk, ensure asset security, and implement regulatory compliance. Traditional business approval models rely on the institution's own data to assess customer conditions, resulting in data silos between financial institutions that cannot fully evaluate the true situation of customers. At the same time, customers' financial privacy information cannot be shared with external parties at will.

[0003] The relevant technology uses federated learning to allow multiple institutions to jointly train the approval model. Only the model parameters are transmitted and the raw data is stored locally. However, in the business approval federated learning scenario, it is vulnerable to poisoning attacks (PA). Participants can damage the global model by tampering with local data or uploading abnormal parameters, resulting in a significant reduction in the model's prediction accuracy. It is impossible to simultaneously protect data privacy, resist PA, and maintain model performance. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a business approval method, apparatus, device, medium, and program product.

[0005] According to a first aspect of this application, a business approval method based on federated learning is provided, applied to a target client. The method includes: in response to a user submitting a business application, obtaining business behavior data associated with the business application; calling a local approval model corresponding to the business application to process the business behavior data and obtain a business approval result; wherein the local approval model is obtained by federated learning training on a global model and a local model based on historical business data using m clients and a server, and a dynamic adjustment strategy based on privacy perturbation parameters; wherein m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius; and returning the business approval result to the user.

[0006] According to an embodiment of this application, the step of using m clients and a server to perform federated learning training on a global model and a local model based on a dynamic adjustment strategy using privacy perturbation parameters and historical business data includes: receiving a global model from a server through the m clients; using the m clients to perform local model training and noise injection based on the local data of the m clients and the global model using the dynamic adjustment strategy using privacy perturbation parameters, obtaining local models corresponding to the m clients, and determining the parameters of the local models corresponding to the m clients; wherein, the local data of the m clients belongs to the historical business data; using the m Each of the m clients uploads its local model parameters to the server. The server receives these local model parameters, performs curvature aggregation and parameter updates on the global model based on the credibility weights and the local model parameters, obtains a trained global model, and distributes the trained global model to the m clients. The m clients receive the trained global model, and, based on the trained global model, fine-tune the local models corresponding to the m clients using the trained global model to obtain a local approval model for each of the m clients.

[0007] According to a second aspect of this application, a business approval method based on federated learning is provided, applied to m clients, where m is an integer greater than 1. The method includes: receiving a global model from a server; performing local model training and noise injection based on local data of the m clients and the global model according to a dynamic adjustment strategy of privacy perturbation parameters, obtaining local models corresponding to the m clients, and determining the parameters of the local models corresponding to the m clients; wherein the privacy perturbation parameters include differential privacy noise and perturbation radius; uploading the parameters of the local models corresponding to the m clients to the server; enabling the server to receive the parameters of the local models corresponding to the m clients, and performing curvature aggregation and parameter update on the global model based on credibility weights and the parameters of the local models corresponding to the m clients to obtain a trained global model, and distributing the trained global model to the m clients; receiving the trained global model; and fine-tuning the local models corresponding to the m clients based on the trained global model to obtain local approval models corresponding to the m clients; wherein the local approval models of the m clients are used by the m clients to approve user business applications.

[0008] According to an embodiment of this application, the dynamic adjustment strategy based on privacy perturbation parameters, which involves local model training and noise injection based on the local data of the m clients and the global model, includes: constructing a local model for the current client based on the global model; iteratively training the local model for the current client using a sharpness-aware minimization method based on the local data of the current client and the perturbation radius; and during the iterative training process, adjusting the privacy perturbation parameters in stages according to the variance of the loss function based on the dynamic adjustment strategy of privacy perturbation parameters, and injecting the differential privacy noise into the local model for the current client.

[0009] According to an embodiment of this application, the dynamic adjustment strategy for the privacy perturbation parameter includes: in response to the loss function variance satisfying the initial training conditions, setting the differential privacy noise to the maximum noise intensity value and setting the perturbation radius to the minimum perturbation radius value; in response to the loss function variance satisfying the mid-training conditions, performing exponential decay adjustment on the differential privacy noise based on the noise exponential decay coefficient, the maximum noise intensity value, and the minimum noise intensity value, and performing linear increase adjustment on the perturbation radius based on the perturbation linear growth factor, the minimum perturbation radius value, and the maximum perturbation radius value; in response to the loss function variance satisfying the late-training conditions, calculating the suppressed noise value based on the noise suppression coefficient and the minimum noise intensity value, setting the differential privacy noise to the suppressed noise value, and setting the perturbation radius to the maximum perturbation radius value.

[0010] According to an embodiment of this application, the initial training condition is that the variance of the loss function is greater than or equal to a first variance threshold; the intermediate training condition is that the variance of the loss function is greater than or equal to a second variance threshold and less than the first variance threshold, or that the variance of the loss function is less than the second variance threshold and the current training round number is less than the later starting round number; the later training condition is that the variance of the loss function is less than the second variance threshold and the current training round number is greater than or equal to the later starting round number.

[0011] According to an embodiment of this application, injecting the differential privacy noise into the local model of the current client includes: calculating the original gradient based on the local data of the current client and the local model parameters of the local model; performing gradient clipping on the original gradient based on a clipping threshold to obtain a local gradient; calculating the noise variance based on the differential privacy noise, the clipping threshold, and the identity matrix, and determining Gaussian noise based on the noise variance and a preset noise mean; and adding the Gaussian noise to the local gradient.

[0012] According to an embodiment of this application, the step of fine-tuning the local models corresponding to the m clients based on the trained global model includes: constructing a regularization term based on the global model parameters corresponding to the trained global model, the local model parameters of the current client, and the personalization intensity parameters; minimizing the loss value of the local model parameters of the current client, and constructing a local objective function of the current client using the regularization term as a constraint; and fine-tuning and updating the local models of the current client based on the local data of the current client and the local objective function, until the local models of the current client converge.

[0013] According to a third aspect of this application, a business approval method based on federated learning is provided, applied to a server. The method includes: distributing a global model to m clients, so that the m clients receive the global model distributed by the server; based on a dynamic adjustment strategy of privacy perturbation parameters, performing local model training and noise injection according to the local data of the m clients and the global model to obtain local models corresponding to the m clients, and determining the parameters of the local models corresponding to the m clients; uploading the parameters of the local models corresponding to the m clients to the server; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius; receiving the parameters of the local models corresponding to the m clients; performing curvature aggregation and parameter update on the global model based on credibility weights and the parameters of the local models corresponding to the m clients to obtain a trained global model; distributing the trained global model to the m clients, so that the m clients receive the trained global model; and fine-tuning the local models corresponding to the m clients based on the trained global model to obtain local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to perform business approval on users' business applications.

[0014] According to an embodiment of this application, the step of performing curvature aggregation and parameter update on the global model based on the credibility weights and the local model parameters corresponding to the m clients includes: calculating the gradient variance of the m clients based on the local model parameters corresponding to the m clients, and calculating the credibility weights of the m clients based on the sensitivity coefficients and the gradient variances of the m clients; calculating the diagonal norm of the Fisher information matrix based on the local model parameters corresponding to the m clients to obtain the model update curvature of the m clients; performing weighted aggregation on the local model parameters corresponding to the m clients based on the credibility weights of the m clients and the model update curvature to obtain the global gradient update amount; and updating the global model parameters of the global model based on the global gradient update amount.

[0015] According to a fourth aspect of this application, a business approval device based on federated learning is provided, applied to a target client. The device includes: a data acquisition module, used to acquire business behavior data associated with the business application in response to a user submitting a business application; a business approval module, used to call a local approval model corresponding to the business application to process the business behavior data and obtain a business approval result; wherein the local approval model is obtained by federated learning training of a global model and a local model based on historical business data using m clients and a server, based on a dynamic adjustment strategy of privacy perturbation parameters; wherein m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius; and a result return module, used to return the business approval result to the user.

[0016] According to a fifth aspect of this application, a business approval device based on federated learning is provided, applied to m clients, where m is an integer greater than 1. The device includes: a first global model receiving module for receiving a global model from a server; a local model training and noise injection module for performing local model training and noise injection based on local data from the m clients and the global model, according to a dynamic adjustment strategy using privacy perturbation parameters, to obtain local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients; wherein the privacy perturbation parameters include differential privacy noise and a perturbation radius; and a local model parameter uploading module for uploading the local models corresponding to the m clients to the server. The system comprises the following components: a global model receiving module for receiving the local model parameters corresponding to the m clients; and a local model fine-tuning training module for fine-tuning the local models corresponding to the m clients based on the trained global model to obtain local approval models corresponding to the m clients. The local approval models for the m clients are used by the m clients to approve user business applications.

[0017] According to a sixth aspect of this application, a business approval device based on federated learning is provided, applied to a server. The device includes: a first global model distribution module, configured to distribute a global model to m clients, enabling the m clients to receive the global model distributed by the server; based on a dynamic adjustment strategy for privacy perturbation parameters, performing local model training and noise injection according to the local data of the m clients and the global model to obtain local models corresponding to the m clients, and determining the local model parameters corresponding to the m clients; uploading the local model parameters corresponding to the m clients to the server; wherein the privacy perturbation parameters include differential privacy noise and perturbation radius; and a local model parameter receiving module. The system is configured to receive local model parameters corresponding to the m clients; a curvature aggregation and parameter update module, used to perform curvature aggregation and parameter update on the global model based on the credibility weight and the local model parameters corresponding to the m clients to obtain a trained global model; and a second global model distribution module, used to distribute the trained global model to the m clients so that the m clients can receive the trained global model; and to fine-tune the local models corresponding to the m clients based on the trained global model to obtain local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve users' business applications.

[0018] According to a seventh aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0019] According to an eighth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0020] According to a ninth aspect of this application, a computer program product is also provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0021] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0022] Figure 1 The illustrations depict application scenarios of business approval methods, apparatus, devices, media, and program products based on federated learning according to embodiments of this application.

[0023] Figure 2 A flowchart illustrating a business approval method based on federated learning according to an embodiment of this application is shown schematically.

[0024] Figure 3 This illustration schematically shows a client training flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0025] Figure 4 This illustration schematically shows a model aggregation flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0026] Figure 5 This illustration schematically shows a local training flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0027] Figure 6 This illustration schematically shows a noise injection flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0028] Figure 7 This illustration schematically shows a local fine-tuning flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0029] Figure 8 This illustration schematically shows a federated learning structure diagram of a credit approval federated learning method according to an embodiment of this application;

[0030] Figure 9 This illustration schematically shows a server training flowchart of a business approval method based on federated learning according to an embodiment of this application;

[0031] Figure 10 This schematic diagram illustrates a structural block diagram of a business approval apparatus based on federated learning according to an embodiment of this application;

[0032] Figure 11 This illustration schematically shows a client-side structural block diagram of a business approval apparatus based on federated learning according to an embodiment of this application;

[0033] Figure 12 This illustration schematically shows a client-side structural block diagram of a business approval apparatus based on federated learning according to an embodiment of this application;

[0034] Figure 13 A block diagram schematically illustrates an electronic device suitable for implementing a federated learning-based business approval method according to an embodiment of this application. Detailed Implementation

[0035] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0036] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0037] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0038] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0039] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0040] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0041] Traditional business approval models rely on internal bank data for decision-making, facing a dual dilemma: First, data silos exist among financial institutions, with limited data coverage from a single institution, making it difficult to comprehensively assess the creditworthiness of new customers, long-tail customers, and cross-bank customers, leading to biases in model training samples. Second, credit data contains highly sensitive personal financial information, and due to compliance constraints, institutions should not share raw data. Federated learning, by employing a federated learning model training approach, allows participating institutions to collaboratively train the model simply by exchanging intermediate model parameters, completing joint modeling without data leaving the domain. However, this framework faces PA (Problem-Avoidance) risk, where malicious nodes can disrupt global model performance by tampering with local training data or submitting malicious model parameters. PA indicates that an abnormal client can cause the global model to produce incorrect predictions by tampering with local data or gradients.

[0042] To protect privacy from PA (Patient Attack), adding fixed-intensity noise during model training can sacrifice model performance for enhanced security, resulting in a significant decrease in model accuracy. Furthermore, existing defense solutions struggle to balance PA defense effectiveness with high model availability while ensuring data privacy and security. They also cannot accurately distinguish between malicious poisoning and normal business data fluctuations, which can lead to critical decision-making risks such as distorted customer credit ratings and altered credit limits.

[0043] This application provides a product recommendation method based on federated learning, applied to a target client, including: responding to a user submitting a business application, obtaining business behavior data associated with the business application; calling the local approval model corresponding to the business application to process the business behavior data and obtain a business approval result; wherein, the local approval model is obtained by federated learning training of the global model and the local model based on historical business data through m clients and a server, using a dynamic adjustment strategy for privacy perturbation parameters; where m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius; and returning the business approval result to the user. In this application's embodiment, federated learning of the global and local models is achieved based on a dynamic adjustment strategy for privacy perturbation parameters. The original business data is retained locally on the client and not transmitted externally, protecting customer financial privacy and breaking down institutional data silos. During the federated learning process, the differential privacy noise and perturbation radius are dynamically adjusted to balance privacy protection and model performance, while mitigating poisoning attacks, suppressing the damage of abnormal parameters and tampered data to the global model, and ensuring the accuracy of model approval prediction. On the target client, a pre-trained local approval model is used for business approval, taking into account compliance and risk control, privacy and security, and anti-attack capabilities, adapting to the regulatory requirements for financial business approval, and stably outputting reliable approval results.

[0044] It should be noted that the business approval method and apparatus based on federated learning in this application can be used in the fields of artificial intelligence and fintech, as well as in any other field. The application fields of the business approval method and apparatus based on federated learning in this application are not limited.

[0045] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user business information, user transaction information, user behavior information, user preference information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0046] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0047] Figure 1The illustration schematically depicts an application scenario of a business approval method, apparatus, device, medium, and program product based on federated learning according to embodiments of this application. For example... Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send information, etc.

[0048] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal devices. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, and it can include, but is not limited to, applications, mini-programs, etc., installed on the terminal devices, and can also be in the form of web pages.

[0049] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and basic cloud computing services such as big data. The server can be the backend server of the aforementioned target application, used to provide backend services to the clients of the target application.

[0050] It should be noted that the business approval method based on federated learning provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-103. Accordingly, the business approval device based on federated learning provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-103.

[0051] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0052] Figure 2 A flowchart illustrating a business approval method based on federated learning according to an embodiment of this application is shown schematically. Figure 2 As shown, the business approval method 200 based on federated learning according to the embodiments of this application is applied to the target client, and the method 200 may include steps S210 to S230.

[0053] In step S210, in response to the user submitting a business application, the business behavior data associated with the business application is obtained.

[0054] Business approval means that financial institutions need to verify the user's qualifications and business materials based on risk control rules and models to determine whether to approve the user's application for account, credit, wealth management, or other services. Business applications include, but are not limited to, credit applications, credit card application, wealth management contract application, and account opening application. Business approval includes, but is not limited to, credit approval, credit card application approval, wealth management contract approval, and account opening approval.

[0055] Business behavior data represents historical time-series records of user behavior during various financial transactions, including application, transaction, performance, and holding information, used for business approval. Business behavior data associated with credit applications includes repayment history, transaction frequency, credit application information, and overdue records; business behavior data associated with credit card applications includes historical card opening information, spending frequency, existing credit card holder information, and billing performance information; business behavior data associated with wealth management contract applications includes historical wealth management subscription and redemption information, fund holding period, risk assessment history, and wealth management transaction frequency; business behavior data associated with account opening applications includes the number of existing accounts, account fund deposit and withdrawal frequency, historical account closures, and lost account information.

[0056] When a user submits any business application such as credit, card opening, wealth management, or account opening, the system identifies the application type, matches the corresponding business behavior dimension, retrieves the user's historical time-series records, such as repayments, transactions, holdings, and account changes, and summarizes the business behavior data associated with the application to complete data acquisition.

[0057] In step S220, the local approval model corresponding to the business application is invoked to process the business behavior data and obtain the business approval result. The local approval model is obtained by federated learning training of the global model and the local model based on historical business data through m clients and servers, and a dynamic adjustment strategy based on privacy perturbation parameters. Here, m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius.

[0058] Federated Learning (FL) is a distributed machine learning framework in which each client keeps the original data locally and does not share the original data. They only exchange model parameters to jointly model the data. Each client and server collaborates to train the global model and the local model, so that the model moves while the data remains stationary, and the data is available but not visible.

[0059] Differential privacy (DP) makes the model output insensitive to changes in a single sample by adding random noise, providing strict privacy guarantees. The privacy budget controls the strength of protection; the smaller the privacy budget, the stronger the privacy.

[0060] The global model is built on a Long Short-Term Memory (LSTM) network. It iteratively mines the temporal features of user business behavior based on multi-layer temporal memory units. Through a gating structure, it retains and filters long-term historical business behavior information layer by layer, correcting the bias in temporal feature extraction. LSTM is highly adaptable to temporal business behavior data in financial business approval scenarios (such as dynamic temporal information such as loan repayment timelines, credit card transaction records, wealth management application and redemption records, and account deposit and withdrawal timelines, as well as risk control information such as overdue records, card holding status, and holding periods). It can effectively capture the complex temporal nonlinear correlation between continuous user business behavior and business risks such as credit, account opening, and wealth management, and accurately extract risk control features for qualification verification.

[0061] The local model and the global model have the same model structure. The global model, initialized by the server, serves as the initial parameters, and each client independently trains its local model based on its local data. Since the local data of each client has non-independent and identically distributed characteristics, the parameters of the local model differ from those of the global model after training.

[0062] The target client refers to any one of the m clients participating in federated learning training, which can call itself to complete the training of its local model.

[0063] Historical business data refers to local data on the historical time-series business behaviors of various clients during the process of users handling various financial transactions. Local data includes credit history data (such as repayment history, transaction frequency, credit application information, and overdue records), credit card opening history data (such as historical card opening information, consumption frequency, existing credit card holder information, and billing performance information), wealth management contract signing history data (such as historical wealth management subscription and redemption information, fund holding period, risk assessment history, and wealth management transaction frequency), and account opening history data (such as the number of existing accounts, account fund deposit and withdrawal frequency, historical account closure information, and lost account information), etc.

[0064] After a user submits a business application and completes the acquisition of business behavior data, the system identifies the approval type of the application, matches and calls the dedicated local approval model, and inputs the user's dedicated time-series business behavior data (credit repayment, consumption performance, wealth management holdings, account deposits and withdrawals, overdue records, etc.) into the local approval model. The local approval model performs serialized time-series feature extraction, explores the time-series non-linear correlation between business behavior and credit risk, completes qualification risk control inference calculation, and outputs business approval results. The business approval results include approval judgment results (such as approval passed, approval rejected), risk score and risk confidence level, etc., and completes risk qualification approval judgment for various businesses such as credit, card opening, wealth management contract signing, and account opening.

[0065] The local approval model is jointly trained by m clients and the server through federated learning. During training, each client retains its original historical business data locally, dynamically adjusts two types of privacy perturbation parameters—differential privacy noise and perturbation radius—to complete local noise-adding training, and each client only uploads the processed local model parameters to the server. The server updates the global model based on the weighted aggregation of the local model parameters from each client and then distributes the global model. Each client performs fine-tuning training locally to obtain the local approval model.

[0066] It should be noted that there are differences between the types of approval applications corresponding to different businesses. We can train dedicated business approval models to adapt to various approval processes, so that the models can match the needs of various approval scenarios.

[0067] In step S230, the business approval result is returned to the user.

[0068] After the local approval model outputs the business approval result, it binds and archives the result with the corresponding business application record, retains risk control logs, and then synchronously displays the approval conclusion on the user's business client interface. If the approval is successful, it indicates subsequent operation instructions such as credit limit, contract signing, and account opening; if rejected, it lists the basis for risk assessment. Simultaneously, approval notifications are pushed through both reserved SMS and in-app messages, allowing users to log in at any time to check the approval result.

[0069] In the embodiments of this application, a dynamic adjustment strategy based on privacy perturbation parameters is used to achieve federated learning of global and local models. Original business data is retained locally on the client and not transmitted externally, protecting customer financial privacy and breaking down institutional data silos. During federated learning, differential privacy noise and perturbation radius are dynamically adjusted to balance privacy protection with model performance. This also mitigates poisoning attacks, suppresses the damage to the global model caused by abnormal parameters and data tampering, and ensures the accuracy of model approval predictions. On the target client, the trained local approval model is used for business approval, balancing compliance risk control, privacy security, and attack resistance, adapting to financial business approval regulatory requirements, and stably outputting reliable approval results.

[0070] According to an embodiment of this application, through m clients and a server, based on a dynamic adjustment strategy using privacy perturbation parameters, federated learning training is performed on a global model and a local model based on historical business data. This includes: receiving a global model from the server through m clients; training and injecting noise into local models based on the local data of the m clients and the global model, using the dynamic adjustment strategy using privacy perturbation parameters, to obtain local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients; wherein the local data of the m clients belongs to historical business data; uploading the local model parameters corresponding to the m clients to the server through the m clients; enabling the server to receive the local model parameters corresponding to the m clients, perform curvature aggregation and parameter updates on the global model based on the credibility weights and the local model parameters corresponding to the m clients, obtain a trained global model, and distribute the trained global model to the m clients; receiving the trained global model through the m clients; and fine-tuning the local models corresponding to the m clients based on the trained global model through the m clients to obtain local approval models corresponding to the m clients.

[0071] Figure 3 The diagram illustrates a client-side training flowchart for a federated learning-based business approval method according to an embodiment of this application. Figure 3 As shown, the business approval method 300 based on federated learning according to the embodiment of this application is applied to m clients, where m is an integer greater than 1. The method 300 may include steps S310 to S360.

[0072] In step S310, m clients receive the global model sent by the server.

[0073] The server builds an LSTM model and completes parameter initialization, generates a global model, and distributes the global model (containing the complete global model structure and initialization parameters) to each client. Each client receives the global model distributed by the server.

[0074] In step S320, based on the dynamic adjustment strategy of privacy perturbation parameters, local model training and noise injection are performed according to the local data of m clients and the global model to obtain the local models corresponding to m clients and determine the parameters of the local models corresponding to m clients; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius.

[0075] The dynamic adjustment strategy for privacy perturbation parameters is a phased adaptive adjustment mechanism for the intensity of differential privacy noise and the size of the perturbation radius. This dynamic adjustment strategy dynamically balances the strength of privacy protection, PA resistance and model prediction accuracy according to the model convergence state. In the early stage of training, privacy protection is emphasized, in the middle stage, anti-attack is taken into account, and in the later stage, the performance of the approval model is prioritized.

[0076] Differential privacy noise refers to random noise superimposed on local gradients after gradient clipping during the local model training phase on the client side. By injecting random noise into the gradients, the gradient features corresponding to the original user business behavior data are blurred, preventing servers or third parties from inferring user privacy information based on uploaded parameters. At the same time, high-intensity noise can reduce the degree of pollution of the global model by data poisoning by participating parties and abnormal parameters. The noise intensity changes dynamically with the training phase.

[0077] The perturbation radius is a core control parameter of the sharpness-aware minimization training mechanism, used to constrain the perturbation search range of local model parameters. By controlling the parameter perturbation boundary, the search space of the model loss surface is adjusted. A smaller perturbation radius can avoid severe oscillations in training loss and prevent the model from getting trapped in sharp local minima, while a larger perturbation radius can expand the flat region of parameter search and improve the model's robustness to covert poisoning perturbations. The value increases linearly with the training process.

[0078] Local data represents historical time-series behavioral records generated during users' handling of various financial transactions, including information on applications, transactions, performance, and holdings, which are used for business approval.

[0079] Based on a dynamic adjustment strategy for privacy perturbation parameters, each client receives the global model from the server and trains its local model using a sharpness-aware minimization method with local data, effectively mitigating the negative impact of introducing differential privacy on model usability. Differential privacy noise is injected into each client's local model, and the noise-added local model parameters are then uploaded to the server, preventing the leakage of original data features from the local model parameters.

[0080] In step S330, m clients upload their respective local model parameters to the server.

[0081] After each client completes local training, each client sends its own trained local model parameters to the server.

[0082] In step S340, the server receives the local model parameters corresponding to m clients, performs curvature aggregation and parameter updates on the global model based on the confidence weight and the local model parameters corresponding to m clients, obtains the trained global model, and distributes the trained global model to m clients.

[0083] Figure 4 The diagram illustrates a model aggregation flowchart of a business approval method based on federated learning according to an embodiment of this application. Figure 4 As shown, the server receives local model parameters uploaded by each client, calculates the global model aggregation weighted credibility (i.e., credibility weights), and updates the global model parameters based on client model curvature weighted aggregation to obtain the global model. Secure aggregation is achieved through a credibility weighting mechanism. Credibility weights are calculated based on client gradient variance to achieve gradient credibility evaluation, suppressing the influence of potentially malicious clients. Curvature-weighted aggregation is performed based on loss function curvature information to improve model convergence stability. The server then distributes the global model to each client.

[0084] According to an embodiment of this application, curvature aggregation and parameter updates are performed on the global model based on the credibility weights and the local model parameters corresponding to m clients. This includes: calculating the gradient variance of the m clients based on the local model parameters corresponding to the m clients, and calculating the credibility weights of the m clients based on the sensitivity coefficients and the gradient variances of the m clients; calculating the diagonal norm of the Fisher information matrix based on the local model parameters corresponding to the m clients to obtain the model update curvature of the m clients; performing weighted aggregation on the local model parameters corresponding to the m clients based on the credibility weights and the model update curvature to obtain the global gradient update amount; and updating the global model parameters of the global model based on the global gradient update amount.

[0085] Server-based client Local model parameters Computing client gradient variance The credibility weight is obtained based on the client's local model parameters, and the calculation formula is as follows:

[0086]

[0087] in, For the client The confidence weights for gradient updates. Sensitivity coefficient Let N be the client gradient variance, and N be the number of clients. The reliability of client updates is determined based on the gradient variance; a smaller gradient variance indicates more stable updates and a greater influence on global model updates. A larger gradient variance indicates a higher potential risk of malicious updates; therefore, the proportion of such clients participating in global model updates is suppressed. Reliability weights are set using client gradient variance to effectively eliminate model parameters updated by abnormal clients.

[0088] The server depends on the client. Update the curvature-weighted aggregation and update the global model. The specific calculation formula is as follows:

[0089]

[0090]

[0091] in, The global gradient update amount of the global model, based on Performing global model parameter updates can reduce computational overhead. These are the global model parameters for the current round t. These are the global model parameters from the previous round t-1. The diagonal approximation curvature of the Fisher information matrix is ​​represented by the diagonal norm of the Fisher information matrix of the model parameters of the i-th client, which characterizes the model update curvature.

[0092] In the embodiments of this application, client credibility weights are calculated using gradient variance to suppress interference from participants' poisoning parameters. Combined with the curvature of the Fisher information matrix, weighted secure aggregation of the model is achieved. Curvature-weighted aggregation preserves abnormal patterns in normal business operations, accurately distinguishing between deliberately tampered parameters and abnormal gradients in normal business, thus preventing the erroneous filtering of long-tail customer business characteristics. Curvature optimization improves the global model's convergence stability, and combined with local personalized fine-tuning, significantly enhances the accuracy of business assessments for new and cross-industry customers, strengthens the inclusiveness and fairness of business decisions, and greatly reduces the probability of missed or incorrect customer assessments.

[0093] In step S350, m clients receive the trained global model.

[0094] In step S360, based on the trained global model, the m clients fine-tune the local models corresponding to the m clients to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve the user's business applications.

[0095] Due to the differences in customer groups among various client services, the local business application data of each client exhibits a non-independent, identically distributed characteristic. The distribution of training data varies significantly across different clients. While a global model can extract general approval features across clients, it struggles to adapt to the unique local data patterns of each client. Directly using the global model for local approval inference will lead to a decline in model performance. To accurately adapt to the business application scenarios of different clients, each client needs to conduct local fine-tuning training based on the distributed global model parameters. Through local fine-tuning training, a local approval model tailored to its own business data can be obtained.

[0096] The locally trained approval model can be deployed to the corresponding client business environment to process various business applications submitted by local users, achieving localized intelligent approval. Based on a federated learning local training mechanism, the local approval model improves the accuracy of approval judgments for single-client business applications without interacting with original business document data, ensuring data privacy and security. It can also assist staff in efficiently identifying business risks.

[0097] In the embodiments of this application, each client and the server collaborate to complete federated learning training. A dynamic adjustment strategy based on privacy perturbation parameters optimizes local model training. Each client only uploads model parameters; its original historical business data remains unexposed, protecting customer privacy data from the source, meeting regulatory compliance requirements, and breaking down data silos between institutions. During training, the client dynamically adjusts differential privacy noise and perturbation radius and injects noise to achieve a dynamic balance between privacy protection and model availability. The server combines the credibility weights of each client to perform curvature aggregation and update the global model. The credibility weights can suppress the destructive effects of poisoning parameters from participating parties on the global model, collaboratively defending against poisoning attacks and ensuring model availability. The updated global model is distributed to each client, combined with personalized fine-tuning of the local model to adapt to the differentiated samples of each institution, improving the accuracy of business approval assessments for different customer groups. This achieves dynamic collaborative optimization of privacy protection, poisoning resistance, and model approval performance, enhancing model generalization ability, reducing approval misjudgments, and ensuring stable and reliable business approval decisions.

[0098] According to embodiments of this application, a dynamic adjustment strategy based on privacy perturbation parameters, which involves training a local model and injecting noise based on local data from m clients and a global model, may include: constructing a local model for the current client based on the global model; iteratively training the local model for the current client using a sharpness-aware minimization method based on the local data and perturbation radius of the current client; and during the iterative training process, adjusting the privacy perturbation parameters in stages according to the variance of the loss function based on the dynamic adjustment strategy for privacy perturbation parameters, and injecting differential privacy noise into the local model for the current client.

[0099] In the Sharpness-Aware Minimization (SAM) method, a perturbation radius is set to constrain the perturbation range. Simultaneously, the sharpness of the loss value and the loss surface is minimized, guiding the model to converge to a relatively flat region of the loss function and avoiding getting trapped in sharp local minima. Sharp local minima are characterized by high curvature; even small perturbations can significantly increase the loss, leading to a decrease in the model's return capability. Conversely, in flat minima, the loss function changes gently near the local minimum, becoming less sensitive to parameter perturbations and enhancing the model's generalization ability.

[0100] Figure 5 The diagram schematically illustrates a local training flowchart of a business approval method based on federated learning according to an embodiment of this application. For example... Figure 5 As shown, the client receives the global model from the server and trains a local model based on local data and the perturbation radius using a sharpness-aware minimization method. The local model training formula is as follows:

[0101]

[0102]

[0103] in, For the first Wheel client model parameters, For the first Wheel model parameters, For learning rate, For model parameters The loss gradient after adding perturbation, For model parameters The loss gradient, This is the perturbation vector in SAM. The SAM perturbation radius is a perturbation parameter that controls the model to converge to a region where the loss function is flatter, thus enhancing its robustness to noise perturbations.

[0104] The privacy perturbation parameters are dynamically adjusted based on the variance of the loss function. These parameters include the perturbation radius and differential privacy noise. The formula for calculating the variance of the loss function is as follows:

[0105]

[0106] in, For the first Round loss function variance, The number of batches in the current round. For the first Loss value for each batch, This represents the average loss value for the current round.

[0107] The perturbation radius of the SAM and the intensity of the differential privacy noise are dynamically adjusted based on the variance of the loss function, and then the dynamically adjusted differential privacy noise is injected into the client's local model.

[0108] In the embodiments of this application, a perturbation radius constraint is introduced through a sharpness-aware minimization method. A dynamic adjustment strategy is used to adaptively adjust the privacy perturbation parameters in stages and inject differential privacy noise through a loss function variance-aware training phase. Combined with dynamic differential privacy noise and dynamic perturbation radius, training oscillations are alleviated and local minima are avoided, thereby enhancing the ability to resist poisoning attacks and achieving a balance between privacy protection, anti-attack performance and model accuracy.

[0109] According to embodiments of this application, a dynamic adjustment strategy for privacy perturbation parameters includes: in response to the loss function variance satisfying the initial training conditions, setting the differential privacy noise to the maximum noise intensity value and setting the perturbation radius to the minimum perturbation radius value; in response to the loss function variance satisfying the mid-training conditions, performing exponential decay adjustment on the differential privacy noise based on the noise exponential decay coefficient, the maximum noise intensity value, and the minimum noise intensity value, and performing linear increase adjustment on the perturbation radius based on the perturbation linear growth factor, the minimum perturbation radius value, and the maximum perturbation radius value; and in response to the loss function variance satisfying the late-training conditions, calculating the suppressed noise value based on the noise suppression coefficient and the minimum noise intensity value, setting the differential privacy noise to the suppressed noise value, and setting the perturbation radius to the maximum perturbation radius value.

[0110] To synergistically optimize defense against PA (Problem-Aware Attack) and model usability, a phased dynamic adjustment strategy for privacy perturbation parameters is adopted. In the early training phase, strong differential privacy noise and weak sharpness perception are used to minimize perturbation. Early model parameters are randomly initialized, resulting in chaotic gradient distribution and vulnerability to PA attacks. Strong differential privacy noise is used to block malicious attacks. Early model loss function surfaces oscillate, and excessively large perturbation radii in the SAM (Sexual Awareness Perturbation) can easily lead to sharp local minima. Smaller perturbation radii can avoid excessive oscillation. In the middle training phase, noise attenuation and enhancement of the SAM perturbation radius are implemented. At this stage, the model has initially converged, the gradient distribution tends to stabilize, and it has learned the core characteristics of business risks. Reducing noise can minimize interference with effective updates. Simultaneously, loss function surface smoothing supports efficient searching of flat regions with large perturbation radii. Increasing the perturbation radius helps avoid getting trapped in sharp local minima and improves robustness against covert poisoning perturbations. In the later stages of training, weak differential privacy noise and strong SAM perturbation radius are used. The model has learned the core features of business risks. Reducing noise can eliminate its inhibition on model accuracy. The strong SAM perturbation radius further compresses the parameter space to a wide and flat region, suppressing targeted attacks and achieving a balance between high accuracy and high robustness.

[0111] Based on the variance of the loss function Dynamically adjust differential privacy noise intensity And sharpness perception minimized perturbation radius The specific parameter conversion formula is as follows:

[0112]

[0113]

[0114] in, For the first Round differential privacy noise intensity, For the first Wheel SAM disturbance radius, This is the high variance threshold (first variance threshold). The low variance threshold (second variance threshold). This represents the maximum noise intensity value for differential privacy. This represents the minimum noise intensity value for differential privacy. This is the minimum perturbation radius value for SAM. This represents the maximum perturbation radius of the SAM. The noise index attenuation coefficient is... This is the perturbation linear growth factor. This is the noise suppression coefficient. To suppress noise levels, The starting round number for the later stage indicates the number of rounds to start in the later stage. It is a pre-set threshold for the number of training iteration rounds used to limit the minimum number of training rounds before the model can enter the later parameter configuration stage. It can be preset according to the dataset size and network depth, such as a lightweight model for small samples. Take 60%–70% of the total number of iterations; big data deep model, Take 70% to 80% of the total number of iterations.

[0115] In the embodiments of this application, the dynamic adjustment strategy for privacy perturbation parameters is implemented in three stages for adaptive adjustment of privacy perturbation. In the early training phase, maximum noise and minimum perturbation radius are set. Strong differential privacy noise protects sensitive information and blocks the effects of poisoning attacks. A small perturbation radius avoids severe oscillations on the loss surface and prevents the model from falling into sharp minima. In the middle training phase, noise decays exponentially, and the perturbation radius increases linearly. After the model initially converges, noise interference on the effective gradient is reduced. A larger perturbation radius searches for flat parameter regions, improving perturbation robustness. In the later training phase, low noise is suppressed, and the maximum perturbation radius is used to reduce noise intensity and ensure model accuracy. A wide perturbation space suppresses targeted poisoning attacks. This phased, progressive optimization prioritizes privacy protection in the early stage, strengthens anti-attack capabilities in the middle stage, and improves model robustness in the later stage, dynamically balancing privacy protection and model usability.

[0116] According to an embodiment of this application, the initial training condition is that the variance of the loss function is greater than or equal to a first variance threshold; the intermediate training condition is that the variance of the loss function is greater than or equal to a second variance threshold and less than the first variance threshold, or the variance of the loss function is less than the second variance threshold and the current training round number is less than the later starting round number; the later training condition is that the variance of the loss function is less than the second variance threshold and the current training round number is greater than or equal to the later starting round number.

[0117] Initial training conditions: Regardless of the training round, a uniform approach of strong noise and minimal perturbation is used. When initial training conditions are met, a configuration combining strong differential privacy noise with weak sharpness perception to minimize perturbation is employed. Fixed at maximum noise intensity To block the malicious gradient impact of poisoning attacks; Fixed as minimum disturbance radius This avoids excessive oscillation of the loss function surface and its getting trapped in sharp local minima.

[0118] Mid-training conditions: The noise exponent decays while the perturbation increases linearly. When the mid-training conditions are met, According to the exponential law, from decay to This reduces interference with effective gradient updates; According to linear law from Gradually increase the size to adapt to the smoothed loss function surface, search for the flat optimal region, and improve the robustness of covert poisoning.

[0119] If during training Decrease to low variance threshold in advance The following is an example, but the current training rounds have not yet reached the starting rounds of the later stage. If the model only converges locally and has not fully learned the global characteristics of business risks, it will not switch to fixed parameters in the later stage, but will continue to use the update strategy of exponentially decaying noise and linearly increasing perturbation radius in the mid-term stage; until the training rounds reach the required number of rounds. And the loss variance remains at Next, switch to the parameter configuration for low noise and maximum disturbance radius in the later stage.

[0120] Conditions for the later stages of training: Fixed low noise, maximum disturbance. Provide a training baseline to prevent the model from converging prematurely and entering the later stages of configuration. When the conditions for later training are met, use a configuration that minimizes perturbations by combining weak differential privacy noise with strong sharpness perception. Fixed as noise suppression value This eliminates the impact of noise on model accuracy. Fixed at the maximum disturbance radius This compresses the parameter space into a wide and flat area, suppressing targeted poisoning attacks.

[0121] In the embodiments of this application, conditions for each training stage are defined, taking into account both the variance of the loss function and the number of training rounds to avoid the premature convergence defect of the model. If the variance of the loss function reaches the target ahead of time but the number of training rounds is insufficient, it is determined that the model has not fully learned the business risk features and is considered to be in the middle stage of training. The mid-stage dynamic parameter tuning strategy is continued to be used. Only when the variance is low and the number of rounds reaches the target is the model switched to the later stage. This avoids the model reducing noise too early, which would lead to the failure of privacy protection and the decline of anti-attack capability, and ensures the privacy protection effect, anti-attack capability and training accuracy of each training stage.

[0122] According to an embodiment of this application, injecting differential privacy noise into the local model of the current client includes: calculating the original gradient based on the local data of the current client and the local model parameters of the local model; performing gradient clipping on the original gradient based on a clipping threshold to obtain the local gradient; calculating the noise variance based on the differential privacy noise, the clipping threshold and the identity matrix, and determining Gaussian noise based on the noise variance and a preset noise mean; and adding Gaussian noise to the local gradient.

[0123] Figure 6 A flowchart illustrating a noise injection process for a federated learning-based business approval method according to an embodiment of this application is shown. Figure 6 As shown, firstly, each client trains a local model using local data, and then performs gradient clipping based on the local model parameters. Specifically, based on the client's local data and local model parameters, noise-free gradient update parameters are obtained, and the original gradient is calculated. The specific formula is as follows:

[0124]

[0125] in, For the original gradient, For loss function, For the client Model parameters, For the client Local data.

[0126] Secondly, gradient clipping is performed based on the original gradient to obtain the local gradient, as shown in the following formula:

[0127]

[0128] in, The cropping threshold, The L2 norm of the original gradient. This is the local gradient.

[0129] Secondly, the client adds adaptive noise to the local model parameters, that is, adds differential privacy adaptive noise to the local gradient. The formula for calculating the gradient after adding noise is as follows:

[0130]

[0131] in, This refers to the gradient after adding noise, i.e., the local model parameters after adding noise. The cropping threshold, It is the identity matrix. This indicates that the preset noise mean is 0 and the noise variance is... Gaussian noise was used initially, with high-intensity noise employed to defend against poisoning attacks, and later gradually reduced to minimum noise level to ensure model accuracy.

[0132] Finally, the client uploads the local model parameters with added differential privacy noise to the server.

[0133] In the embodiments of this application, the original gradient is clipped to obtain the local gradient, and the local gradient is superimposed with adaptive Gaussian noise. The Gaussian noise is used to avoid privacy leakage. Combined with the phased noise intensity change, the high noise in the early stage of training blocks the original business features from leaking out of the local model parameters and resists the global pollution caused by data poisoning and tampering. In the middle and later stages of training, the noise gradually decays, reducing the interference of noise on the effective update of the gradient. There is no need to transmit the original customer privacy data, and the noise weakens the loss of business approval prediction accuracy.

[0134] According to an embodiment of this application, fine-tuning training is performed on the local models corresponding to m clients based on the trained global model, including: constructing a regularization term based on the global model parameters corresponding to the trained global model, the local model parameters of the current client, and the personalization intensity parameters; minimizing the loss value of the local model parameters of the current client, and constructing the local objective function of the current client with the regularization term as a constraint; and fine-tuning training and parameter updating of the local model of the current client based on the local data and the local objective function of the current client until the local model of the current client converges.

[0135] Figure 7 A schematic diagram illustrates a locally fine-tuned flowchart of a business approval method based on federated learning according to an embodiment of this application. For example... Figure 7 As shown, the client receives the global model from the server and performs personalized fine-tuning of the local model based on the global model. The local objective function for personalized fine-tuning of the local model is as follows:

[0136]

[0137] in, For the local model parameters of client i, For global model parameters, Personalized strength parameters Loss function. The regularization term measures the difference between the local model parameters and the global model parameters. The larger the value of the personalization strength parameter, the higher the degree of personalization in the local model training.

[0138] Repeat the fine-tuning steps to fine-tune the local model and update its parameters until the local model converges, at which point the fine-tuning training ends.

[0139] It's worth noting that in real-world business scenarios, the local training data of each client is not independent and identically distributed. If the client's training data differs significantly, the global model may learn common features across clients, but it might not be able to fully adapt to the local data distribution of each client. Directly using the global model for inference locally will degrade model performance. Adding regularization constraints to the global model distributed by the server prevents local model parameters from deviating from the global optimum while ensuring the global model fits the local data distribution, thus improving the model accuracy for local data.

[0140] In the embodiments of this application, global model parameters, local model parameters, and personalized strength are integrated to construct a regularization term. This regularization term serves as a constraint for constructing a local objective function. Based on the objective function and local data, fine-tuning training and parameter updates of the local model are performed, completing the local model fine-tuning. This adapts to the differences in customer groups across financial institutions, providing personalized adaptation for different customer groups and addressing the insufficient adaptability of the unified global model. By balancing globally common risk characteristics with locally unique customer behavior characteristics using regularization constraints, the accuracy of cross-bank and new customer business approval and assessment is significantly improved, reducing the probability of misjudging high-risk customers.

[0141] Based on the above-mentioned federated learning-based business approval method, taking credit approval as an example, embodiments of this application also provide a federated learning method for credit approval based on dynamic collaborative defense against poisoning attacks. Figure 8 A schematic diagram illustrating the federated learning architecture of the credit approval federated learning method according to an embodiment of this application is shown. Figure 8As shown, the system receives the global model from the server. The client trains its local model using local data through a sharpness-aware minimization method, enabling local training. Clients can be of any number. To prevent the leakage of sensitive user information, differential privacy noise is injected into the client's local model. The noise-added local model parameters are then uploaded to the server, preventing the leakage of original data features. The server receives the updated parameters from the client's local model and performs weighted aggregation based on gradient curvature analysis using a credibility-weighted mechanism to update the global model. This secure aggregation accurately distinguishes between malicious updates and genuine business anomalies, addressing core issues in credit federated learning scenarios such as poisoning attacks threatening model security, privacy leaks, and the difficulty in co-optimizing model performance. Clients receive the global model; different clients contain different user groups. Each client performs local fine-tuning of the received global model, adapting it to regional credit feature differences through personalized adjustments.

[0142] The server-side weighted aggregation module and the client-side local fine-tuning module dynamically add gradient compression on the client side and send local model parameters to the server. This effectively defends against privacy inference attacks while reducing communication bandwidth and improving communication efficiency. To address the issue of global model deviation from the global optimum caused by heterogeneous statistical data, the server aggregates local model parameters, adds regularization constraints to the global model, and then distributes the aggregated and trained global model to the client. Since different regions and customers have different financial preferences, to better align with local data distribution and accurately characterize user profiles, the client performs personalized fine-tuning training of its local model based on the distributed global model, achieving personalized learning from local data. Through joint modeling, the client obtains the fine-tuned local model. The input of this model is the user's business characteristic data, and the output is the product matching prediction result corresponding to the user. This enables prediction of customer matching degree for financial products, characterizing customer profiles from multiple dimensions and in depth, accurately identifying potential product buyers, and recommending financial products based on the product matching prediction results, significantly improving the success rate of financial product recommendations.

[0143] Figure 9 The diagram illustrates a server training flowchart for a federated learning-based business approval method according to an embodiment of this application. Figure 9 As shown, the business approval method 500 based on federated learning according to the embodiments of this application is applied to a server, and the method 500 may include steps S510 to S560.

[0144] In step S510, the server sends the global model to m clients.

[0145] The server initializes the global model and sets hyperparameters such as learning rate and training epochs. The server then retrieves the global model and distributes it to each client, which in turn receives the global model from the server.

[0146] In step S520, m clients receive the global model sent by the server; based on the dynamic adjustment strategy of privacy perturbation parameters, local model training and noise injection are performed according to the local data of m clients and the global model to obtain the local models corresponding to m clients and determine the local model parameters corresponding to m clients; the local model parameters corresponding to m clients are uploaded to the server; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius.

[0147] It should be noted that step S520 is the same as the method described in steps S310 to S330 above, and will not be repeated here.

[0148] In step S530, the server receives local model parameters corresponding to m clients.

[0149] In step S540, the server performs curvature aggregation and parameter updates on the global model based on the confidence weights and the local model parameters corresponding to m clients, thereby obtaining the trained global model.

[0150] In step S550, the server distributes the trained global model to m clients.

[0151] It should be noted that steps S530 to S550 are the same as the method described in step S340 above, and will not be repeated here.

[0152] In step S560, m clients receive the trained global model; based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve the user's business applications.

[0153] It should be noted that step S560 is the same as the method described in steps S350 to S360 above, and will not be repeated here.

[0154] Based on the above-described federated learning-based business approval method, embodiments of this application also provide a federated learning-based business approval device, applied to a target client. The following will combine... Figure 10 The device is described in detail.

[0155] Figure 10 A schematic diagram illustrating the structure of a business approval apparatus based on federated learning according to an embodiment of this application is shown.

[0156] like Figure 10 As shown, the business approval device 1000 based on federated learning in this embodiment includes a data acquisition module 1010, a business approval module 1020, and a result return module 1030.

[0157] The data acquisition module 1010 is used to acquire business behavior data associated with a user's submitted business application. In one embodiment, the data acquisition module 1010 can be used to execute step S210 described above, which will not be repeated here.

[0158] The business approval module 1020 is used to call the local approval model corresponding to the business application to process the business behavior data and obtain the business approval result. The local approval model is obtained by federated learning training of the global model and the local model based on historical business data, using m clients and servers and a dynamic adjustment strategy for privacy perturbation parameters. Here, m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius. In one embodiment, the business approval module 1020 can be used to execute step S220 described above, which will not be repeated here.

[0159] The result return module 1030 is used to return the business approval result to the user. In one embodiment, the result return module 1030 can be used to execute the step S230 described above, which will not be repeated here.

[0160] According to an embodiment of this application, the apparatus 1000 further includes: a federated training module, configured to receive a global model from a server through the m clients; through the m clients, based on a dynamic adjustment strategy for privacy perturbation parameters, perform local model training and noise injection based on the local data of the m clients and the global model to obtain local models corresponding to the m clients, and determine the parameters of the local models corresponding to the m clients; wherein the local data of the m clients belongs to the historical business data; through the m clients, upload the parameters of the local models corresponding to the m clients to the server; so that the server receives the parameters of the local models corresponding to the m clients, performs curvature aggregation and parameter updates on the global model based on the credibility weight and the parameters of the local models corresponding to the m clients, obtains a trained global model, and distributes the trained global model to the m clients; through the m clients, receive the trained global model; and through the m clients, fine-tune the training of the local models corresponding to the m clients based on the trained global model to obtain local approval models corresponding to the m clients.

[0161] According to embodiments of this application, any multiple modules among the data acquisition module 1010, business approval module 1020, result return module 1030, and federated training module can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 1010, business approval module 1020, result return module 1030, and federated training module can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuits, or implemented in hardware or firmware, or in any one of software, hardware, and firmware implementations, or in a suitable combination of any of these. Alternatively, at least one of the data acquisition module 1010, business approval module 1020, result return module 1030, and federated training module can be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0162] Based on the above-described federated learning-based business approval method, embodiments of this application also provide a federated learning-based business approval device, applied to m clients, where m is an integer greater than 1. The following will combine... Figure 11 The device is described in detail.

[0163] Figure 11 The diagram illustrates a client-side structural block diagram of a business approval apparatus based on federated learning according to an embodiment of this application.

[0164] like Figure 11 As shown, the business approval device 1100 based on federated learning in this embodiment includes a first global model receiving module 1110, a local model training and noise injection module 1120, a local model parameter uploading module 1130, a second global model receiving module 1140, and a local model fine-tuning training module 1150.

[0165] The first global model receiving module 1110 is used to receive the global model sent by the server. In one embodiment, the first global model receiving module 1110 can be used to execute the step S310 described above, which will not be repeated here.

[0166] The local model training and noise injection module 1120 is used for a dynamic adjustment strategy based on privacy perturbation parameters. It performs local model training and noise injection based on the local data of the m clients and the global model to obtain the local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients. The privacy perturbation parameters include differential privacy noise and perturbation radius. In one embodiment, the local model training and noise injection module 1120 can be used to execute step S320 described above, which will not be repeated here.

[0167] The local model parameter upload module 1130 is used to upload the local model parameters corresponding to the m clients to the server; so that the server receives the local model parameters corresponding to the m clients, performs curvature aggregation and parameter updates on the global model based on the confidence weight and the local model parameters corresponding to the m clients, obtains the trained global model, and distributes the trained global model to the m clients. In one embodiment, the local model parameter upload module 1130 can be used to execute steps S330~S340 described above, which will not be repeated here.

[0168] The second global model receiving module 1140 is used to receive the trained global model. In one embodiment, the second global model receiving module 1140 can be used to execute step S350 described above, which will not be repeated here.

[0169] The local model fine-tuning training module 1150 is used to fine-tune the local models corresponding to the m clients based on the trained global model, thereby obtaining local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve user business applications. In one embodiment, the local model fine-tuning training module 1150 can be used to execute step S360 described above, which will not be repeated here.

[0170] According to an embodiment of this application, the local model training and noise injection module 1120 is further configured to construct a local model for the current client based on the global model; employ a sharpness-aware minimization method to iteratively train the local model for the current client based on the local data of the current client and the perturbation radius; during the iterative training process, based on the dynamic adjustment strategy of the privacy perturbation parameters, the privacy perturbation parameters are adjusted in stages according to the variance of the loss function, and the differential privacy noise is injected into the local model for the current client.

[0171] According to an embodiment of this application, the dynamic adjustment strategy for the privacy perturbation parameter includes: in response to the loss function variance satisfying the initial training conditions, setting the differential privacy noise to the maximum noise intensity value and setting the perturbation radius to the minimum perturbation radius value; in response to the loss function variance satisfying the mid-training conditions, performing exponential decay adjustment on the differential privacy noise based on the noise exponential decay coefficient, the maximum noise intensity value, and the minimum noise intensity value, and performing linear increase adjustment on the perturbation radius based on the perturbation linear growth factor, the minimum perturbation radius value, and the maximum perturbation radius value; in response to the loss function variance satisfying the late-training conditions, calculating the suppressed noise value based on the noise suppression coefficient and the minimum noise intensity value, setting the differential privacy noise to the suppressed noise value, and setting the perturbation radius to the maximum perturbation radius value.

[0172] According to an embodiment of this application, the initial training condition is that the variance of the loss function is greater than or equal to a first variance threshold; the intermediate training condition is that the variance of the loss function is greater than or equal to a second variance threshold and less than the first variance threshold, or that the variance of the loss function is less than the second variance threshold and the current training round number is less than the later starting round number; the later training condition is that the variance of the loss function is less than the second variance threshold and the current training round number is greater than or equal to the later starting round number.

[0173] According to an embodiment of this application, the local model training and noise injection module 1120 is further configured to: calculate the original gradient based on the local data of the current client and the local model parameters of the local model; perform gradient clipping on the original gradient based on a clipping threshold to obtain a local gradient; calculate the noise variance based on the differential privacy noise, the clipping threshold and the identity matrix, and determine Gaussian noise based on the noise variance and a preset noise mean; and add the Gaussian noise to the local gradient.

[0174] According to an embodiment of this application, the local model fine-tuning training module 1150 is further configured to: construct a regularization term based on the global model parameters corresponding to the trained global model, the local model parameters of the current client, and the personalized intensity parameters; minimize the loss value of the local model parameters of the current client; construct the local objective function of the current client using the regularization term as a constraint; and fine-tune and update the local model of the current client according to the local data of the current client and the local objective function, until the local model of the current client converges.

[0175] According to embodiments of this application, any multiple modules among the first global model receiving module 1110, local model training and noise injection module 1120, local model parameter uploading module 1130, second global model receiving module 1140, and local model fine-tuning training module 1150 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the first global model receiving module 1110, local model training and noise injection module 1120, local model parameter uploading module 1130, second global model receiving module 1140, and local model fine-tuning training module 1150 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, and firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first global model receiving module 1110, the local model training and noise injection module 1120, the local model parameter uploading module 1130, the second global model receiving module 1140, and the local model fine-tuning training module 1150 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0176] Based on the above-described federated learning-based business approval method, embodiments of this application also provide a federated learning-based business approval device applied to a server. The following will combine... Figure 12 The device is described in detail.

[0177] Figure 12 The diagram illustrates a client-side structural block diagram of a business approval apparatus based on federated learning according to an embodiment of this application.

[0178] like Figure 12 As shown, the business approval device 1300 based on federated learning in this embodiment includes a first global model distribution module 1310, a local model parameter receiving module 1320, a curvature aggregation and parameter update module 1330, and a second global model distribution module 1340.

[0179] The first global model distribution module 1310 is used to distribute a global model to m clients, so that the m clients can receive the global model distributed by the server; based on a dynamic adjustment strategy of privacy perturbation parameters, local model training and noise injection are performed according to the local data of the m clients and the global model to obtain the local models corresponding to the m clients, and the parameters of the local models corresponding to the m clients are determined; the parameters of the local models corresponding to the m clients are uploaded to the server; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius. In one embodiment, the first global model distribution module 1310 can be used to execute steps S510~S520 described above, which will not be repeated here.

[0180] The local model parameter receiving module 1320 is used to receive the local model parameters corresponding to the m clients. In one embodiment, the local model parameter receiving module 1320 can be used to execute step S530 described above, which will not be repeated here.

[0181] The curvature aggregation and parameter update module 1330 is used to perform curvature aggregation and parameter update on the global model based on the confidence weights and the local model parameters corresponding to the m clients, to obtain the trained global model. In one embodiment, the curvature aggregation and parameter update module 1330 can be used to execute step S540 described above, which will not be repeated here.

[0182] The second global model distribution module 1340 is used to distribute the trained global model to the m clients, so that the m clients receive the trained global model; based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve user business applications. In one embodiment, the second global model distribution module 1340 can be used to execute steps S550~S560 described above, which will not be repeated here.

[0183] According to an embodiment of this application, the curvature aggregation and parameter update module 1330 is further configured to: calculate the gradient variance of the m clients based on the local model parameters corresponding to the m clients; calculate the confidence weights of the m clients based on the sensitivity coefficient and the gradient variances of the m clients; calculate the diagonal norm of the Fisher information matrix based on the local model parameters corresponding to the m clients to obtain the model update curvature of the m clients; perform weighted aggregation on the local model parameters corresponding to the m clients based on the confidence weights of the m clients and the model update curvature to obtain the global gradient update amount; and update the global model parameters of the global model based on the global gradient update amount.

[0184] According to embodiments of this application, any multiple modules among the first global model distribution module 1310, the local model parameter receiving module 1320, the curvature aggregation and parameter update module 1330, and the second global model distribution module 1340 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the first global model distribution module 1310, the local model parameter receiving module 1320, the curvature aggregation and parameter update module 1330, and the second global model distribution module 1340 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, and firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first global model distribution module 1310, the local model parameter receiving module 1320, the curvature aggregation and parameter update module 1330, and the second global model distribution module 1340 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0185] Figure 13 A block diagram schematically illustrates an electronic device suitable for implementing a federated learning-based business approval method according to an embodiment of this application.

[0186] like Figure 13 As shown, an electronic device 1200 according to an embodiment of this application includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory 1202 or a program loaded from a storage portion 1208 into a random access memory 1203. The processor 1201 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 1201 may also include onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0187] Random access memory 1203 stores various programs and data required for the operation of electronic device 1200. Processor 1201, read-only memory 1202, and random access memory 1203 are interconnected via bus 1204. Processor 1201 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 1202 and / or random access memory 1203. It should be noted that the programs may also be stored in one or more memories other than read-only memory 1202 and random access memory 1203. Processor 1201 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0188] According to embodiments of this application, the electronic device 1200 may further include an input / output interface 1205, which is also connected to the bus 1204. The electronic device 1200 may also include one or more of the following components connected to the input / output interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card, such as a local area network card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the input / output interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1210 as needed so that computer programs read from it can be installed into the storage section 1208 as needed.

[0189] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0190] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 1202, and / or random access memory 1203, and / or one or more memories other than read-only memory 1202 and random access memory 1203 described above.

[0191] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0192] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1209, and / or installed from the removable medium 1211. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0193] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by processor 1201, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0194] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0195] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0196] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A business approval method based on federated learning, characterized in that, Applied to the target client, the method includes: In response to a user submitting a service request, obtain the service behavior data associated with the service request; The local approval model corresponding to the business application is invoked to process the business behavior data and obtain the business approval result. The local approval model is obtained by federated learning of the global model and the local model based on historical business data, using m clients and servers and a dynamic adjustment strategy based on privacy perturbation parameters. Here, m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius. Return the business approval result to the user.

2. The method according to claim 1, characterized in that, The method involves using m clients and servers, and a dynamic adjustment strategy based on privacy perturbation parameters to perform federated learning training on the global and local models based on historical business data. This includes: The m clients receive the global model sent by the server. Using the m clients, a dynamic adjustment strategy based on privacy perturbation parameters is employed. Local model training and noise injection are performed based on the local data of the m clients and the global model to obtain the local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients. The local data of the m clients belongs to the historical business data. The m clients upload their local model parameters to the server; the server receives the local model parameters from the m clients, performs curvature aggregation and parameter updates on the global model based on the confidence weight and the local model parameters from the m clients, obtains a trained global model, and distributes the trained global model to the m clients. The trained global model is received through the m clients; and Using the m clients, and based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients.

3. A business approval method based on federated learning, characterized in that, The method, applied to m clients where m is an integer greater than 1, includes: Receive the global model sent by the server; A dynamic adjustment strategy based on privacy perturbation parameters is adopted, which involves training local models and injecting noise based on the local data of the m clients and the global model to obtain the local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius; Upload the local model parameters corresponding to the m clients to the server; so that the server receives the local model parameters corresponding to the m clients, performs curvature aggregation and parameter update on the global model based on the confidence weight and the local model parameters corresponding to the m clients, obtains the trained global model, and distributes the trained global model to the m clients; Receive the trained global model; and Based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve the user's business applications.

4. The method according to claim 3, characterized in that, The dynamic adjustment strategy based on privacy perturbation parameters, which performs local model training and noise injection based on the local data of the m clients and the global model, includes: Based on the global model, construct the local model for the current client; The sharpness-aware minimization method is used to iteratively train the local model of the current client based on the local data of the current client and the perturbation radius; During iterative training, based on the dynamic adjustment strategy of the privacy perturbation parameters, the privacy perturbation parameters are adjusted in stages according to the variance of the loss function, and the differential privacy noise is injected into the local model of the current client.

5. The method according to claim 4, characterized in that, The dynamic adjustment strategy for the privacy perturbation parameters includes: In response to the loss function variance satisfying the initial training conditions, the differential privacy noise is set to the maximum noise intensity value, and the perturbation radius is set to the minimum perturbation radius value; In response to the loss function variance satisfying the mid-training condition, the differential privacy noise is adjusted by exponential decay based on the noise exponential decay coefficient, the maximum noise intensity value, and the minimum noise intensity value, and the perturbation radius is adjusted by linear increase based on the perturbation linear growth factor, the minimum perturbation radius value, and the maximum perturbation radius value; In response to the loss function variance satisfying the late-stage training condition, a suppressed noise value is calculated based on the noise suppression coefficient and the minimum noise intensity value. The differential privacy noise is set as the suppressed noise value, and the perturbation radius is set as the maximum perturbation radius value.

6. The method according to claim 5, characterized in that, The initial training condition is that the variance of the loss function is greater than or equal to the first variance threshold. The mid-training condition is that the variance of the loss function is greater than or equal to the second variance threshold and less than the first variance threshold, or that the variance of the loss function is less than the second variance threshold and the current training round number is less than the later starting round number. The late training condition is that the variance of the loss function is less than the second variance threshold and the current training round number is greater than or equal to the late training starting round number.

7. The method according to claim 4, characterized in that, The step of injecting the differential privacy noise into the local model of the current client includes: Calculate the original gradient based on the local data of the current client and the local model parameters of the local model; Based on the clipping threshold, the original gradient is clipped to obtain the local gradient; Based on the differential privacy noise, the pruning threshold, and the identity matrix, the noise variance is calculated, and Gaussian noise is determined based on the noise variance and a preset noise mean; and Add the Gaussian noise to the local gradient.

8. The method according to claim 4, characterized in that, The step of fine-tuning the local models corresponding to the m clients based on the trained global model includes: Based on the global model parameters corresponding to the trained global model, the local model parameters of the current client, and the personalization intensity parameters, a regularization term is constructed; Minimize the loss value of the local model parameters of the current client, and construct the local objective function of the current client using the regularization term as a constraint; and Based on the local data of the current client and the local objective function, the local model of the current client is fine-tuned and its parameters are updated until the local model of the current client converges.

9. A business approval method based on federated learning, characterized in that, Applied to a server, the method includes: A global model is distributed to m clients so that the m clients receive the global model distributed by the server; based on a dynamic adjustment strategy of privacy perturbation parameters, local model training and noise injection are performed according to the local data of the m clients and the global model to obtain the local models corresponding to the m clients, and the parameters of the local models corresponding to the m clients are determined; the parameters of the local models corresponding to the m clients are uploaded to the server; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius; Receive the local model parameters corresponding to the m clients; Based on the credibility weights and the local model parameters corresponding to the m clients, the global model is subjected to curvature aggregation and parameter updates to obtain a trained global model. The trained global model is distributed to the m clients so that the m clients receive the trained global model; based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to approve the user's business applications.

10. The method according to claim 9, characterized in that, The curvature aggregation and parameter update of the global model based on the credibility weight and the local model parameters corresponding to the m clients includes: Based on the local model parameters corresponding to the m clients, calculate the gradient variance of the m clients, and calculate the credibility weight of the m clients based on the sensitivity coefficient and the gradient variance of the m clients. Based on the local model parameters corresponding to the m clients, calculate the diagonal norm of the Fisher information matrix to obtain the model update curvature of the m clients; Based on the credibility weights of the m clients and the model update curvature, the local model parameters corresponding to the m clients are weighted and aggregated to obtain the global gradient update amount; and Based on the global gradient update amount, update the global model parameters of the global model.

11. A business approval device based on federated learning, characterized in that, Applied to a target client, the device includes: The data acquisition module is used to acquire business behavior data associated with the business application in response to the user's submission of a business application; The business approval module is used to call the local approval model corresponding to the business application to process the business behavior data and obtain the business approval result. The local approval model is obtained by federated learning of the global model and the local model based on historical business data, using m clients and servers and a dynamic adjustment strategy based on privacy perturbation parameters. Here, m is an integer greater than 1, the m clients include the target client, and the privacy perturbation parameters include differential privacy noise and perturbation radius. The result return module is used to return the business approval result to the user.

12. A business approval device based on federated learning, characterized in that, The device, applicable to m clients where m is an integer greater than 1, comprises: The first global model receiving module is used to receive the global model sent by the server; The local model training and noise injection module is used for dynamic adjustment strategies based on privacy perturbation parameters. It performs local model training and noise injection based on the local data of the m clients and the global model to obtain the local models corresponding to the m clients and determine the parameters of the local models corresponding to the m clients. The privacy perturbation parameters include differential privacy noise and perturbation radius. The local model parameter upload module is used to upload the local model parameters corresponding to the m clients to the server; so that the server receives the local model parameters corresponding to the m clients, performs curvature aggregation and parameter update on the global model based on the confidence weight and the local model parameters corresponding to the m clients, obtains the trained global model, and distributes the trained global model to the m clients. The second global model receiving module is used to receive the trained global model; and The local model fine-tuning training module is used to fine-tune the local models corresponding to the m clients based on the trained global model, so as to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to conduct business approval on the user's business applications.

13. A business approval device based on federated learning, characterized in that, Applied to a server, the device includes: The first global model distribution module is used to distribute a global model to m clients, so that the m clients can receive the global model distributed by the server; based on a dynamic adjustment strategy of privacy perturbation parameters, local model training and noise injection are performed according to the local data of the m clients and the global model to obtain the local models corresponding to the m clients, and the parameters of the local models corresponding to the m clients are determined; the local model parameters corresponding to the m clients are uploaded to the server; wherein, the privacy perturbation parameters include differential privacy noise and perturbation radius; The local model parameter receiving module is used to receive the local model parameters corresponding to the m clients; The curvature aggregation and parameter update module is used to perform curvature aggregation and parameter update on the global model based on the confidence weights and the local model parameters corresponding to the m clients, to obtain a trained global model; and The second global model distribution module is used to distribute the trained global model to the m clients so that the m clients can receive the trained global model; based on the trained global model, the local models corresponding to the m clients are fine-tuned and trained to obtain the local approval models corresponding to the m clients; wherein, the local approval models of the m clients are used by the m clients to conduct business approval on the user's business applications.

14. An electronic device comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 10.

15. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

16. A computer program product comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

Images