Financial fraud detection model training method, system, device, medium and equipment

Abstract

The application discloses a financial fraud detection model training method, system, device, medium and equipment. A central server constructs an initial financial fraud detection model, each client constructs a local detection model with the same structure as the initial financial fraud detection model locally, divides the local detection model into a shared layer and a personalized layer, and trains the local detection model by using local data. A client uploads part of shared layer parameters to the central server, the central server aggregates the received shared layer parameters uploaded by each client according to layers, obtains updated shared layer parameters, and distributes the updated shared layer parameters to each client. Each client splices the updated shared layer parameters and local personalized layer parameters, updates model parameters, and iteratively trains to obtain a final financial fraud detection model. The application avoids data leakage risks and improves the accuracy and robustness of financial fraud detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and financial technology, and in particular to a financial fraud detection model training method, system, device, medium and equipment. BACKGROUND

[0002] Financial fraud detection models are used to identify and intercept potential fraudulent transactions by modeling and analyzing financial transaction data, with the aim of protecting the financial security of financial institutions and users, and maintaining the stability and compliance of the financial system. Current financial fraud detection methods are mainly divided into three types: rule-based, traditional machine learning and deep learning.

[0003] Rule-based methods quickly identify fraudulent behavior by setting thresholds, blacklists, and transaction amount monitoring, but lack flexibility and are difficult to deal with complex and changing fraud methods. Traditional machine learning-based methods, such as logistic regression and decision trees, can learn fraud patterns from data, but have limited processing capacity for high-dimensional sparse data and rely on the quality of feature engineering. Deep learning-based methods use techniques such as graph neural networks and Transformers to automatically extract deep features from transaction data, significantly improving detection accuracy, but face privacy protection and compliance challenges.

[0004] Although deep learning can improve the performance of financial fraud detection, financial transaction data involves sensitive information such as account identity and consumption habits, and is subject to legal and regulatory requirements, so financial institutions cannot share raw data, which not only limits the coverage of the data, but also makes the model prone to bias and overfitting, and in practical applications, the model lacks flexibility in adapting to different business scenarios, further restricting the effectiveness of the detection. SUMMARY

[0005] Therefore, it is necessary to propose a financial fraud detection model training method, system, device, medium and equipment to address the problems of limited detection accuracy and insufficient scene adaptability of existing financial fraud detection models under data privacy constraints.

[0006] The first aspect of the present application provides a financial fraud detection model training method, which is executed on a central server, comprising:

[0007] An initial financial fraud detection model and a server super network are constructed on the central server, and the model parameters of the initial financial fraud detection model and the initial parameters of the server super network are sent to a plurality of clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data;

[0008] receive the shared layer parameters uploaded by each of the clients, wherein the shared layer parameters are selected by each client after training a local detection model using local data, calculating upload weights of each shared layer parameter through a client super network, and filtering the shared layer parameters to be uploaded;

[0009] aggregate the received shared layer parameters uploaded by each of the clients by layer to obtain updated shared layer parameters;

[0010] distribute the updated shared layer parameters to each of the clients, so that each of the clients splices the updated shared layer parameters with local personalized layer parameters to obtain updated model parameters and construct a local detection model for the next round of training;

[0011] re-enter the step of receiving the shared layer parameters uploaded by each of the clients, and drive each of the clients to iteratively train the local detection model until a preset convergence condition is met and the training is terminated, and determine the local detection model obtained by each client at the time of convergence as the final financial fraud detection model of the corresponding client.

[0012] Further, the central server constructs an initial financial fraud detection model and a server super network, and sends model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters locally, and divides the local detection model into a shared layer and a personalized layer, and the step of training the local detection model by each client using local data comprises:

[0013] construct an initial financial fraud detection model and a server super network at the central server end to obtain structure data of the initial financial fraud detection model and structure data of the server super network; wherein the initial financial fraud detection model comprises an embedding layer, an encoder layer, a residual layer, a normalization layer and an output layer connected in turn;

[0014] initialize the weight parameters of the initial financial fraud detection model and the initial parameters of the server super network;

[0015] pack the structure data, weight parameters of the initial financial fraud detection model and the structure data, and initial parameters of the server super network to obtain model parameters;

[0016] The model parameters are distributed to each of the clients; after each of the clients receives the model parameters of the initial financial fraud detection model, a local detection model with the same structure as the initial financial fraud detection model is constructed locally based on the structure data and weight parameters of the initial financial fraud detection model, a shared layer and a personalized layer are constructed for the local detection model of each client, and each client instantiates a client super network locally based on the structure data and initial parameters of the server super network; each client trains the local detection model using local data and calculates upload weights of each shared layer parameter using the client super network, wherein the shared layer parameters are selectively uploaded to the central server during training, and the personalized layer parameters are only iteratively updated locally during training.

[0017] Further, the step of aggregating the received shared layer parameters uploaded by each of the clients by layer to obtain updated shared layer parameters comprises:

[0018] After receiving the shared layer parameters uploaded by all the clients, the same shared layers uploaded by different clients are classified according to the identification of the shared layers;

[0019] The classified same shared layer parameters are processed by layer-by-layer aggregation through a preset aggregation algorithm, and the parameters of the corresponding layer are kept unchanged for the layer for which no client uploaded parameters;

[0020] The parameter sets corresponding to all the shared layers are integrated, and the parameter sets are taken as the updated shared layer parameters.

[0021] The second aspect of the application provides a financial fraud detection model training method, which is executed on a client and comprises:

[0022] The model parameters of an initial financial fraud detection model and initial parameters of a server super network sent by a central server are obtained, and a local detection model with the same structure as the initial financial fraud detection model is constructed locally based on the model parameters, and a client super network is instantiated locally based on the initial parameters of the server super network;

[0023] The local detection model is divided into a shared layer and a personalized layer;

[0024] The local detection model is trained using local data, and the model parameters of the local detection model are updated by calculating loss and performing back propagation;

[0025] After each round of training, the uploading weights of the shared layer parameters in the local detection model are calculated by the client super network, the shared layer parameters to be uploaded to the central server are screened out, and the screened shared layer parameters are uploaded to the central server, and the central server aggregates the shared layer parameters uploaded by each client to obtain updated shared layer parameters;

[0026] The updated shared layer parameters issued by the central server are received, the updated shared layer parameters are spliced with the local personalized layer parameters to obtain updated model parameters, and a local detection model for the next round of training is constructed;

[0027] The local detection model is iteratively trained until a preset convergence condition is met and the training is terminated, and the local detection model obtained at the time of convergence is determined as the final financial fraud detection model used by the local client.

[0028] Further, the step of dividing the local detection model into shared layers and personalized layers comprises:

[0029] determining the layer position c of the model segmentation point based on a preset division strategy;

[0030] the first layer to the c-th layer in the local detection model are determined as shared layers, and the shared layers are represented as , wherein is the i-th client, and the shared layer parameters are selectively uploaded to the central server in training;

[0031] the c+1-th layer to the L-th layer in the local detection model are determined as personalized layers, and the personalized layers are represented as , and the personalized layer parameters are only locally iteratively updated in training.

[0032] Further, the step of, after each round of training, calculating the uploading weights of the shared layer parameters in the local detection model by the client super network, screening out the shared layer parameters to be uploaded to the central server, uploading the screened shared layer parameters to the central server, and aggregating the shared layer parameters uploaded by each client by the central server to obtain updated shared layer parameters, comprises:

[0033] After each round of training, the parameters of each shared layer in the local detection model are inputted, and the uploading weights of each shared layer are calculated by the client super network;

[0034] It is respectively judged whether the uploading weight of each shared layer exceeds a preset dynamic threshold;

[0035] If yes, the shared layer parameters whose upload weights are higher than the preset dynamic threshold are uploaded to the central server for aggregation; the central server aggregates the shared layer parameters uploaded by each client to obtain updated shared layer parameters.

[0036] The third aspect of the present application provides a financial fraud detection model training system, comprising a central server and a plurality of clients.

[0037] The central server is configured to build an initial financial fraud detection model and a server super network, and send model parameters of the initial financial fraud detection model and initial parameters of the server super network to the plurality of clients.

[0038] Each client obtains the model parameters of the initial financial fraud detection model and the initial parameters of the server super network sent by the central server, builds a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and instantiates a client super network based on the initial parameters of the server super network; the local detection model is divided into a shared layer and a personalized layer; the local detection model is trained using local data, the model parameters of the local detection model are updated by calculating loss and performing back propagation; after each round of training, the upload weights of each shared layer parameter in the local detection model are calculated by the client super network, the shared layer parameters to be uploaded to the central server are screened out, and the screened shared layer parameters are uploaded to the central server.

[0039] The central server receives the shared layer parameters uploaded by each client; aggregates the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters; and sends the updated shared layer parameters to each client.

[0040] Each client receives the updated shared layer parameters sent by the central server, splices the updated shared layer parameters with the local personalized layer parameters to obtain updated model parameters, and builds a local detection model for the next round of training.

[0041] The central server re-enters the step of receiving the shared layer parameters uploaded by each client, and each client iteratively trains the local detection model until a preset convergence condition is met and the training is terminated, and each client determines the local detection model obtained at convergence as the final financial fraud detection model used by the local client.

[0042] The fourth aspect of the present application provides a financial fraud detection model training device, which is arranged on a central server and comprises:

[0043] An initial model building module is configured to build an initial financial fraud detection model and a server super network on the central server, and send model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; wherein each client builds a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data.

[0044] A data receiving module is configured to receive shared layer parameters uploaded by each client, wherein the shared layer parameters are selected by each client to upload after uploading weights of each shared layer parameter calculated by a client super network after training a local detection model using local data.

[0045] A data aggregation module is configured to aggregate the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters.

[0046] A data issuing module is configured to issue the updated shared layer parameters to each client, so that each client splices the updated shared layer parameters with local personalized layer parameters to obtain updated model parameters and build a local detection model for the next round of training.

[0047] An iterative training module is configured to re-receive shared layer parameters uploaded by each client, drive each client to perform iterative training on the local detection model until a preset convergence condition is met and the training is terminated, and determine the local detection model obtained by each client when converging as a final financial fraud detection model of the corresponding client.

[0048] A fifth aspect of the present application provides a computer readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above financial fraud detection model training method.

[0049] A sixth aspect of the present application provides a computer device, which comprises a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the above financial fraud detection model training method.

[0050] The financial fraud detection model training method, system, device, medium and equipment of the application, the server issues an initial model and super network parameters, does not collect client original data, clearly retains client data locally, only receives filtered shared layer parameters, thereby avoiding the leakage risk of cross-end transmission of sensitive financial data; the server initially and uniformly issues the same financial fraud detection model and super network parameters, converts dispersed client local knowledge into globally optimal shared features by layer-aggregated shared layer parameters uploaded by each client, realizes server-centered global optimization of parameters, and avoids deviation of the client training direction; the server indirectly dominates the filtering standard of the shared layer parameters through pre-constructed super network parameters, so that the client only uploads parameters critical to the global model improvement, greatly reduces the data volume of cross-end transmission, reduces communication overhead, effectively fuses multiple client knowledge to form global shared features, while retaining the personalized model characteristics of the client, so that the trained model has global generalization ability and can adapt to different client business scenarios, improving the accuracy and robustness of financial fraud detection. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Among them:

[0053] Figure 1 The flowchart of the financial fraud detection model training method in one embodiment;

[0054] Figure 2 The flowchart of the financial fraud detection model training method in another embodiment;

[0055] Figure 3 The structural schematic diagram of the financial fraud detection model training system in one embodiment;

[0056] Figure 4 The structural block diagram of the financial fraud detection model training device in one embodiment;

[0057] Figure 5 The structural block diagram of the computer device in one embodiment. DETAILED DESCRIPTION

[0058] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0059] As shown in Figure 1 In one embodiment, a financial fraud detection model training method is provided, which is executed on a central server, comprising:

[0060] S1: constructing an initial financial fraud detection model and a server super network on the central server, and sending model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters locally, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data;

[0061] S2: receiving shared layer parameters uploaded by each client, wherein the shared layer parameters are selected to be uploaded by each client after training the local detection model using local data, and uploading the shared layer parameters by the client super network to calculate the uploading weight of each shared layer parameter;

[0062] S3: aggregating the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters;

[0063] S4: distributing the updated shared layer parameters to each client, so that each client splices the updated shared layer parameters with the local personalized layer parameters to obtain updated model parameters and construct a local detection model for the next round of training;

[0064] S5: re-entering the step of receiving the shared layer parameters uploaded by each client, driving each client to iteratively train the local detection model, until a preset convergence condition is met and the training is terminated, and determining the local detection model obtained by each client at the time of convergence as the final financial fraud detection model of the corresponding client.

[0065] In the above step S1, the initial financial fraud detection model adopts a deep learning architecture including an embedding layer, an encoder layer, a residual layer, a normalization layer, and an output layer. The embedding layer is responsible for converting discrete financial features into continuous vectors. The encoder layer captures complex correlations between features. The residual layer alleviates gradient vanishing in deep networks. The normalization layer stabilizes the training process. The output layer outputs fraud prediction results. The server super network can be a small neural network used for subsequent learning parameter screening strategies. The server initializes the weight parameters of the initial financial fraud detection model and the parameters of the server super network. Random initialization can be used, or pre-trained parameters based on historical financial data can be used to accelerate convergence. Subsequently, the structure data, weight parameters of the initial financial fraud detection model, and the structure data, initial parameters of the server super network are packaged into a data packet and distributed to multiple clients. After receiving the data packet, the client constructs a local detection model based on the model structure data and weight parameters in the data packet. The client instantiates a client super network based on the structure data and initial parameters of the super network. The client divides the local detection model into a shared layer and a personalized layer. The shared layer is used to learn fraud features common across clients, and the personalized layer is used to learn fraud patterns specific to the client's scenario. After the division is completed, the client trains the local detection model using local financial data. By calculating loss functions such as cross-entropy and executing backpropagation algorithms, the client updates all parameters in the shared layer and the personalized layer of the local detection model.

[0066] In the above step S2, after each round of local training at the client, the client evaluates the parameters of each layer of the shared layer of the local detection model using the locally instantiated client super network. The client super network takes the shared layer parameters as input and calculates upload weights for each layer. The upload weight reflects the contribution of the corresponding shared layer parameter to the performance of the global model. Then, the client filters out shared layer parameters with upload weights greater than the threshold and uploads them to the central server according to the pre-set dynamic threshold. Shared layer parameters with upload weights less than the threshold are not uploaded in this round, effectively reducing communication overhead and preserving local sensitive information, improving personalized training effectiveness. While ensuring cross-client knowledge sharing, the model's adaptability and performance in different client scenarios are improved.

[0067] In the above step S3, after receiving the shared layer parameters uploaded by different clients, the central server classifies the same layer parameters uploaded by different clients according to the shared layer identifier. Then, the server processes the classified same shared layer parameters using aggregation algorithms such as weighted average or sample quantity-based weighted average. For shared layers without client uploaded parameters in this round, the server keeps their parameters unchanged from the previous round. Finally, the global updated shared layer parameters are integrated.

[0068] In step S4, the central server distributes the globally updated shared layer parameters to all clients. Each client replaces the original shared layer parameters in the local detection model with the new shared layer parameters, and then splices the updated shared layer parameters with the locally retained personalized layer parameters to form a complete local detection model for the next round of local training.

[0069] In step S5, the client trains again using the new local detection model assembled in step S4, filters and uploads the shared layer parameters through the client super network, and the central server reenters the process of step S2. The central server continues to perform the aggregation operation of step S3 and the parameter distribution of step S4, and thus drives the iterative training of the local detection model of each client. In the iteration process, it is monitored in real time whether the preset convergence condition is met. The convergence condition can include, for example, that the central server evaluates the model performance (such as fraud detection accuracy) of the globally shared layer on the independent validation set, and the performance improvement amplitude of consecutive multiple rounds is lower than a preset threshold; or the training iteration round reaches a preset maximum round; or the local training loss of each client does not decrease significantly for consecutive multiple rounds. When any of the preset convergence conditions is met, the training is terminated. At this time, the local detection model of each client, which is composed of the final global shared layer parameters and the personalized layer parameters optimized by the client, is determined as the final financial fraud detection model for the actual scene of the corresponding client.

[0070] The method of the embodiment avoids the risk of leakage of sensitive financial data across terminals by having the server distribute the initial model and the super network parameters, not collecting the original data of the clients, explicitly retaining the data of the clients locally, and only receiving the filtered shared layer parameters. The server initially distributes the same financial fraud detection model and super network parameters, converts the dispersed local knowledge of the clients into globally optimal shared features by aggregating the shared layer parameters uploaded by each client by layer, realizes global optimization of parameters centered on the server, and avoids deviation of the training direction of the clients. The server indirectly dominates the filtering standard of the shared layer parameters through the pre-constructed super network parameters, so that the clients only upload parameters that are critical to the improvement of the global model, greatly reduces the amount of data transmitted across terminals, reduces communication overhead, effectively fuses the knowledge of multiple clients to form global shared features, while retaining the personalized model characteristics of the clients, so that the trained model has both global generalization ability and adaptability to different business scenarios of the clients, and improves the accuracy and robustness of financial fraud detection.

[0071] In some embodiments, the central server constructs an initial financial fraud detection model and a server super network, and sends model parameters of the initial financial fraud detection model and initial parameters of the server super network to multiple clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, divides the local detection model into a shared layer and a personalized layer, and trains the local detection model using local data.

[0072] S101: Constructing an initial financial fraud detection model and a server super network at the central server end to obtain structure data of the initial financial fraud detection model and structure data of the server super network; wherein the initial financial fraud detection model comprises an embedding layer, an encoder layer, a residual layer, a normalization layer and an output layer connected in turn;

[0073] S102: Initializing weight parameters of the initial financial fraud detection model and initial parameters of the server super network;

[0074] S103: Packaging the structure data, weight parameters of the initial financial fraud detection model and the structure data, initial parameters of the server super network to obtain model parameters;

[0075] S104: Distributing the model parameters to each client; after receiving the model parameters of the initial financial fraud detection model, each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the structure data and weight parameters of the initial financial fraud detection model, constructs a shared layer and a personalized layer for the local detection model of each client, and instantiates a client super network locally based on the structure data and initial parameters of the server super network; each client trains the local detection model using local data and calculates upload weights of each shared layer parameter using the client super network, wherein the shared layer parameters are selectively uploaded to the central server during training, and the personalized layer parameters are only iteratively updated locally during training.

[0076] In this embodiment, in step S101, the central server first completes the architecture design and construction of the initial financial fraud detection model and the server super network on its side. The initial financial fraud detection model adopts a deep learning architecture including an embedding layer, an encoder layer, a residual layer, a normalization layer, and an output layer connected in sequence. The embedding layer is responsible for converting discrete financial features into continuous low-dimensional vector representations; the encoder layer (e.g., a Transformer encoder) is used to capture long-distance dependencies and complex associations between features; the residual layer alleviates the gradient vanishing problem in deep network training through a skip connection; the normalization layer is used to standardize the input of each layer, accelerate model convergence, and improve training stability; the output layer outputs fraud probability or classification results. At the same time, the server constructs a server super network, which is usually a lightweight neural network. The core function of the server super network is to learn how to evaluate the importance of shared layer parameters, providing a basis for subsequent parameter selection strategies.

[0077] In step S102, the server initializes all weight parameters of the constructed initial financial fraud detection model and the parameters of the server super network. The initialization method can be random initialization; or the parameters obtained by pre-training on a large-scale public or historical financial dataset can be used for initialization. During the initialization process, the server also defines the division boundary between the shared layer and the personalized layer, for example, the first c layers of the model are defined as the shared layer, which is responsible for learning fraud features common to all clients; the c+1 layer to the output layer is the personalized layer, which is used to adapt to the local scene of the client. This division rule is packaged together with the structure data of the initial model and then distributed to the client, so that the client can logically divide its local detection model into a shared layer and a personalized layer when constructing the model locally.

[0078] In step S103, the server packs and integrates the structure data (including input dimension, residual block number, activation function type, etc.) of the initial financial fraud detection model and its initialized weight parameters, as well as the structure data of the server super network and its initial parameters, into a complete model parameter data package.

[0079] In step S104 above, the server sends the packaged model parameter data packet to each client participating in the training. Upon receiving the data packet, each client, based on the initial financial fraud detection model structure data and weight parameters contained within, constructs a local detection model on its local hardware environment (such as a PC or edge server) that is completely identical to the server's initial model structure. Simultaneously, the client instantiates a client-side hypernetwork locally based on the server hypernetwork structure data and initial parameters contained in the data packet. Subsequently, the client logically divides its local detection model into a shared layer and a personalized layer: the shared layer is used to learn fraud patterns common across clients, and the personalized layer is used to learn fraud features specific to the client's local environment. After completing model construction and partitioning, the client uses its local financial transaction data to perform the first round of training on the local detection model, updating all model parameters by calculating the loss function and performing backpropagation. After training, the client starts the locally instantiated client-side hypernetwork, calculates the upload weights of the shared layer parameters in the current local detection model, and prepares for the next round of parameter selection and upload. During this process, shared layer parameters will be selectively uploaded to the server in subsequent rounds, while personalized layer parameters will always be iteratively updated locally and will not participate in global upload and aggregation.

[0080] In some specific embodiments, step S3, which involves aggregating the received shared layer parameters uploaded by each of the clients layer by layer to obtain the updated shared layer parameters, includes:

[0081] S301: After receiving the shared layer parameters uploaded by all clients, classify the same shared layer uploaded by different clients according to the shared layer identifier;

[0082] S302: After classification, the same shared layer parameters are aggregated layer by layer using a preset aggregation algorithm. For layers that have not received parameters uploaded by the client, the parameters of the corresponding layer remain unchanged.

[0083] S303: Integrate all parameter sets corresponding to the shared layer, and use the parameter sets as the updated shared layer parameters.

[0084] In this embodiment, in step S301 above, after receiving the shared layer parameters uploaded by all clients, the central server first categorizes the parameters from different clients according to a preset shared layer identifier (such as a layer index). For example, all the second-layer shared layer parameters uploaded by all clients are grouped into one group, and all the fifth-layer shared layer parameters uploaded by all clients are grouped into another group.

[0085] In step S302, the server aggregates the same shared layer parameters by using a preset aggregation algorithm. Common aggregation algorithms include weighted average, in which the weights can be determined based on the local data sample size of each client, the importance of data distribution, or the upload weight, to ensure that clients with large sample size and high data quality have more influence in aggregation. For shared layers that have no parameters uploaded by any client in the current round of training, the server keeps the parameters unchanged from the previous round.

[0086] In step S303, the server integrates the updated shared layer parameters obtained by aggregation and the unchanged shared layer parameters to form a complete set of updated global shared layer parameters.

[0087] While the updated shared layer parameters are being aggregated, the central server also updates the parameters of the server super network. Specifically, the server collects the model performance feedback corresponding to the shared layer parameters updated by aggregation in the current round, combines the shared layer parameters uploaded by each client and the corresponding upload weight, calculates the loss function of the super network, and updates the parameters of the server super network through the back propagation algorithm and the optimizer. The updated server super network parameters will be sent to each client together with the updated shared layer parameters, so that the parameter selection strategy of each client can be optimized in the next round of training, thereby continuously improving the efficiency and model performance of the entire federated learning system.

[0088] Referring to Figure 2 One embodiment of the present application discloses a financial fraud detection model training method, which is executed on a client and includes:

[0089] T1: obtaining the model parameters of an initial financial fraud detection model and the initial parameters of a server super network sent by a central server, and constructing a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and instantiating a client super network based on the initial parameters of the server super network locally;

[0090] T2: dividing the local detection model into shared layers and personalized layers;

[0091] T3: training the local detection model using local data, updating the model parameters of the local detection model by calculating the loss and performing back propagation;

[0092] T4: After each round of training, the uploading weight of each shared layer parameter in the local detection model is calculated through the client super network, the shared layer parameter to be uploaded to the central server is screened out, and the screened shared layer parameter is uploaded to the central server. The central server aggregates each shared layer parameter uploaded by the client to obtain updated shared layer parameters;

[0093] T5: The updated shared layer parameters are received, the updated shared layer parameters are spliced with the local personalized layer parameters to obtain updated model parameters, and a local detection model for the next round of training is constructed;

[0094] T6: The local detection model is iteratively trained until a preset convergence condition is met and the training is terminated, and the local detection model obtained at the time of convergence is determined as the final financial fraud detection model used by the local client.

[0095] In the embodiment, in the step T1, the client first obtains a data packet containing the initial financial fraud detection model parameters and the initial parameters of the server super network from the central server. The client parses the data packet, extracts the structure data and weight parameters of the initial financial fraud detection model, and constructs a local detection model on the local hardware (such as PC, edge server, etc.) which is completely consistent with the structure of the initial server model. At the same time, the client instantiates a client super network on the local using the server super network structure data and initial parameters in the data packet.

[0096] In the step T2, the client divides the local detection model into a shared layer and a personalized layer according to the division rule attached in the data packet. The shared layer is responsible for learning general financial fraud features, and the personalized layer is used to capture local fraud patterns specific to the client.

[0097] Specifically, the step T2 of dividing the local detection model into a shared layer and a personalized layer comprises:

[0098] T201: determining the layer position c of the model segmentation point based on a preset division strategy;

[0099] T202: determining the first layer to the c-th layer in the local detection model as the shared layer, and the shared layer is represented as , wherein is the i-th client, and the shared layer parameter is selectively uploaded to the central server in the training;

[0100] T203: determining the c+1-th layer to the L-th layer in the local detection model as the personalized layer, and the personalized layer is represented as , and the personalized layer parameter is only locally updated in the training.

[0101] In step T3, the client trains its local detection model using its own local financial data. The training process includes: inputting the pre-processed financial transaction data into the model to obtain fraud prediction results; comparing the prediction results with the true labels to calculate the loss (such as cross-entropy loss); then performing a backpropagation algorithm to calculate the gradient layer by layer from the output layer to the input layer, and using an optimizer to update the parameters of all layers in the local detection model, including the parameters of the shared layers and the personalized layers.

[0102] Specifically, N clients collaboratively train a model while not sharing each other's private data. Each client has a local dataset , where represents the input dataset, represents the corresponding label. The total number of samples of all clients is denoted as , where is the number of samples of the i-th client, the model contains L Transformer layers, and the model parameters are denoted as . Each client maintains its own model .

[0103] The specific process of calculating the loss is that the goal of all clients is to minimize the weighted local loss function , where , and is the cross-entropy loss function associated with its local dataset . The difference between the predicted value of the data sample and the true label is measured by the cross-entropy loss .

[0104] In step T4, after each round of local training, the client calculates the upload weight through its locally instantiated client super network. The super network takes the parameters of the shared layers in the current local detection model as input and calculates an upload weight for each shared layer. The client sets a dynamic threshold and filters out the shared layer parameters whose upload weight is greater than the threshold. Then, the client uploads these filtered shared layer parameters to the central server. The central server collects these parameters from multiple clients and aggregates them by layer to generate updated global shared layer parameters.

[0105] In step T5, the client receives the updated shared layer parameters issued by the central server after global aggregation. The client replaces the original shared layer parameters in the local detection model with the new shared layer parameters. Meanwhile, the client retains the personalized layer parameters optimized in the local training. The client splices the updated shared layer parameters and the locally retained personalized layer parameters to form a complete local detection model with updated parameters, which is used for the next round of training.

[0106] In step T6, the client repeats the process of steps T3 to T5, i.e., training a new local detection model using local data, screening and uploading new shared layer parameters through the client super network, receiving the latest shared layer parameters issued by the server and updating the local model, based on the new local detection model constructed in the last round. This iterative process continues until a preset convergence condition is met. The convergence condition can be that the local training loss of the client no longer decreases significantly, the performance of the model on the local validation set no longer improves, or a termination instruction is received from the server. When the training is terminated, the local detection model currently possessed by the client, i.e., the model composed of the final global shared layer parameters and the personalized layer parameters optimized through multiple rounds of local iterations, is determined as the final model of the client for the actual financial fraud detection task.

[0107] The method of the embodiment reduces the amount of data transmitted across the terminals and the communication overhead by uploading only the parameters critical to the improvement of the global model. While protecting the privacy of the data, the method efficiently contributes valuable knowledge to the global model, so that the server can effectively fuse the knowledge of multiple clients to form global shared features. By receiving and fusing the global shared layer parameters while retaining the personalized model characteristics of the client, the trained model has both global generalization ability and adaptability to different client business scenarios, thereby improving the accuracy and robustness of financial fraud detection.

[0108] In some specific embodiments, the step T4 of calculating the upload weight of each shared layer parameter in the local detection model through the client super network after the end of each round of training, screening the shared layer parameters to be uploaded to the central server, uploading the screened shared layer parameters to the central server, and aggregating the received shared layer parameters uploaded by each client by the central server to obtain updated shared layer parameters, comprises:

[0109] T401: After the end of each round of training, inputting the parameters of each shared layer in the local detection model into the client super network to calculate the upload weight of each shared layer;

[0110] T402: Respectively judging whether the upload weight of each shared layer exceeds a preset dynamic threshold;

[0111] T403: If so, the shared layer parameters whose upload weight is higher than the preset dynamic threshold are uploaded to the central server for aggregation; the central server aggregates the shared layer parameters uploaded by each client layer to obtain updated shared layer parameters.

[0112] In this embodiment, the parameters of the shared layer are represented by a multi-dimensional vector. Specifically, each client i is input with a d-dimensional vector. This vector represents the data distribution characteristics and training state. It is composed of the label distribution, gradient statistics, and input entropy, and is used as input to the client hypernetwork to generate an importance score for each client. The output of the client hypernetwork is a c-dimensional matrix. This is used to score the upload weight of each shared layer, denoted as... ,in, It is a hypernetwork. For the parameters of the hypernetwork, each Indicates client i to the first Layer weight scoring.

[0113] After local training, the client calculates the actual parameter update magnitude for each shared layer. Then, the weights of the supernetwork are merged to obtain the final weight values ​​for each shared layer. Combined with dynamic threshold The set of shared layer parameters to be uploaded to the central server is obtained. ,in, and These are threshold control parameters set during the central server initialization process (included in the initial parameters of the server's hypernetwork). It can be seen that the parameter update magnitude is small and... Lower weight score of shared layer Low-weight parameters, with a weight of zero, will not be uploaded to the central server. The central server aggregates the shared layer parameters uploaded by each client, and sends the aggregated shared parameters back to each client, comparing them with their local parameters. Reassemble and continue with the next round of training.

[0114] This embodiment significantly reduces the amount of communication data between the client and the server, alleviating network bandwidth pressure and latency, thus making the federated learning process more efficient. Furthermore, by uploading only a subset of parameters, it also enhances the security of client data to some extent.

[0115] This application also provides a financial fraud detection model training system, including a central server and several clients;

[0116] The central server is configured to build an initial financial fraud detection model and a server super network, and send model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients;

[0117] Each of the clients obtains the model parameters of the initial financial fraud detection model and the initial parameters of the server super network sent by the central server, and locally builds a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and instantiates a client super network locally based on the initial parameters of the server super network; the local detection model is divided into a shared layer and a personalized layer; the local detection model is trained using local data, the model parameters of the local detection model are updated by calculating a loss and performing back propagation; after each round of training ends, the upload weight of each shared layer parameter in the local detection model is calculated by the client super network, the shared layer parameters that need to be uploaded to the central server are screened out, and the screened shared layer parameters are uploaded to the central server;

[0118] The central server receives the shared layer parameters uploaded by each of the clients; the received shared layer parameters uploaded by each of the clients are aggregated by layer to obtain updated shared layer parameters; and the updated shared layer parameters are distributed to each of the clients;

[0119] Each of the clients receives the updated shared layer parameters distributed by the central server, splices the updated shared layer parameters with the local personalized layer parameters to obtain updated model parameters, and builds a local detection model for the next round of training;

[0120] The central server reenters the step of receiving the shared layer parameters uploaded by each of the clients, and each of the clients iteratively trains the local detection model until a preset convergence condition is met and the training is terminated, and each of the clients determines the local detection model obtained at the time of convergence as a final financial fraud detection model used by the local client.

[0121] Figure 3 FIG. 1 is a schematic diagram of a financial fraud detection model training system in one specific embodiment. Referring to FIG. 1, Figure 3 The financial fraud detection model training method is applied to a financial fraud detection model training system. The financial fraud detection model training system includes a central server and a plurality of clients. The central server and the plurality of clients are connected through a network, the central server can be implemented by an independent server, and the client can be at least one of a computer, an edge server, etc., and the client scenarios include, for example, bank A, payment platform B, securities company C, insurance company D, etc.

[0122] In some embodiments, the central server is configured to build an initial financial fraud detection model and a server super network, and send model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients, including: the server builds an initial financial fraud detection model and a server super network at the central server end, to obtain structure data of the initial financial fraud detection model and structure data of the server super network; wherein the initial financial fraud detection model comprises an embedding layer, an encoder layer, a residual layer, a normalization layer and an output layer connected in sequence; the weight parameters of the initial financial fraud detection model and the initial parameters of the server super network are initialized; the structure data of the initial financial fraud detection model, the weight parameters and the structure data and initial parameters of the server super network are packaged to obtain model parameters; and the model parameters are sent to each of the clients.

[0123] In some embodiments, the central server aggregates the shared layer parameters uploaded by each of the clients according to layers to obtain updated shared layer parameters, including: after the server receives the shared layer parameters uploaded by all the clients, the same shared layers uploaded by different clients are classified according to the identification of the shared layers; the classified same shared layer parameters are processed by layer-by-layer aggregation according to a preset aggregation algorithm, and the parameters of the layers for which no client uploaded parameters are kept unchanged; and the parameter sets corresponding to all the shared layers are integrated, and the parameter sets are taken as the updated shared layer parameters.

[0124] In some embodiments, the client divides the local detection model into shared layers and personalized layers, including determining the layer position c of the model segmentation point based on a preset division strategy; determining the first layer to the c-th layer in the local detection model as the shared layers, and the shared layers are represented as , wherein for the i-th client, the shared layer parameters are selectively uploaded to the central server in training; determining the c+1-th layer to the L-th layer in the local detection model as the personalized layers, and the personalized layers are represented as , and the personalized layer parameters are only iteratively updated locally in training.

[0125] In some embodiments, the step of screening and uploading the shared layer parameters to the central server by the client after each round of training includes: the client inputs the parameters of each shared layer in the local detection model after each round of training, and calculates the upload weight of each shared layer by the client super network; it is judged whether the upload weight of each shared layer exceeds the preset dynamic threshold respectively; if yes, the shared layer parameters with the upload weight higher than the preset dynamic threshold are uploaded to the central server for aggregation; the central server aggregates the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters.

[0126] Referring to Figure 4 The embodiments of the present application provide a financial fraud detection model training device, which is arranged on a central server and comprises:

[0127] An initial model building module 10 is configured to build an initial financial fraud detection model and a server super network on the central server, and send model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; wherein each client builds a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into shared layers and personalized layers, and each client trains the local detection model using local data.

[0128] A data receiving module 20 is configured to receive shared layer parameters uploaded by each client, wherein the shared layer parameters are the shared layer parameters selected by each client to be uploaded after the client calculates the upload weight of each shared layer parameter by a client super network after training the local detection model using local data.

[0129] A data aggregation module 30 is configured to aggregate the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters.

[0130] A data delivery module 40 is configured to deliver the updated shared layer parameters to each client, so that each client splices the updated shared layer parameters with local personalized layer parameters to obtain updated model parameters and build a local detection model for the next round of training.

[0131] An iterative training module 50 is configured to receive the shared layer parameters uploaded by each client again, drive each client to iteratively train the local detection model respectively, until a preset convergence condition is met and the training is terminated, and determine the local detection model obtained by each client when the convergence is met as the final financial fraud detection model of the corresponding client.

[0132] In this embodiment, the server distributes the initial model and hypernetwork parameters without collecting raw client data, explicitly retaining client data locally and only receiving filtered shared layer parameters. This mitigates the risk of leakage during cross-platform transmission of sensitive financial data. The server initially distributes the same financial fraud detection model and hypernetwork parameters uniformly. By aggregating the shared layer parameters uploaded by each client layer by layer, the scattered local client knowledge is transformed into globally optimal shared features, achieving server-centric global parameter optimization and preventing client training from deviating from the correct direction. The server indirectly controls the selection criteria for shared layer parameters through pre-constructed hypernetwork parameters, ensuring that clients only upload parameters that are crucial for improving the global model. This significantly reduces the amount of cross-platform data transmission, lowers communication overhead, effectively integrates knowledge from multiple clients to form globally shared features, and retains the personalized model characteristics of each client. This allows the trained model to possess both global generalization capabilities and adaptability to different client business scenarios, improving the accuracy and robustness of financial fraud detection.

[0133] Figure 5 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 5 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for training a financial fraud detection model. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the method for training a financial fraud detection model. Those skilled in the art will understand that… Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0134] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0135] constructing an initial financial fraud detection model and a server super network on the central server, and sending model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data;

[0136] receiving shared layer parameters uploaded by each client, wherein the shared layer parameters are selected by each client to upload after uploading weights of each shared layer parameter calculated by a client super network after training the local detection model using local data;

[0137] aggregating the received shared layer parameters uploaded by each client by layer to obtain updated shared layer parameters;

[0138] downloading the updated shared layer parameters to each client, so that each client splices the updated shared layer parameters with local personalized layer parameters to obtain updated model parameters and construct a local detection model for the next round of training;

[0139] re-entering the step of receiving shared layer parameters uploaded by each client, driving each client to iteratively train the local detection model, until a preset convergence condition is met and the training is terminated, and determining the local detection model obtained by each client at the time of convergence as the final financial fraud detection model of the corresponding client.

[0140] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0141] A financial fraud detection model training method, which is performed on a central server, includes:

[0142] constructing an initial financial fraud detection model and a server super network on the central server, and sending model parameters of the initial financial fraud detection model and initial parameters of the server super network to a plurality of clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data;

[0143] receive the shared layer parameters uploaded by each of the clients, wherein the shared layer parameters are selected by each client after training a local detection model using local data, calculating upload weights of each shared layer parameter through a client super network, and filtering the shared layer parameters to be uploaded;

[0144] aggregate the received shared layer parameters uploaded by each of the clients by layer to obtain updated shared layer parameters;

[0145] distribute the updated shared layer parameters to each of the clients, so that each of the clients splices the updated shared layer parameters with local personalized layer parameters to obtain updated model parameters and construct a local detection model for the next round of training;

[0146] re-enter the step of receiving the shared layer parameters uploaded by each of the clients, and drive each of the clients to iteratively train the local detection model until a preset convergence condition is met and the training is terminated, and determine the local detection model obtained by each client at the time of convergence as the final financial fraud detection model of the corresponding client.

[0147] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The program can be stored in a non-volatile computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM).

[0148] Each technical feature of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of the technical features do not exist, they should be considered as the scope of the present disclosure.

[0149] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a more specific and detailed manner, but should not be construed as limiting the scope of the patent of the present application. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for training a financial fraud detection model, characterized in that, The method is executed on a central server and includes: An initial financial fraud detection model and a server hypernet are constructed on the central server, and the model parameters of the initial financial fraud detection model and the initial parameters of the server hypernet are sent to multiple clients. Each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer. Each client trains the local detection model using local data. The shared layer parameters uploaded by each client are received. The shared layer parameters are the parameters that need to be uploaded after each client trains a local detection model using local data and calculates the upload weight of each shared layer parameter through the client hypernetwork. Client i shares layers with each layer Calculate the actual parameter update magnitude Integrating super network weights This yields the final weight values ​​for each shared layer. ; Determine each shared layer separately Upload weight Does it exceed the preset dynamic threshold? , and The threshold control parameters set during the initialization process of the central server are used to upload shared layer parameters whose upload weight is higher than the preset dynamic threshold to the central server for aggregation. The shared layer parameters uploaded by each of the clients are aggregated layer by layer to obtain the updated shared layer parameters; The updated shared layer parameters are sent to each client, and each client concatenates the updated shared layer parameters with the local personalized layer parameters to obtain the updated model parameters and build the local detection model for the next round of training. The process re-enters the step of receiving shared layer parameters uploaded by each client, drives each client to iteratively train its local detection model until a preset convergence condition is met and training is terminated, and the local detection model obtained by each client at the convergence point is determined as the final financial fraud detection model for the corresponding client.

2. The financial fraud detection model training method according to claim 1, characterized in that, The steps include: constructing an initial financial fraud detection model and a server hypernet on the central server, and sending the model parameters of the initial financial fraud detection model and the initial parameters of the server hypernet to multiple clients; wherein each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer; and each client trains the local detection model using local data. An initial financial fraud detection model and a server hypernet are constructed on the central server to obtain the structural data of the initial financial fraud detection model and the structural data of the server hypernet; wherein, the initial financial fraud detection model includes an embedding layer, an encoder layer, a residual layer, a normalization layer and an output layer connected in sequence; Initialize the weight parameters of the initial financial fraud detection model and the initial parameters of the server hypernet; The structural data and weight parameters of the initial financial fraud detection model, and the structural data and initial parameters of the server hypernetwork are packaged together to obtain the model parameters; The model parameters are distributed to each of the clients. After receiving the model parameters of the initial financial fraud detection model, each client constructs a local detection model with the same structure as the initial financial fraud detection model based on the structure data and weight parameters of the initial financial fraud detection model. A shared layer and a personalized layer are constructed for each client's local detection model. Each client instantiates a client hypernetwork locally based on the structure data and initial parameters of the server hypernetwork. Each client trains the local detection model using local data and calculates the upload weights of each shared layer parameter using the client hypernetwork. The shared layer parameters are selectively uploaded to the central server during training, while the personalized layer parameters are iteratively updated locally only during training.

3. The financial fraud detection model training method according to claim 1, characterized in that, The step of aggregating the received shared layer parameters uploaded by each of the clients layer by layer to obtain the updated shared layer parameters includes: After receiving the shared layer parameters uploaded by all clients, the shared layer is categorized according to the shared layer identifier. The parameters of the same shared layer after being classified are aggregated layer by layer using a preset aggregation algorithm. For layers that have not received parameters uploaded by the client, the parameters of the corresponding layer remain unchanged. Integrate all parameter sets corresponding to the shared layer, and use the parameter sets as the updated shared layer parameters.

4. A method for training a financial fraud detection model, characterized in that, The method is executed on the client and includes: Obtain the model parameters of the initial financial fraud detection model sent by the central server and the initial parameters of the server hypernetwork, and construct a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and instantiate the client hypernetwork locally based on the initial parameters of the server hypernetwork. The local detection model is divided into a shared layer and a personalized layer; The local detection model is trained using local data, and the model parameters of the local detection model are updated by calculating the loss and performing backpropagation. After each round of training, the client hypernetwork calculates the upload weights of each shared layer parameter in the local detection model, filters out the shared layer parameters that need to be uploaded to the central server, and uploads the filtered shared layer parameters to the central server. The central server then aggregates the shared layer parameters uploaded by each client layer to obtain updated shared layer parameters; wherein, client i for each shared layer... Calculate the actual parameter update magnitude Integrating super network weights This yields the final weight values ​​for each shared layer. ; Determine each shared layer separately Upload weight Does it exceed the preset dynamic threshold? , and The threshold control parameters set during the initialization process of the central server are used to upload shared layer parameters whose upload weight is higher than the preset dynamic threshold to the central server for aggregation. The system receives updated shared layer parameters from the central server, concatenates the updated shared layer parameters with the local personalized layer parameters to obtain updated model parameters, and constructs the local detection model for the next round of training. The local detection model is iteratively trained until a preset convergence condition is met and training is terminated. The local detection model obtained at convergence is determined as the final financial fraud detection model used by the local client.

5. The financial fraud detection model training method according to claim 4, characterized in that, The step of dividing the local detection model into a shared layer and a personalized layer includes: The layer position of the model segmentation point is determined based on a preset partitioning strategy. ; The first to the second layers of the local detection model The layer is determined to be a shared layer, and the shared layer is represented as: ,in For the first Each client has its own shared layer parameters, which are selectively uploaded to the central server during training. The first in the local detection model +1 layer to the total number of layers in the model The layer is defined as a personalization layer, and the personalization layer is represented as... The parameters of the personalized layer are updated iteratively only locally during training.

6. The financial fraud detection model training method according to claim 4, characterized in that, The step of calculating the upload weights of each shared layer parameter in the local detection model through the client hypernetwork after each round of training, filtering out the shared layer parameters that need to be uploaded to the central server, and uploading the filtered shared layer parameters to the central server, and then having the central server aggregate the received shared layer parameters uploaded by each client layer to obtain updated shared layer parameters, includes: After each round of training, the parameters of each shared layer in the local detection model are used as input, and the upload weight of each shared layer is calculated through the client hypernetwork. Determine whether the upload weight of each shared layer exceeds the preset dynamic threshold; If so, the shared layer parameters with upload weights higher than the preset dynamic threshold are uploaded to the central server for aggregation; the central server then aggregates the shared layer parameters uploaded by each client layer to obtain updated shared layer parameters.

7. A training system for a financial fraud detection model, characterized in that, Includes a central server and several clients; The central server is used to execute the construction of an initial financial fraud detection model and a server hypernet, and to send the model parameters of the initial financial fraud detection model and the initial parameters of the server hypernet to multiple clients; Each client obtains the model parameters of the initial financial fraud detection model and the initial parameters of the server hypernetwork sent by the central server, and constructs a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and instantiates the client hypernetwork locally based on the initial parameters of the server hypernetwork. The local detection model is divided into a shared layer and a personalized layer; the local detection model is trained using local data, and the model parameters of the local detection model are updated by calculating the loss and performing backpropagation. After each round of training, the client hypernetwork calculates the upload weights of each shared layer parameter in the local detection model, filters out the shared layer parameters that need to be uploaded to the central server, and uploads the filtered shared layer parameters to the central server; wherein, client i assigns weights to each shared layer parameter. Calculate the actual parameter update magnitude Integrating super network weights This yields the final weight values ​​for each shared layer. ; Determine each shared layer separately Upload weight Does it exceed the preset dynamic threshold? , and The threshold control parameters set during the initialization process of the central server are used to upload shared layer parameters whose upload weight is higher than the preset dynamic threshold to the central server for aggregation. The central server receives the shared layer parameters uploaded by each client; aggregates the received shared layer parameters by layer to obtain updated shared layer parameters; and sends the updated shared layer parameters to each client. Each client receives the updated shared layer parameters from the central server, concatenates the updated shared layer parameters with the local personalized layer parameters to obtain the updated model parameters, and constructs the local detection model for the next round of training. The central server re-enters the step of receiving the shared layer parameters uploaded by each client. Each client iteratively trains its local detection model until a preset convergence condition is met and training is terminated. Each client determines the local detection model obtained at the convergence point as the final financial fraud detection model used by the local client.

8. A training device for a financial fraud detection model, characterized in that, The device is located on a central server and includes: An initial model building module is used to build an initial financial fraud detection model and a server hypernet on the central server, and send the model parameters of the initial financial fraud detection model and the initial parameters of the server hypernet to multiple clients; wherein, each client builds a local detection model with the same structure as the initial financial fraud detection model based on the model parameters, and divides the local detection model into a shared layer and a personalized layer, and each client trains the local detection model using local data; The data receiving module is used to receive the shared layer parameters uploaded by each client. These shared layer parameters are selected by each client after training a local detection model using local data and calculating the upload weights of each shared layer parameter through the client's hypernetwork. Client i handles each shared layer... Calculate the actual parameter update magnitude Integrating super network weights This yields the final weight values ​​for each shared layer. ; Determine each shared layer separately Upload weight Does it exceed the preset dynamic threshold? , and The threshold control parameters set during the initialization process of the central server are used to upload shared layer parameters whose upload weight is higher than the preset dynamic threshold to the central server for aggregation. The data aggregation module is used to aggregate the shared layer parameters uploaded by each of the clients by layer to obtain updated shared layer parameters; The data delivery module is used to send the updated shared layer parameters to each of the clients, so that each client can concatenate the updated shared layer parameters with the local personalized layer parameters to obtain the updated model parameters and build the local detection model for the next round of training. The iterative training module is used to re-receive the shared layer parameters uploaded by each client, drive each client to iteratively train the local detection model until the preset convergence condition is met and the training is terminated, and determine the local detection model obtained by each client at the convergence time as the final financial fraud detection model of the corresponding client.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the financial fraud detection model training method as described in any one of claims 1 to 3.

10. A computer device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the financial fraud detection model training method as described in any one of claims 1 to 3.

Images