Federal learning-based credit risk assessment method, system and device, and medium

Abstract

The invention relates to the technical field of artificial intelligence and finance, and provides a federated learning-based credit risk assessment method, system and device and a medium, which can be used for training a model based on federated learning when each participant uploads model parameters of a respective local model to a central server for the first time. The central server performs grading processing on each participant according to the model parameters uploaded by each participant to obtain the grade of each participant, and after each participant receives the respective grade and the corresponding quantitative compression strategy, the model parameters are firstly subjected to quantitative compression processing when local model parameters are uploaded to the central server subsequently. Therefore, the volume of the uploading model can be reduced, the communication overhead and the calculation overhead can be reduced, and while the model detection accuracy is ensured, the model training efficiency and the processing efficiency during customer credit risk assessment by using the model are improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and financial technology, and particularly to a credit risk assessment method, system, device, and medium based on federated learning. Background Art

[0002] In the fields of data privacy protection and secure computing, traditional centralized learning methods require aggregating multi-party data to a single location for training, but this often poses a risk of data privacy leakage. In addition, multi-party collaborative computing tasks also pose a series of challenges in terms of performance and privacy protection. Therefore, to address the problems brought by traditional machine learning, the existing technology has introduced the concept of federated learning.

[0003] Federated learning is an emerging fundamental technology of artificial intelligence, designed to carry out efficient machine learning among multiple participants or computing nodes while ensuring information security during big data exchange, protecting terminal data and personal data privacy, and ensuring legal compliance.

[0004] However, in financial scenarios, traditional federated learning frameworks (such as cross-institutional joint credit risk assessment model training) usually use various encryption methods to protect privacy data, which will additionally increase a large amount of computational overhead and communication burden in large-scale distributed computing. Summary of the Invention

[0005] In view of the above, it is necessary to provide a credit risk assessment method, system, device, and medium based on federated learning, aiming to solve the problems of large computational overhead, heavy communication burden, and low processing efficiency when assessing customer credit risk based on federated learning in the financial field.

[0006] A customer credit risk assessment method based on federated learning, which is applied to a customer credit risk assessment system based on federated learning. The customer credit risk assessment system based on federated learning includes a central server and multiple participants, and the central server communicates with the multiple participants. The customer credit risk assessment method based on federated learning includes:

[0007] The central server initializes a global model for customer credit risk assessment and distributes the global model to each participant;

[0008] Each participant obtains local data and trains the global model based on the corresponding local data to obtain a local model for each participant;

[0009] Each participant encrypts and uploads the model parameters of their respective local models to the central server;

[0010] The central server aggregates the model parameters uploaded by each participant to obtain target global model parameters, and classifies each participant based on the model parameters uploaded by each participant to obtain the level of each participant;

[0011] The central server determines the quantization compression strategy corresponding to each level, and distributes the level of each participant, the corresponding quantization compression strategy, and the target global model parameters to the corresponding participant;

[0012] Each participant optimizes and trains its local model according to the target global model parameters and the local data, performs quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypts and uploads the obtained model parameters after quantization compression to the central server respectively;

[0013] The central server and each participant perform federated learning training of the model based on the level of each participant and the corresponding quantization compression strategy. When the accuracy of the global model trained by the central server reaches the preset accuracy, the federated learning training is stopped to obtain a customer credit risk assessment model;

[0014] The central server evaluates the credit risk of each customer according to the customer credit risk assessment model.

[0015] A customer credit risk assessment system based on federated learning, the customer credit risk assessment system based on federated learning includes a central server and multiple participants, the central server communicates with the multiple participants, and the customer credit risk assessment system based on federated learning includes:

[0016] The central server is used to initialize a global model for customer credit risk assessment and distribute the global model to each participant;

[0017] Each participant is used to obtain local data and train the global model based on the corresponding local data respectively to obtain the local model of each participant;

[0018] Each participant is also used to encrypt and upload the model parameters of its local model to the central server respectively;

[0019] The central server is also used to aggregate the model parameters uploaded by each participant to obtain target global model parameters, and classify each participant based on the model parameters uploaded by each participant to obtain the level of each participant;

[0020] The central server is also used to determine the quantization compression strategy corresponding to each level, and distribute the level of each participant, the corresponding quantization compression strategy, and the target global model parameters to the corresponding participant;

[0021] Each participating party is also used to optimize and train its respective local model according to the target global model parameters and the local data, perform quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypt and upload the model parameters obtained after quantization compression to the central server respectively;

[0022] The central server and each participating party are also used to perform federated learning training of the model based on the level of each participating party and the corresponding quantization compression strategy, and stop the federated learning training until the accuracy of the global model trained by the central server reaches the preset accuracy, so as to obtain a customer credit risk assessment model;

[0023] The central server is also used to evaluate the credit risk of each customer according to the customer credit risk assessment model.

[0024] A computer device, the computer device includes:

[0025] A memory that stores at least one instruction; and

[0026] A processor that executes the instructions stored in the memory to implement the customer credit risk assessment method based on federated learning.

[0027] A computer-readable storage medium stores at least one instruction, and the at least one instruction is executed by a processor in a computer device to implement the customer credit risk assessment method based on federated learning.

[0028] It can be seen from the above technical solutions that when the present invention trains a model based on federated learning, when each participating party first uploads the model parameters of its respective local model to the central server, the central server performs grading processing on each participating party according to the model parameters uploaded by each participating party to obtain the level of each participating party. After each participating party receives its respective level and the corresponding quantization compression strategy, when uploading the local model parameters to the central server subsequently, the model parameters are first subjected to quantization compression processing, so that the volume of the uploaded model can be reduced, and the communication overhead and calculation overhead can be reduced. While ensuring the accuracy of model detection, the training efficiency of the model and the processing efficiency when using the model to evaluate customer credit risk are also improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of a preferred embodiment of the customer credit risk assessment method based on federated learning of the present invention.

[0030] Figure 2 is a functional module diagram of a preferred embodiment of the customer credit risk assessment system based on federated learning of the present invention.

[0031] Figure 3 It is a schematic structural diagram of a computer device which is a preferred embodiment for the present invention to implement the method for customer credit risk assessment based on federated learning. Detailed implementation manners

[0032] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] As Figure 1 shown, it is a flowchart of a preferred embodiment of the method for customer credit risk assessment based on federated learning according to the present invention. According to different requirements, the order of steps in this flowchart can be changed and some steps can be omitted.

[0034] The method for customer credit risk assessment based on federated learning is applied to one or more computer devices. The computer device is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0035] The computer device can be any electronic product capable of human-computer interaction with users. For example, personal computers, tablet computers, smart phones, personal digital assistants (PDAs), game consoles, Internet protocol televisions (IPTVs), smart wearable devices, etc.

[0036] The computer device may further include network devices and / or user devices. Among them, the network devices include but are not limited to a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of hosts or network servers based on cloud computing.

[0037] The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0038] Among them, Artificial Intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.

[0039] The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, robotics, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0040] The network where the computer device is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, Virtual Private Network (VPN), etc.

[0041] This embodiment is applied to a customer credit risk assessment system based on federated learning. The customer credit risk assessment system based on federated learning includes a central server and multiple participants. The central server communicates with the multiple participants. The customer credit risk assessment method based on federated learning includes:

[0042] S10, the central server initializes a global model for customer credit risk assessment and distributes the global model to each participant.

[0043] In this embodiment, the global model may include, but is not limited to: logistic regression model, decision tree model, neural network model, etc.

[0044] In this embodiment, in the federated learning scenario, the central server mainly has the following functions:

[0045] Coordinator: Overall planning of the whole, initializing a global model for customer credit risk assessment, and responsible for distributing and collecting model data among participants. For example, determining when to send the initial model to each institutional terminal (i.e., each participant) and when to receive the trained model parameters, etc.

[0046] Aggregator: Securely aggregate the encrypted model parameters uploaded by each participant and update the global model. Just like integrating the local credit risk assessment information provided by each institution to form a more comprehensive assessment model.

[0047] Supervisor: Score and rate the models of participants, regularly review and dynamically adjust. Based on the contribution of each institutional participant's model to the overall credit risk assessment, determine its subsequent participation method.

[0048] In this embodiment, in the context of federated learning, each participant mainly has the following functions:

[0049] Data provider: Each institution has its own customer credit data and uses this data locally to train models, providing data support for the global credit risk assessment model. For example, a bank provides data such as customer loan records and repayment situations for training.

[0050] Local trainer: Independently trains the credit risk assessment model on local data and updates the model parameters. For example, a fintech company uses its own algorithms and data to locally adjust the initial global model.

[0051] Participant: Quantifies and compresses the model according to the rating requirements of the central server, uploads the encrypted parameters, and receives the updated global model for continued training, participating in the construction of the joint credit risk assessment model.

[0052] The specific functions of the central server and each participant will be elaborated in detail below and will not be repeated here.

[0053] S11, each participant obtains local data and respectively trains the global model based on the corresponding local data to obtain the local model of each participant.

[0054] In this embodiment, each participant independently trains based on the local data on the basis of the global model issued by the central server and updates the model parameters according to the local data.

[0055] S12, each participant encrypts and uploads the model parameters of its own local model to the central server respectively.

[0056] In this embodiment, homomorphic encryption algorithms, differential privacy techniques, etc. can be used to encrypt and upload the model parameters of their respective local models to the central server respectively to ensure the security of data during transmission.

[0057] S13, the central server aggregates the model parameters uploaded by each participant to obtain the target global model parameters, and classifies each participant according to the model parameters uploaded by each participant to obtain the level of each participant.

[0058] In this embodiment, the central server can perform secure aggregation on the model parameters uploaded by each participant based on secure multi-party computing protocols, homomorphic encryption protocols, differential privacy protocols, etc., so as to obtain the target global model parameters.

[0059] In this embodiment, the classifying each participant according to the model parameters uploaded by each participant to obtain the level of each participant includes:

[0060] The central server obtains the accuracy rate, model volume, and processing efficiency of the model corresponding to each participant from the model parameters uploaded by each participant.

[0061] The central server obtains a pre-configured accuracy rate scoring mapping table, a model volume scoring mapping table, and a processing efficiency scoring mapping table.

[0062] The central server uses the accuracy rate of the model corresponding to each participant to perform matching in the accuracy rate scoring mapping table to obtain the accuracy rate score of the model corresponding to each participant.

[0063] The central server uses the model volume of the model corresponding to each participant to perform matching in the model volume scoring mapping table to obtain the model volume score of the model corresponding to each participant.

[0064] The central server uses the processing efficiency of the model corresponding to each participant to perform matching in the processing efficiency scoring mapping table to obtain the processing efficiency score of the model corresponding to each participant.

[0065] The central server obtains a first weight corresponding to the accuracy rate, a second weight corresponding to the model volume, and a third weight corresponding to the processing efficiency.

[0066] The central server calculates the weighted score of the model corresponding to each participant according to the first weight, the second weight, the third weight, the accuracy rate score of the model corresponding to each participant, the model volume score of the model corresponding to each participant, and the processing efficiency score of the model corresponding to each participant.

[0067] The central server obtains a participant level mapping table; wherein, the participant level mapping table is used to record the mapping relationship between the participant level and the weighted score.

[0068] The central server uses the weighted score of the model corresponding to each participant to perform matching in the participant level mapping table to obtain the level of each participant.

[0069] For example: For participant X, if the accuracy rate is between 90% and 100% in the accuracy rate scoring mapping table, the corresponding accuracy rate score is 100. Similarly, the model volume score can be obtained as 90, and the processing efficiency score is 90. When the first weight is 0.5, the second weight is 0.2, and the third weight is 0.3, the weighted score corresponding to participant X is 100 * 0.5 + 90 * 0.2 + 90 * 0.3 = 95. After performing matching in the participant level mapping table, the level corresponding to participant X can be obtained, such as level one.

[0070] Through the above embodiments, it is possible to calculate the level of each participant by combining the parameters of the local models of each participant.

[0071] S14. The central server determines the quantization compression strategy corresponding to each level, and sends the level of each participant, the corresponding quantization compression strategy, and the target global model parameters to the corresponding participant.

[0072] In this embodiment, the central server can also encrypt the data sent down based on a certain encryption algorithm to improve the security of the data.

[0073] S15. Each participant optimizes and trains its local model according to the target global model parameters and the local data, performs quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypts and uploads the model parameters obtained after quantization compression to the central server respectively.

[0074] In this embodiment, the performing quantization compression processing on the trained model according to the corresponding quantization compression strategy includes:

[0075] Each participant obtains the parameter precision adjustment strategy in the corresponding quantization strategy, and adjusts the floating-point number bits of the corresponding local model to a preset number of bits according to the parameter precision adjustment strategy;

[0076] Each participant obtains the parameter sparsification strategy in the corresponding quantization strategy, and adjusts the value of the specified parameter of the corresponding local model to 0 according to the parameter sparsification strategy, and prohibits uploading the specified parameter to the central server;

[0077] Each participant obtains the model structure simplification strategy in the corresponding quantization strategy, and removes redundant layers and / or redundant neurons of the corresponding local model according to the model structure simplification strategy.

[0078] Specifically, for high-rated participants, since their model quality is high and they are important, a relatively small degree of quantization compression is allowed, such as only slightly reducing the precision of the model parameters to try to maintain the model performance as much as possible; for medium-rated participants, a moderate quantization compression strategy can be adopted, moderately quantizing some less critical model parameters to reduce the model size while ensuring a certain model performance; for low-rated participants, a more aggressive quantization compression strategy can be adopted, quantizing the model parameters to a greater extent, and even some parameters with little impact on the overall model performance can be discarded to minimize the model transmission size.

[0079] For example, the precision of model parameters can be adjusted according to the compression strategy corresponding to the rating. For instance, for low-rating participants, the parameters originally represented as 32-bit floating-point numbers can be converted to 8-bit integer representation to reduce the number of bytes required for data storage and transmission.

[0080] Another example: The model parameters can be made sparse through a specific algorithm (such as a pruning algorithm), that is, making some parameter values become 0. When storing and transmitting, only the positions and values of non-zero parameters are recorded, reducing the actual amount of data transmitted. For medium-rating participants, some parameter sparsification processing can be appropriately carried out.

[0081] Another example: For the case of low-rating participants with a relatively complex model structure, the model structure can be simplified to a certain extent, such as removing some redundant layers or neurons, but this operation needs to be cautious to avoid having too much impact on the model performance.

[0082] Through the above embodiments, it is possible to perform targeted quantization and compression processing on the respective local model parameters according to different participant levels, thereby reducing the size of the uploaded model to reduce communication overhead.

[0083] In this embodiment, after the model obtained by training is quantized and compressed according to the corresponding quantization and compression strategy, the method further includes:

[0084] Each participant obtains a local test set and uses the local test set to verify the performance of the quantized and compressed model to obtain a verification result;

[0085] When the verification result of a participant shows that the performance of the model obtained after quantization and compression is lower than the performance threshold, the participant adjusts the corresponding quantization and compression strategy and reports the adjusted quantization and compression strategy and the corresponding verification result to the central server.

[0086] Specifically, after completing the quantization and compression operation, each participant can perform a simple performance verification on the compressed model locally, such as performing a small number of inferences on the local test set or calculating metrics such as accuracy, to ensure that the performance of the compressed model is still within an acceptable range. If the performance drops too much, the degree of quantization and compression can be adjusted appropriately according to the situation or relevant information can be fed back to the central server.

[0087] Through the above embodiments, it is possible to ensure the compression scale and avoid affecting the model performance due to excessive compression.

[0088] S16. The central server and each participant perform federated learning training of the model based on the level of each participant and the corresponding quantization and compression strategy until the accuracy of the global model obtained by the central server reaches the preset accuracy, and then stop the federated learning training to obtain a customer credit risk assessment model.

[0089] In this embodiment, the preset accuracy rate can be configured according to the requirements in the actual application scenario. In this way, when the performance of the model itself has reached the expectation, the training of federated learning is stopped, and at this time, a global model with high performance can be obtained.

[0090] In this embodiment, the method further includes:

[0091] During the process of federated learning training, when the number of iterations reaches the preset number, each participant obtains the performance change curve, memory occupancy, and data transmission volume of the corresponding local model as model verification data;

[0092] Each participant uploads its respective model verification data to the central server;

[0093] The central server adjusts the level of each participant based on the model verification data corresponding to each participant, and issues the adjusted level and the corresponding quantization and compression strategy to the corresponding participant.

[0094] Specifically, the central server adjusting the level of each participant based on the model verification data corresponding to each participant includes:

[0095] When the performance change curve of the local model corresponding to the first participant shows that the model quality improvement efficiency is greater than the preset efficiency, the central server raises the level of the first participant;

[0096] When the memory occupancy of the local model corresponding to the second participant is greater than the first threshold and less than or equal to the second threshold, the central server lowers the level of the second participant;

[0097] When the data transmission volume of the local model corresponding to the third participant is greater than the first transmission volume and less than or equal to the second transmission volume, the central server lowers the level of the third participant.

[0098] Among them, the preset efficiency, the first threshold, the second threshold, the first transmission volume, and the second transmission volume can be custom-configured.

[0099] Specifically, for participants with relatively fast and stable improvement in model quality, their ratings can be appropriately increased to reduce the degree of compression of their models, so as to make full use of their high-quality model parameters to promote the optimization of the global model; similarly, for participants with slow or unstable improvement in model quality, their ratings can be lowered and the compression intensity can be increased.

[0100] Specifically, lower the ratings of some participants with large memory occupancy, so that they adopt a more aggressive quantization and compression strategy, thereby reducing the memory occupancy during model transmission and aggregation.

[0101] Specifically, for parties with frequent communication or large amounts of transmitted data, their ratings are lowered and the model compression ratio is increased, thereby reducing the communication data volume.

[0102] Through the above embodiments, a mechanism for re-audit is introduced. After a certain number of iterations or a certain period of time, the central server will perform an audit operation again and dynamically modify the ratings of each party.

[0103] S17, the central server evaluates the credit risk of each customer according to the customer credit risk assessment model.

[0104] It can be understood that in large-scale training in the financial field, the number of models will increase sharply, which will lead to an increase in communication overhead and a large amount of memory occupation during the aggregation process, having a great impact on the efficiency of the entire training.

[0105] To address the above problems, in this embodiment, a scoring mechanism is also introduced under the federated learning training mechanism, so as to perform targeted quantization compression on the local models of different parties according to different party ratings, effectively reducing the size of the uploaded models to reduce communication overhead.

[0106] Moreover, after introducing the rating system, the models of less important parties are processed by a quantization compression mechanism, so as to ensure that the communication volume during the entire training process is in a controllable stage, dynamically allocate rating requirements according to the current required model quality, the maximum memory overhead that can be tolerated, and the communication overhead, so as to achieve the final learning and training, having a more efficient training process compared with traditional federated learning.

[0107] In this embodiment, the method further includes:

[0108] When the data of the fourth party is private data, the fourth party uses the local model to evaluate the credit risk of the corresponding customer.

[0109] For example: If a party pays attention to data privacy and the local data is unique, it is suitable to use the local model for prediction.

[0110] Of course, for scenarios with high requirements for real-time performance and where the party needs to make frequent predictions, using the local model for processing can also reduce communication with the central server, reducing latency and communication costs. For example, in personalized recommendations on mobile devices, the local model can quickly respond to user requests.

[0111] Of course, this embodiment can also be used in the medical field. For example, a hospital can use the local model after federated learning to predict the disease risk of patients. For example, it can use relatively private data such as the medical records, age, work situation, and gender of local patients in each hospital, as well as data with local characteristics such as local eating habits and local climate characteristics to train the local model, and use the trained local model to predict the disease risk of local patients. In this way, since the private data is not leaked and the training data has local characteristics, it not only protects the privacy of patient data but also optimizes the prediction results according to the characteristics of local patients.

[0112] As can be seen from the above technical solutions, when the present invention trains a model based on federated learning, when each participating party first uploads the model parameters of its local model to the central server, the central server performs a grading process on each participating party according to the model parameters uploaded by each participating party to obtain the level of each participating party. After each participating party receives its corresponding level and quantization compression strategy, when uploading the local model parameters to the central server subsequently, it first performs quantization compression processing on the model parameters, thereby reducing the volume of the uploaded model, reducing communication overhead and computational overhead, and while ensuring the accuracy of model detection, also improving the training efficiency of the model and the processing efficiency when using the model for customer credit risk assessment.

[0113] As Figure 2 shown, it is a functional module diagram of a preferred embodiment of the customer credit risk assessment system based on federated learning of the present invention. The customer credit risk assessment system 11 based on federated learning includes a central server 110 and a plurality of participating parties 111. The central server 110 communicates with the plurality of participating parties 111. The modules / units referred to in the present invention refer to a series of computer program segments that can be executed by a processor and can complete fixed functions, and are stored in a memory. In this embodiment, the functions of each module / unit will be described in detail in subsequent embodiments.

[0114] The central server 110 is used to initialize a global model for customer credit risk assessment and send the global model to each participating party 111.

[0115] In this embodiment, the global model may include, but is not limited to: a logistic regression model, a decision tree model, a neural network model, etc.

[0116] In this embodiment, in the federated learning scenario, the central server 110 mainly has the following functions:

[0117] Coordinator: Overall planning of the whole situation, initializing the global model for customer credit risk assessment, and responsible for distributing and collecting model data among each Party 111. For example, determining when to send the initial model to each institutional terminal (i.e., each Party 111), and when to receive the trained model parameters, etc.

[0118] Aggregator: Securely aggregate the encrypted model parameters uploaded by each Party 111 and update the global model. It is like integrating the local credit risk assessment information provided by each institution to form a more comprehensive assessment model.

[0119] Supervisor: Rate and grade the models of Party 111, conduct regular audits and dynamically adjust. Based on the contribution of the models of each institutional Party 111 to the overall credit risk assessment, determine their subsequent participation methods.

[0120] In this embodiment, in the federated learning scenario, each Party 111 mainly has the following functions:

[0121] Data provider: Each institution has its own customer credit data and uses this data locally to train the model, providing data support for the global credit risk assessment model. For example, a bank provides data such as customer loan records and repayment situations for training.

[0122] Local trainer: Independently train the credit risk assessment model on local data and update the model parameters. For example, a fintech company uses its own algorithms and data to make local adjustments to the initial global model.

[0123] Participant: Quantize and compress the model according to the rating requirements of the central server 110, upload the encrypted parameters and receive the updated global model to continue training, and participate in building the joint credit risk assessment model.

[0124] The specific functions of the central server 110 and each Party 111 will be elaborated in detail below and will not be repeated here.

[0125] Each Party 111 is used to obtain local data and respectively train the global model based on the corresponding local data to obtain the local model of each Party 111.

[0126] In this embodiment, each Party 111 independently trains based on the local data on the basis of the global model issued by the central server 110 and updates the model parameters according to the local data.

[0127] Each Party 111 is further used to encrypt and upload the model parameters of its respective local model to the central server 110.

[0128] In this embodiment, homomorphic encryption algorithms, differential privacy techniques, etc. can be used to encrypt and upload the model parameters of their respective local models to the central server 110 respectively to ensure the security of data during the transmission process.

[0129] The central server 110 is further configured to aggregate the model parameters uploaded by each participant 111 to obtain target global model parameters, and perform hierarchical processing on each participant 111 according to the model parameters uploaded by each participant 111 to obtain the level of each participant 111.

[0130] In this embodiment, the central server 110 can perform secure aggregation on the model parameters uploaded by each participant 111 based on secure multi-party computation protocols, homomorphic encryption protocols, differential privacy protocols, etc., so as to obtain the target global model parameters.

[0131] In this embodiment, the hierarchical processing of each participant 111 according to the model parameters uploaded by each participant 111 to obtain the level of each participant 111 includes:

[0132] The central server 110 obtains the accuracy rate, model volume, and processing efficiency of the model corresponding to each participant 111 from the model parameters uploaded by each participant 111;

[0133] The central server 110 obtains a pre-configured accuracy rate scoring mapping table, a model volume scoring mapping table, and a processing efficiency scoring mapping table;

[0134] The central server 110 uses the accuracy rate of the model corresponding to each participant 111 to perform matching in the accuracy rate scoring mapping table to obtain the accuracy rate score of the model corresponding to each participant 111;

[0135] The central server 110 uses the model volume of the model corresponding to each participant 111 to perform matching in the model volume scoring mapping table to obtain the model volume score of the model corresponding to each participant 111;

[0136] The central server 110 uses the processing efficiency of the model corresponding to each participant 111 to perform matching in the processing efficiency scoring mapping table to obtain the processing efficiency score of the model corresponding to each participant 111;

[0137] The central server 110 obtains a first weight corresponding to the accuracy rate, a second weight corresponding to the model volume, and a third weight corresponding to the processing efficiency;

[0138] The central server 110 calculates the weighted score of the model corresponding to each participant 111 according to the first weight, the second weight, the third weight, the accuracy score of the model corresponding to each participant 111, the model volume score of the model corresponding to each participant 111, and the processing efficiency score of the model corresponding to each participant 111;

[0139] The central server 110 obtains a participant level mapping table; wherein, the participant level mapping table is used to record the mapping relationship between participant levels and weighted scores;

[0140] The central server 110 uses the weighted score of the model corresponding to each participant 111 to perform matching in the participant level mapping table to obtain the level of each participant 111.

[0141] For example: For participant X, if the accuracy rate is between 90% and 100% in the accuracy score mapping table, the corresponding accuracy score is 100. Similarly, the model volume score can be obtained as 90, and the processing efficiency score is 90. When the first weight is 0.5, the second weight is 0.2, and the third weight is 0.3, the weighted score corresponding to participant X is 100 * 0.5 + 90 * 0.2 + 90 * 0.3 = 95. After performing matching in the participant level mapping table, the level corresponding to participant X can be obtained, such as level one.

[0142] Through the above embodiments, the level of each participant 111 can be calculated in combination with the parameters of the local models of each participant 111.

[0143] The central server 110 is further configured to determine a quantization compression strategy corresponding to each level, and send the level of each participant 111, the corresponding quantization compression strategy, and the target global model parameters to the corresponding participant 111.

[0144] In this embodiment, the central server 110 may also perform encryption processing on the sent data based on a certain encryption algorithm to improve the security of the data.

[0145] Each participant 111 is further configured to optimize and train its own local model according to the target global model parameters and the local data, perform quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypt and upload the obtained model parameters after quantization compression to the central server 110 respectively.

[0146] In this embodiment, the performing quantization compression processing on the trained model according to the corresponding quantization compression strategy includes:

[0147] Each participant 111 obtains a parameter precision adjustment strategy in a corresponding quantization strategy, and adjusts the number of floating-point digits of a corresponding local model to a preset number of digits according to the parameter precision adjustment strategy;

[0148] Each participant 111 obtains a parameter sparsification strategy in a corresponding quantization strategy, and adjusts the value of a designated parameter of a corresponding local model to 0 according to the parameter sparsification strategy, and prohibits uploading the designated parameter to the central server 110;

[0149] Each participant 111 obtains the model structure simplification strategy in the corresponding quantization strategy, and removes redundant layers and / or redundant neurons of the corresponding local model according to the model structure simplification strategy.

[0150] Specifically, for participants 111 with high ratings, due to their high model quality and great importance, a smaller degree of quantization compression is allowed, such as only slightly reducing the precision of model parameters to maintain model performance as much as possible; for participants 111 with medium ratings, a moderate quantization compression strategy can be adopted to quantize some less critical model parameters to a moderate degree, thereby reducing the model size while ensuring a certain model performance; for participants 111 with low ratings, a more aggressive quantization compression strategy can be adopted to quantize the model parameters to a greater extent, and even some parameters that have little impact on the overall model performance can be discarded to minimize the model transmission size.

[0151] For example, the accuracy of the model parameters can be adjusted according to the compression strategy corresponding to the rating. For example, for the low-rated participant 111, the original 32-bit floating point number parameters can be converted to 8-bit integer representation to reduce the number of bytes required for data storage and transmission.

[0152] For another example, a specific algorithm (such as a pruning algorithm) can be used to make the model parameters sparse, that is, to make some parameter values ​​​​zero. During storage and transmission, only the position and value of non-zero parameters are recorded to reduce the amount of data actually transmitted. For the medium-rated participant 111, some parameter sparse processing can be appropriately performed.

[0153] For another example: For low-rated participant 111 and a relatively complex model structure, the model structure can be simplified to a certain extent, such as removing some redundant layers or neurons, but this operation must be done with caution to avoid excessive impact on model performance.

[0154] Through the above embodiments, the local model parameters of each participant can be quantized and compressed in a targeted manner according to the different levels of the participants, thereby reducing the size of the uploaded model and lowering the communication overhead.

[0155] In this embodiment, after the model obtained through training is quantized and compressed according to the corresponding quantization compression strategy, each participant 111 obtains a local test set, and uses the local test set to verify the performance of the quantized and compressed model to obtain a verification result;

[0156] When the verification result of a participant 111 shows that the performance of the model obtained after quantization and compression is lower than the performance threshold, the participant 111 adjusts the corresponding quantization compression strategy, and reports the adjusted quantization compression strategy and the corresponding verification result to the central server 110.

[0157] Specifically, after the quantization compression operation is completed, each participant 111 can perform a simple performance verification on the compressed model locally, such as performing a small number of inferences on the local test set or calculating metrics such as accuracy, to ensure that the performance of the compressed model is still within an acceptable range. If the performance drops too much, the degree of quantization compression can be adjusted appropriately according to the situation or relevant information can be fed back to the central server 110.

[0158] Through the above embodiments, the compression scale can be ensured, and the model performance can be prevented from being affected due to excessive compression.

[0159] The central server 110 and each participant 111 are also used to perform federated learning training of the model based on the level of each participant 111 and the corresponding quantization compression strategy, and stop the federated learning training until the accuracy of the global model trained by the central server 110 reaches the preset accuracy, so as to obtain a customer credit risk assessment model.

[0160] In this embodiment, the preset accuracy can be configured according to the requirements in the actual application scenario. In this way, when the performance of the model itself has reached the expectation, the federated learning training is stopped, and at this time, a global model with high performance can be obtained.

[0161] In this embodiment, during the process of performing federated learning training, when the number of iterations reaches the preset number, each participant 111 obtains the performance change curve, memory occupancy, and data transmission volume of the corresponding local model as model verification data;

[0162] Each participant 111 uploads its respective model verification data to the central server 110;

[0163] The central server 110 adjusts the level of each participant 111 based on the model verification data corresponding to each participant 111, and issues the adjusted level and the corresponding quantization compression strategy to the corresponding participant 111.

[0164] Specifically, the central server 110 adjusts the level of each participant 111 based on the model verification data corresponding to each participant 111, including:

[0165] When the performance change curve of the local model corresponding to the first participant shows that the model quality improvement efficiency is greater than the preset efficiency, the central server 110 raises the level of the first participant;

[0166] When the memory occupancy of the local model corresponding to the second participant is greater than the first threshold and less than or equal to the second threshold, the central server 110 lowers the level of the second participant;

[0167] When the data transmission volume of the local model corresponding to the third participant is greater than the first transmission volume and less than or equal to the second transmission volume, the central server 110 lowers the level of the third participant.

[0168] Among them, the preset efficiency, the first threshold, the second threshold, the first transmission volume, and the second transmission volume can be custom-configured.

[0169] Specifically, for a participant 111 with a relatively fast and stable improvement in model quality, its rating can be appropriately increased to reduce the compression degree of its model, so as to make full use of its high-quality model parameters to promote the optimization of the global model; similarly, for a participant 111 with a slow or unstable improvement in model quality, its rating can be lowered and the compression intensity can be increased.

[0170] Specifically, the rating of some participants 111 that occupy a large amount of memory is lowered, so that they adopt a more aggressive quantization compression strategy, thereby reducing the memory occupancy during model transmission and aggregation.

[0171] Specifically, the rating of participants 111 with frequent communication or a large amount of transmitted data is lowered, and the model compression ratio is increased, thereby reducing the communication data volume.

[0172] Through the above embodiments, a mechanism for re-audit is introduced. After a certain number of iterations or a certain period of time, the central server 110 will perform an audit operation again and dynamically modify the ratings of each participant 111.

[0173] The central server 110 is further configured to evaluate the credit risk of each customer according to the customer credit risk assessment model.

[0174] It can be understood that in large-scale training in the financial field, the number of models will increase sharply, which will cause an increase in communication overhead and a large amount of memory occupancy during the aggregation process, and has a great impact on the efficiency of the entire training.

[0175] In view of the above problems, in this embodiment, a scoring mechanism is further introduced under the federated learning training mechanism, so as to perform targeted quantization compression on the local models of different participants 111 according to different participant ratings, effectively reducing the size of the uploaded models to reduce communication overhead.

[0176] Moreover, after introducing the rating system, the models of less important participants 111 are processed by means of quantization compression, so as to ensure that the communication volume during the entire training process is in a controllable stage. The rating requirements are dynamically allocated according to the current required model quality, the maximum memory overhead that can be tolerated, and the communication overhead, so as to achieve the final learning and training, which has a more efficient training process compared with traditional federated learning.

[0177] In this embodiment, when the data of the fourth participant is private data, the fourth participant uses the local model to evaluate the credit risk of the corresponding customers.

[0178] For example: If participant 111 pays attention to data privacy and the local data is unique, it is suitable to use the local model for prediction.

[0179] Of course, for scenarios with high requirements for real-time performance and where participant 111 needs to make frequent predictions, using the local model for processing can also reduce the communication with the central server 110, reducing latency and communication costs. For example, in personalized recommendations on mobile devices, the local model can quickly respond to user requests.

[0180] Of course, this embodiment can also be used in the medical field. For example: Hospitals can use the local model after federated learning to predict the disease risks of patients. For example, they can use relatively private data such as the local patient medical records, patient age, patient work situation, and patient gender of each hospital, as well as local characteristic data such as local eating habits and local climate characteristics to train the local model, and use the trained local model to predict the disease risks of local patients. In this way, since the private data is not leaked and the training data has local characteristics, both the privacy of patient data is protected and the prediction results can be optimized according to the characteristics of local patients.

[0181] As can be seen from the above technical solutions, when the present invention trains a model based on federated learning, when each participant first uploads the model parameters of its local model to the central server, the central server performs hierarchical processing on each participant according to the model parameters uploaded by each participant to obtain the level of each participant. After each participant receives its corresponding level and quantization compression strategy, when uploading the local model parameters to the central server subsequently, the model parameters are first subjected to quantization compression processing, thereby reducing the volume of the uploaded model and reducing the communication overhead and computational overhead. While ensuring the accuracy of model detection, the training efficiency of the model and the processing efficiency when using the model for customer credit risk assessment are also improved.

[0182] As Figure 3 shown, it is a schematic structural diagram of a computer device of a preferred embodiment for implementing the method for customer credit risk assessment based on federated learning of the present invention.

[0183] The computer device 1 may include a memory 12, a processor 13, and a bus, and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a customer credit risk assessment program based on federated learning.

[0184] Those skilled in the art can understand that the schematic diagram is only an example of the computer device 1 and does not constitute a limitation on the computer device 1. The computer device 1 may be a bus structure or a star structure. The computer device 1 may also include more or fewer other hardware or software than shown, or different component arrangements. For example, the computer device 1 may also include input / output devices, network access devices, etc.

[0185] It should be noted that the computer device 1 is only an example. Other existing or future possible electronic products that can be adapted to the present invention should also be included within the protection scope of the present invention and are hereby incorporated by reference.

[0186] Among them, the memory 12 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (such as SD or DX memory, etc.), magnetic memory, magnetic disk, optical disc, etc. In some embodiments, the memory 12 can be an internal storage unit of the computer device 1, such as the mobile hard disk of the computer device 1. In other embodiments, the memory 12 can also be an external storage device of the computer device 1, such as a plug-in mobile hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. equipped on the computer device 1. Further, the memory 12 can also include both the internal storage unit and the external storage device of the computer device 1. The memory 12 can be used not only to store application software installed on the computer device 1 and various types of data, such as the code of the customer credit risk assessment program based on federated learning, etc., but also to temporarily store data that has been output or will be output.

[0187] In some embodiments, the processor 13 can be composed of integrated circuits. For example, it can be composed of a single packaged integrated circuit, or can be composed of multiple integrated circuits with the same or different functions packaged, including a combination of one or more central processing units (CPU), microprocessors, digital processing chips, graphics processors, and various control chips, etc. The processor 13 is the control core (Control Unit) of the computer device 1, connecting various components of the entire computer device 1 through various interfaces and lines, and by running or executing programs or modules stored in the memory 12 (such as executing the customer credit risk assessment program based on federated learning, etc.), and calling data stored in the memory 12, to perform various functions of the computer device 1 and process data.

[0188] The processor 13 executes the operating system of the computer device 1 and various installed application programs. The processor 13 executes the application program to implement the steps in the above-mentioned embodiments of various customer credit risk assessment methods based on federated learning, such as Figure 1 the steps shown.

[0189] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory 12 and executed by the processor 13 to implement the present invention. The one or more modules / units may be a series of computer-readable instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the computer device 1. For example, the computer program may be divided into a central server 110 and multiple participating parties 111.

[0190] The integrated units implemented in the form of software function modules as described above may be stored in a computer-readable storage medium. The above software function modules stored in a storage medium include several instructions for causing a computer device (which may be a personal computer, a computer device, or a network device, etc.) or a processor to execute a part of the method for customer credit risk assessment based on federated learning according to each embodiment of the present invention.

[0191] If the modules / units integrated in the computer device 1 are implemented in the form of software function units and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above method embodiments of the present invention, it may also be completed by a computer program instructing relevant hardware devices. The computer program may be stored in a computer-readable storage medium, and when the computer program is executed by a processor, the steps of the above method embodiments may be implemented.

[0192] Among them, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disc, a computer memory, a read-only memory (ROM, Read-Only Memory), a random access memory, etc.

[0193] Further, the computer-readable storage medium mainly includes a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the blockchain node, etc.

[0194] The blockchain referred to in the present invention is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithm. Blockchain, in essence, is a decentralized database, a string of data blocks generated by using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (anti-counterfeiting) and generate the next block. The blockchain can include a blockchain underlying platform, a platform product service layer, an application service layer, etc.

[0195] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience in representation, in Figure 3 it is only represented by a single straight line, but it does not mean that there is only one bus or one type of bus. The bus is arranged to achieve connection and communication between the memory 12 and at least one processor 13, etc.

[0196] Although not shown, the computer device 1 may further include a power supply (such as a battery) for powering each component. Preferably, the power supply can be logically connected to the at least one processor 13 through a power management device, so as to implement functions such as charging management, discharging management, and power consumption management through the power management device. The power supply may also include any components such as one or more DC or AC power supplies, a recharge device, a power failure detection circuit, a power converter or inverter, a power status indicator, etc. The computer device 1 may also include various sensors, a Bluetooth module, a Wi-Fi module, etc., which will not be elaborated here.

[0197] Furthermore, the computer device 1 may further include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is usually used to establish a communication connection between the computer device 1 and other computer devices.

[0198] Optionally, the computer device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, the user interface may also be a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch liquid crystal display, and an OLED (Organic Light-Emitting Diode) toucher, etc. Among them, the display may also be appropriately referred to as a display screen or a display unit, which is used to display the information processed in the computer device 1 and to display a visual user interface.

[0199] It should be understood that the above embodiments are only for illustration purposes and are not limited by this structure in the scope of the patent application.

[0200] Those skilled in the art can understand that Figure 3 the shown structure does not limit the computer device 1, and it may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0201] In combination with Figure 1 , the memory 12 in the computer device 1 stores a plurality of instructions to implement a method for customer credit risk assessment based on federated learning, and the processor 13 can execute the plurality of instructions to implement:

[0202] The central server initializes a global model for customer credit risk assessment and distributes the global model to each participant;

[0203] Each participant obtains local data and trains the global model respectively based on the corresponding local data to obtain a local model for each participant;

[0204] Each participant encrypts and uploads the model parameters of their respective local models to the central server;

[0205] The central server aggregates the model parameters uploaded by each participant to obtain target global model parameters, and performs grading processing on each participant according to the model parameters uploaded by each participant to obtain the level of each participant;

[0206] The central server determines a quantization compression strategy corresponding to each level, and distributes the level of each participant, the corresponding quantization compression strategy, and the target global model parameters to the corresponding participant;

[0207] Each participating party optimizes and trains its respective local model based on the target global model parameters and the local data, performs quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypts and uploads the model parameters obtained after quantization compression to the central server respectively;

[0208] The central server and each participating party perform federated learning training of the model based on the level of each participating party and the corresponding quantization compression strategy until the accuracy of the global model trained by the central server reaches the preset accuracy, and then stop the federated learning training to obtain a customer credit risk assessment model;

[0209] The central server evaluates the credit risk of each customer according to the customer credit risk assessment model.

[0210] Specifically, the specific implementation method of the above instructions by the processor 13 can refer to Figure 1 the description of the relevant steps in the corresponding embodiment, which will not be elaborated here.

[0211] It should be noted that all the data involved in this case are legally obtained. The non-company software tools or components appearing in the embodiments of this application are only for illustrative introduction and do not represent actual use.

[0212] In several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0213] The present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present invention can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0214] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical unit, that is, it may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0215] In addition, in each embodiment of the present invention, each functional module may be integrated in a processing unit, may exist separately as individual physical units, or two or more units may be integrated in one unit. The above integrated unit may be implemented in the form of hardware or in the form of a combination of hardware and software functional modules.

[0216] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0217] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be construed as limiting the claimed invention.

[0218] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices described in the present invention can also be implemented by one unit or device through software or hardware. Words such as "first" and "second" are used to denote names and do not denote any particular order.

[0219] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A customer credit risk assessment method based on federated learning, characterized in that: Applied to a customer credit risk assessment system based on federated learning, the customer credit risk assessment system based on federated learning includes a central server and multiple participants, the central server communicates with the multiple participants, and the customer credit risk assessment method based on federated learning includes: The central server initializes a global model for customer credit risk assessment and sends the global model to each participant; Each participant obtains local data, and trains the global model based on the corresponding local data to obtain a local model of each participant; Each participant encrypts and uploads the model parameters of its own local model to the central server; The central server aggregates the model parameters uploaded by each participant to obtain the target global model parameters, and performs grading processing on each participant according to the model parameters uploaded by each participant to obtain the level of each participant; The central server determines the quantitative compression strategy corresponding to each level, and sends the level of each participant, the corresponding quantitative compression strategy and the target global model parameters to the corresponding participant; Each participant optimizes and trains their own local model according to the target global model parameters and the local data, performs quantization compression processing on the trained model according to the corresponding quantization compression strategy, and encrypts and uploads the model parameters obtained after quantization compression to the central server; The central server and each participant perform federated learning training of the model based on the level of each participant and the corresponding quantization compression strategy until the accuracy of the global model trained by the central server reaches a preset accuracy, then the federated learning training is stopped to obtain a customer credit risk assessment model; The central server evaluates the credit risk of each customer according to the customer credit risk evaluation model.

2. The method for assessing customer credit risk based on federated learning according to claim 1, characterized in that: The step of performing grading processing on each participant according to the model parameters uploaded by each participant to obtain the level of each participant includes: The central server obtains the accuracy, model volume and processing efficiency of the model corresponding to each participant from the model parameters uploaded by each participant; The central server obtains a pre-configured accuracy score mapping table, a model volume score mapping table, and a processing efficiency score mapping table; The central server uses the accuracy of the model corresponding to each participant to match the accuracy score mapping table to obtain the accuracy score of the model corresponding to each participant; The central server uses the model volume of the model corresponding to each participant to match the model volume score mapping table to obtain the model volume score of the model corresponding to each participant; The central server uses the processing efficiency of the model corresponding to each participant to match the processing efficiency score mapping table to obtain the processing efficiency score of the model corresponding to each participant; The central server obtains a first weight corresponding to the accuracy rate, a second weight corresponding to the model volume, and a third weight corresponding to the processing efficiency; The central server calculates a weighted score of the model corresponding to each participant according to the first weight, the second weight, the third weight, the accuracy score of the model corresponding to each participant, the model volume score of the model corresponding to each participant, and the processing efficiency score of the model corresponding to each participant; The central server obtains a participant level mapping table; wherein the participant level mapping table is used to record the mapping relationship between the participant level and the weighted score; The central server uses the weighted score of the model corresponding to each participant to match in the participant level mapping table to obtain the level of each participant.

3. The method for assessing customer credit risk based on federated learning according to claim 1, characterized in that: The quantization compression processing of the trained model according to the corresponding quantization compression strategy includes: Each participant obtains a parameter precision adjustment strategy in a corresponding quantization strategy, and adjusts the number of floating-point digits of a corresponding local model to a preset number of digits according to the parameter precision adjustment strategy; Each participant obtains a parameter sparsification strategy in a corresponding quantization strategy, and adjusts the value of a designated parameter of a corresponding local model to 0 according to the parameter sparsification strategy, and prohibits uploading the designated parameter to the central server; Each participant obtains the model structure simplification strategy in the corresponding quantization strategy, and removes redundant layers and / or redundant neurons of the corresponding local model according to the model structure simplification strategy.

4. The method for assessing customer credit risk based on federated learning according to claim 1, characterized in that: After the trained model is quantized and compressed according to the corresponding quantization compression strategy, the method further includes: Each participant obtains a local test set, and uses the local test set to perform performance verification on the model obtained after quantization and compression to obtain a verification result; When the verification result of a participant shows that the model performance obtained after quantization compression is lower than the performance threshold, the participant adjusts the corresponding quantization compression strategy and reports the adjusted quantization compression strategy and the corresponding verification result to the central server.

5. The method for assessing customer credit risk based on federated learning according to claim 1, characterized in that: The method further comprises: During the federated learning training process, when the number of iterations reaches the preset number, each participant obtains the performance change curve, memory usage, and data transmission volume of the corresponding local model as model verification data; Each participant uploads their own model verification data to the central server; The central server adjusts the level of each participant based on the model verification data corresponding to each participant, and sends the adjusted level and the corresponding quantitative compression strategy to the corresponding participant.

6. The method for assessing customer credit risk based on federated learning according to claim 5, characterized in that: The central server adjusts the level of each participant based on the model verification data corresponding to each participant, including: When the performance change curve of the local model corresponding to the first participant shows that the model quality improvement efficiency is greater than the preset efficiency, the central server upgrades the level of the first participant; When the memory usage of the local model corresponding to the second participant is greater than the first threshold and less than or equal to the second threshold, the central server lowers the level of the second participant; When the data transmission volume of the local model corresponding to a third participant is greater than the first transmission volume and less than or equal to the second transmission volume, the central server lowers the level of the third participant.

7. The method for assessing customer credit risk based on federated learning according to claim 1, characterized in that: The method further comprises: When the data of a fourth party is private data, the fourth party uses a local model to assess the credit risk of the corresponding customer.

8. A customer credit risk assessment system based on federated learning, characterized in that: The customer credit risk assessment system based on federated learning includes a central server and multiple participants, the central server communicates with the multiple participants, and the customer credit risk assessment system based on federated learning includes: The central server is used to initialize a global model for customer credit risk assessment and send the global model to each participant; Each participant is configured to obtain local data and train the global model based on the corresponding local data to obtain a local model of each participant; Each participant is also used to encrypt and upload the model parameters of the respective local models to the central server; The central server is further used to aggregate the model parameters uploaded by each participant to obtain the target global model parameters, and to grade each participant according to the model parameters uploaded by each participant to obtain the level of each participant; The central server is further used to determine the quantitative compression strategy corresponding to each level, and send the level of each participant, the corresponding quantitative compression strategy and the target global model parameters to the corresponding participant; Each participant is further configured to optimize and train their respective local models according to the target global model parameters and the local data, quantize and compress the trained models according to the corresponding quantization and compression strategies, and encrypt and upload the model parameters obtained after quantization and compression to the central server; The central server and each participant are further used to perform federated learning training of the model based on the level of each participant and the corresponding quantization compression strategy, until the accuracy of the global model trained by the central server reaches a preset accuracy, then stop the federated learning training and obtain a customer credit risk assessment model; The central server is also used to evaluate the credit risk of each customer according to the customer credit risk evaluation model.

9. A computer device, characterized in that: The computer device comprises: a memory storing at least one instruction; and A processor executes instructions stored in the memory to implement the customer credit risk assessment method based on federated learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and the at least one instruction is executed by a processor in a computer device to implement the customer credit risk assessment method based on federated learning as described in any one of claims 1 to 7.

Images