Credit assessment method, system, device, equipment, medium and product

Abstract

The invention provides a credit evaluation method, system, device, equipment, medium and product, and the method comprises the steps: responding to a credit evaluation request which is transmitted to a target user by a target federation participant, and transmitting a user data obtaining request for the target user to each federation participant; under the condition that user encrypted data returned by each federated participant for the user data acquisition request is received, integrating the user encrypted data returned by each federated participant to obtain a user feature set; sending a credit prediction instruction and a user feature set to each federated participant; and under the condition of receiving an encryption result of a credit prediction score returned by each federated participant after processing the user feature set according to the credit prediction instruction, performing fusion processing on the encryption result returned by each federated participant to obtain a credit score of the target user, and sending the credit score to the target federated participant.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present specification relate to the technical field of computer technology, and in particular to a credit evaluation method, system, device, equipment, medium and product. BACKGROUND

[0002] In the scenarios of credit approval, lease without deposit, qualification review, etc., in order to quantify risks and establish trust, credit evaluation is usually performed on applicants (such as individual users or enterprises) to predict the future performance ability and willingness of the applicants, thereby reducing the default risk and identifying high-quality applicants to match them with better services and resources, achieving a fair and efficient social trust system construction, and laying a foundation for orderly economic operation.

[0003] To realize credit evaluation, a credit inquiry API (Application Programming Interface) is usually connected to a credit inquiry system in a credit approval platform, a lease without deposit platform, or a qualification review platform to query a credit report of an applicant. The credit report records the private data of the applicant, and neither the applicant nor the above platforms expect the private data to be leaked, but the above way of querying the credit report through the credit inquiry API is likely to cause data leakage due to vulnerabilities in the API, which shows that the related technology has a high risk of data leakage. SUMMARY

[0004] Therefore, to at least solve the technical problem of the related technology that has a high risk of data leakage in the credit evaluation process, one or more embodiments of the present specification provide technical solutions as follows. According to a first aspect of one or more embodiments of the present specification, a credit evaluation method is provided, applied to a federation server, comprising: In response to a credit evaluation request issued by a target federation participant for a target user, sending a user data acquisition request for the target user to each federation participant; the each federation participant includes the target federation participant and other federation participants; In the case of receiving user encrypted data returned by the each federation participant in response to the user data acquisition request, performing integration processing on the user encrypted data returned by the each federation participant to obtain a user feature set; Sending a credit prediction instruction and the user feature set to the each federation participant; the credit prediction instruction is used to trigger the each federation participant to call a locally deployed credit prediction model to process the user feature set to obtain a credit prediction score; In the case that the encrypted results returned by the federal participants are received, the encrypted results returned by the federal participants are fused to obtain the credit score of the target user, and the credit score is sent to the target federal participant.

[0005] According to a second aspect of one or more embodiments of the present specification, another credit evaluation method is provided, which is applied to a target federal participant, and includes: sending a credit evaluation request for a target user to a federal server; the credit evaluation request is used to trigger the federal server to send a user data acquisition request for the target user to federal participants; the federal participants include the target federal participant and other federal participants; in response to the user data acquisition request sent by the federal server, obtaining user data of the target user, encrypting the user data, and returning encrypted user data to the federal server; in response to a credit prediction instruction for the user encrypted data sent by the federal server, calling a locally deployed credit prediction model to process a user feature set sent by the federal server to obtain a credit prediction score, and returning an encrypted result of the credit prediction score to the federal server; the user feature set is obtained by the federal server by integrating the user encrypted data returned by the federal participants; in the case that the credit score returned by the federal server is received, the credit score is sent to a user terminal associated with the target user; the credit score is obtained by the federal server by fusing the encrypted results returned by the federal participants.

[0006] According to a third aspect of one or more embodiments of the present specification, a credit evaluation system is provided, which includes: a federal server, configured to, in the case that a credit evaluation request for a target user is received from a target federal participant, obtain a credit score of the target user by executing the steps of the credit evaluation method provided in the first aspect, and send the credit score to the target federal participant; and a plurality of federal participants, including the target federal participant and other federal participants; the target federal participant is configured to cooperate with the federal server to obtain the credit score of the target user by executing the steps of the credit evaluation method provided in the second aspect, and send the credit score to a user terminal associated with the target user; the other federal participants are configured to: In response to a user data acquisition request sent by the federal server, user data of the target user is acquired and encrypted, and the encrypted user encrypted data is returned to the federal server; In response to a credit prediction instruction sent by the federal server, a locally deployed credit prediction model is called to process a user feature set sent by the federal server to obtain a credit prediction score, and the credit prediction score is sent to the federal server.

[0007] According to a fourth aspect of one or more embodiments of the present specification, a credit evaluation device is provided, applied to a federal server, comprising: The federal request module is configured to: in response to a credit evaluation request issued by a target federal participant for a target user, send a user data acquisition request for the target user to each federal participant; the federal participants include the target federal participant and other federal participants; The feature processing module is configured to: in the case of receiving user encrypted data returned by the federal participants in response to the user data acquisition request, integrate and process the user encrypted data returned by the federal participants to obtain a user feature set; The federal call module is configured to: send a credit prediction instruction and the user feature set to the federal participants; the credit prediction instruction is used to trigger the federal participants to call a locally deployed credit prediction model to process the user feature set to obtain a credit prediction score; The credit score module is configured to: in the case of receiving an encrypted result of the credit prediction score returned by the federal participants in response to the credit prediction instruction, fuse and process the encrypted result returned by the federal participants to obtain a credit score of the target user, and send the credit score to the target federal participant.

[0008] According to a fifth aspect of one or more embodiments of the present specification, another credit evaluation device is provided, applied to a target federal participant, comprising: The credit request module is configured to: send a credit evaluation request for a target user to a federal server; the credit evaluation request is used to trigger the federal server to send a user data acquisition request for the target user to each federal participant; the federal participants include the target federal participant and other federal participants; The data processing module is configured to: in response to the user data acquisition request sent by the federal server, acquire user data of the target user, encrypt the user data, and return the encrypted user encrypted data to the federal server; The credit prediction module is configured to: respond to the credit prediction instruction from the federal server for the encrypted user data, call the locally deployed credit prediction model to process the user feature set sent by the federal server, obtain a credit prediction score, and return the encrypted result of the credit prediction score to the federal server; the user feature set is obtained by the federal server integrating and processing the encrypted user data returned by each federal participant. The credit score sending module is configured to: upon receiving a credit score returned by the federal server, send the credit score to the user terminal associated with the target user; the credit score is obtained by the federal server through fusion processing based on the encrypted results returned by each federal participant.

[0009] According to a sixth aspect of one or more embodiments of this specification, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor implements the steps of the credit assessment method provided in the first and / or second aspects by executing the executable instructions.

[0010] According to a seventh aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the credit assessment method provided in the first and / or second aspects described above.

[0011] According to an eighth aspect of one or more embodiments of this specification, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of the credit assessment method provided in the first and / or second aspects described above.

[0012] As can be seen from the above embodiments, in the process of credit assessment of target users, this specification sends a request to each federal participant to obtain user data of the target user through the federal server, triggering each federal participant to return encrypted user data. This ensures that the user data recorded by each federal participant is encrypted before being transmitted to the federal server, guaranteeing the confidentiality and security of the data during transmission, preventing theft on the communication link, and ensuring data privacy. Next, the federal server continues to integrate and process all encrypted user data to obtain a user feature set, which is then sent to each federal participant. This triggers each federal participant to call its locally deployed credit prediction model to process the user feature set and obtain a credit prediction score. This ensures that the user feature set not only includes the features of the encrypted user data returned by each federal participant (i.e., the user feature set includes multimodal data), but also remains encrypted, thus guaranteeing the confidentiality and security of the data during transmission. Furthermore, once the data arrives at each federal participant, the participants cannot know the original data, thus ensuring the confidentiality and security of the data during processing by each federal participant. In addition, since the credit prediction model of each federal participant is based on multimodal data processing to obtain the credit prediction score, the data silo problem can be avoided, improving the accuracy of the credit prediction score and the coverage of applicants suitable for credit assessment (e.g., increasing the loan approval rate for those with no credit history from 30% to about 54%). Next, the encrypted credit prediction scores returned by each participating party are further processed and merged by the federated server to obtain the target user's credit score. On the one hand, since the data returned by each participating party is an encrypted result of the credit prediction score, the confidentiality and security of the encrypted result during data transmission can still be guaranteed. On the other hand, the federated server's fusion processing based on all encrypted results to obtain the credit score is beneficial for comprehensively considering multimodal data, improving the reliability of the final credit score, and better reducing the risk of default. In summary, the technical solution provided by one or more embodiments of this specification, by combining multimodal data and data privacy processing in a federated manner, can effectively solve the problems of high privacy leakage risk and data silos in related technologies, and can be widely applied to risk control decision support in scenarios such as credit approval, lease deposit waiver, and qualification review. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the architecture of a credit assessment service system provided in an exemplary embodiment.

[0014] Figure 2 This is a flowchart of an exemplary embodiment of a credit assessment method applied to a federated server.

[0015] Figure 3 This is a flowchart of an exemplary embodiment of a credit assessment method applied to a target federal participant.

[0016] Figure 4 This is a schematic structural diagram of a device provided in an exemplary embodiment.

[0017] Figure 5 This is a structural block diagram of a credit assessment device applied to a federal server, provided as an exemplary embodiment.

[0018] Figure 6 This is a structural block diagram of a credit assessment device applied to a target federal participant, provided as an exemplary embodiment. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this manual are all information and data authorized by the user or fully authorized by all parties. The collection, use and processing of related data shall comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals shall be provided for users to choose to authorize or refuse.

[0021] The following is an explanation of some technical terms used in this manual: Federated learning: a distributed machine learning paradigm in which multiple federated participants jointly train a machine learning model without needing to centralize the original data.

[0022] Multimodal data refers to data with different structures, forms, or sources, such as tabular financial data, sequential user behavior data, and graph-structured social data.

[0023] Privacy-preserving computation: a general term for a class of technologies, including secure multi-party computation, trusted execution environments, homomorphic encryption, etc., to protect data privacy during processing and achieve the goal of data being usable but not visible.

[0024] Homomorphic encryption refers to a special type of encryption that allows the federated server to perform mathematical operations, such as addition or multiplication, directly on the ciphertext. The decrypted result is identical to the result of performing the same operation on the plaintext. Throughout the entire process, the federated server never accesses the plaintext.

[0025] TEE (Trusted Execution Environment): is a secure zone configured within a federated server that ensures the confidentiality and integrity of code and data loaded within it.

[0026] Federal participants: refers to an entity in a federal collaborative framework that has independent data resources and is willing to participate in collaborative tasks, such as banks, e-commerce platforms, and government platforms.

[0027] Privacy processing refers to a technology that, after processing raw data, can produce encrypted data with encryption effects. This includes encryption processing and federated feature processing.

[0028] Data silos refer to a situation where data is isolated within an organization or between different organizations, making it impossible or difficult to interconnect, share, and integrate. For example, in credit assessment scenarios, data from financial institutions, internet platforms, and government platforms is scattered and fragmented, making it impossible to share.

[0029] Credit history length: refers to the length of time since the applicant opened their first credit account (such as a credit card or loan).

[0030] Average order price: This refers to the average amount paid by the applicant for orders placed within a certain period. For example, if the applicant placed three orders within a month, and the total amount spent on these three orders was 90 yuan, then the average order price would be 90 / 3 = 30 yuan.

[0031] To address the technical issue of high data leakage risks associated with related technologies during credit assessment, this specification proposes a method whereby the federal server sends user data acquisition requests for the target user to each participating party during the credit assessment process. This triggers each participating party to return encrypted user data, ensuring that all user data recorded by each participating party is encrypted before being transmitted to the federal server. This guarantees the confidentiality and security of the data during transmission, prevents theft on the communication link, and protects data privacy. Next, the federal server continues to integrate and process all encrypted user data to obtain a user feature set, which is then sent to each federal participant. This triggers each federal participant to call its locally deployed credit prediction model to process the user feature set and obtain a credit prediction score. This ensures that the user feature set not only includes the features of the encrypted user data returned by each federal participant (i.e., the user feature set includes multimodal data), but also remains encrypted, thus guaranteeing the confidentiality and security of the data during transmission. Furthermore, once the data arrives at each federal participant, the participants cannot know the original data, thus ensuring the confidentiality and security of the data during processing by each federal participant. In addition, since the credit prediction model of each federal participant is based on multimodal data processing to obtain the credit prediction score, the data silo problem can be avoided, improving the accuracy of the credit prediction score and the coverage of applicants suitable for credit assessment (e.g., increasing the loan approval rate for those with no credit history from 30% to about 54%). Next, the encrypted credit prediction scores returned by each participating party are further processed and merged by the federated server to obtain the target user's credit score. On the one hand, since the data returned by each participating party is an encrypted result of the credit prediction score, the confidentiality and security of the encrypted result during data transmission can still be guaranteed. On the other hand, the federated server's fusion processing based on all encrypted results to obtain the credit score is beneficial for comprehensively considering multimodal data, improving the reliability of the final credit score, and better reducing the risk of default. In summary, the technical solution provided by one or more embodiments of this specification, by combining multimodal data and data privacy processing in a federated manner, can effectively solve the problems of high privacy leakage risk and data silos in related technologies, and can be widely applied to risk control decision support in scenarios such as credit approval, lease deposit waiver, and qualification review.

[0032] Based on this, let's first introduce an example of an application scenario for the credit assessment method provided in this manual: Figure 1 This is a schematic diagram of the architecture of a credit assessment service system provided in an exemplary embodiment. Figure 1 As shown, the system may include a server 11, a network 12, and several electronic devices, such as a PC (Personal Computer) 13, a mobile phone 14, etc.

[0033] Server 11 can be a physical server containing an independent host, or it can be a virtual server hosted in a host cluster. During operation, server 11 can run server-side programs for a certain application to implement the relevant functions of that application. For example, when server 11 runs a credit assessment service program, it can function as a corresponding credit assessment service platform.

[0034] PC13 and mobile phone14 are just some of the types of electronic devices that users can use. In reality, users can obviously also use electronic devices such as tablets, laptops, PDAs (Personal Digital Assistants), wearable devices (such as smart glasses, smartwatches, etc.), etc., and one or more embodiments in this specification do not limit this. During operation, the electronic device can run a client-side program of an application to achieve the relevant functions of that application. For example, when the electronic device runs a credit assessment service program, it can act as a client for that credit assessment service. The aforementioned credit assessment service client application can be launched and run on the electronic device. This client-side program can be a native application installed on the electronic device, or it can be a mini-program, quick app, or other similar form. Of course, when using web technologies such as HTML5 or similar, the relevant functions can be achieved through a page displayed by a browser. This browser can be a standalone browser application or a browser module embedded in some applications.

[0035] As for the network 12 that enables interaction between electronic devices such as PC13 and mobile phone 14 and server 11, communication can be achieved using either wired or wireless networks, depending on the communication methods supported by the respective electronic devices. This specification does not impose any restrictions on this. For example, PC13 can support both wired and wireless communication, so it can use either wired or wireless networks as needed. Mobile phone 14 typically only supports wireless communication, so it can use a wireless network for communication.

[0036] The credit assessment method provided in one or more embodiments of this specification can be applied to a federated server, which can be server 11 described above. Please refer to [link / reference needed]. Figure 2 , Figure 2 This is a flowchart of an exemplary embodiment of a credit assessment method applied to a federated server, the method comprising the following steps: In step S100, in response to the credit assessment request issued by the target federal participant for the target user, a user data acquisition request for the target user is sent to each federal participant; the federal participants include the target federal participant and other federal participants. In step S200, upon receiving encrypted user data returned by each of the federation participants in response to the user data acquisition request, the encrypted user data returned by each of the federation participants is integrated to obtain a user feature set. In step S400, a credit prediction instruction and the user feature set are sent to each of the federal participants; the credit prediction instruction is used to trigger each of the federal participants to call the locally deployed credit prediction model to process the user feature set and obtain a credit prediction score. In step S500, upon receiving the encrypted results of the credit prediction scores returned by each of the federal participants in response to the credit prediction instruction, the encrypted results returned by each of the federal participants are fused to obtain the credit score of the target user, and the credit score is sent to the target federal participant.

[0037] In scenarios such as loan approval, mortgage waiver for leases, or eligibility verification, if an applicant submits a request for a loan, mortgage waiver for leases, or eligibility verification through the relevant systems provided in these scenarios, the relevant platform will generate a credit assessment request for the applicant upon receiving the request. For example, if applicant A submits a loan application to bank A, after receiving the loan application, the bank needs to know applicant A's credit score, thus triggering a credit assessment request and sending it to the federated server. It is clear that the target user in this case is applicant A, and the target federated participant is bank A. The credit assessment request can carry applicant A's identity information. This identity information is used to distinguish the applicant's identity and can be any of the following: a mobile phone number verified with real-name authentication, a registered email address, or a device ID verified with real-name authentication, but it is not limited to these, as long as it ensures that the unique user identity can be confirmed based on the identity information.

[0038] Upon receiving a credit assessment request, the federal server executes step S100, responding to the request by sending a user data acquisition request for the target user to each federal participant, thereby triggering each participant to acquire the target user's user data. This user data acquisition request carries the target user's identity information to ensure that each federal participant can accurately acquire the target user's data. Federal participants may include, but are not limited to, banks, e-commerce platforms, and government platforms. Furthermore, not all federal participants necessarily record the target user's user data. For federal participants that do not record the target user's user data, the data returned after responding to the aforementioned user data acquisition request can be an empty set or a feedback message indicating that no user data for the target user is recorded.

[0039] Taking the user data of target users recorded by banks and e-commerce platforms as an example, the user data obtained by banks for target users may include, but is not limited to: the target user's identity information, monthly income, credit history length, number of historical overdue payments, account transaction records, and credit records; the user data obtained by e-commerce platforms for target users may include, but is not limited to: the target user's identity information, user activity level, average order price, annual total consumption, shopping behavior data, browsing history, return records, and review behavior data.

[0040] After receiving the aforementioned user data acquisition request, each participating party in the federation obtains the target user's data. To ensure the confidentiality and security of the data during transmission, they perform privacy processing on the user data before returning it to the federation server. It is evident that the data received by the federation server from each participating party is encrypted user data. This privacy processing can be implemented using encryption techniques or federated feature processing techniques, which will not be detailed here.

[0041] The following examples illustrate an exemplary representation of encrypted user data provided by two federal parties: user data obtained by banks, including target user identity information, monthly income, credit history length, and number of past delinquencies; and user data obtained by e-commerce platforms, including target user identity information, user activity, average order price, and annual total spending. The encrypted user data provided by the bank is: {"user_id_hash": "a1b2c3", "feature_band": [6500, 12, 0]}; where "a1b2c3" represents identity information, and "feature_band": [6500, 12, 0] represents a monthly income of 6500 yuan, a credit history length of 12 months, and 0 historical overdue payments.

[0042] The e-commerce platform provides encrypted user data: {“user_id_hash”: “a1b2c3”, “feature_mall”: [0.85, 45, 1200]}; where “a1b2c3” represents identity information, and “feature_mall”: [0.85, 45, 1200] indicates that the user's activity level is 0.85, the average order price is 45 yuan, and the total annual consumption is 1200 yuan.

[0043] After receiving the encrypted user data returned by each federation participant, the federation server executes step S200 to integrate all the encrypted user data and obtain a user feature set. As an example of integration processing, the encrypted user data returned by each federation participant can be merged to form a single user feature set. Continuing with the above example, the user feature set at this time is: {"feature_band": [6500, 12, 0], "feature_mall": [0.85, 45, 1200]}.

[0044] Although the above integration process can yield a user feature set, different federation participants may record user data using different data formats, and there may be duplicate user data. Furthermore, as mentioned above, encrypted user data includes large numerical values, which may increase the model's processing complexity and computational load, and decrease the model's output accuracy. Therefore, to address these technical problems, in some embodiments, the credit assessment method provided by one or more embodiments of this specification also provides another integration processing scheme. That is, the step S200 above, which involves integrating the encrypted user data returned by each federation participant to obtain the user feature set, may include the following steps: In step S210, the remaining encrypted data other than the identity information of the target user is extracted from the user encrypted data returned by each federation participant, and the remaining encrypted data is deduplicated to obtain the target encrypted data of the target user. In step S220, the target encrypted data is subjected to privacy processing to obtain re-encrypted data; In step S230, the re-encrypted data is subjected to feature alignment processing to obtain the user feature set.

[0045] Understandably, during the execution of step S200, the encrypted user data returned by each federation participant can be integrated by executing steps S210 to S230.

[0046] To avoid processing duplicate user features and affecting processing efficiency and output accuracy, step S210 is executed first to extract the remaining encrypted data (excluding identity information) from all encrypted user data returned by each federated participant. Then, the remaining encrypted data is deduplicated to obtain the target encrypted data. The data processing procedure described in step S210 is also equivalent to a privacy-preserving intersection process.

[0047] Next, in preparation for subsequent privacy computations, to ensure that the data remains encrypted during processing by each federal participant based on the user feature set, step S220 is executed to perform secondary privacy processing on the target encrypted data, resulting in re-encrypted data. This secondary privacy processing can be implemented using related technologies such as homomorphic encryption, secure multi-party computation, or de-identification techniques.

[0048] After obtaining the re-encrypted data, in order to enable data from different sources and in different formats to be processed under a unified language and standard, so as to release the true value of the data, enhance data consistency and comparability, and thus improve the processing performance and prediction accuracy of the credit prediction model, step S230 is executed to perform feature alignment processing on the re-encrypted data to obtain a user feature set, so that features from different sources and in different formats in the user feature set are standardized under a unified language and standard.

[0049] The feature alignment processing described above can be implemented using relevant technologies. However, to make the user feature set more suitable for credit assessment scenarios and thus better improve the processing performance and prediction accuracy of the credit prediction model for the user feature set, in some embodiments, the credit assessment method provided by one or more embodiments of this specification also provides another feature alignment scheme. That is, the step of performing feature alignment processing on the re-encrypted data to obtain the user feature set in step S230 above may include the following steps: In step S231, the re-encrypted data is standardized to obtain standardized feature data after eliminating the differences in feature dimensions among different federal participants. In step S232, the standardized feature data is normalized and missing value processing is performed to obtain the user feature set.

[0050] Understandably, during the execution of step S230, feature alignment processing of the re-encrypted data can be achieved by executing steps S231 to S232, thereby obtaining the user feature set.

[0051] First, step S231 is executed to standardize the re-encrypted data. For example, the data is converted into a distribution with a mean of 0 and a standard deviation of 1, thereby obtaining standardized feature data that eliminates the differences in feature dimensions among different federation participants. For example, assuming that the monthly income in the encrypted user data provided by the bank is 20,000 yuan, and the annual consumption amount in the encrypted user data provided by the e-commerce platform is 150,000 yuan, then after standardizing the re-encrypted data, the standardized monthly income can be 1.2, and the standardized annual consumption amount can be 0.8.

[0052] Next, step S232 is executed to perform feature normalization and missing value processing on the standardized feature data, resulting in a user feature set. This can be done by first normalizing the standardized feature data and then processing the missing values. Feature normalization maps each feature data point to a unified numerical range [0, 1]. During the missing value processing of the normalized feature data, a unified strategy can be used to fill in the missing feature data from each participating party. For example, if the bank lacks monthly income data for the target user, and the e-commerce platform lacks annual total consumption data for the target user, then the average monthly income per person, the median monthly income per person, or a specific value configured based on actual needs or experience can be used to fill in the target user's monthly income data. The same applies to filling in the annual total consumption data. This transforms incomplete and flawed feature data into complete, usable, and high-quality feature data, ensuring the data integrity of the user feature set and improving the reliability and accuracy of subsequent model predictions.

[0053] When a credit assessment request is initiated for multiple target users, the federated server may send a user data retrieval request to each federated participant for each target user. Therefore, all encrypted user data received by the federated server may include encrypted user data from multiple target users. This encrypted user data may include the encrypted ID set returned by each federated participant and the privacy-processed user data described above. In this regard, during step S210, the common target user set of different federated participants can be determined first based on all encrypted ID sets. For example, assuming there are four target users, the encrypted ID set provided by the bank federated participant is {E(a1), E(b2), E(d4)}, and the encrypted ID set provided by the e-commerce platform federated participant is {E(a1), E(b2), E(c3)}. To obtain the common target user set of these encrypted ID sets, the intersection of these encrypted ID sets can be calculated using privacy intersection techniques. The intersection is then {E(a1), E(b2)}, meaning that target user a1 and target user b2 are common users of the bank federated participant and the e-commerce platform federated participant. It is evident that during the intersection calculation process, non-intersection IDs are not exposed. Next, based on the identity information E(a1) of target user a1, the encrypted user data of target user a1 can be obtained from all user encrypted data. Similarly, the encrypted user data of target user b2 can also be obtained. Subsequently, for each target user, the remaining operations of step S210, as well as steps S220 to S230, are executed to obtain the user feature set of each target user.

[0054] After obtaining the user feature set through any of the above embodiments, step S400 is executed to send a credit prediction instruction and the user feature set to each federal participant. The credit prediction instruction triggers each federal participant to call its locally deployed credit prediction model to process the user feature set and obtain a credit prediction score.

[0055] Understandably, after receiving the credit prediction instruction and the user feature set, each federal participant responds to the credit prediction instruction by inputting the user feature set into a pre-trained credit prediction model to obtain the credit prediction score output by the credit prediction model.

[0056] In the above, credit prediction models deployed locally by different federal participants can be trained based on the same initial network model or on different initial network models. In some embodiments, each federal participant can choose a more suitable initial network model to train its credit prediction model based on its own data characteristics. For example, banks can use a logistic regression model to train their credit prediction model, while e-commerce platforms can use an XGBoost model. Different models can complement each other, which is beneficial to improving the accuracy of subsequent credit scoring and also protects data privacy because the federal participants do not need to share the same model structure.

[0057] In the above, the bank's initial network model can be trained based on a training dataset provided locally by the bank. The training samples in the bank's training dataset can include the user data provided by the bank as described above. The label for each training sample can include a first true credit score, which can be calculated based on the user's overdue payment records and / or default status. The calculation principle can be found in relevant technologies, or it can be obtained through manual annotation. Therefore, the bank's initial network model can be iteratively trained using the bank's training dataset until the initial network model converges, thus obtaining the corresponding credit prediction model.

[0058] Similarly, the credit prediction model of an e-commerce platform can also be based on the e-commerce training dataset provided locally by the e-commerce platform. The training samples in the e-commerce training dataset can include the user data provided by the e-commerce platform as described above. The sample label of each training sample can include a second true credit score, which can be calculated based on the user's payment behavior. The calculation principle can be found in relevant technologies, or it can be obtained through manual annotation. Therefore, the initial network model of the e-commerce platform can be iteratively trained using the e-commerce training dataset until the initial network model converges, thus obtaining the corresponding credit prediction model.

[0059] In some embodiments, to better improve the efficiency and effectiveness of credit prediction, the credit assessment method provided in one or more embodiments of this specification also provides a model training scheme. Based on this, an initial meta-model is deployed on the federated server. The initial network model used to train the credit prediction model includes an initial local model and an initial local meta-model, wherein the initial meta-model and the initial local meta-model have the same model structure. Accordingly, the training process of the initial network model includes: In step S011, a model training request is initiated to each federation participant. The model training request is used to trigger the federation participant to call the initial local model to process the training dataset obtained from the local model, and obtain a training intermediate result set and a training prediction result set. The training intermediate result set includes the results output by the intermediate layer of the initial local model for each sample in the training dataset, such as the gradient vector of the intermediate layer of the model or the features generated after the training dataset is partially processed inside the model. The number of prediction scores included in the training prediction result set is the same as the number of training samples contained in the corresponding training dataset. In step S012, upon receiving the training intermediate result set and training prediction result set returned by each federation participant in response to the model training request, the initial meta-model is trained using the training intermediate result set and training prediction result set returned by each federation participant until the initial meta-model converges, thus obtaining the global meta-model. In step S013, with the global meta-model obtained, the model parameters of the global meta-model and parameter update instructions are sent to each federation participant. The parameter update instructions are used to trigger each federation participant to update the model parameters of the initial local meta-model to the model parameters of the global meta-model in order to obtain the local meta-model. The model parameters of the initial local model are adjusted through the local meta-model, or the first layer of the local meta-model is connected to the last layer of the initial local model to obtain the credit prediction model.

[0060] Understandably, during the model training phase, the initial meta-model in the federated server and the initial network models in each federated participant can be trained through the steps S011 to S013 described above.

[0061] During the model training phase, developers can trigger a model training request through the human-computer interaction interface configured on the federation server. At this time, the federation server executes step S011 to send a model training request to each federation participant.

[0062] After receiving a model training request, each federation participant calls its initial local model to process the training dataset obtained locally, resulting in a set of intermediate training results and a set of prediction results for training. The training dataset is described above and will not be repeated here. It is evident that each federation participant only needs to obtain its own training dataset locally during the training of its initial local model, without needing to obtain it from the federation server. This effectively reduces the amount of communication data during training, and each participant only needs to focus on its own local data and initial local model, thus improving training efficiency to some extent. Furthermore, the training of the initial local model by each federation participant using its local training dataset allows the initial local model to achieve a certain level of convergence.

[0063] To improve the convergence efficiency, convergence effect, and output accuracy of the initial local model, each federation participant sends the training intermediate result set and training prediction result set output by the initial local model to the federation server. In some embodiments, to protect data privacy, each federation participant first performs privacy processing on the training intermediate result set and training prediction result set before sending them to the federation server.

[0064] After receiving the training intermediate result sets and training prediction result sets from each federated participant, the federated server executes step S012. First, it performs feature alignment processing on all training intermediate results and training prediction result sets. For example, based on the user ID, it merges the training intermediate results and training prediction result sets for the same user to obtain the corresponding input feature set. Then, it inputs the input feature set and the corresponding global credit score label into the initial meta-model and iteratively trains the initial meta-model until it converges, thus obtaining the global meta-model. The global credit score label can be obtained manually, but is not limited to this method. It is evident that the initial meta-model can learn how to combine the output features of each federated participant into a better global credit score prediction, achieving effective fusion of multimodal data.

[0065] After obtaining the global metamodel, step S013 is executed to send the model parameters and parameter update instructions of the global metamodel to each federation participant.

[0066] Upon receiving the parameter update instruction, each participating party in the federation responds by updating the model parameters of the initial local meta-model to the model parameters of the global meta-model, thus obtaining the local meta-model. Subsequently, to improve the prediction accuracy of the initial local model, each participating party can further utilize the local meta-model in conjunction with the initial local model to obtain a credit prediction model.

[0067] One approach is to use a local meta-model to adjust the parameters of the initial local model. Once the parameters of the initial local model are adjusted, the local prediction model is obtained. As an example of parameter adjustment, user features obtained locally can be input into both the initial local model and the local meta-model. The loss between the predicted scores output by the initial local model and the local meta-model is then calculated. Based on this loss, the parameters of the initial local model are adjusted. This process is repeated multiple times until the initial local model converges. At this point, the parameters of the initial local model are considered to have been adjusted, and the initial local model at this stage is the local prediction model.

[0068] Based on this, during the model application phase, when each federal participant receives a credit prediction instruction from the federal server, they can input the user feature set sent by the federal server into their local prediction model and use the output of the local prediction model as the credit prediction score.

[0069] As another approach, the first layer of the local meta-model can be connected to the last layer of the initial local model to obtain a credit prediction model. In essence, the current initial local model can be used as the local prediction model, and the last layer of the local prediction model can be connected to the first layer of the local meta-model to construct the credit prediction model.

[0070] Based on this, during the model application phase, when each federal participant receives a credit prediction instruction from the federal server, they can input the user feature set sent by the federal server into the local prediction model and the local meta-model. This allows the local meta-model to process the output of the local prediction model and the user feature set, and output a more accurate credit prediction score.

[0071] As another approach, the first layer of the local meta-model can be connected to the middle and last layers of the initial local model to obtain a credit prediction model. Similarly, the current initial local model can be considered as a local prediction model.

[0072] Based on this, during the model application phase, when each federal participant receives a credit prediction instruction from the federal server, they can input the user feature set sent by the federal server into the local prediction model. The intermediate and output results obtained by the local prediction model from processing the user feature set are then input into the local meta-model to trigger the local meta-model to output a more accurate credit prediction score.

[0073] As can be seen from the above, the model training scheme provided in one or more embodiments of this specification is a model training scheme that combines hierarchical federated learning and multimodal data fusion. By having each federated participant train the model based on its local training dataset, it can not only reduce the amount of communication data between each federated participant and the federated server, but also ensure that user data does not leave the domain, which is conducive to ensuring data privacy and security during the training phase. Furthermore, by using a unified framework meta-model to fuse multimodal data, it achieves a dual improvement in model training efficiency and training effect.

[0074] Based on the embodiments shown in steps S011 to S013 above, to ensure that the model maintains its original or even higher output accuracy over time and to improve the robustness of the model, in some embodiments, the credit assessment method provided in one or more embodiments of this specification also provides a model update scheme. This model update scheme can be executed by timed triggering, for example, executing the model update scheme every week or every month. Based on this, the credit assessment method may further include the following steps: In step S021, a first model update instruction is sent to each of the federation participants. The first model update instruction is used to trigger the federation participants to call their local prediction models to process the user feature sample set obtained locally, and obtain an intermediate result set and a credit prediction result set. The intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set. The number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the corresponding user feature sample set. In step S022, upon receiving the intermediate result set and credit prediction result set returned by each of the federation participants for the model update instruction, the global meta-model is trained using the intermediate result set and credit prediction result set returned by each of the federation participants until the global meta-model converges. In step S023, if the global meta-model converges, the target model parameters and the second model update instruction of the converged global meta-model are sent to each federation participant. The second model update instruction is used to trigger each federation participant to update the model parameters of the local meta-model to the target model parameters, and adjust the model parameters of the local prediction model through the updated local meta-model, or connect the first layer of the updated local meta-model to the last layer of the local prediction model.

[0075] The technical principle of updating the parameters of the global meta-model and the credit prediction models of each federal participant by executing the above steps S021 to S023 can be found in the relevant records of the above steps S011 to S013, which will not be detailed here.

[0076] After each federal participant processes the user feature set using the credit prediction model of any of the above embodiments to obtain a credit prediction score, in order to ensure the confidentiality and security of the data during transmission, the credit prediction score is processed for privacy, and the corresponding encrypted result is obtained before being returned to the federal server.

[0077] After receiving the encrypted results returned by each federation participant, the federation server executes step S500 to merge all the encrypted results, obtain the credit score of the target user, and send the credit score to the target federation participant.

[0078] In the above, all encrypted results can be decrypted before fusion processing. However, to better ensure the confidentiality and security of data during fusion processing, in some embodiments, all encrypted results are first fused, such as by weighted summation, to obtain an encrypted fusion result. Then, the encrypted fusion result is decrypted to obtain the credit score. That is, the step S500 above, which involves fusing the encrypted results returned by each federated participant to obtain the target user's credit score, may include: In step S510, the encryption results returned by each federation participant are fused to obtain an encryption fusion result; In step S520, the encrypted fusion result is decrypted to obtain the credit score.

[0079] As can be seen from the above, the data received by the federated server is always processed and not in its original plaintext form, thus effectively protecting the confidentiality and security of the data and preventing the leakage of private data.

[0080] Furthermore, in order for the federated server to be able to fuse all encrypted results, the encrypted results can be obtained by the corresponding federated participants through homomorphic encryption technology. Thus, by utilizing the special mathematical properties of homomorphic encryption technology, the federated server can directly perform weighted fusion processing on all encrypted results in the encrypted state, so that the resulting encrypted fusion result, after decryption, is completely consistent with the result of performing the same calculation on the credit prediction score in the unencrypted state.

[0081] The weights used in the weighted fusion processing can be pre-configured based on experience or actual needs, and the sum of all weights is 1. However, this weight configuration scheme may result in encrypted results with higher accuracy being assigned smaller weights, while encrypted results with lower accuracy are assigned larger weights, thus affecting the accuracy of the credit score calculated based on this. Therefore, to solve this technical problem, in some embodiments, the credit assessment method provided by one or more embodiments of this specification also provides another fusion processing scheme, in which the weight coefficients (i.e., weight values) are dynamically adjusted during the fusion processing, and these weight systems are calculated based on the performance of the credit prediction models of each federal participant. That is, the step S510 above, which involves fusing the encrypted results returned by each federal participant to obtain the encrypted fusion result, may include the following steps: In step S511, the performance scores of the credit prediction models of each federal participant are obtained, and the weight coefficients of the corresponding encryption results are determined based on the performance scores of each credit prediction model. In step S512, the encryption results are weighted and calculated according to the weight coefficients of each encryption result to obtain the encryption fusion result.

[0082] Understandably, the fusion processing of the encryption results can be achieved by executing the above steps S511 to S512: first execute step S511 to obtain the performance scores of the credit prediction models of each federal participant, and then determine the weight system of the corresponding encryption results based on the performance scores of each credit prediction model.

[0083] In the above, model performance evaluation techniques from related technologies can be used to calculate one or more of the accuracy, recall, and sensitivity of each credit prediction model. Then, the weighted sum or weighted product of one or more of them can be used as the performance score of the credit prediction model. Then, according to the condition that the sum of all weights is 1, the weight coefficient of the corresponding encryption result can be calculated based on the performance score of each credit prediction model.

[0084] While the above-mentioned weighting coefficient calculation scheme can configure appropriate weighting coefficients for the encrypted results corresponding to each credit prediction model, in order to better evaluate the performance of each credit prediction model in the credit assessment scenario and obtain more accurate performance scores, thereby improving the accuracy of subsequent credit scoring based on encrypted fusion results, in some embodiments, the credit assessment method provided by one or more embodiments of this specification also provides another scheme for obtaining model performance scores. That is, the step of obtaining the performance scores of the credit prediction models of each federal participant in step S511 above may include the following steps: In steps S511-11, the historical performance score of each credit prediction model is calculated based on at least one of the long-term accuracy, stability coefficient, and timeliness score of each credit prediction model; the long-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a first set time period, the stability coefficient is used to characterize the fluctuation of the prediction results output by the credit prediction model, and the timeliness score is used to characterize the update time of the credit prediction model. In steps S511-12, the input data quality score of each credit prediction model is calculated based on at least one of the completeness score, freshness score, consistency score, and relevance score of the input data provided by the federal participants in this credit prediction. The completeness score is used to characterize the completeness of the input data, the freshness score is used to characterize the update time of the input data, the consistency score is used to characterize the degree of matching between the input data and the set verification rules, and the relevance score is used to characterize the degree of correlation between the input data and the credit score. In steps S511-13, the computational performance score of each credit prediction model is calculated based on at least one of the following: short-term accuracy, scenario adaptability, and response performance score. The short-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a second set time period, where the second set time period is shorter than the first set time period. The scenario adaptability is used to characterize the degree of adaptability of the credit prediction model to the business scenario in which the credit assessment request is located. The response performance score is used to characterize the response time of the credit prediction model in this credit prediction. In steps S511-14, the performance score of each credit prediction model is calculated based on at least one of the historical performance score, input data quality score, and computational performance score of each credit prediction model.

[0085] The following explains the technical principle behind obtaining the performance scores of each credit prediction model through the technical solutions described in steps S511-11 to S511-14: The technical solution described in step S511-11 is used to evaluate the model's historical performance score, considering one or more of three aspects: long-term accuracy, performance stability, and timeliness, to accurately evaluate the model's historical performance score. The calculation methods for the indicators involved in these three aspects are as follows: Long-term accuracy is characterized by long-term accuracy rate, which is calculated as: Long-term accuracy rate = Number of correct predictions made by the model within a first set time period / Total number of predictions. For example, if a credit prediction model makes 81 correct predictions within 90 days and makes 90 total predictions, then the long-term accuracy rate is 81 / 90 = 0.9.

[0086] The stability of the model is characterized by the stability coefficient, which is calculated as 1 - the standard deviation of the model's predicted scores over a set number of times / the average score of the model's predicted scores over a set number of times. For example, if the standard deviation of the model's predicted scores over 30 times is s and the average score is a, then the stability coefficient is 1 - s / a.

[0087] Timeliness is represented by a timeliness score, calculated as: Timeliness Score = max(0, 1 - (Current Time - Last Model Update Time) / Maximum Model Validity Period). The maximum model validity period represents the longest period during which the model can remain unupdated, and can be configured based on experience or experimentation, for example, 90 days. As an example of calculating the timeliness score, assuming the maximum model validity period is 90 days, the current time is October 10th, and the last model update was on October 7th, the last update was 3 days ago. Therefore, current time - last model update time = 3, and the timeliness score = 1 - 3 / 90. As another example, assuming the maximum model validity period is 90 days, and the model has not been updated for more than 90 days (current time - last model update time > 90), the timeliness score is 0. It is evident that by employing a linear decay function, we can ensure that the newer the model, the higher the score, which encourages each federal party to update the model regularly to guarantee the accuracy and robustness of the model's output.

[0088] During step S511-11, if only one of the target items—long-term accuracy, stability coefficient, and timeliness score—is considered when calculating the historical performance score, then the value of that target item can be calculated using only its formula, and this calculated value can be used as the model's historical performance score. Understandably, the value of each considered item is calculated to achieve the historical performance score.

[0089] However, to more comprehensively evaluate the model's historical performance, some embodiments can combine long-term accuracy, stability coefficient, and timeliness score to calculate a more accurate historical performance score. The combination method can be a weighted summation. The weights of long-term accuracy, stability coefficient, and timeliness score can be configured according to preset importance or based on actual needs or experience, but the sum of the weights of the three must be 1. For example, the weight of long-term accuracy can be set to 0.4, the weight of stability coefficient can be set to 0.3, and the weight of timeliness score can be set to 0.3. Based on this, the historical performance score = 0.4 × long-term accuracy + 0.3 × stability coefficient + 0.3 × timeliness score.

[0090] The technical solution described in steps S511-12 is used to evaluate the quality score of the model's input data, considering one or more of four aspects: data integrity, data freshness, data consistency, and data relevance, to accurately assess the quality of the model's input data. The calculation methods for the indicators involved in these four aspects are as follows: Data integrity is represented by an integrity score, calculated as: Integrity Score = Number of non-empty fields in the current input data / Total number of fields. The total number of fields is a pre-agreed fixed value. In the early stages of federated learning collaboration, each participating party pre-agrees on a set of standard feature fields. For example, the agreement might be that the bank provides 15 standard fields and the e-commerce platform provides 12. Based on this, the total number of fields for the bank's credit prediction model is 15, and for the e-commerce platform's credit prediction model, it's 12. Taking the bank's credit prediction model as an example, assuming the total number of fields is 15, but the bank only provided 12 fields of user feature data in this prediction, then the integrity score would be 12 / 15 = 0.8.

[0091] Data freshness is represented by a freshness score, which is the reciprocal of the time difference between the data update time and the current time. The reciprocal is used to standardize the freshness score, ensuring it's calculated on the same standard as other scores. Therefore, assuming an item in the input data was updated on October 1st and the current time is October 11th, the freshness score would be 1 / (11-1) = 0.1. Data freshness is considered when evaluating model performance because user data can change. For example, a user's monthly income may change over time. Continuing to use outdated monthly income data could affect the accuracy of the final credit score or loan amount assessment. Therefore, data freshness can be considered when evaluating model performance.

[0092] Data consistency is represented by a consistency score, which is calculated as: (Number of user data items that pass validation based on the defined business rules) / (Total number of user data items). For example, suppose the business rules include the following validation items: 1. Age ≥ 18 years old (meaning only adults can apply for loans or rent without a deposit); 2. Monthly income ≥ twice the monthly repayment amount (indicating that the user has a good debt repayment ability); 3. Years of work experience ≤ age - 18 (indicating the user has a reasonable career path); 4. Credit card limit ≥ used limit (to ensure data consistency).

[0093] Based on the above business rules, assuming the bank provides 1000 user data records, and 950 of these records pass all the validation checks according to the business rules, then the consistency score = 950 / 1000 = 0.95.

[0094] Data relevance is represented by a relevance score, which is the statistical relevance between the input data and the credit score. This statistical relevance can be expressed using the Pearson coefficient. Based on this, the statistical relevance between the input data and the credit score can be calculated using the relevance principle of the Pearson coefficient, which will not be detailed here.

[0095] During steps S511-12, if only one of the target items (completeness score, freshness score, consistency score, and relevance score) is considered when calculating the historical performance score, the value of the target item can be calculated using only its formula, and this calculated value can be used as the input data quality score for the model. Understandably, the value of each considered item is calculated to achieve the input data quality score.

[0096] However, to more comprehensively evaluate the quality of the model's input data, some embodiments can fuse four factors—completeness score, freshness score, consistency score, and relevance score—to calculate a more accurate input data quality score. The fusion method can be a weighted summation. The weights of the completeness score, freshness score, consistency score, and relevance score can be configured according to preset importance levels, or based on actual needs or experience, but the sum of the weights of the four factors must be 1. For example, the weight of the completeness score can be set to 0.25, the weight of the freshness score can be set to 0.2, the weight of the consistency score can be set to 0.2, and the weight of the relevance score can be set to 0.3. Based on this, the input data quality score = 0.25 × completeness score + 0.25 × freshness score + 0.2 × consistency score + 0.3 × relevance score.

[0097] The technical solution described in steps S511-13 is used to evaluate the computational performance score of the model. It can check the current state of the model and considers one or more of three aspects—short-term accuracy, scenario adaptability, and system performance—to accurately evaluate the model's computational performance. The calculation methods for the indicators involved in these three aspects are as follows: Short-term accuracy is characterized by short-term accuracy rate, which is calculated as: short-term accuracy rate = number of correct predictions made by the model within a second set time period / total number of predictions. For example, if the credit prediction model makes 90 correct predictions within 24 hours and the total number of predictions is 100, then the short-term accuracy rate is 90 / 100 = 0.9.

[0098] Scenario adaptability is characterized by scenario fit, which is the average historical accuracy of the model in similar business scenarios. This involves pre-summarizing and categorizing potential business scenarios that may generate credit assessment requests. These categorized scenarios can include, but are not limited to, credit card approval, personal loans, and SME loans. Based on this, assuming the current prediction by the bank's credit prediction model corresponds to the credit card approval scenario, the historical accuracy of all predictions in this scenario can be extracted. The average of these historical accuracy rates is then used as the scenario fit. For example, if the bank's average accuracy in historical credit card approval scenarios is 88%, the scenario fit would be 0.88.

[0099] System performance is characterized by a response performance score, which can be calculated based on at least one of response time and response success rate. In some embodiments, to more accurately and comprehensively evaluate system performance, the response performance score can be calculated using both response time and response success rate. That is, Response Performance Score = a × Response Time Score + b × Response Success Rate Score, where a and b are weighting coefficients, and the sum of a and b is 1. a and b can be set based on experience or actual needs; they can be different or the same. For example, a and b can both be set to 0.5. Response Time Score = max(0, 1 - (Actual Response Time / Timeout Threshold)), where the timeout threshold is a fixed value set based on experience or actual needs, for example, 2 seconds. Based on this, if the actual response time is 0.5 seconds, then the response time score = 1 - 0.5 / 2 = 0.75; if the actual response time exceeds 2 seconds, then the response time score is 0. The response success rate score is calculated as follows: Response Success Rate Score = Number of Successful Responses / Total Number of Requests. The total number of requests can be a fixed value preset based on experience or actual needs, for example, 100. Therefore, the number of successful responses refers to the number of times the model successfully returned an encrypted result in the last 100 requests for credit prediction. Assuming 95 successful responses, the response success rate score = 95 / 100 = 0.95. After obtaining the response time score and response success rate score, these scores can be substituted into the above formula for calculating the response performance score to obtain the final response performance score.

[0100] During the execution of steps S511-13, if only one of the target items—short-term accuracy, scene fit, and response performance score—is considered when calculating the historical performance score, then the value of the target item can be calculated using only its formula, and this calculated value can be used as the model's computational performance score. Understandably, the value of each considered item is calculated to achieve the computational performance score.

[0101] However, to more comprehensively evaluate the model's computational performance score, some embodiments can calculate a more accurate score using three metrics: short-term accuracy, scene fit, and response performance score. The fusion method can be a weighted summation. The weights of short-term accuracy, scene fit, and response performance score can be configured according to preset importance or based on actual needs or experience, but the sum of their weights must be 1. For example, the weight of short-term accuracy can be set to 0.5, the weight of scene fit to 0.3, and the weight of response performance score to 0.2. Based on this, the input data quality score = 0.5 × short-term accuracy + 0.3 × scene fit + 0.2 × response performance score.

[0102] It should be noted that the steps in S511-11 to S511-13 can be executed in parallel or in sequence. In the case of serial execution, the order of the three steps is not limited.

[0103] After obtaining the historical performance score, input data quality score, and computational performance score of each credit prediction model through steps S511-11 to S511-13, step S511-14 is executed to calculate the performance score of each credit prediction model based on at least one of the historical performance score, input data quality score, and computational performance score. Understandably, when calculating the performance score of a credit prediction model, one or more of the following can be used: historical performance score, input data quality score, and computational performance score. In some implementations, to comprehensively consider the model's performance from more aspects and obtain a more accurate performance score, the historical performance score, input data quality score, and computational performance score can be combined to calculate the performance score. The combination method can be a weighted summation, and the weights of each item can be pre-configured according to their importance, experience, or actual needs. For example, the weight of the historical performance score can be configured as 0.4, the weight of the input data quality score as 0.35, and the weight of the computational performance score as 0.25. Based on this, the performance score = 0.4 × historical performance score + 0.35 × input data quality score + 0.25 × computational performance score.

[0104] After obtaining the performance scores of all credit prediction models, proceed to step S511 to determine the weight coefficients of the corresponding encryption results based on the performance scores of all credit prediction models. The determination process may include: In steps S511-21, the performance scores of each credit prediction model are normalized to obtain the weight coefficients of the corresponding encryption results; wherein the sum of all weight coefficients is 1.

[0105] Understandably, steps S511-21 can be used to calculate the weight coefficients of each encryption result based on the performance scores of each credit prediction model. The normalization process is to ensure that the weight coefficients of each encryption result are relative and comparable values, and that the sum of the weight coefficients of all encryption results is 1.

[0106] After obtaining the weight coefficients of all encryption results, step S512 is executed. Based on the weight coefficients of each encryption result, a weighted sum is calculated for all encryption results to obtain the encryption fusion result. For example, assuming there are two participants in the federation, a bank and an e-commerce platform, there are also two encryption results: one provided by the bank and the other by the e-commerce platform. The encryption fusion result = s1 × w1 + s2 × w2; where s1 is the encryption result provided by the bank, w1 is the weight coefficient of encryption result s1, s2 is the encryption result provided by the e-commerce platform, and w2 is the weight coefficient of encryption result s2.

[0107] After obtaining the encrypted fusion result, in order to obtain the credit score, step S520 is executed. First, the encrypted fusion result is decrypted to obtain the decrypted fusion result. Then, the decrypted fusion result is mapped to the interval [0, 1] to obtain a normalized mapping value. Finally, the normalized mapping value is converted into a credit score (e.g., 0~1000). An activation function, such as the sigmoid function, can be used to map the decrypted fusion result to a value in the [0, 1] range. Alternatively, a lookup table or other methods from related technologies can be used to convert the normalized mapping value into a credit score, which will not be detailed here.

[0108] After obtaining the target user's credit score, the federal server can send the score to the target federal participants. It can also further determine the approval outcome based on the credit score. For example, assuming the target user's credit assessment request is triggered by a loan application, after obtaining the target user's credit score, the federal server can determine whether to allow the loan application based on the credit score and a set credit score threshold or range. For instance, if the target user's credit score is higher than the set threshold, the user is considered to have the ability to repay the loan, and the approval result indicates approval. Conversely, if the target user's credit score is lower than the set threshold, the user is considered to lack the ability to repay the loan, and the approval result indicates disapproval, with the reason for disapproval also recorded. Based on this, the federal server can also send the credit score and approval result together to the target federal participants.

[0109] After receiving the credit score and approval result, the target federal participants can send the credit score and approval result to the user terminal linked to the target user, so that the target user can understand their own credit score and whether the loan has been approved.

[0110] In some embodiments, to further improve the data security of the federated server in integrating and processing user encrypted data and in merging the encrypted results, a TEE can be configured in the federated server, and the integration and merging processes can be performed in the TEE.

[0111] In some cases, a target user's credit assessment request may be triggered by fraudulent activity. For example, the target user's device or account information may be stolen and used to apply for loans. Continuing to respond to the credit assessment request could result in financial loss and privacy breaches for the target user. To address this technical problem, in some embodiments, the credit assessment method provided in one or more embodiments of this specification also provides an anti-fraud solution. That is, after step S200, the credit assessment method may further include: In step S310, the user feature set is processed by a pre-trained fraud detection model to obtain a fraud probability; the fraud probability is used to assess whether the credit assessment request is triggered by fraudulent behavior. In step S320, if the fraud probability exceeds a set probability threshold, a credit assessment rejection message and a warning message are sent to the target federal participant.

[0112] Accordingly, in step S400 above, the step of sending the credit prediction instruction and the user feature set to each of the federal participants includes: sending the credit prediction instruction and the user feature set to each of the federal participants when the fraud probability does not exceed the probability threshold.

[0113] Understandably, before triggering the credit prediction operation of each federal participant, step S310 is executed first, which processes the user feature set through a pre-trained fraud detection model to obtain the fraud probability; the fraud detection model can be constructed using the construction principle of fraud detection models of related technologies, which will not be detailed here.

[0114] After obtaining the fraud probability, it is compared with a pre-set fraud probability threshold. If the fraud probability is greater than the threshold, it indicates that the current credit assessment request was triggered by fraudulent behavior. In this case, step S320 is executed to return a credit assessment rejection message and a warning message to the target federal participant. The fraud probability threshold can be pre-configured based on experience or actual needs.

[0115] After receiving the credit assessment rejection message and warning message, the target federal participants will directly reject the relevant application of the target user and perform warning actions, such as sending fraud alert messages to the target user to alert them to financial risks.

[0116] Conversely, if the fraud probability is not greater than the fraud probability threshold, it means that the current credit assessment request is triggered by a security action. In this case, step S400 is executed to complete the credit assessment of the target user.

[0117] As can be seen from the above description, during the execution of the credit assessment method provided in one or more embodiments of this specification by the federal server, the target federal participant will cooperate with the work of the federal server. Therefore, this specification also provides a credit assessment method applied to the target federal participant, such as... Figure 3 As shown, Figure 3 This is a flowchart illustrating an exemplary embodiment of a credit assessment method applied to a target federal participant, the method comprising: In step S010, a credit assessment request for the target user is sent to the federal server; the credit assessment request is used to trigger the federal server to send user data acquisition requests for the target user to each federal participant; the federal participants include the target federal participant and other federal participants. In step S020, in response to the user data acquisition request issued by the federated server, the user data of the target user is acquired, the user data is encrypted, and the encrypted user data is returned to the federated server. In step S030, in response to the credit prediction instruction from the federal server for the encrypted user data, the locally deployed credit prediction model is invoked to process the user feature set sent by the federal server to obtain a credit prediction score, and the encrypted result of the credit prediction score is returned to the federal server; the user feature set is obtained by the federal server integrating and processing the encrypted user data returned by each federal participant. In step S040, upon receiving the credit score returned by the federal server, the credit score is sent to the user terminal associated with the target user; the credit score is obtained by the federal server through fusion processing based on the encrypted results returned by each federal participant.

[0118] For the technical principles and explanations of steps S010 to S040 above, please refer to the relevant records of the credit assessment method applied to the federated server, which will not be repeated here.

[0119] In some embodiments, to update model parameters and improve the robustness and generalization ability of the model, the credit assessment method for target federal participants provided in one or more embodiments of this specification may further include: In step S050, in response to the first model update instruction sent by the federated server, the local prediction model is invoked to process the user feature sample set obtained locally, obtaining an intermediate result set and a credit prediction result set, which are then sent to the federated server; wherein, the intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set; the number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the user feature sample set; In step S060, in response to the second model update instruction sent by the federated server, the model parameters of the local meta-model are updated to the target model parameters sent by the federated server, and the model parameters of the local prediction model are adjusted through the updated local meta-model, or the first layer of the updated local meta-model is connected to the last layer of the local prediction model; wherein, the target model parameters are obtained by the federated server training the global meta-model based on the intermediate result set and credit prediction result set returned by each federated participant.

[0120] Similarly, for the technical principles and explanations of steps S050 to S060 above, please refer to the relevant records of the credit assessment method applied to the federal server, which will not be repeated here.

[0121] Figure 4 This is a schematic structural diagram of a device provided in an exemplary embodiment. For example... Figure 4 As shown, device 400 mainly consists of a communication interface 402, a user interface 404, a processor 406, and a data storage 408. These components are interconnected and communicate with each other via a system bus, network, or other connection mechanism 410. The communication interface 402 enables device 400 to communicate with other devices, access networks, and transmission networks via analog or digital modulation. For example, the communication interface 402 may include a chipset and antenna for wireless communication with a radio access network or access point. Furthermore, the communication interface 402 can be a wired interface such as Ethernet, Token Ring, or a USB port, or a wireless interface such as Wi-Fi, Bluetooth, Global Positioning System (GPS), or a wide-area wireless interface (e.g., WiMAX or LTE). Of course, the communication interface 402 can also support other forms of physical layer interfaces and standard or proprietary communication protocols. The communication interface 402 may also include multiple physical communication interfaces, such as Wi-Fi, Bluetooth, and wide-area wireless interfaces.

[0122] User interface 404 includes receiving user input and providing output to the user. Therefore, user interface 404 may include input components such as a keypad, keyboard, touch-sensitive or presence-sensitive panel, computer mouse, trackball, joystick, microphone, still camera, and video camera, and output components such as a display screen (which may be combined with a touch-sensitive panel), CRT, LCD, LED, display using DLP technology, printer, and other similar devices known or developed in the future. User interface 404 may also generate auditory output via speakers, speaker jacks, audio output ports, audio output devices, headphones, and other similar devices known or developed in the future. In some embodiments, user interface 404 may include software, circuitry, or other forms of logic capable of transmitting and receiving data from external user input / output devices. Additionally or alternatively, device 400 may support remote access from other devices via communication interface 402 or another physical interface (not shown). User interface 404 may be configured to receive user input, the position and movement of which may be indicated by an indicator or cursor described herein. User interface 404 may also be configured as a display device for rendering or displaying text fragments.

[0123] Processor 406 may contain one or more general-purpose processors and / or special-purpose processors.

[0124] Data storage 408 may include one or more volatile and / or non-volatile storage components and may be integrated wholly or partially with processor 406. Data storage 408 may include removable and non-removable components.

[0125] Processor 406 is capable of executing program instructions 418 (e.g., compiled or uncompiled program logic and / or machine code) stored in data storage 408 to perform the various functions described herein. Data storage 408 may comprise a non-transitory computer-readable medium on which program instructions are stored, which, when executed by device 400, enable device 400 to perform any methods, processes, or functions disclosed in this specification and / or the accompanying drawings. Processor 406 executing program instructions 418 may result in processor 406 using data 412.

[0126] For example, program instructions 418 may include an operating system 422 (e.g., an operating system kernel, device drivers, and / or other modules) installed on device 400 and one or more applications 420 (e.g., a browser, social application, or game application). Similarly, data 412 may include operating system data 416 and application data 414. Operating system data 416 is primarily accessible to the operating system 422, while application data 414 is primarily accessible to one or more applications 420. Application data 414 may reside in a file system visible or hidden from the user of device 400.

[0127] Application 420 can communicate with operating system 422 through one or more application programming interfaces (APIs). These APIs help application 420 read and / or write application data 414, transmit or receive information via communication interface 402, receive or display information on user interface 404, etc.

[0128] In some terminology, application 420 may be simply referred to as "app". Furthermore, application 420 can be downloaded to device 400 through one or more online app stores or app markets. However, applications can also be installed on device 400 in other ways, such as through a web browser or a physical interface on device 400 (e.g., a USB port).

[0129] In some embodiments, device 400 may function as a federation server or as a device among federation participants.

[0130] Please refer to Figure 5 , Figure 5 This is a structural block diagram of a credit assessment device applied to a federated server, provided as an exemplary embodiment. The credit assessment device can be applied to, for example... Figure 4 The device shown implements the technical solution described in this specification. The credit assessment device 500 applied to the federal server may include: The federation request module 510 is configured to: in response to a credit assessment request issued by a target federation participant for a target user, send a user data acquisition request for the target user to each federation participant; the each federation participant includes the target federation participant and other federation participants; The feature processing module 520 is configured to: upon receiving encrypted user data returned by each of the federation participants in response to the user data acquisition request, integrate and process the encrypted user data returned by each of the federation participants to obtain a set of user features. The federated invocation module 530 is configured to send a credit prediction instruction and the user feature set to each federated participant; the credit prediction instruction is used to trigger each federated participant to invoke a locally deployed credit prediction model to process the user feature set and obtain a credit prediction score. The credit scoring module 540 is configured to: upon receiving the encrypted results of the credit prediction scores returned by each of the federal participants in response to the credit prediction instruction, perform fusion processing on the encrypted results returned by each of the federal participants to obtain the credit score of the target user, and send the credit score to the target federal participant.

[0131] In some embodiments, the process by which the feature processing module 520 integrates and processes the encrypted user data returned by each federation participant to obtain a user feature set is configured to include: From the encrypted user data returned by each federation participant, extract the remaining encrypted data other than the target user's identity information, and perform deduplication on the remaining encrypted data to obtain the target encrypted data of the target user; The target encrypted data is subjected to privacy processing to obtain re-encrypted data; The re-encrypted data is subjected to feature alignment processing to obtain the user feature set.

[0132] In some embodiments, the process by which the feature processing module 520 performs feature alignment processing on the re-encrypted data to obtain the user feature set is configured to include: The re-encrypted data is standardized to obtain standardized feature data after eliminating the differences in feature dimensions among different federal participants. The standardized feature data is subjected to feature normalization and missing value processing to obtain the user feature set.

[0133] In some embodiments, the process by which the credit scoring module 540 fuses the encrypted results returned by each federal participant to obtain the credit score of the target user is configured to include: The encryption results returned by each of the federation participants are merged to obtain the encryption fusion result; The encrypted fusion result is decrypted to obtain the credit score.

[0134] In some embodiments, the process by which the credit scoring module 540 fuses the encrypted results returned by each federal participant to obtain an encrypted fusion result is configured to include: Obtain the performance scores of the credit prediction models of each federal participant, and determine the weight coefficients of the corresponding encryption results based on the performance scores of each credit prediction model. Based on the weight coefficients of each encryption result, a weighted sum is calculated for all encryption results to obtain the encryption fusion result.

[0135] In some embodiments, the process by which the credit scoring module 540 obtains the performance scores of the credit prediction models of each federal participant is configured to include: The historical performance score of each credit prediction model is calculated based on at least one of the following: long-term accuracy, stability coefficient, and timeliness score. The long-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a first set time period, the stability coefficient is used to characterize the volatility of the prediction results output by the credit prediction model, and the timeliness score is used to characterize the update time of the credit prediction model. The input data quality score for each credit prediction model is calculated based on at least one of the following: completeness score, freshness score, consistency score, and relevance score of the input data provided by the federal participants in this credit prediction. The completeness score is used to characterize the completeness of the input data, the freshness score is used to characterize the update time of the input data, the consistency score is used to characterize the degree of matching between the input data and the set verification rules, and the relevance score is used to characterize the degree of correlation between the input data and the credit score. The computational performance score of each credit prediction model is calculated based on at least one of the following: short-term accuracy, scenario adaptability, and response performance score. The short-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a second set time period, where the second set time period is shorter than the first set time period. The scenario adaptability is used to characterize the degree of adaptability of the credit prediction model to the business scenario in which the credit assessment request is located. The response performance score is used to characterize the response time of the credit prediction model in this credit prediction. The performance score of each credit prediction model is calculated based on at least one of the historical performance score, input data quality score, and computational performance score.

[0136] In some embodiments, the process by which the credit scoring module 540 determines the weighting coefficients of the corresponding encryption results based on the performance scores of each credit prediction model is configured to include: The performance scores of each credit prediction model are normalized to obtain the corresponding weight coefficients of the encryption results; where the sum of all weight coefficients is 1.

[0137] In some embodiments, the credit assessment device 500 applied to the federated server may further include: The anti-fraud module is configured to: process the user feature set using a pre-trained fraud detection model to obtain a fraud probability; the fraud probability is used to assess whether the credit assessment request is triggered by fraudulent behavior; and if the fraud probability exceeds a set probability threshold, send a credit assessment rejection message and a warning message to the target federated party.

[0138] Accordingly, the process by which the federated invocation module 530 sends the credit prediction instruction and the user feature set to each federated participant is configured to: send the credit prediction instruction and the user feature set to each federated participant when the fraud probability does not exceed the probability threshold.

[0139] In some embodiments, the credit assessment device 500 applied to the federated server may further include: The model update control module is configured as follows: A first model update instruction is sent to each of the federation participants; the first model update instruction is used to trigger the federation participants to call the local prediction model to process the user feature sample set obtained from the local model, and obtain an intermediate result set and a credit prediction result set; wherein, the intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set; the number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the corresponding user feature sample set; Upon receiving the intermediate result set and credit prediction result set returned by each of the federation participants in response to the model update instruction, the global meta-model is trained using the intermediate result set and credit prediction result set returned by each of the federation participants until the global meta-model converges. When the global meta-model converges, the target model parameters and the second model update instruction of the converged global meta-model are sent to each federation participant. The second model update instruction is used to trigger each federation participant to update the model parameters of the local meta-model to the target model parameters, and adjust the model parameters of the local prediction model through the updated local meta-model, or connect the first layer of the updated local meta-model to the last layer of the local prediction model.

[0140] Please refer to Figure 6 , Figure 6 This is a structural block diagram of an exemplary embodiment of a credit assessment device applied to a target federal participant. The credit assessment device can be applied to, for example... Figure 4 The device shown implements the technical solution described in this specification. The credit assessment device 600 applied to a target federal participant includes: The credit request module 610 is configured to: send a credit assessment request for a target user to the federation server; the credit assessment request is used to trigger the federation server to send user data acquisition requests for the target user to each federation participant; the federation participants include the target federation participant and other federation participants; The data processing module 620 is configured to: in response to the user data acquisition request issued by the federated server, acquire the user data of the target user, encrypt the user data, and return the encrypted user data to the federated server; The credit prediction module 630 is configured to: respond to the credit prediction instruction from the federal server for the encrypted user data, call a locally deployed credit prediction model to process the user feature set sent by the federal server, obtain a credit prediction score, and return the encrypted result of the credit prediction score to the federal server; the user feature set is obtained by the federal server integrating and processing the encrypted user data returned by each federal participant. The credit score sending module 640 is configured to: upon receiving a credit score returned by the federal server, send the credit score to the user terminal associated with the target user; the credit score is obtained by the federal server through fusion processing based on the encrypted results returned by each federal participant.

[0141] In some embodiments, the credit assessment device 600 applied to the target federal participant may further include: The model update response module is configured as follows: In response to the first model update instruction sent by the federated server, the local prediction model is invoked to process the user feature sample set obtained locally, obtaining an intermediate result set and a credit prediction result set, which are then sent to the federated server; wherein, the intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set; the number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the user feature sample set; In response to the second model update instruction sent by the federated server, the model parameters of the local meta-model are updated to the target model parameters sent by the federated server, and the model parameters of the local prediction model are adjusted through the updated local meta-model, or the first layer of the updated local meta-model is connected to the last layer of the local prediction model; wherein, the target model parameters are obtained by the federated server training the global meta-model based on the intermediate result set and credit prediction result set returned by each federated participant.

[0142] For ease of description, the above devices are described by dividing them into various modules or units based on their functions. Of course, when implementing one or more of these specifications, the functions of each module or unit can be implemented in the same or different software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0143] Based on the same concept as the methods described above, this specification also provides a credit assessment system, including: A federated server, upon receiving a credit assessment request for a target user from a target federated party, executes the steps of the credit assessment method applied to the federated server provided in one or more of the above embodiments to obtain the target user's credit score and sends the credit score to the target federated party; and Multiple federal participants, including the target federal participant and other federal participants; The target federal participant is used to obtain the credit score of the target user by performing the steps of the credit assessment method applied to the target federal participant provided in one or more of the above embodiments, in cooperation with the federal server, and to send the credit score to the user terminal associated with the target user; The other federal participants are used for: In response to a user data acquisition request sent by the federation server, the system acquires and encrypts the user data of the target user, and returns the encrypted user data to the federation server. In response to the credit prediction instruction sent by the federal server, the locally deployed credit prediction model is invoked to process the user feature set sent by the federal server, obtain a credit prediction score, and send it to the federal server.

[0144] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor executes the executable instructions to implement the steps of the credit assessment method applied to a federal server and / or the credit assessment method applied to a target federal participant as described in any of the above embodiments.

[0145] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the credit assessment method applied to a federal server and / or the credit assessment method applied to a target federal participant as described in any of the above embodiments.

[0146] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the credit assessment method applied to a federal server and / or the credit assessment method applied to a target federal participant as described in any of the above embodiments.

[0147] What those skilled in the art will understand is: In this specification, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of additional identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded.

[0148] In this specification, “a,” “an,” and “the” do not specifically refer to the singular, but may also include the plural.

[0149] In this specification, ordinal numbers such as "first," "second," etc., do not necessarily indicate order; they are often used to distinguish between objects. For example, "first server" and "second server" usually refer to two servers. To differentiate between these two servers, they are described as "first server" and "second server." Of course, sometimes these two servers may be the same server.

[0150] In this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can also mean indirect receiving and sending. For example, A receiving data sent by B can be understood as A directly receiving the data sent by B, or it can be understood as A indirectly receiving the data sent by B through other entities such as C. Similarly, B sending data to A can be understood as B sending the data directly to A, or it can be understood as B indirectly sending the data to A through other entities such as C. Here, C can be one entity, or it can be two or more entities.

[0151] In this specification, unless explicitly stated otherwise, the relationships between structures can be direct or indirect. For example, when describing "A is connected to B," unless it is explicitly stated that A and B are directly connected, it should be understood that A can be directly connected to B or indirectly connected to B. Similarly, when describing "A is on top of B," unless it is explicitly stated that A is directly above B (AB is adjacent and A is above B), it should be understood that A can be directly above B or indirectly above B (AB is separated by other elements, and A is above B). And so on.

[0152] This specification uses specific terms to describe embodiments thereof. Terms such as "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of those different embodiments or examples, without contradiction.

[0153] Although one or more embodiments of this specification provide method steps as described in the embodiments or flowcharts, it is understood that the order of steps listed in the embodiments or flowcharts is only one of many possible execution orders and does not represent the only execution order. Therefore, when the claims involve method steps, any changes or adjustments to the order of such steps, or the parallelism between steps, are also within the scope of protection of the claims.

Claims

1. A credit assessment method applied to a federated server, the method comprising: In response to a credit assessment request issued by a target federal participant for a target user, a user data acquisition request for the target user is sent to each federal participant. The federal participants include the target federal participant and other federal participants; Upon receiving encrypted user data returned by each of the federation participants in response to the user data acquisition request, the encrypted user data returned by each of the federation participants is integrated and processed to obtain a user feature set. Send a credit prediction instruction and the user feature set to each of the federal participants; the credit prediction instruction is used to trigger each of the federal participants to call the locally deployed credit prediction model to process the user feature set and obtain a credit prediction score; Upon receiving the encrypted results of the credit prediction scores returned by each of the federal participants in response to the credit prediction instruction, the encrypted results returned by each of the federal participants are merged to obtain the credit score of the target user, and the credit score is sent to the target federal participant.

2. The method according to claim 1, wherein the step of integrating and processing the encrypted user data returned by each federated participant to obtain a user feature set includes: From the encrypted user data returned by each federation participant, extract the remaining encrypted data other than the target user's identity information, and perform deduplication on the remaining encrypted data to obtain the target encrypted data of the target user; The target encrypted data is subjected to privacy processing to obtain re-encrypted data; The re-encrypted data is subjected to feature alignment processing to obtain the user feature set.

3. The method according to claim 2, wherein the step of performing feature alignment processing on the re-encrypted data to obtain the user feature set includes: The re-encrypted data is standardized to obtain standardized feature data after eliminating the differences in feature dimensions among different federal participants. The standardized feature data is subjected to feature normalization and missing value processing to obtain the user feature set.

4. The method according to claim 1, wherein the step of fusing the encrypted results returned by each federal participant to obtain the target user's credit score includes: The encryption results returned by each of the federation participants are merged to obtain the encryption fusion result; The encrypted fusion result is decrypted to obtain the credit score.

5. The method according to claim 4, wherein the step of processing the encryption results returned by each federated participant to obtain the encryption fusion result includes: Obtain the performance scores of the credit prediction models of each federal participant, and determine the weight coefficients of the corresponding encryption results based on the performance scores of each credit prediction model. Based on the weight coefficients of each encryption result, a weighted sum is calculated for all encryption results to obtain the encryption fusion result.

6. The method of claim 5, wherein the step of obtaining the performance score of the credit prediction model for each federal participant includes: The historical performance score of each credit prediction model is calculated based on at least one of the following: long-term accuracy, stability coefficient, and timeliness score. The long-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a first set time period, the stability coefficient is used to characterize the volatility of the prediction results output by the credit prediction model, and the timeliness score is used to characterize the update time of the credit prediction model. Calculate the input data quality score for each credit forecasting model based on at least one of the following: completeness score, freshness score, consistency score, and relevance score of the input data provided by the federal participants to which each credit forecasting model belongs in this credit forecasting. The completeness score is used to characterize the completeness of the input data, the freshness score is used to characterize the update time of the input data, the consistency score is used to characterize the degree of matching between the input data and the set verification rules, and the relevance score is used to characterize the degree of correlation between the input data and the credit score. The computational performance score of each credit prediction model is calculated based on at least one of the following: short-term accuracy, scenario adaptability, and response performance score. The short-term accuracy is used to characterize the prediction accuracy of the credit prediction model within a second set time period, where the second set time period is shorter than the first set time period. The scenario adaptability is used to characterize the degree of adaptability of the credit prediction model to the business scenario in which the credit assessment request is located. The response performance score is used to characterize the response time of the credit prediction model in this credit prediction. The performance score of each credit prediction model is calculated based on at least one of the historical performance score, input data quality score, and computational performance score.

7. The method according to claim 5, wherein the step of determining the weight coefficient of the corresponding encryption result based on the performance score of each credit prediction model includes: The performance scores of each credit prediction model are normalized to obtain the corresponding weight coefficients of the encryption results; where the sum of all weight coefficients is 1.

8. The method according to claim 1, further comprising, after obtaining the user feature set: The fraud probability is obtained by processing the user feature set using a pre-trained fraud detection model. The fraud probability is used to assess whether the credit assessment request was triggered by fraudulent behavior; If the fraud probability exceeds a set probability threshold, a credit assessment rejection message and a warning message will be sent to the target federal participant. The step of sending credit prediction instructions and the user feature set to each of the federal participants includes: If the fraud probability does not exceed the probability threshold, a credit prediction instruction and the user feature set are sent to each of the federal participants.

9. The method according to any one of claims 1 to 8, wherein the federated server is deployed with a global meta-model; the credit prediction model includes a local prediction model and a local meta-model; the method further includes: Send a first model update instruction to each of the aforementioned federal participants; The first model update instruction is used to trigger the federated participants to call the local prediction model to process the user feature sample set obtained locally, and obtain an intermediate result set and a credit prediction result set; wherein, the intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set; the number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the corresponding user feature sample set. Upon receiving the intermediate result set and credit prediction result set returned by each of the federation participants in response to the model update instruction, the global meta-model is trained using the intermediate result set and credit prediction result set returned by each of the federation participants until the global meta-model converges. When the global meta-model converges, the target model parameters and the second model update instruction of the converged global meta-model are sent to each federation participant. The second model update instruction is used to trigger each federation participant to update the model parameters of the local meta-model to the target model parameters, and adjust the model parameters of the local prediction model through the updated local meta-model, or connect the first layer of the updated local meta-model to the last layer of the local prediction model.

10. A credit assessment method applied to a target federal participant, the method comprising: Send a credit assessment request for the target user to the federal server; The credit assessment request is used to trigger the federal server to send a user data acquisition request for the target user to each federal participant; the federal participants include the target federal participant and other federal participants; In response to the user data acquisition request issued by the federated server, the system acquires the user data of the target user, encrypts the user data, and returns the encrypted user data to the federated server. In response to the credit prediction instruction from the federal server for the encrypted user data, the locally deployed credit prediction model is invoked to process the user feature set sent by the federal server to obtain a credit prediction score, and the encrypted result of the credit prediction score is returned to the federal server; the user feature set is obtained by the federal server integrating and processing the encrypted user data returned by each federal participant. Upon receiving the credit score returned by the federal server, the credit score is sent to the user terminal associated with the target user; the credit score is obtained by the federal server through fusion processing based on the encrypted results returned by each federal participant.

11. The method according to claim 10, wherein the credit prediction model comprises a local prediction model and a local meta-model; the federated server is deployed with a global meta-model; the method further comprises: In response to the first model update instruction sent by the federated server, the local prediction model is invoked to process the user feature sample set obtained locally, obtaining an intermediate result set and a credit prediction result set, which are then sent to the federated server; wherein, the intermediate result set includes the results output by the intermediate layer of the local prediction model for each sample in the user feature sample set; the number of credit prediction results included in the credit prediction result set is the same as the number of samples included in the user feature sample set; In response to the second model update instruction sent by the federated server, the model parameters of the local meta-model are updated to the target model parameters sent by the federated server, and the model parameters of the local prediction model are adjusted through the updated local meta-model, or the first layer of the updated local meta-model is connected to the last layer of the local prediction model; wherein, the target model parameters are obtained by the federated server training the global meta-model based on the intermediate result set and credit prediction result set returned by each federated participant.

12. A credit assessment system, comprising: A federal server, upon receiving a credit assessment request for a target user from a target federal participant, obtains the credit score of the target user by performing the steps of the method described in any one of claims 1 to 9, and sends the credit score to the target federal participant. as well as Multiple federal participants, including the target federal participant and other federal participants; The target federal participant is used to cooperate with the federal server to obtain the credit score of the target user by performing the steps of the method described in any one of claims 10-11, and to send the credit score to the user terminal associated with the target user. The other federal participants are used for: In response to a user data acquisition request sent by the federation server, the system acquires and encrypts the user data of the target user, and returns the encrypted user data to the federation server. In response to the credit prediction instruction sent by the federal server, the locally deployed credit prediction model is invoked to process the user feature set sent by the federal server, obtain a credit prediction score, and send it to the federal server.

13. A credit assessment device, applied to a federated server, comprising: The federation request module is configured to: in response to a credit assessment request issued by a target federation participant for a target user, send a user data acquisition request for the target user to each federation participant; the each federation participant includes the target federation participant and other federation participants; The feature processing module is configured to: upon receiving encrypted user data returned by each of the federation participants in response to the user data acquisition request, integrate and process the encrypted user data returned by each of the federation participants to obtain a set of user features. The federated invocation module is configured to send a credit prediction instruction and the user feature set to each federated participant; the credit prediction instruction is used to trigger each federated participant to invoke a locally deployed credit prediction model to process the user feature set and obtain a credit prediction score. The credit scoring module is configured to: upon receiving the encrypted results of the credit prediction scores returned by each of the federation participants in response to the credit prediction instruction, merge the encrypted results returned by each of the federation participants to obtain the credit score of the target user, and send the credit score to the target federation participant.

14. A credit assessment device applied to a target federal participant, the device comprising: The credit request module is configured to send a credit assessment request for the target user to the federated server. The credit assessment request is used to trigger the federal server to send a user data acquisition request for the target user to each federal participant; the federal participants include the target federal participant and other federal participants; The data processing module is configured to: in response to the user data acquisition request issued by the federated server, acquire the user data of the target user, encrypt the user data, and return the encrypted user data to the federated server; The credit prediction module is configured to: respond to the credit prediction instruction from the federal server for the encrypted user data, call the locally deployed credit prediction model to process the user feature set sent by the federal server, obtain a credit prediction score, and return the encrypted result of the credit prediction score to the federal server; the user feature set is obtained by the federal server integrating and processing the encrypted user data returned by each federal participant. The credit score sending module is configured to: upon receiving a credit score returned by the federal server, send the credit score to the user terminal associated with the target user; the credit score is obtained by the federal server through fusion processing based on the encrypted results returned by each federal participant.

15. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as described in any one of claims 1 to 9 and / or 10 to 11 by executing the executable instructions.

16. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 9 and / or 10 to 11.

17. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1-9 and / or 10-11.

Images