Risk assessment method, apparatus, device, medium, and program product

Abstract

This application provides a risk assessment method, apparatus, device, medium, and program product, which can be applied to the fields of artificial intelligence and fintech. The method includes: extracting a first feature vector based on obtained operational data of a target object; obtaining a second feature vector generated locally by at least one collaborating node, wherein the second feature vector is extracted based on private data about the target object from at least one collaborating node using locally stored risk assessment model parameters, and the private data characterizes the target object's ability to cope with risks; generating a public opinion risk feature vector based on financial risk semantic rules according to obtained public opinion about the target object; and inputting the first feature vector, the second feature vector, and the public opinion risk feature vector into a pre-trained risk assessment model to obtain a risk assessment result for the target object.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and financial technology, specifically to financial risk assessment, and more specifically to a risk assessment method, apparatus, equipment, medium, and program product. Background Technology

[0002] As the requirements for data dimensions in the financial risk control field continue to increase, the accuracy of risk assessment models increasingly relies on the fusion and analysis of multi-source heterogeneous data. The key data required for financial risk assessment is scattered across multiple independent entities' internal network systems. In practical applications, data silos and data isolation across departments and government / enterprise institutions have become core bottlenecks restricting the improvement of risk control model effectiveness. There is a need to optimize new methods for credit risk assessment. Summary of the Invention

[0003] In view of the above problems, embodiments of this application provide a risk assessment method, apparatus, device, medium, and program product.

[0004] According to a first aspect of this application, a risk assessment method is provided, comprising: extracting a first feature vector based on the operating data of a target object; obtaining a second feature vector generated locally by at least one collaborative node, wherein the second feature vector is extracted based on private data of the target object from at least one collaborative node using locally stored risk assessment model parameters, and the private data characterizes the target object's ability to cope with risks; generating a public opinion risk feature vector based on financial risk semantic rules according to the public opinion data of the target object; and inputting the first feature vector, the second feature vector, and the public opinion risk feature vector into a pre-trained risk assessment model to obtain a risk assessment result for the target object.

[0005] According to an embodiment of this application, a public opinion risk feature vector generated based on financial risk semantic rules, based on the public opinion data of the target object, includes: obtaining publicly available text information of the target object; performing entity association parsing on the publicly available text information to identify target events related to the risk; and determining the risk score based on the target events according to preset rules to obtain the public opinion risk feature vector.

[0006] According to an embodiment of this application, a risk score is determined based on a target event according to preset rules, including: determining the corresponding risk score based on the correlation between the target event and the company's debt repayment ability; using the risk score as a public opinion risk feature vector; wherein, if the impact of the target event on the debt repayment ability exceeds a preset threshold, a trigger signal is generated and the target risk assessment result is output.

[0007] According to an embodiment of this application, the method for obtaining the risk assessment model includes: obtaining the encryption gradient of at least one collaborative node, wherein the encryption gradient is obtained by at least one collaborative node calculating the gradient based on the current global model parameters and the second feature vector and encrypting it; performing secure aggregation processing on the received encryption gradient of at least one collaborative node to obtain the global encryption gradient; and distributing the global encryption gradient to at least one collaborative node to update the current global model parameters until a preset convergence condition is met.

[0008] According to an embodiment of this application, a secure aggregation process is performed on the received encryption gradient of at least one cooperative node to obtain a global encryption gradient, including: decrypting the encryption gradient of at least one cooperative node in a hardware security module or trusted execution environment to obtain a decryption gradient; determining the weight of each node according to the sample size of at least one cooperative node; obtaining an aggregation result using the weights according to the decryption gradient; and encrypting the aggregation result to obtain a global encryption gradient.

[0009] According to an embodiment of this application, the encryption gradient is calculated by the collaborating node as follows: the predicted default probability is calculated based on the current global model parameters and the second feature vector; the sample-level gradient is calculated based on the error between the predicted default probability and the true label; and the encryption gradient of the collaborating node is obtained by averaging all sample-level gradients within the same collaborating node.

[0010] According to an embodiment of this application, the encryption gradient is calculated by the collaborating node as follows: for any collaborating node, the gradient magnitude distribution of all samples within that collaborating node is statistically analyzed; based on the statistical characteristics of the gradient magnitude distribution, the gradient pruning threshold corresponding to that collaborating node is determined, wherein the statistical characteristics include the variance or quantile of the gradient magnitude; after calculating the gradient locally, if the absolute value of any gradient component is greater than the gradient pruning threshold, the gradient component is truncated to the boundary range corresponding to the gradient pruning threshold; and an encryption gradient is generated based on the truncated gradient.

[0011] According to a second aspect of this application, a risk assessment apparatus is provided, comprising: a first feature vector extraction module, configured to extract a first feature vector based on the operating data of a target object; a second feature vector acquisition module, configured to acquire a second feature vector generated locally by at least one collaborative node, wherein the second feature vector is extracted based on private data of the at least one collaborative node regarding the target object using locally stored risk assessment model parameters, and the private data characterizes the target object's ability to cope with risks; a public opinion risk feature vector acquisition module, configured to generate a public opinion risk feature vector based on financial risk semantic rules according to the public opinion data of the target object; and a risk assessment result output module, configured to input the first feature vector, the second feature vector, and the public opinion risk feature vector into a pre-trained risk assessment model to obtain a risk assessment result for the target object.

[0012] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0013] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0014] According to a fifth aspect of this application, a computer program product is also provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0015] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0016] Figure 1 The illustrations depict application scenarios of risk assessment methods, apparatus, devices, media, and program products according to embodiments of this application.

[0017] Figure 2 A flowchart illustrating a risk assessment method according to an embodiment of this application is shown schematically.

[0018] Figure 3 The schematic diagram illustrates the principle of obtaining public opinion risk feature vectors according to an embodiment of this application;

[0019] Figure 4 This schematic diagram illustrates the principle of hierarchical processing of target events according to an embodiment of this application.

[0020] Figure 5 The schematic diagram illustrates the principle of obtaining the global encryption gradient according to an embodiment of this application;

[0021] Figure 6 This schematic diagram illustrates the structural block diagram of a risk assessment apparatus according to an embodiment of this application;

[0022] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a risk assessment method according to an embodiment of this application. Detailed Implementation

[0023] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0026] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0027] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary security measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0028] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0029] In the process of digital implementation, some risk control systems of relevant commercial banks face the following three technical bottlenecks, which restrict the accuracy of approval, the timeliness of risk warning, and the ability to manage asset quality.

[0030] First, data isolation across departments and government / enterprise organizations poses compliance restrictions to traditional centralized modeling methods.

[0031] Core accounting data, tax compliance data, and legal ownership risk data of banks are stored in different institutions' internal network systems. Constrained by enterprise data compliance requirements, the raw plaintext data is prohibited from cross-domain aggregation, centralized storage, or unauthorized cross-platform sharing. Traditional risk control modeling typically requires centralized computation of the entire raw dataset. This approach poses compliance risks under the current regulatory framework and is difficult to implement on a large scale.

[0032] Second, general public opinion analysis technology lacks specific judgment standards for financial risk control, resulting in omissions and misjudgments in risk identification.

[0033] Many relevant public opinion monitoring systems rely on keyword matching, failing to distinguish the authority of the entities issuing the information, verify the substantive connection between the public opinion events and the company's debt repayment ability, or effectively filter out noise information unrelated to the company's operational risks. In practice, a large number of warning requests unrelated to credit risk are pushed out, while key negative signals such as hidden tax arrears, guarantee chain risks, and tight cash flow may be identified with delays, affecting the timeliness and accuracy of risk warnings.

[0034] Third, the relevant credit models have a long update cycle, making it difficult to reflect changes in business operations in real time.

[0035] Most bank risk control models update parameters offline in batches on a monthly or quarterly basis, which is essentially a post-event analysis-based risk control approach. When a company experiences sudden large-scale fund transfers, temporary administrative penalties, credit events involving the actual controller, or concentrated supply chain disruptions, the model cannot adjust its rating or credit granting strategies in sync. The risk control response speed may lag behind the changes in risk, and there is room for improvement in post-loan management.

[0036] In summary, the relevant technologies have shortcomings in cross-institutional compliance modeling, financial-specific public opinion quantification, and dynamic model iteration. A new technology solution for corporate credit risk assessment that improves compliance, accuracy, and real-time performance is needed.

[0037] Figure 1 The illustrations depict application scenarios of risk assessment methods, apparatus, devices, media, and program products according to embodiments of this application. Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send information, etc.

[0038] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal devices. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, and it can include, but is not limited to, applications, mini-programs, etc., installed on the terminal devices, and can also be in the form of web pages.

[0039] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data. The server can be the backend server of the aforementioned target application, used to provide backend services to the clients of the target application.

[0040] It should be noted that the risk assessment method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-103. Accordingly, the risk assessment device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-103.

[0041] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0042] Figure 2 A flowchart illustrating a risk assessment method according to an embodiment of this application is shown.

[0043] like Figure 2 As shown, the risk assessment method according to the embodiments of this application may include steps S210 to S230.

[0044] In step S210, the first feature vector is extracted based on the operating data of the target object.

[0045] In embodiments of this application, user consent or authorization can be obtained before acquiring user information. For example, a request to acquire user information can be sent to the user before step S210. If the user consents or authorizes the acquisition of user information, step S210 is executed.

[0046] In the embodiments of this application, a corresponding operation entry can be provided to the user, allowing the user to choose to agree to or reject the automated decision result. That is, before processing / making a decision on user information in step S210, the user can provide an instruction to agree to or reject the processing / decision through the corresponding operation entry. If the user agrees to the processing / decision, then the user information is processed / made a decision in step S210, i.e., step S210 is executed. If the user refuses to process / make a decision, then the expert decision-making process is initiated.

[0047] The target entities can be corporate entities, non-corporate entities, and micro and small enterprises that have opened corporate settlement accounts with commercial banks and handle business such as working capital, fixed assets, bank acceptance bills, letters of guarantee, and supply chain credit.

[0048] Operational data can be data related to business activities from different business systems with different data structures and storage formats. For example, it can include time-series transaction records of corporate accounts within banks, structured indicators of corporate financial statements, tax payment ledger data, judicial litigation and enforcement disclosure data, upstream and downstream supply chain performance transaction ledgers, and other cross-institutional and cross-format compliant operational data collections.

[0049] The embodiments of this application can be applied to a central node. Each node, including the central node and collaborating nodes, extracts numerical features based on its own data characteristics using a preset feature engineering method, forming a feature vector. , where d is the feature dimension and i is the node index.

[0050] The first feature vector can be an ordered set of numerical values ​​obtained after cleaning, normalizing, and numericalizing the operating data. For example, it can be a numerical array of length d, where each position corresponds to an operating indicator, such as the average daily deposit balance over the past 30 days or the average actual tax paid over the past 12 months. The central node can be an institutional computing node with a risk assessment model and some operating data, such as a bank's risk control computing cluster, which can deploy a feature extraction engine and model inference services.

[0051] For example, the central node could be a bank node, and the extracted features could include: fund settlement behavior characteristics, such as the average daily deposit balance over the past 30 days, the standard deviation of transaction flow, the proportion of nighttime transactions, and the frequency of transactions with upstream and downstream counterparties; and business history characteristics, such as the number of historical overdue payments, the maximum number of overdue days, and the proportion of the current credit limit to the total credit limit. The extraction can be implemented using methods such as sliding window statistics, ratio calculation, and time-series differencing.

[0052] In step S220, a second feature vector is obtained by at least one collaborative node locally, wherein the second feature vector is obtained by extracting risk assessment model parameters stored locally based on the private data of at least one collaborative node about the target object, and the private data characterizes the target object's ability to cope with risks.

[0053] Collaborating nodes can be other institutional computing nodes that participate in model training or inference with the central node but do not share the original data, such as the internal computing nodes of tax authorities, the data processing nodes of judicial institutions, and third-party public opinion service nodes.

[0054] Private data can be original or anonymized business-related information stored locally on the collaborative node, subject to data security regulations, and unable to be directly transmitted externally, such as corporate tax declaration details on the tax node and case file summaries on the judicial node.

[0055] The second feature vector can be a set of numerical features obtained by the collaborating node using the current model parameters stored locally to perform feature engineering on private data. For example, it can be a numerical array of length d', covering indicators such as tax stability indicators and litigation risk indicators.

[0056] Risk assessment model parameters can be parameters used in the second feature vector generated from private data. Alternatively, risk assessment model parameters can be trainable variables such as weight matrices and bias terms of each layer in the model, used to output prediction results given input features.

[0057] For example, the process of generating the second feature vector by the collaborative node may include: the tax node reads the enterprise's tax declaration records for the past 24 months from local storage, calculates the year-on-year growth rate of the monthly actual tax payment and the number of overdue declarations, forming tax-related features; the judicial node reads the enterprise's litigation information from local storage, counts the number of cases in which the enterprise is the defendant, the ratio of the total amount involved to the registered capital, forming judicial-related features; and the various features are concatenated in a preset order to obtain the second feature vector.

[0058] In embodiments of this application, data collection modules can be deployed at banking, tax, judicial, and public opinion nodes, respectively. Each node performs data cleaning, de-identification, and normalization formatting only within its internal network environment, and does not send raw plaintext data externally.

[0059] For example, tax node feature extraction can include: tax payment characteristics, such as the average, standard deviation, and year-on-year or month-on-month growth rate of actual tax paid over the past 12 months; compliance characteristics, such as the number of overdue declarations or tax audit markers. Extraction methods can include aggregate statistics and / or time series decomposition.

[0060] Judicial node feature extraction can include: litigation characteristics, such as the number of cases in which the defendant is located, the total amount involved, and the ratio of the amount of enforcement proceeds to registered capital; and creditworthiness characteristics, such as whether the individual is listed as a dishonest judgment debtor (0 / 1) and the number of times high-level consumption has been restricted. Extraction methods can include counting and normalization.

[0061] In step S230, a public opinion risk feature vector is generated based on the public opinion data of the target object and the semantic rules of financial risk.

[0062] Public opinion can include publicly available textual information related to the target entity's business operations, equity structure, and compliance status, released through the internet, government information disclosure platforms, and judicial disclosure channels. This includes publicly available textual information strongly related to corporate operations, legal representatives, actual controllers, equity assets, and compliant operations released by authoritative media, government information disclosure platforms, and judicial disclosure channels, but excludes privacy-related confidential data and non-business entertainment information.

[0063] Financial risk semantic rules can be a predefined set of judgment logic used to map unstructured text into a quantitative risk score. For example, it can be a set of rules for classifying risk levels based on event type, the authority of the issuing entity, and the degree of relevance to debt repayment ability.

[0064] A public opinion risk feature vector can be a numerical vector obtained by identifying entities, classifying events, and scoring risks in public opinion information. For example, it can be a vector of length k, with each element corresponding to a public opinion risk score of a different dimension.

[0065] For example, the process of generating a public opinion risk feature vector may include: collecting publicly available corporate texts from the entire network, using a pre-trained language model in the financial field to identify corporate entities and event types in the text, filtering out content that is irrelevant to the company's debt repayment ability, and then assigning risk scores based on the degree of impact of the event on operational stability, such as assigning a tax arrears announcement a score of 0.7 and a normal operation announcement a score of 0, ultimately forming a public opinion risk feature vector.

[0066] In step S240, the first feature vector, the second feature vector, and the public opinion risk feature vector are input into the pre-trained risk assessment model to obtain the risk assessment result of the target object.

[0067] The risk assessment model is built on a distributed collaborative training framework, and the original data of each collaborative node does not leave its local machine during the training process. Under the premise that the original data does not leave its local machine, does not transmit plaintext, and is not centrally stored, it achieves secure joint modeling of multi-source data and avoids the compliance risks of traditional centralized modeling.

[0068] A distributed collaborative training framework can be an architecture that supports multiple nodes in jointly training machine learning models without exchanging raw data. Examples include horizontal federated learning frameworks and vertical federated learning frameworks. A federated learning framework can be a distributed collaborative model training framework involving multiple institutions, where raw data does not leave the local machine, and only encrypted model gradients and weight parameters are transmitted, achieving "usable but invisible" data across institutions. "Raw data not leaving the local machine" means that each collaborating node only transmits encrypted intermediate calculation results such as model gradients and weight updates, without sending any unencrypted raw sample data, feature details, or label information. Through the federated learning framework, raw data remains within the private network, and only encrypted parameters are exchanged, meeting the data security requirements of the financial industry.

[0069] The embodiments of this application can improve the comprehensiveness of risk assessment by combining distributed collaborative training with local feature extraction, while meeting data security and compliance requirements, and reducing the complexity of cross-institutional data integration through standardized feature vectors and unified model interfaces.

[0070] The first and second feature vectors mentioned above, as well as the public opinion risk feature vector, can be standardized or normalized. The standardized parameters are calculated and saved locally and are not uploaded.

[0071] In some embodiments, the first feature vector, the second feature vector, and the public opinion risk feature vector can be concatenated along the channel dimension and input into a logistic regression model or a neural network model. Real-time collection of enterprise characteristics for the current day is possible. Input solidification model Calculate the probability of default The mapping from 'p' to five credit rating levels can be configured based on the bank's risk appetite, as shown in the example configuration below:

[0072] p<0.02 indicates AAA grade (high quality, extremely low risk).

[0073] A score of 0.02 ≤ p < 0.05 is classified as AA (good, low risk).

[0074] 0.05≤p<0.10 is classified as Grade A (general, low to medium risk).

[0075] 0.10≤p<0.20 is classified as Grade B (attention, medium risk).

[0076] A p ≥ 0.20 indicates a Grade D (high risk of default, or default has already occurred).

[0077] The model can output a default probability value (i.e., the p-value).

[0078] Risk control strategies can be automatically matched based on the above ratings. The risk control strategies are matched as follows:

[0079] AAA / AA level: Automatic renewal via green channel, simplified approval process.

[0080] Level A: Routine review, amount remains unchanged.

[0081] Category B: Manual review triggered; it is recommended to reduce credit or add collateral.

[0082] Category D: Credit granting will be immediately suspended, and the area will be placed under close monitoring.

[0083] The embodiments of this application can persistently store data interaction records, model parameter versions, rating results, and risk control operation logs, and support the export of audit reports according to regulatory requirements.

[0084] In other embodiments, to further enhance the robustness of feature vectors, adversarial training mechanisms can be introduced during the feature extraction stage. A generator network can be built on local nodes to simulate differences in data distribution across different institutions, and adversarial discriminant loss can be added to the feature extractor. This allows the extracted first and second feature vectors to eliminate institution-specific distribution biases, thereby improving the model's generalization ability during cross-institutional collaborative training. For example, features extracted by a bank node may have statistical characteristics specific to banking operations. Adversarial training can make these features more representative of general business operations, reducing model bias caused by differences in data distribution.

[0085] In the embodiments of this application, by extracting features from the central node and collaborating nodes respectively and retaining the original data locally, the security risks brought about by cross-institutional data flow can be reduced. At the same time, the fusion of multi-source features can enable the risk assessment results to cover more risk dimensions and improve the reference value of the assessment results.

[0086] The embodiments of this application can support daily automated incremental training in the background, enabling credit ratings to reflect the latest business operations and public opinion changes of enterprises in real time, solving the risk lag problem of traditional models. It can replace manual financial statement collection, public opinion screening, and scoring, reducing the manual costs of post-event risk control. Through multi-dimensional data fusion modeling, compared to models relying on only a single data source, it can capture more comprehensive risk signals, helping banks to identify potential defaulting customers in advance and take risk control measures, thereby reducing non-performing assets.

[0087] Figure 3 The schematic diagram illustrates the principle of obtaining public opinion risk feature vectors according to an embodiment of this application.

[0088] like Figure 3As shown in the embodiments of this application, the public opinion risk feature vector generated based on the public opinion data of the target object and the semantic rules of financial risk may include: obtaining the publicly available text information of the target object; performing entity association parsing on the publicly available text information to identify target events related to the risk; and determining the risk score based on the target events according to preset rules to obtain the public opinion risk feature vector.

[0089] Publicly available text information can be textual content related to business operations published on official media, government information disclosure platforms, and judicial information disclosure websites, such as administrative penalty decisions, court judgments, and announcements of business changes by market supervision departments.

[0090] Entity association resolution can be the process of identifying entities such as company names, legal representatives, shareholders, and branches mentioned in text and matching them with target objects. For example, it can be accurately associated using the company's unified social credit code or business registration number.

[0091] For example, such as Figure 3 As shown, the process involves obtaining publicly available text information about the target object, followed by sequential steps of text entity extraction, entity relationship matching, and semantic association analysis. After these processes, the output can be routed to the risk score calculation unit for quantitative accounting, ultimately generating a structured public opinion risk feature vector. For example, a pre-trained financial language model can be used to identify the corporate entity, event type, and sentiment in the text, filtering out information irrelevant to the company's solvency.

[0092] The target event can be a specific matter related to changes in the company's operating status as described in publicly available text information, such as equity freeze, large administrative penalties, bankruptcy filing, major contract breach, etc.

[0093] The preset rules can be pre-configured event classification and assignment logic. For example, events can be divided into four levels: normal operation, minor changes, negative risks, and major crises, each corresponding to a different risk score. The risk score can be a numerical value used to quantify the degree of risk of an event, and the value range can be set to a continuous or discrete value between 0 and 1. For example, 0 represents no risk, and 1 represents extremely high risk.

[0094] The public opinion risk feature vector can take the form of a multi-dimensional vector. For example, the first dimension represents the number of recent negative events, the second dimension represents the highest risk level, the third dimension represents the frequency of event occurrence, the fourth dimension represents the scope of event impact, and the fifth dimension represents the change in risk trend. Each dimension can be calculated using statistical methods or models.

[0095] In some embodiments, all publicly available texts published within the past 180 days can be retrieved, and entity association analysis can be performed on each text to filter out texts directly related to the target object. Each text is then categorized by event type, determining whether it falls under the categories of business registration changes, tax issues, legal proceedings, or operational anomalies. A basic risk score is assigned to each event type according to preset rules, and the score weight is adjusted based on the authority of the publishing entity; for example, penalty announcements issued by relevant departments have a higher weight than ordinary media reports. The risk scores of all relevant events are then weighted and summed using a time decay factor to obtain the final public opinion risk feature vector.

[0096] In other embodiments, the public opinion risk feature vector can be further refined into a dynamic time-series vector. This not only considers the risk score at the current moment but also incorporates time series analysis to calculate indicators such as volatility and acceleration of public opinion risk. For example, a sliding time window can be constructed, and the variance of the risk score within the window can be statistically analyzed. A larger variance indicates that the company is at the center of public opinion, with high risk uncertainty. This dynamic feature helps the model distinguish between occasional public opinion events and ongoing business deterioration, avoiding misjudging short-term noise as long-term risk.

[0097] In the embodiments of this application, a public opinion risk feature vector can be generated based on financial risk semantic rules. This can transform unstructured text information into quantifiable model input, reduce interference from irrelevant information, and enable public opinion factors to be integrated into the risk assessment process in an orderly manner, thereby improving the targeting of risk identification.

[0098] According to an embodiment of this application, determining a risk score based on a target event according to preset rules may include: determining the corresponding risk score based on the correlation between the target event and the company's debt repayment ability; using the risk score as a public opinion risk feature vector; wherein, if the impact of the target event on the debt repayment ability exceeds a preset threshold, a trigger signal is generated and the target risk assessment result is output.

[0099] The degree of correlation between a target event and a company's debt repayment ability can be an indicator of the impact of the target event on the company's future repayment ability. For example, events that directly affect cash flow have a higher degree of correlation, while events that only affect the company's image have a lower degree of correlation.

[0100] Risk scores can be coefficients assigned based on the degree of correlation, used to adjust the level of risk. Preset thresholds can be pre-defined cutoff values ​​used to determine whether special handling is triggered; for example, a risk score exceeding 0.8 might be considered a significant risk. Trigger signals can be indicators used to activate specific risk control processes, such as suspending credit granting or initiating manual review. The target risk assessment result can be a rating directly determined based on the trigger signal, such as being directly classified as the lowest level.

[0101] Figure 4A schematic diagram illustrating the principle of hierarchical processing of target events according to an embodiment of this application is shown.

[0102] For example, such as Figure 4 As shown, four risk levels can be set, and target events can be categorized according to preset rules:

[0103] Level 1 (Neutral): Normal business announcements, favorable policies, etc., risk score = 0, not participating in model training.

[0104] Level 2 (Mildly Related): Ordinary business registration changes and routine personnel adjustments, risk score = 0.1, archived only in the back office.

[0105] Level 3 (Substantial Negative): Tax arrears audit, contract breach, equity freeze, large administrative penalty, risk score = 0.7, included in the core feature pool.

[0106] Level 4 (Major Negative): Bankruptcy liquidation, loss of contact with the actual controller, criminal case filing, forced seizure of core assets, triggering the circuit breaker mechanism, directly adjusting the credit rating to the lowest level and issuing a red alert.

[0107] The embodiments of this application can employ a four-level risk classification rule to filter irrelevant noise and quantify implicit negative public opinion. Compared with the pure keyword matching method, it can significantly reduce the false positive rate and false negative rate.

[0108] In some embodiments, the process of determining the risk score may include: first, assessing the direct impact of the target event on the target's cash flow, such as a large fine directly reducing cash reserves, with a weight value of 0.8; second, assessing the indirect impact on financing channels, such as negative public opinion potentially leading to banks tightening credit, with a weight value of 0.3; and then weighting and summing the direct and indirect impacts to obtain a comprehensive risk score.

[0109] In some embodiments, if the target event is a bankruptcy liquidation application, its correlation with solvency is at the highest level, and the risk score can be set to 1.0. If it exceeds the preset threshold of 0.9, a trigger signal is generated, the target risk assessment result is directly determined to be the lowest credit rating, and a red warning is output. If the target event is a common business registration change, the correlation is low, the risk score can be set to 0.1, and no trigger signal is generated.

[0110] In some embodiments, the potential impact of a target event on its upstream and downstream supply chain partners can be analyzed. If the target is a core enterprise and defaults on its debt, its risk score can be increased by a certain percentage to reflect the cascading risk in supply chain finance. Conversely, if the target enterprise is merely a peripheral supplier with sufficient alternative sources, the risk score can be appropriately lowered. This risk adjustment based on supply chain linkages can make the assessment results more consistent with the complex scenarios of actual financial transactions.

[0111] In the embodiments of this application, the risk score can be determined based on the correlation between the target event and the company's debt repayment ability. This makes the quantification of public opinion risk more in line with the actual needs of financial risk control, avoids irrelevant public opinion from interfering with the assessment results, and enables a rapid response to major risks by setting thresholds and trigger signals.

[0112] According to an embodiment of this application, the method of obtaining the risk assessment model may include: obtaining the encryption gradient of at least one collaborative node, wherein the encryption gradient is obtained by at least one collaborative node based on the current global model parameters and the gradient calculated by the second feature vector and encrypted; performing secure aggregation processing on the received encryption gradient of at least one collaborative node to obtain the global encryption gradient; distributing the global encryption gradient to at least one collaborative node; updating the current global model parameters until a preset convergence condition is met.

[0113] Encrypted gradients can be ciphertext data generated by collaborating nodes processing locally computed gradient data using symmetric or asymmetric encryption algorithms, such as a byte stream processed by a block cipher algorithm. Current global model parameters can be a set of model parameters maintained by the central node and shared by all participating nodes in the current round, denoted as θ.

[0114] Secure aggregation processing can be a process of calculating the encrypted gradients of multiple nodes without decrypting the gradients of individual nodes, to obtain a global update containing information about all nodes.

[0115] The global encrypted gradient can be the encrypted gradient update amount that represents the training direction of all nodes, obtained after secure aggregation.

[0116] The embodiments of this application can iteratively update the local model parameters of each node. After each round of updates, the central node evaluates whether the convergence condition is met. The convergence condition can be any one of the following:

[0117] a. Change in loss function: Calculate the global loss for this round. Compared to the previous round of overall losses The absolute difference, if less than the threshold If so, then it is considered convergent.

[0118] b. Gradient Norm: If the L2 norm of the global gradient is... If the value is less than the threshold, it means that the parameters are close to optimal.

[0119] c. Model parameter changes: If the parameter update amount The value is very small, indicating that the model has stabilized.

[0120] d. Maximum number of rounds: To prevent infinite training, a maximum number of rounds is set. (For example, 100 rounds), and then it will be forcibly stopped.

[0121] If the convergence condition is met, the central node will finalize the model parameters. The encrypted model is distributed to each collaborating node, which then saves and solidifies the model for subsequent predictions. If the conditions are not met, the next iteration continues.

[0122] In some embodiments, the training process of the risk assessment model can employ a horizontal federated learning architecture. Before training begins, the central node initializes global model parameters and distributes them to each collaborating node. In each round of training, each collaborating node calculates gradients using local data, encrypts the gradients, and uploads them to the central node. The central node decrypts and aggregates the encrypted gradients within a hardware security module or trusted execution environment to obtain the global encrypted gradient before distributing it. Each collaborating node updates its local model parameters using the global encrypted gradient, repeating the above process until the model converges.

[0123] In some embodiments, the encryption gradient can be generated using a national standard symmetric encryption algorithm. The collaborating nodes and the central node pre-negotiate a pair of session keys K. i Node i uses key K i For gradient G i Encryption is performed by converting each floating-point number in the gradient vector into a byte stream, then encrypting it in blocks to generate ciphertext C. i After receiving all the ciphertext, the central node decrypts it internally within secure hardware and calculates a weighted average. This prevents the central node or other eavesdroppers from retrieving the original data from the gradient, providing financial-grade security. Constraints on the original features can be set. ,Label Unencrypted gradient None of them may leave the local intranet of the node.

[0124] In some embodiments, to prevent gradient leakage attacks, differential privacy processing can be performed on the gradients before encryption. Random noise conforming to a Laplace distribution or other distributions can be added to the local gradients, with the scale of the noise adjusted according to the privacy budget ε. This way, even if an attacker attempts to deduce the original data from the gradients, the noise will prevent them from obtaining accurate information, thus further protecting data privacy during federated learning.

[0125] In the embodiments of this application, by transmitting encrypted gradients during training and aggregating them in a secure environment, it can be ensured that the original data of each collaborative node never leaves its local machine, while meeting the needs of multi-party joint modeling and improving the generalization ability and evaluation accuracy of the model.

[0126] like Figure 5As shown, according to an embodiment of this application, performing secure aggregation processing on the received encryption gradients of at least one collaborative node to obtain a global encryption gradient may include: decrypting the encryption gradients of at least one collaborative node in a hardware security module or trusted execution environment to obtain a decryption gradient; determining the weights corresponding to each node based on the sample size of at least one collaborative node and the central node; obtaining an aggregation result using the weights based on the decryption gradient; and encrypting the aggregation result to obtain a global encryption gradient.

[0127] A hardware security module can be a physical computing device used to protect and manage encryption keys. It can perform key generation, storage, destruction, and encryption operations, all of which can be completed within a tamper-proof hardware module, ensuring that key materials are not exposed to the external environment.

[0128] A trusted execution environment can be a hardware-based secure isolation zone that provides "safe"-like protection for sensitive code and data through processor instruction sets and encryption technology, ensuring that the computation process and data are "usable but not visible".

[0129] The sample size can be the number of training samples held by each node, used as the weighting basis for the weighted average. The aggregation result can be the combined gradient obtained by summing the gradients of all nodes according to their weights.

[0130] In some embodiments, after receiving the encryption gradients uploaded by each collaborating node, the central node can transmit these ciphertext data to the hardware security module. In the hardware security module, each ciphertext is decrypted using a pre-stored decryption key to obtain the plaintext gradient data. Then, the number of samples n reported by each node is read. i The total sample size N is calculated to be equal to the sum of the n samples. i The sum. Then, calculate the weight of each node, which equals n. i Divide by N. Multiply the plaintext gradient of each node by its corresponding weight and sum them to obtain the aggregated result. Finally, encrypt the aggregated result using the global public key to generate a globally encrypted gradient.

[0131] The central node receives the ciphertext gradients from all participants. Then, aggregation operations are performed under secure conditions. The central node is equipped with a hardware security module or a trusted execution environment. Within the hardware security module or trusted execution environment, the gradients of each node are decrypted, weighted average aggregation is performed, and then the output is encrypted. This ensures that the decrypted plaintext gradients are not exposed to the host operating system or other processes.

[0132] Aggregation function: uses a weighted average, with the weights equal to the sample size of each node. :

[0133] ;

[0134] Aggregation results The global gradient update value was initially presented as plaintext and was subsequently re-encrypted in the hardware security module. The central node can Distribute to all participating collaborative nodes.

[0135] Each collaborating node can use the locally stored encryption algorithm key to decrypt. Obtain the plaintext global gradient Then update the local model parameters:

[0136]

[0137] Where, θ new For the updated model parameters, θ old These are the model parameters before the update. The learning rate is a hyperparameter, typically set to 0.01~0.1. The learning rate controls the step size of parameter updates; too large a rate leads to oscillations, while too small a rate results in slow convergence.

[0138] In some embodiments, an enclave can be created in the processor to load the decryption and aggregation code for execution. External operating systems and virtual machine monitors cannot access the memory data within the enclave, thereby achieving the same level of security.

[0139] In some embodiments, the central node can split its private key into multiple fragments and distribute them to different collaborating nodes. During decryption, multiple collaborating nodes need to work together to recover the private key; a single node cannot complete the decryption operation independently. This embodiment avoids the risk of the central node becoming a single point of failure or being compromised, achieving decentralized secure aggregation.

[0140] In the embodiments of this application, performing decryption and aggregation in a hardware security module or trusted execution environment can prevent gradients from being acquired or tampered with during transmission and processing, thus ensuring the security and reliability of the model training process.

[0141] According to an embodiment of this application, the encryption gradient can be calculated by the collaborating node as follows: based on the current global model parameters and the second feature vector, the predicted default probability is calculated; based on the error between the predicted default probability and the true label, the sample-level gradient is calculated; and the encryption gradient of the collaborating node is obtained by averaging all sample-level gradients within the same collaborating node.

[0142] The predicted default probability can be a numerical value, usually denoted as p, representing the likelihood that a target entity will be unable to repay its debts within a specific future timeframe, based on current parameters and input features. The true label can be a binary identifier reflecting the target entity's actual creditworthiness; for example, 0 represents no default, and 1 represents default.

[0143] A sample-level gradient can be a gradient vector calculated for a single training sample. The gradient magnitude distribution can be a statistical law describing the magnitude of the absolute values ​​of each element in the gradient vector. The gradient clipping threshold can be a boundary parameter used to limit the maximum value of the gradient.

[0144] Assume the current global model is The parameter set is θ. Each node Have local sample set ,in For the label (whether a breach has occurred, 0 or 1).

[0145] The process of cooperating nodes calculating the encryption gradient may include: loading the current global model parameters θ, and iterating through each sample (X) in the local training set. i ,y i For each sample, first calculate the predicted default probability p. i The calculation formula is p i Equals the logical function acting on the transpose of θ multiplied by X i The result. Next, the prediction error e is calculated. i equals p i Subtract y i Then calculate the sample-level gradient g of the sample. i equals e i Multiply by X i After calculating all samples, the average gradient gi of all samples within the same collaborative node is arithmetically averaged to obtain the average gradient of that node. Finally, the average gradient is encrypted to obtain the encrypted gradient.

[0146] The loss function in the above process can be the cross-entropy loss, which is suitable for binary classification tasks. Its mathematical expression is:

[0147] ;

[0148] Where n is the total number of samples, y i For real labels, The default probability predicted by the model. For logical functions, θ T This is the transpose of the parameter vector.

[0149] Gradient calculation: Node i calculates the loss function L i For model parameters gradient G i :

[0150] .

[0151] For logistic regression, the closed-form solution of the gradient can be expressed as:

[0152] ;

[0153] That is, the prediction error of each sample is multiplied by the feature vector and then averaged.

[0154] In some embodiments, to prevent gradient explosion, a gradient clipping operation can be introduced after calculating the sample-level gradients. The magnitude distribution of all sample-level gradients within a node can be statistically analyzed, for example, by calculating the L2 norm of all gradient vectors. Based on the statistical characteristics of the magnitude distribution, for example, the 95th percentile of the L2 norm of all sample gradients can be used as the gradient clipping threshold. After locally calculating the gradients, if the absolute value of any gradient component exceeds this threshold, the gradient component is truncated to the specified boundary.

[0155] In some embodiments, the pruning threshold can be dynamically adjusted based on the number of training epochs. In the early stages of training, when model parameters are unstable and gradients fluctuate significantly, a higher pruning threshold can be set. As the number of training epochs increases, the model gradually converges, and the gradients tend to stabilize, allowing the pruning threshold to be gradually reduced. This dynamic adjustment mechanism can ensure the model's convergence speed while avoiding training instability caused by a fixed threshold that is too high or too low.

[0156] In the embodiments of this application, the contribution direction of local data to model optimization can be reflected by calculating and averaging the sample-level gradient. Combined with gradient pruning technology, the stability and convergence efficiency of model training can be effectively improved.

[0157] According to an embodiment of this application, the encrypted gradient can be calculated by the collaborating node as follows: for any collaborating node, the gradient magnitude distribution of all samples within the collaborating node is statistically analyzed; based on the statistical characteristics of the gradient magnitude distribution, the gradient pruning threshold corresponding to the collaborating node is determined, wherein the statistical characteristics include the variance or quantile of the gradient magnitude; after calculating the gradient locally, if the absolute value of any gradient component is greater than the gradient pruning threshold, the gradient component is truncated to the boundary range corresponding to the gradient pruning threshold; and an encrypted gradient is generated based on the truncated gradient.

[0158] The gradient magnitude distribution can be a statistical law describing the magnitude of the absolute values ​​of each element in the gradient vector. Statistical features can be numerical values ​​used to characterize the distribution, such as mean, variance, maximum value, minimum value, or specific quantiles. The gradient clipping threshold can be a boundary parameter used to limit the maximum value of the gradient. Truncation can be the process of forcibly setting values ​​exceeding the boundary to the boundary value.

[0159] In some embodiments, the process of statistically analyzing the gradient magnitude distribution of all samples within a given collaborative node may include: after calculating the gradients at all sample levels, calculating the L2 norm of each gradient vector to obtain a set of norm values; performing statistical analysis on these norm values ​​to calculate their variance or quantiles; for example, selecting the 90th quantile as a statistical feature; determining a gradient clipping threshold based on this statistical feature, for example, setting the threshold to 1.5 times the statistical feature value; and before generating encrypted gradients, checking each gradient component and, if its absolute value is greater than the gradient clipping threshold, setting it as the sign of the gradient clipping threshold multiplied by the gradient clipping threshold.

[0160] Gradient clipping can be applied not only to individual components but also to the entire gradient vector. The L2 norm of the entire gradient vector can be calculated, and if it exceeds a preset global clipping threshold, the entire vector is scaled down proportionally until its norm equals the threshold.

[0161] In some embodiments, a first-order momentum estimate and a second-order momentum estimate of the gradient can be maintained. During pruning, not only the magnitude of the current gradient is considered, but also the cumulative effect of historical gradients. If the current gradient is in the same direction as the historical momentum and its magnitude is too large, a stronger pruning is performed; if the directions are opposite, a larger gradient may be retained to accelerate escape from the saddle point. This pruning method can better balance the convergence speed and stability of the model.

[0162] In the embodiments of this application, the pruning threshold can be dynamically adjusted based on the gradient magnitude distribution and the gradient can be truncated, which can effectively prevent the extreme gradients generated by individual abnormal samples from damaging the global model and improve the robustness of federated learning training.

[0163] Based on the above risk assessment method, embodiments of this application also provide a risk assessment device. The following will be combined with... Figure 6 The device is described in detail.

[0164] Figure 6 A schematic block diagram of a risk assessment apparatus according to an embodiment of this application is shown.

[0165] like Figure 6 As shown, the risk assessment device 600 of this embodiment includes a first feature vector extraction module 610, a second feature vector acquisition module 620, a public opinion risk feature vector acquisition module 630, and a risk assessment result output module 640.

[0166] The first feature vector extraction module 610 is used to extract the first feature vector based on the operating data of the target object.

[0167] The second feature vector acquisition module 620 is used to obtain a second feature vector generated locally by at least one collaborative node, wherein the second feature vector is obtained by extracting risk assessment model parameters stored locally based on private data of at least one collaborative node about the target object, and the private data characterizes the target object's ability to cope with risks.

[0168] The public opinion risk feature vector acquisition module 630 is used to generate a public opinion risk feature vector based on financial risk semantic rules according to the public opinion data of the target object.

[0169] The risk assessment result output module 640 is used to input the first feature vector, the second feature vector, and the public opinion risk feature vector into the pre-trained risk assessment model to obtain the risk assessment result of the target object.

[0170] According to an embodiment of this application, the public opinion risk feature vector acquisition module 630 is further used to acquire publicly available text information of the target object; perform entity association parsing on the publicly available text information to identify target events related to the risk; and determine the risk score based on the target events according to preset rules to obtain the public opinion risk feature vector.

[0171] According to an embodiment of this application, the public opinion risk feature vector acquisition module 630 is further used to determine the corresponding risk score based on the correlation between the target event and the enterprise's debt repayment ability; and to use the risk score as the public opinion risk feature vector; wherein, if the impact of the target event on the debt repayment ability exceeds a preset threshold, a trigger signal is generated and the target risk assessment result is output.

[0172] According to an embodiment of this application, the method for obtaining the risk assessment model of the risk assessment result output module 640 includes: obtaining the encryption gradient of at least one collaborative node, wherein the encryption gradient is obtained by at least one collaborative node based on the current global model parameters and the gradient calculated according to the second feature vector and encrypted; performing secure aggregation processing on the received encryption gradient of at least one collaborative node to obtain the global encryption gradient; and sending the global encryption gradient to at least one collaborative node to update the current global model parameters until the preset convergence condition is met.

[0173] According to an embodiment of this application, the risk assessment result output module 640 is further configured to, in a hardware security module or trusted execution environment, decrypt the encryption gradient of at least one collaborative node to obtain a decryption gradient; determine the weight of each node based on the sample size of at least one collaborative node and the central node; obtain an aggregation result using the weights based on the decryption gradient; and encrypt the aggregation result to obtain a global encryption gradient.

[0174] According to an embodiment of this application, the risk assessment result output module 640 is further configured to calculate the predicted default probability based on the current global model parameters and the second feature vector; calculate the sample-level gradient based on the error between the predicted default probability and the true label; and average all sample-level gradients within the same collaborative node to obtain the encrypted gradient of that collaborative node.

[0175] According to an embodiment of this application, the risk assessment result output module 640 is further configured to: statistically analyze the gradient magnitude distribution of all samples within any collaborative node; determine the gradient pruning threshold corresponding to the collaborative node based on the statistical characteristics of the gradient magnitude distribution, wherein the statistical characteristics include the variance or quantile of the gradient magnitude; after calculating the gradient locally, if the absolute value of any gradient component is greater than the gradient pruning threshold, then truncate the gradient component to the boundary range corresponding to the gradient pruning threshold; and generate an encrypted gradient based on the truncated gradient.

[0176] According to embodiments of this application, any multiple modules among the first feature vector extraction module 610, the second feature vector acquisition module 620, the public opinion risk feature vector acquisition module 630, and the risk assessment result output module 640 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the first feature vector extraction module 610, the second feature vector acquisition module 620, the public opinion risk feature vector acquisition module 630, and the risk assessment result output module 640 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first feature vector extraction module 610, the second feature vector acquisition module 620, the public opinion risk feature vector acquisition module 630, and the risk assessment result output module 640 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0177] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a risk assessment method according to an embodiment of this application.

[0178] like Figure 7As shown, an electronic device 700 according to an embodiment of this application includes a processor 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory 702 or a program loaded from a storage portion 708 into a random access memory 703. The processor 701 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 701 may also include onboard memory for caching purposes. The processor 701 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0179] Random access memory 703 stores various programs and data required for the operation of electronic device 700. Processor 701, read-only memory 702, and random access memory 703 are interconnected via bus 704. Processor 701 executes various steps of the method flow according to embodiments of this application by executing programs stored in read-only memory 702 and / or random access memory 703. It should be noted that programs may also be stored in one or more memories other than read-only memory 702 and random access memory 703. Processor 701 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0180] According to embodiments of this application, the electronic device 700 may further include an input / output interface 705, which is also connected to a bus 704. The electronic device 700 may also include one or more of the following components connected to the input / output interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card, such as a local area network card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the input / output interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 708 as needed.

[0181] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0182] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 702, and / or random access memory 703, and / or one or more memories other than read-only memory 702 and random access memory 703 described above.

[0183] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0184] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 709, and / or installed from a removable medium 711. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0185] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by processor 701, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0186] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0187] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0188] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A risk assessment method, characterized in that, include: Based on the target object's operational data, the first feature vector is extracted; A second feature vector is obtained by at least one collaborative node locally, wherein the second feature vector is extracted based on the private data of the at least one collaborative node about the target object and using locally stored risk assessment model parameters, the private data characterizing the target object's ability to cope with risks; Based on the public opinion data of the target object, a public opinion risk feature vector is generated according to financial risk semantic rules; The first feature vector, the second feature vector, and the public opinion risk feature vector are input into a pre-trained risk assessment model to obtain the risk assessment result of the target object.

2. The method according to claim 1, characterized in that, The public opinion risk feature vector generated based on the public opinion data of the target object and the semantic rules of financial risk includes: Obtain the publicly available text information of the target object; Perform entity association parsing on publicly available text information to identify risk-related target events; Based on preset rules, risk scores are determined according to the target events to obtain the public opinion risk feature vector.

3. The method according to claim 2, characterized in that, The step of determining a risk score based on the target event according to preset rules includes: The corresponding risk score is determined based on the degree of correlation between the target event and the company's debt repayment ability; The risk score is used as the public opinion risk feature vector. If the impact of the target event on the debt repayment ability exceeds a preset threshold, a trigger signal is generated and the target risk assessment result is output.

4. The method according to claim 1, characterized in that, The methods for obtaining the risk assessment model include: Obtain the encryption gradient of the at least one collaborative node, wherein the encryption gradient is obtained by the at least one collaborative node through encryption of the gradient calculated by the at least one collaborative node based on the current global model parameters and the second feature vector; A secure aggregation process is performed on the received cryptographic gradients of at least one cooperating node to obtain a global cryptographic gradient; The global encryption gradient is sent to at least one collaborating node to update the current global model parameters until the preset convergence condition is met.

5. The method according to claim 4, characterized in that, The step of performing secure aggregation processing on the received encryption gradients of the at least one cooperative node to obtain a global encryption gradient includes: In a hardware security module or trusted execution environment, the encryption gradient of the at least one cooperating node is decrypted to obtain the decryption gradient; The weights of each node are determined based on the sample size of the at least one cooperating node. Based on the decryption gradient, the aggregation result is obtained using the weights; The aggregation result is encrypted to obtain a global encryption gradient.

6. The method according to claim 4, characterized in that, The encryption gradient is calculated by the collaborating node in the following manner: Calculate the predicted default probability based on the current global model parameters and the second feature vector; Calculate the sample-level gradient based on the error between the predicted default probability and the actual label; The encrypted gradient of a collaborative node is obtained by averaging all sample-level gradients within the same collaborative node.

7. The method according to claim 4, characterized in that, The encryption gradient is calculated by the collaborating node in the following manner: For any of the aforementioned cooperative nodes, statistically analyze the gradient magnitude distribution of all samples within that cooperative node; Based on the statistical characteristics of the gradient magnitude distribution, the gradient clipping threshold corresponding to the cooperative node is determined, wherein the statistical characteristics include the variance or quantile of the gradient magnitude. After calculating the gradient locally, if the absolute value of any gradient component is greater than the gradient clipping threshold, then the gradient component is truncated to the boundary range corresponding to the gradient clipping threshold. The encrypted gradient is generated based on the truncated gradient.

8. A risk assessment device, characterized in that, include: The first feature vector extraction module is used to extract the first feature vector based on the operating data of the target object; The second feature vector acquisition module is used to obtain a second feature vector generated locally by at least one collaborative node, wherein the second feature vector is obtained by extracting risk assessment model parameters stored locally based on the private data of the at least one collaborative node about the target object, and the private data characterizes the target object's ability to cope with risks; A public opinion risk feature vector acquisition module is used to generate a public opinion risk feature vector based on financial risk semantic rules according to the public opinion data of the target object; and The risk assessment result output module is used to input the first feature vector, the second feature vector, and the public opinion risk feature vector into a pre-trained risk assessment model to obtain the risk assessment result of the target object.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images