Business risk analysis methods, devices, equipment and media for multi-party secure computation
By encrypting cross-party anchor point identification data and extracting anchor point-related business subgraphs, and combining this with a pre-trained model for multi-party business risk assessment, the problems of data leakage and difficulty in capturing topological relationships in traditional methods are solved, thus achieving reliability and privacy protection in cross-platform business risk analysis.
Patent Information
- Application Number
- CN202611341385.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional business risk analysis methods are prone to data leakage risks when performing cross-platform collaborative computing, and they are difficult to capture the topological relationships of the related business graphs among the participants, resulting in poor analysis reliability.
By encrypting the cross-party anchor point identifier data of the target business object, extracting the anchor point-related business subgraph, and using a pre-trained target business risk prediction model to conduct multi-party business risk assessment, cross-platform secure computing and joint computing of deep topology structures are achieved.
While protecting the privacy of business data of all parties, it significantly improves the reliability of cross-platform joint business risk analysis, avoids the risk of data leakage, and can capture the hidden relationships between the same user and different participants.
Smart Images

Figure CN123122932A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, applicable to fintech and medical technology scenarios, and particularly to a business risk analysis method, apparatus, equipment, and medium for multi-party secure computation. Background Technology
[0002] Traditional business risk analysis methods typically match business data from different participants based on the plaintext identity of the target business entity. The matched data is then converted into structured features and used for risk prediction through secure aggregation. For example, in cross-platform insurance fraud analysis, to jointly identify whether an insured person is at risk of cross-platform fraud, the first participant (e.g., insurance institution A), the second participant (e.g., insurance institution B), and the third participant (e.g., insurance institution C) can directly use the insured person's real ID number or mobile phone number, among other plaintext identifiers, to perform cross-institutional data matching and alignment. After successful matching, features such as "historical number of medical visits" and "claim amount" are extracted and merged to perform a joint risk score for the insured person's insurance business risk. For example, in the identification of violations in healthcare scenarios, the first participant (e.g., Hospital D), the second participant (e.g., Hospital E), and the third participant (Pharmaceutical Platform F) typically perform alignment and matching among the three institutions (D, E, and F) based on plaintext identifiers such as the patient's real medical insurance card number or ID card number. They then extract features such as "historical number of visits" and "prescription frequency" to jointly score the patient's medical behavior risk. However, this method, which directly matches across parties based on plaintext identifiers, makes it difficult to perform multi-party joint calculations without exposing the original data. This leads to data leakage risks during multi-party business risk analysis. Furthermore, this method only performs joint calculations on the discrete features of each participant, making it difficult to capture the interconnected business graph topology relationships between them (such as the hidden relationships between the same user and different participants), resulting in poor reliability of business risk analysis. Therefore, how to achieve reliable cross-platform joint risk analysis while protecting the business data of all parties has become an urgent technical problem to be solved. Summary of the Invention
[0003] This invention provides a business risk analysis method, apparatus, computer equipment, and medium for multi-party secure computation, in order to solve the technical problem of how to achieve reliable cross-platform joint risk analysis while protecting the business data of all parties.
[0004] Firstly, a business risk analysis method for multi-party secure computation is provided, including: In response to a business risk analysis request for a target business object, obtain the participant business graph of at least two target participants; Obtain cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, and is used to locate anchor point nodes belonging to the same target business object in the business graphs of different participants; The cross-anchor point identification data is encrypted to obtain encrypted cross-anchor point data; Based on the cross-party anchor point encrypted data, the anchor point associated business sub-graphs of each target participant are extracted from the business graphs of each participant. The pre-trained target business risk prediction model is used to conduct multi-party business risk assessments on the business subgraphs associated with each anchor point, thereby obtaining the business risk assessment data of the target business object.
[0005] Secondly, a business risk analysis device for multi-party secure computation is provided, including: The participant business graph acquisition module is used to obtain the participant business graphs of at least two target participants in response to a business risk analysis request for a target business object. The cross-party anchor point identification data module is used to obtain the cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, and is used to locate the anchor point node belonging to the same target business object in the business graph of different participants; An encryption processing module is used to encrypt the cross-anchor point identification data to obtain encrypted cross-anchor point data. Anchor point associated business subgraph extraction module is used to extract the anchor point associated business subgraph of each target participant from the business graph of each participant based on the cross-party anchor point encrypted data; The multi-party business risk assessment module is used to perform multi-party business risk assessment on each anchor point-related business subgraph using a pre-trained target business risk prediction model, and obtain the business risk assessment data of the target business object.
[0006] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned business risk analysis method for multi-party secure computation.
[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned business risk analysis method for multi-party secure computation.
[0008] In the above-mentioned solution implemented by the business risk analysis method, apparatus, computer equipment, and storage medium for multi-party secure computation, in response to a business risk analysis request for a target business object, at least two target participants' participant business graphs can be obtained; cross-party anchor point identification data used to locate anchor point nodes belonging to the same target business object in different participant business graphs can be obtained, and the cross-party anchor point identification data can be encrypted to obtain cross-party anchor point encrypted data; based on the cross-party anchor point encrypted data, anchor point-related business subgraphs of each target participant can be extracted from each participant business graph; finally, a multi-party business risk assessment of each anchor point-related business subgraph can be performed using a pre-trained target business risk prediction model to obtain business risk assessment data of the target business object. In this application, encrypted cross-party anchor data is obtained by encrypting the cross-party anchor identifier data. This enables secure cross-domain anchoring of multi-party data without exposing the original business object's identity, avoiding the data leakage risk caused by traditional methods that directly interact based on plaintext identity identifiers. Furthermore, by extracting the anchor-related business subgraphs of each party and inputting them into a pre-trained model for joint evaluation, the application achieves complete preservation and joint computation of the deep topology of the business networks of each participating party. It no longer relies solely on discrete features. Even in the face of hidden and complex business risks in multi-party scenarios, the application can capture the hidden relationships of the same user among different participating parties through the topological features between subgraphs. This effectively avoids the shortcomings of traditional discrete feature fusion, which struggles to capture the topological structure of the associated graph, leading to poor analytical reliability. This significantly improves the reliability of cross-platform joint business risk analysis while fully protecting the privacy of all parties' business data. Attached Figure Description
[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 This is a schematic diagram of an application environment for a business risk analysis method for multi-party secure computation according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a business risk analysis method for multi-party secure computation according to an embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation of step S104; Figure 4 This is another flowchart illustrating the business risk analysis method for multi-party secure computation in one embodiment of the present invention; Figure 5yes Figure 4 A schematic diagram of a specific implementation of step S301; Figure 6 yes Figure 5 A flowchart illustrating a specific implementation of step S404; Figure 7 yes Figure 4 A schematic diagram of a specific implementation of step S304; Figure 8 yes Figure 2 A schematic diagram of a specific implementation of step S105; Figure 9 This is a flowchart illustrating a business risk analysis method for multi-party secure computation according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 11 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation
[0011] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.
[0012] The business risk analysis method, apparatus, equipment, and medium for multi-party secure computation provided in this application are specifically described through the following embodiments. First, the business risk analysis method for multi-party secure computation in this application is described.
[0013] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0014] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.
[0015] The business risk analysis method for multi-party secure computation provided in this application relates to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing a policy claims data filtering method, but is not limited to the above forms.
[0016] Traditional business risk analysis methods typically match business data from different participants based on the plaintext identity of the target business entity. The matched data is then converted into structured features and used for risk prediction through secure aggregation. For example, in cross-platform insurance fraud analysis, to jointly identify whether an insured person is at risk of cross-platform fraud, the first participant (e.g., insurance institution A), the second participant (e.g., insurance institution B), and the third participant (e.g., insurance institution C) can directly use the insured person's real ID number or mobile phone number, among other plaintext identifiers, to perform cross-institutional data matching and alignment. After successful matching, features such as "historical number of medical visits" and "claim amount" are extracted and merged to perform a joint risk score for the insured person's insurance business risk. For example, in the identification of violations in healthcare scenarios, the first participant (e.g., Hospital D), the second participant (e.g., Hospital E), and the third participant (Pharmaceutical Platform F) typically perform alignment and matching among the three institutions (D, E, and F) based on plaintext identifiers such as the patient's real medical insurance card number or ID card number. They then extract features such as "historical number of visits" and "prescription frequency" to jointly score the patient's medical behavior risk. However, this method, which directly matches across parties based on plaintext identifiers, makes it difficult to perform multi-party joint calculations without exposing the original data. It also lacks privacy protection capabilities for institutional graph data, leading to data leakage risks during multi-party business risk analysis. Furthermore, this method only performs joint calculations on the discrete features of each participant, failing to capture the interconnected business graph topology relationships between them (such as the hidden relationships between the same user and different participants), resulting in poor reliability of business risk analysis. Therefore, how to achieve reliable cross-platform joint risk analysis while protecting the business data of all parties has become an urgent technical problem to be solved.
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] The business risk analysis method for multi-party secure computation provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server, in response to a business risk analysis request targeting a specific business object, can obtain the participant business graphs of at least two target participants; acquire cross-party anchor point identification data to characterize anchor nodes belonging to the same target business object within different participant business graphs, and encrypt this cross-party anchor point identification data to obtain encrypted cross-party anchor point data; based on this encrypted cross-party anchor point data, extract the anchor point-related business subgraphs of each target participant from each participant business graph; and finally, perform multi-party business risk assessment on each anchor point-related business subgraph using a pre-trained target business risk prediction model to obtain the business risk assessment data for the target business object. In this application, encrypted cross-party anchor data is obtained by encrypting the cross-party anchor identifier data. This enables secure cross-domain anchoring of multi-party data without exposing the original business object's identity, avoiding the data leakage risk caused by traditional methods that directly interact based on plaintext identity identifiers. Furthermore, by extracting the anchor-related business subgraphs of each party and inputting them into a pre-trained model for joint evaluation, the deep topology of the business networks of each participating party is fully preserved and jointly calculated. It no longer relies solely on discrete features. Even in the face of hidden and complex business risks in multi-party scenarios, the hidden relationships between the same user and different participating parties can be captured through the topological features between subgraphs. This effectively avoids the shortcomings of traditional discrete feature fusion, which struggles to capture the topological structure of the associated graph, leading to poor analytical reliability. This significantly improves the reliability of cross-platform joint business risk analysis while fully protecting the privacy of all parties' business data. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a separate server or a server cluster composed of multiple servers. The invention will be described in detail below through specific embodiments.
[0019] like Figure 2 As shown, Figure 2 A flowchart illustrating the business risk analysis method for multi-party secure computation provided in this embodiment of the invention may include, but is not limited to, steps S101 to S105: Step S101: In response to a business risk analysis request for the target business object, obtain the participant business map of at least two target participants.
[0020] Step S102: Obtain cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, which is used to locate the anchor point node belonging to the same target business object in the business graph of different participants.
[0021] Step S103: Encrypt the cross-anchor point identification data to obtain encrypted cross-anchor point data.
[0022] Step S104: Extract the anchor-related business sub-graphs of each target participant from the business graphs of each participant based on the cross-party anchor encrypted data.
[0023] Step S105: Use the pre-trained target business risk prediction model to perform multi-party business risk assessment on the business subgraphs associated with each anchor point, and obtain the business risk assessment data of the target business object.
[0024] Steps S101 to S105 as illustrated in this embodiment involve obtaining participant business graphs of at least two target participants in response to a business risk analysis request for a target business object; obtaining cross-party anchor point identification data to characterize anchor point nodes belonging to the same target business object in different participant business graphs, and encrypting the cross-party anchor point identification data to obtain encrypted cross-party anchor point data; extracting anchor point-related business subgraphs of each target participant from each participant business graph based on the encrypted cross-party anchor point data; and finally performing multi-party business risk assessment on each anchor point-related business subgraph using a pre-trained target business risk prediction model to obtain business risk assessment data for the target business object. In this application, encrypted cross-party anchor data is obtained by encrypting the cross-party anchor identifier data. This enables secure cross-domain anchoring of multi-party data without exposing the original business object's identity, avoiding the data leakage risk caused by traditional methods that directly interact based on plaintext identity identifiers. Furthermore, by extracting the anchor-related business subgraphs of each party and inputting them into a pre-trained model for joint evaluation, the application achieves complete preservation and joint computation of the deep topology of the business networks of each participating party. It no longer relies solely on discrete features. Even in the face of hidden and complex business risks in multi-party scenarios, the application can capture the hidden relationships of the same user among different participating parties through the topological features between subgraphs. This effectively avoids the shortcomings of traditional discrete feature fusion, which struggles to capture the topological structure of the associated graph, leading to poor analytical reliability. This significantly improves the reliability of cross-platform joint business risk analysis while fully protecting the privacy of all parties' business data.
[0025] In step S101 of some embodiments, specifically, in response to a business risk analysis request for a target business object, a participant business map of at least two target participants is obtained.
[0026] Specifically, the target business object refers to the natural person, business account, or business event to be subject to business risk analysis.
[0027] Specifically, target participants refer to institutions that hold business data related to the target business entity and participate in joint risk assessment.
[0028] Specifically, the participant business graph refers to graph structure data consisting of graph business nodes, node connection edges between graph business nodes, and corresponding business attributes. Graph business nodes are used to represent business objects, business events, or related business entities, and node connection edges are used to represent the business relationships between different entities.
[0029] For example, in an insurance scenario, the target business object can be the policyholder, the insured, or the claim applicant, and the target participants can include insurance institutions, medical institutions, and physical examination institutions. The participant business graph can include personnel nodes, policy nodes, claim event nodes, medical institution nodes, and their insurance relationships, claim relationships, and treatment relationships, etc. In a medical scenario, the target business object can be the patient, and the target participants can include different hospitals, primary healthcare institutions, and pharmaceutical service institutions. The participant business graph can include patient nodes, medical visit event nodes, diagnosis result nodes, doctor nodes, drug nodes, and their medical visit relationships, diagnosis relationships, and medication relationships, etc.
[0030] Specifically, a business risk analysis request can carry object information of the target business object, the type of business to be analyzed, and the scope of participants. In response to the business risk analysis request, at least two target participants can be identified based on the scope of participants, and the participant business map of each target participant can be obtained through a preset data interface, either pre-built or in real-time.
[0031] For example, in an insurance scenario, in response to a request for anti-fraud risk verification and analysis against a claims applicant, the insurance business map of different insurance institutions can be obtained; in a medical scenario, in response to a request for abnormal diagnosis and treatment risk analysis against a patient, the diagnosis and treatment business map held by different medical institutions can be obtained.
[0032] In this embodiment, by acquiring the participant business graphs of at least two target participants, the business data held by each target participant can be organized into graph structure data that retains entity attributes and entity relationships, providing a data foundation for subsequent cross-platform object location and topology risk analysis.
[0033] In step S102 of some embodiments, specifically, cross-party anchor identification data of the target business object is obtained; wherein, the cross-party anchor identification data is the identity identifier shared by the target business object among the target participants, and is used to locate the anchor node belonging to the same target business object in the business graph of different participants.
[0034] Specifically, cross-party anchor point identification data refers to the shared identity identifier of the target business object among various target participants. That is, the shared identity identifier information that can uniquely identify the same business object entity among different target participants. Anchor point node refers to the specific graph vertex object that the entity is mapped to in the corresponding graph data structure.
[0035] For example, in the insurance scenario, cross-party anchor identification data can be the ID card number used by the policyholder (i.e., the target business object) when registering, and the anchor node is the "user node" representing that ID card number in the insurance business graph of each insurance institution; in the medical scenario, cross-party anchor identification data can be the same ID card number registered by the patient when seeking medical treatment, and the anchor node is the "patient node" representing that patient in the medical business graph of each medical institution.
[0036] Specifically, the key identity field corresponding to the target business object can be extracted from the parameter context of the business risk analysis request, and this key identity field can be used as a unique key value for data alignment between different business domains, thereby determining the cross-party anchor point identification data.
[0037] In this embodiment, by obtaining the cross-anchor point identification data of the target business object, the unified standard key for node alignment between heterogeneous business graphs can be clearly identified, ensuring the accuracy and consistency of identifying the same business object across platforms.
[0038] In step S103 of some embodiments, specifically, the cross-anchor point identification data is encrypted to obtain cross-anchor point encrypted data.
[0039] Specifically, cross-anchor encrypted data refers to ciphertext identifiers that lose plaintext readability after cryptographic transformation but retain comparison equivalence.
[0040] Specifically, the cross-anchor identifier data can be encrypted using the SHA-256 hash algorithm to obtain encrypted cross-anchor data.
[0041] For example, in an insurance scenario, encryption can be performed using the SHA-256 hash algorithm, and the cross-anchor encrypted data can be a 64-bit hexadecimal string generated by hashing the insured's ID number; in a medical scenario, encryption is also performed using the SHA-256 hash algorithm, and the cross-anchor encrypted data is the same hexadecimal string converted from the patient's ID number.
[0042] In this embodiment, by encrypting the cross-party anchor point identification data, each target participant can locate the same target business object based on the encrypted cross-party anchor point data, thereby reducing the risk of the target business object's identity information being leaked or illegally associated during the cross-party node matching process.
[0043] In step S104 of some embodiments, specifically, anchor-related business sub-graphs of each target participant are extracted from the business graphs of each participant based on cross-party anchor encrypted data.
[0044] Specifically, the anchor-related business subgraph refers to the local network topology structure formed by radiating a certain number of neighboring nodes and their connecting edges outward from the anchor node.
[0045] For example, in a fintech scenario, an anchor-linked business subgraph can be a local relationship network of a participating insurance institution, including the policyholder's mobile phone number node and its associated "historical policy node", "claims case node", and "claims time edge"; similarly, in a healthcare scenario, an anchor-linked business subgraph can be a local relationship network of a participating hospital, including the patient's ID number node and its associated "attending physician node", "prescription node", and "visit frequency edge".
[0046] Specifically, in this map database, each target participant can locate the corresponding central node by performing encrypted index comparison based on cross-party anchor point encrypted data, and then perform a breadth-first search based on the central node to extract node and edge data within a preset depth range, generating the anchor point associated business sub-graph for that target participant.
[0047] In this embodiment, anchor-related business subgraphs of each target participant are extracted from the business graphs of each participant based on cross-party anchor-encrypted data. This enables the formation of local business networks related to the target business objects on each target participant's local premises. This reduces the involvement of irrelevant business data in the calculation while preserving the topology between the target business objects and related entities.
[0048] like Figure 3 As shown, in some embodiments, step S104 includes, but is not limited to, steps S201 to S203: Step S201: In the node index table preset by each target participant, perform a matching node query on the cross-party anchor point encrypted data to obtain matching node query data; wherein, the node index table is used to store the mapping relationship between the graph business nodes of each participant's business graph and the node encryption identifier of the graph business nodes.
[0049] Step S202: If the matching node query data is used to represent the successful matching between the cross-party anchor encrypted data and the graph business node, then the target business node corresponding to the cross-party anchor encrypted data is obtained.
[0050] Step S203: Using the target business node as the central node, perform a multi-hop topological traversal of the business graph of each participant to obtain the anchor point associated business subgraph of each target participant.
[0051] In step S201 of some embodiments, specifically, the node index table refers to a hash table or mapping dictionary that each target participant pre-establishes locally for quickly retrieving graph service nodes.
[0052] Specifically, the matching node query data is used to indicate whether the cross-party anchor encrypted data and the graph business node are successfully matched or not.
[0053] Specifically, user identity identifiers (such as mobile phone numbers and ID card numbers) in the business data of each target participant can be irreversibly encrypted using a secure hash algorithm (such as the SHA-256 algorithm), and the encrypted ciphertext identifiers can be bound to the corresponding node IDs in the constructed participant business graph to generate a node index table.
[0054] For example, in a cross-institutional joint claims fraud scenario, the target participants are insurance institutions A, B, and C. Each party pre-hashes and encrypts the ID numbers of all users, and stores the encrypted string as a relation to the "user identity node" in each insurance business graph in its local node index table. When receiving cross-party anchor encrypted data, each insurance institution directly performs a ciphertext comparison query in its index table. Similarly, in a medical misconduct analysis scenario, the target participants are hospitals, community health centers, and pharmaceutical platforms. Each party pre-hashes and encrypts the medical insurance card numbers or ID numbers of all patients, and stores the encrypted string as a relation to the "patient identity node" in each medical business graph in its local node index table. When receiving cross-party anchor encrypted data, each institution directly performs a ciphertext comparison query in its respective index table.
[0055] In step S202 of some embodiments, specifically, the target business node refers to a specific graph node in the participant business graph that successfully matches the cross-party anchor encrypted data of the target business object.
[0056] Specifically, when the matching node returns a successful query result, the node ID of the encrypted data mapping of the cross-anchor point in the node index table is read, thereby locating the target business node in the graph database.
[0057] In step S203 of some embodiments, specifically, multi-hop topology traversal refers to starting from a node and searching and expanding to the outer nodes along the edges in the graph structure at a preset depth (such as 1 hop, 2 hops, etc.).
[0058] Specifically, breadth-first search (BFS) or depth-first search (DFS) algorithms can be used to obtain all associated nodes and connecting edges within a preset number of hops, with the target business node as the central node, and extract the anchor point associated business subgraph to form the target participant.
[0059] For example, in a cross-institutional joint claims anti-fraud scenario, by performing a two-hop topological traversal with the "insured user's ID number node" as the central node, each participating insurance institution can extract an insurance-related business subgraph containing the "insured user's ID number," "associated transfer account node," "login device node," "IP address node," and "transfer relationship edge," as well as the connections between them. Similarly, in a joint violation identification scenario in healthcare, by performing a two-hop topological traversal with the "patient's ID number node" as the central node, each medical institution and pharmaceutical platform can simultaneously extract a medical-related business subgraph containing connections such as the "patient's ID number node," "attending physician node," "medical record node," "prescription node," and "prescription edge."
[0060] Through steps S201 to S203, the corresponding nodes of the target business object in the graphs of each platform can be accurately located without decrypting cross-party anchor data or revealing the real identities of business participants. By traversing the topology of multiple hops to extract local graphs, a data foundation containing rich business association structures is provided for subsequent multi-party joint graph calculations, breaking down information silos with structured features.
[0061] like Figure 4 As shown, in some embodiments, the business risk analysis of multi-party secure computation also includes, but is not limited to, steps S301 to S304: Step S301: Perform graph feature privacy processing on the business subgraph associated with each anchor point to obtain the target subgraph features.
[0062] Step S302: The target subgraph features are used to predict business risks through a preset original business risk prediction model to obtain the predicted business risk assessment data of the target business object in each target participant.
[0063] Step S303: Obtain the actual business risk assessment data of the target business object, and calculate the loss based on the predicted business risk assessment data and the actual business risk assessment data to obtain the target loss value of each target participant.
[0064] Step S304: Based on the target loss values of each target participant, perform multi-party reverse training on the original business risk prediction model to obtain the target business risk prediction model.
[0065] like Figure 5 As shown, in some embodiments, the anchor point associated service subgraph includes subgraph service nodes and node connection edges between subgraph service nodes; the target subgraph features include a fused node edge feature matrix and a privacy-enhanced node adjacency matrix, and step S301 includes, but is not limited to, steps S401 to S404: Step S401: Extract features from the subgraph business nodes to obtain the node feature matrix, and extract features from the node connecting edges to obtain the node connecting edge feature matrix.
[0066] Step S402: Perform feature fusion on the node feature matrix and the node connection edge feature matrix to obtain the fused node edge feature matrix.
[0067] Step S403: Construct an initial node adjacency matrix based on the subgraph business nodes and node connection edges; wherein, the initial node adjacency matrix is used to characterize the edge connection structure relationship between the subgraph business nodes.
[0068] Step S404: Perform differential privacy processing on the initial node adjacency matrix to obtain the privacy-preserving node adjacency matrix.
[0069] In step S401 of some embodiments, specifically, the node feature matrix refers to the matrix formed by arranging the node feature vectors of each subgraph business node according to a preset node order. This matrix can represent the node's own attribute information such as node type, business amount, number of events, time information, and business status.
[0070] Specifically, the node connection edge feature matrix refers to the matrix formed by arranging the edge feature vectors of each node connection edge according to a preset node connection edge order. This matrix can represent business interaction attribute information, including relationship type, association direction, occurrence time, association frequency, and relationship weight.
[0071] For example, in fintech scenarios, the node feature matrix of participating insurance institution A or B can represent the node's own attribute information such as user account balance, user credit score, and number of user logins, while the node connection edge feature matrix can represent business interaction attribute information such as transfer relationship type, transfer time, and transfer amount. In healthcare scenarios, the node feature matrix of hospitals or pharmaceutical platforms can represent the node's own attribute information such as treatment costs, diagnosis type, and patient age, while the node connection edge feature matrix can represent business interaction attribute information such as the type of medical relationship, prescription frequency, and treatment time.
[0072] Specifically, attribute fields (such as age, account balance, etc.) can be extracted from the subgraph business nodes and embedded to obtain the node feature matrix. Interaction attribute fields (such as transaction time, compensation amount, etc.) can be extracted from the node connection edges and embedded to obtain the node connection edge feature matrix.
[0073] In step S402 of some embodiments, specifically, the fused node edge feature matrix refers to the feature vector representation that combines the node feature matrix and the node connection edge feature matrix.
[0074] Specifically, a multilayer perceptron (MLP) can be used to map the feature dimensions of the node connection edge feature matrix to be consistent with the node feature matrix, and then a concatenation operation is performed to obtain a node edge feature matrix that integrates the information of the nodes and the node connection edges.
[0075] In step S403 of some embodiments, specifically, the initial node adjacency matrix refers to a matrix constructed based on whether there are node connection edges between the subgraph business nodes in the anchor-associated business subgraph. That is, it is composed of Boolean values of "0" and "1" and is used to reflect whether there is a real connection relationship between the subgraph business nodes in the anchor-associated business subgraph. "0" indicates that there is no real connection relationship between the subgraph business nodes and "1" indicates that there is a real connection relationship between the subgraph business nodes.
[0076] Specifically, traverse the business nodes of the subgraph. If there is a node connection edge between two nodes, set the element at the corresponding position of the initial node adjacency matrix to 1 or the corresponding edge weight; if there is no edge connection edge, set the element at the corresponding position of the matrix to 0.
[0077] like Figure 6 As shown, in some embodiments, step S404 includes, but is not limited to, steps S501 to S504: Step S501: Obtain the matrix elements of the initial node adjacency matrix.
[0078] Step S502: Obtain the element flipping probability or element retention probability of the matrix element according to the preset element random response logic.
[0079] Step S503: Based on the element flipping probability, the matrix elements are flipped to obtain flipped matrix elements, and based on the element retention probability, the matrix elements are retained to obtain retained matrix elements.
[0080] Step S504: Determine the privacy-preserving node adjacency matrix based on flipped matrix elements and preserved matrix elements.
[0081] In step S501 of some embodiments, specifically, the matrix element refers to the value in the initial node adjacency matrix used to characterize whether there is a node connection edge between two subgraph business nodes.
[0082] Specifically, the matrix elements in the initial node adjacency matrix can be read sequentially according to the matrix row index and column index, and each matrix element can be used as a pending element for random response processing.
[0083] For example, in an insurance scenario, matrix elements can be used to represent whether there is an insurance relationship between the policyholder and the policy; in a medical scenario, matrix elements can be used to represent whether there is a prescription relationship between the patient and the prescription drug.
[0084] In step S502 of some embodiments, specifically, the element random response logic is a probabilistic perturbation algorithm based on a differential privacy protection mechanism, used to introduce random noise into the real data.
[0085] Specifically, the element flipping probability is used to characterize the probability of changing a matrix element from a first preset value (e.g., 1) to a second preset value (e.g., 0) or from a second preset value to a first preset value, while the element retention probability is used to characterize the probability of maintaining the original value of the matrix element.
[0086] Specifically, the probability of preserving the actual connected edges (i.e., the probability of preserving elements) p refers to the matrix elements of the adjacency matrix of the initial node. The perturbation matrix elements obtained after element random response processing under the condition of 1. The conditional probability is still 1, i.e., p = P( =1| =1), the non-edge addition probability (i.e., element flipping probability) q refers to the matrix element in the adjacency matrix of the initial node. The perturbation matrix elements obtained after element random response processing under the condition that the value is 0. The conditional probability that is 1, i.e., q=P( =1| =0). Therefore, when matrix elements When =1, it is kept at 1 with probability p, and flipped to 0 with probability 1-p; when matrix elements When =0, it is flipped to 1 with probability q, and kept at 0 with probability 1-q.
[0087] Where p and q satisfy the following privacy budget formula:
[0088] in, This represents the privacy budget, used to constrain the strength of privacy protection for edges connecting nodes. A smaller privacy budget value indicates stronger privacy protection but lower usability. represents the probability of flipping an element, and q represents the probability of keeping an element.
[0089] For example, in an insurance scenario, if the privacy budget for an insurance relationship is ε=1, the element retention probability p is approximately 0.73 and the element flip probability q is approximately 0.27, determined by the element random response logic. When a matrix element is 1, it remains 1 with a probability of 0.73; when a matrix element is 0, it flips to 1 with a probability of 0.27. In a healthcare scenario, a smaller privacy budget ε=0.5 can be determined for highly sensitive diagnostic relationships, with an element retention probability p of approximately 0.62 and an element flip probability q of approximately 0.38. When a matrix element is 1, it remains 1 with a probability of 0.62; when a matrix element is 0, it flips to 1 with a probability of 0.38. Therefore, the degree of random perturbation of the matrix elements corresponding to diagnostic relationships is higher than that of insurance relationships, thereby improving the privacy protection strength of sensitive medical relationships.
[0090] In step S503 of some embodiments, specifically, the flipping process refers to the process of changing a matrix element from 0 to 1 or from 1 to 0, and the preservation process refers to the process of maintaining the original value of the matrix element.
[0091] For example, in a fintech scenario, for matrix element 1 on the insurance institution A's side, which represents the login association between user nodes and high-risk IP nodes, the element can be kept at a probability p of 1, or at a probability of 1 p is flipped to 0; In healthcare scenarios, for matrix element 0 used by participating hospitals to represent the absence of a diagnostic relationship between patient nodes and specific critical illness indicator nodes, it can be flipped to 1 according to the element flipping probability q set by the medical side, or according to probability 1. q remains at 0.
[0092] In step S504 of some embodiments, specifically, the privacy-preserving node adjacency matrix refers to the perturbed adjacency matrix obtained by performing random response processing on each matrix element in the initial node adjacency matrix.
[0093] Specifically, the corresponding matrix elements can be written into the corresponding matrix positions according to the row and column indices of each flip matrix element and each maintain matrix element in the initial node adjacency matrix, so as to obtain the privacy-preserving node adjacency matrix.
[0094] Furthermore, when the initial node adjacency matrix is used to represent an undirected service graph, the matrix elements can be... Matrix elements at positions symmetrical to it As a single processing unit, random response processing is performed, and the same processing results are written to their corresponding symmetrical positions to maintain the symmetry of the privacy-preserving node adjacency matrix. When the initial node adjacency matrix is used to represent a directed service graph, the matrix elements can be processed separately. and matrix elements Perform random response processing to preserve the directional representation of node connection edges. For the diagonal elements of the initial node adjacency matrix, the original values can be maintained according to the configuration requirements of self-connection edges, or they can be set to preset diagonal values.
[0095] Through steps S501 to S504, the matrix elements in the initial node adjacency matrix are subjected to probabilistic flipping or preservation processing according to the preset element random response logic to obtain a privacy-preserving node adjacency matrix. This allows each target participant to complete graph convolution message transmission without exposing the original edge relationships of the graph, while preserving the overall topological statistical features of the anchor-related business subgraph. This effectively reduces the risk of a single node connection edge being identified or inferred, ensuring that attackers cannot reverse-engineer specific business objects from the exported subgraph. This protects the topological privacy of the graph while maintaining the statistical characteristics of the graph structure.
[0096] Through steps S401 to S404, node and edge features are deeply fused to enrich feature representation. At the same time, differential privacy technology is combined to transform the deterministic topology into a privacy-oriented node adjacency matrix. This achieves high-quality graph feature extraction while ensuring the privacy and security of the underlying graph structure data of each participating party, providing reliable and compliant graph data features for subsequent secure aggregation.
[0097] In step S302 of some embodiments, specifically, the original business risk prediction model refers to a local graph neural network model deployed locally by each target participant without being sufficiently trained by multiple parties.
[0098] Specifically, the predicted business risk assessment data refers to the predicted risk score, predicted risk level, or predicted risk label output by the original business risk prediction model based on the characteristics of the target subgraph.
[0099] Specifically, the node edge feature matrix and privacy node adjacency matrix corresponding to each target participant can be input into the original business risk prediction model. The node feature aggregation range is determined by the privacy node adjacency matrix, and multi-level graph feature aggregation is performed on the node edge feature matrix to obtain the predicted risk score of the target business object in each target participant.
[0100] In step S303 of some embodiments, specifically, real business risk assessment data refers to risk labels generated based on historical business results, manual review results, or confirmed risk events.
[0101] Specifically, the target loss value refers to the numerical value used to characterize the difference between the predicted business risk assessment data and the actual business risk assessment data.
[0102] Specifically, the cross-entropy loss function can be used to calculate the target loss value for each target participant by comparing the predicted business risk assessment data of each target participant with the corresponding actual business risk assessment data.
[0103] like Figure 7 As shown, in some embodiments, step S304 includes, but is not limited to, steps S601 to S604: Step S601: Based on the target loss value of each target participant, perform back gradient calculation on the original model weight parameters of the original business risk prediction model to obtain the model weight gradient.
[0104] Step S602: Encrypt the model weight gradient to obtain the encrypted model weight gradient of each target participant.
[0105] Step S603: Perform a weighted calculation based on the encryption model weight gradient of each target participant to obtain the global model gradient of all target participants.
[0106] Step S604: Train the weight parameters of the original business risk prediction model based on the global model gradient to obtain the target business risk prediction model.
[0107] In step S601 of some embodiments, specifically, the original model weight parameters refer to the trainable parameters in the original business risk prediction model used to perform feature transformation, feature aggregation, and risk scoring.
[0108] Specifically, the model weight gradient refers to the partial derivative of the target loss value with respect to the original model weight parameters.
[0109] Specifically, each target participant can perform error backpropagation locally based on the target loss value, and use the chain rule to calculate the gradient of the target loss value relative to the original model weight parameters of each layer, thus obtaining the model weight gradient.
[0110] For example, in a cross-institutional joint claims fraud scenario, after insurance institutions A, B, and C calculate the target loss value based on the cross-party joint prediction results, they can independently calculate their local model weight gradients based on their respective local insurance business graphs. Similarly, in a joint violation identification scenario in the healthcare industry, after participating hospitals, community health centers, and pharmaceutical platforms calculate the target loss value based on the joint prediction results, they can independently calculate their local model weight gradients based on their respective local healthcare business graphs.
[0111] This embodiment avoids the direct transmission of anchor-related service subgraphs and their original service data by performing reverse gradient calculations locally on each target participant.
[0112] In step S602 of some embodiments, specifically, the encrypted model weight gradient refers to the ciphertext gradient obtained after encrypting the model weight gradient using a preset encryption algorithm.
[0113] Specifically, each participating party can use the aggregated public key to homomorphically encrypt the model weight gradient, or split the model weight gradient into multiple secret shares and send the encrypted model weight gradient or secret shares to the secure aggregation node.
[0114] For example, in a cross-institutional joint claims fraud prevention scenario, insurance institutions A, B, and C can each use the public key issued by the multi-party secure computation coordination node to homomorphically encrypt the local model weight gradient, ensuring that when the insurance data gradient of each party goes out of domain, it is not used by other insurance institutions to reverse-engineer the original business data. As another example, in joint violation identification in the healthcare scenario, participating hospitals, community health centers, and pharmaceutical platforms can each perform secret sharing processing on the local model weight gradient, thereby completing the interaction of encrypted gradients while protecting patient privacy.
[0115] This embodiment reduces the risk of deducing participant business data from plaintext gradients during model training by encrypting the model weight gradients.
[0116] In step S603 of some embodiments, specifically, the global model gradient refers to the joint gradient obtained by integrating the weight gradients of all target participants' models.
[0117] Specifically, the gradient aggregation weights for each target participant can be determined based on the number of training samples, data quality, or preset participant weights. The federated learning coordination server receives the encrypted model weight gradients of each target participant and performs summation or weighted average calculations directly in the encrypted state to obtain the global model gradients of all target participants.
[0118] This embodiment calculates the weight gradient of the encrypted model by weighting it, which can integrate the training information of each target participant and avoid the security aggregation node from obtaining the plaintext model weight gradient of a single target participant.
[0119] In step S604 of some embodiments, specifically, the target business risk prediction model refers to a graph neural network model that is locally deployed by each target participant and has been fully trained jointly by multiple parties.
[0120] Specifically, after obtaining the global model gradient, each participating party decrypts it using its private key and uses the decrypted global gradient to update the weight parameters of its local original business risk prediction model, completing one round of multi-party reverse training. If the target loss value of the updated model does not meet the preset convergence condition (e.g., less than 0.7) or the number of training iterations reaches the preset number (e.g., 10), the original model weight parameters of the original business risk prediction model are iteratively updated. When the model loss value meets the preset convergence condition or the number of training iterations reaches the preset number, the updated original business risk prediction model is determined as the target business risk prediction model.
[0121] This embodiment updates the model weight parameters through global model gradient, enabling joint model training without aggregating the original business data of each target participant.
[0122] Through steps S601 to S604, by performing encrypted interaction and global aggregation of the model gradients of each target participant in the backpropagation process, the aggregation weight of the federated model is adaptively adjusted according to the data quality and contribution of each participant. This avoids low-quality data from polluting the global model and effectively ensures that the original business subgraph data and loss results of each target participant do not leave the domain, thereby avoiding the risk of data quality degradation and reverse data leakage during model training.
[0123] Through steps S301 to S304, a multi-party joint training based on graph feature privacy protection and federated gradients can be fully constructed. Differential privacy protects the graph features of each participant, and encrypted aggregation protects the model parameter output of each participant. Multi-party joint model training can be completed without directly exchanging the original business graphs of each target participant, and the ability of the target business risk prediction model to identify cross-party related risks can be improved.
[0124] In step S105 of some embodiments, specifically, a multi-party business risk assessment is performed on the business subgraph associated with each anchor point through a pre-trained target business risk prediction model to obtain business risk assessment data of the target business object.
[0125] Specifically, business risk assessment data refers to the probability values or classification labels output by the model to quantify the degree of risk.
[0126] Specifically, each target participant can input the extracted anchor point-related business subgraph into the pre-trained target business risk prediction model. After extracting local graph features through the model's graph convolutional layer, the model performs encrypted aggregation of cross-party features and classification calculations through fully connected layers under a secure multi-party computation protocol, ultimately outputting the business risk assessment data for the target business object.
[0127] For example, in cross-institutional joint claims fraud prevention scenarios, the model processes the local business subgraphs of each insurance institution to output local feature vectors in the financial domain. After encrypted aggregation, it outputs a cross-platform fraud risk score for the user. In joint violation identification in the medical and health scenario, the model processes the business subgraphs of each hospital, community health center, and pharmaceutical platform to output local feature vectors in the medical domain. After encrypted aggregation, it outputs a risk score for the patient's possible illegal medication use or medical insurance fraud.
[0128] like Figure 8 As shown, in some embodiments, step S105 includes, but is not limited to, steps S701 to S705: Step S701: Perform matrix transformation on the edge feature matrix of the fusion node using the target model weight parameters of the target business risk prediction model to obtain the transformed feature matrix.
[0129] Step S702: Obtain the target business node corresponding to the cross-party anchor encrypted data, and query the neighbor node features of the target business node from the transformation feature matrix based on the privacy node adjacency matrix, and aggregate the neighbor node features to obtain the neighbor context features. Step S703: Extract features from the target business node to obtain the target business node features, and fuse the target business node features with the neighbor context features to obtain the target fusion features of each target participant.
[0130] Step S704: Perform feature fusion on the target fusion features of each target participant to obtain cross-party global features.
[0131] Step S705: Perform business risk scoring on cross-party global features to obtain business risk assessment data for the target business object.
[0132] In step S701 of some embodiments, specifically, the target model weight parameters refer to the model parameters obtained after the target business risk prediction model has undergone multi-party reverse training.
[0133] Specifically, the transformed feature matrix refers to the matrix obtained by mapping the node edge feature matrix to a preset feature space using the target model weight parameters.
[0134] Specifically, the node edge feature matrix can be multiplied with the target model weight parameters, and the result of the multiplication can be subjected to bias superposition and nonlinear activation processing to obtain the transformed feature matrix.
[0135] For example, in fintech scenarios, insurance institution A or insurance institution B can map the node edge features of users, transfer accounts, and login devices to insurance risk features of a unified dimension; in healthcare scenarios, hospitals or pharmaceutical platforms can map the node edge features of patients, diagnosis results, and drugs to medical violation risk features of a unified dimension.
[0136] This embodiment transforms the node edge feature matrix, enabling the conversion of different types of business features into feature representations that can be aggregated for graph features.
[0137] In step S702 of some embodiments, specifically, a neighboring node refers to a subgraph service node that is directly or indirectly connected to the target service node through a node connection edge.
[0138] Specifically, neighbor context features refer to the features obtained by aggregating the features of neighbor nodes, which are used to characterize the local business environment in which the target business node is located.
[0139] Specifically, based on the node index of the target business node, the corresponding matrix row or column can be read from the privacy node adjacency matrix, and the neighbor node index can be determined based on the matrix elements that meet the preset connection conditions. Then, based on the neighbor node index, the neighbor node features can be read from the transformation feature matrix, and aggregated by summation, mean, maximum value or attention weighting to obtain the neighbor context features.
[0140] For example, in a fintech scenario, insurance institution A can locally aggregate features of transfer account nodes, login device nodes, and IP address nodes associated with user nodes; in a healthcare scenario, participating hospitals can locally aggregate features of consultation department nodes, diagnosis result nodes, and prescription drug nodes associated with patient nodes.
[0141] This embodiment can extract topological context information between the target business object and surrounding related business entities by aggregating the features of neighboring nodes.
[0142] In step S703 of some embodiments, specifically, the target service node feature refers to the feature vector corresponding to the target service node in the transformation feature matrix.
[0143] Specifically, target fusion features refer to the local risk feature vector representation obtained by fusing the target business node's own attribute features with the neighbor's context features.
[0144] Specifically, the target business node features can be read from the transformation feature matrix based on the node index of the target business node, and the target business node features can be concatenated with the neighbor context features to obtain the target fusion features of the corresponding target participants.
[0145] This embodiment, by fusing the characteristics of the target service node and the contextual characteristics of its neighbors, can simultaneously retain the risk information of the target service object's own attributes and its associated service network.
[0146] In step S704 of some embodiments, specifically, the cross-party global feature refers to the cross-party risk feature vector representation of the target business object obtained by fusing the target fusion features of each target participant.
[0147] Specifically, each target participant can homomorphically encrypt, secretly share, or securely encode the locally generated target fusion features, and perform concatenation, weighted summation, or attention fusion on the target fusion features of different target participants through a secure aggregation protocol to obtain cross-party global features.
[0148] For example, in fintech scenarios, secure multi-party computation protocols can be used to encrypt and fuse the local financial risk features of the same user generated by insurance institutions A, B, and C respectively, generating cross-party global features. Similarly, in healthcare scenarios, secure multi-party computation protocols can be used to encrypt and fuse the local medical risk features of the same patient generated by hospitals, community health centers, and pharmaceutical platforms respectively, generating cross-party global features.
[0149] This embodiment integrates the target fusion features of each target participant to form a global risk representation of target business objects covering different business data domains, while avoiding the direct exchange of the original business data of each target participant.
[0150] In step S705 of some embodiments, specifically, business risk scoring refers to the process of mapping cross-party global features into business risk assessment data using the risk scoring layer of the target business risk prediction model.
[0151] Specifically, cross-regional global features can be input into a fully connected layer for linear transformation, and then the risk score can be obtained through the Sigmoid function. Based on a preset risk threshold, the corresponding risk level and risk warning result can be determined.
[0152] For example, in fintech scenarios, when the joint anti-fraud risk score calculated based on cross-domain global features exceeds a preset threshold, the user can be identified as a high-risk individual and a risk warning can be generated. This effectively identifies group operation risks or fraud risks that appear normal within a single insurance institution but have logical conflicts across platforms (such as dispersed transfers into institution A and concentrated transfers out of institution B). Similarly, in healthcare scenarios, when the joint violation risk score calculated based on cross-domain global features exceeds a preset threshold, the patient can be identified as a high-risk individual and a warning can be generated. This effectively identifies illegal medication use or medical insurance fraud risks that appear normal within a single hospital but have logical conflicts across hospitals (such as repeated prescriptions of the same limited-quantity prescription drugs in hospitals A and B within a short period).
[0153] This embodiment can convert multi-party graph features into business risk assessment data that can be used for business review or risk control by scoring cross-party global features for business risk assessment.
[0154] Through steps S701 to S705, while protecting the original business data and node connection structure of each target participant, the system comprehensively preserves and jointly calculates the deep topology of the business network of each participant by integrating the target business object's own attributes, local topological context, and cross-participant risk information. It no longer relies solely on discrete features. Even when facing hidden and complex business risks in multi-party scenarios, it can capture the hidden relationships of the same user among different participants through the topological features between subgraphs. This effectively avoids the shortcomings of traditional discrete feature fusion, which makes it difficult to capture the topological structure of the associated graph, resulting in poor analysis reliability. Under the premise of fully protecting the privacy of business data of all parties, it significantly improves the reliability of cross-platform joint business risk analysis, thereby improving the reliability of business risk assessment.
[0155] As can be seen, in the above scheme, in response to a business risk analysis request for the target business object, at least two target participants' business graphs are obtained; cross-party anchor point identification data is obtained to characterize anchor point nodes belonging to the same target business object in different participants' business graphs, and this cross-party anchor point identification data is encrypted to obtain cross-party anchor point encrypted data; based on this cross-party anchor point encrypted data, anchor point-related business subgraphs of each target participant are extracted from each participant's business graph; finally, a pre-trained target business risk prediction model is used to perform multi-party business risk assessment on each anchor point-related business subgraph to obtain business risk assessment data for the target business object. In this application, encrypted cross-party anchor data is obtained by encrypting the cross-party anchor identifier data. This enables secure cross-domain anchoring of multi-party data without exposing the original business object's identity, avoiding the data leakage risk caused by traditional methods that directly interact based on plaintext identity identifiers. Furthermore, by extracting the anchor-related business subgraphs of each party and inputting them into a pre-trained model for joint evaluation, the application achieves complete preservation and joint computation of the deep topology of the business networks of each participating party. It no longer relies solely on discrete features. Even in the face of hidden and complex business risks in multi-party scenarios, the application can capture the hidden relationships of the same user among different participating parties through the topological features between subgraphs. This effectively avoids the shortcomings of traditional discrete feature fusion, which struggles to capture the topological structure of the associated graph, leading to poor analytical reliability. This significantly improves the reliability of cross-platform joint business risk analysis while fully protecting the privacy of all parties' business data.
[0156] Please see Figure 9This application also provides a business risk analysis apparatus for multi-party secure computation, which can implement the above-mentioned business risk analysis method for multi-party secure computation. The apparatus includes: The participant business graph acquisition module is used to obtain the participant business graphs of at least two target participants in response to a business risk analysis request for a target business object. The cross-party anchor point identification data module is used to obtain the cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, and is used to locate the anchor point node belonging to the same target business object in the business graph of different participants; The encryption processing module is used to encrypt the cross-anchor point identification data to obtain encrypted cross-anchor point data. The anchor point association business subgraph extraction module is used to extract the anchor point association business subgraph of each target participant from the business graph of each participant based on the cross-party anchor point encrypted data. The multi-party business risk assessment module is used to conduct multi-party business risk assessments on the business subgraphs associated with each anchor point through a pre-trained target business risk prediction model, and obtain business risk assessment data for the target business object.
[0157] In one embodiment, the anchor point associated business subgraph extraction module is specifically used for: In the node index table preset by each target participant, the cross-party anchor point encrypted data is matched with node query to obtain the matching node query data; the node index table is used to store the mapping relationship between the graph business nodes of each participant's business graph and the node encryption identifier of the graph business nodes. If the matching node query data is used to represent that the cross-party anchor encrypted data and the graph business node are successfully matched, then the target business node corresponding to the cross-party anchor encrypted data is obtained. Using the target business node as the central node, a multi-hop topological traversal is performed on the business graphs of each participant to obtain the anchor point associated business subgraphs of each target participant.
[0158] In one embodiment, the business risk analysis apparatus for multi-party secure computation further includes: The graph feature privacy processing module is used to perform graph feature privacy processing on the business subgraphs associated with each anchor point to obtain the target subgraph features. The subgraph business risk prediction module is used to predict the business risks of the target subgraph features through a preset original business risk prediction model, and obtain the predicted business risk assessment data of the target business object in each target participant. The loss calculation module is used to obtain the actual business risk assessment data of the target business object, and to perform loss calculation based on the predicted business risk assessment data and the actual business risk assessment data to obtain the target loss value of each target participant. The multi-party reverse training module is used to perform multi-party reverse training on the original business risk prediction model based on the target loss values of each target participant, so as to obtain the target business risk prediction model.
[0159] In one embodiment, if the anchor-associated service subgraph includes subgraph service nodes and node connection edges between subgraph service nodes; and the target subgraph features include a fused node edge feature matrix and a privacy-enhanced node adjacency matrix, then the graph feature privacy processing module is specifically used for: The feature extraction submodule is used to extract features from the business nodes of the subgraph to obtain the node feature matrix, and to extract features from the connecting edges of the nodes to obtain the connecting edge feature matrix. The feature fusion submodule is used to fuse the node feature matrix and the node connection edge feature matrix to obtain the fused node edge feature matrix. The adjacency matrix construction submodule is used to construct an initial node adjacency matrix based on the subgraph business nodes and node connection edges; wherein, the initial node adjacency matrix is used to represent the edge connection structure relationship between the business nodes in the subgraph; The differential privacy processing submodule is used to perform differential privacy processing on the initial node adjacency matrix to obtain a privacy-preserving node adjacency matrix.
[0160] In one embodiment, the differential privacy processing submodule is specifically used for: Get the elements of the initial node adjacency matrix; The element flipping probability or element retention probability of the matrix element is obtained according to the preset element random response logic; The matrix elements are flipped based on the element flipping probability to obtain the flipped matrix elements, and the matrix elements are kept based on the element keeping probability to obtain the kept matrix elements. The privacy-preserving node adjacency matrix is determined by flipping and preserving matrix elements.
[0161] In one embodiment, the multi-party reverse training module is specifically used for: Based on the target loss values of each target participant, the inverse gradient of the original model weight parameters of the original business risk prediction model is calculated to obtain the model weight gradient. The model weight gradients are encrypted to obtain the encrypted model weight gradients of each target participant. The global model gradient of all target participants is obtained by weighted calculation based on the encryption model weight gradient of each target participant. The target business risk prediction model is obtained by training the weight parameters of the original business risk prediction model based on the global model gradient.
[0162] In one embodiment, the multi-party business risk assessment module is specifically used for: The transformed feature matrix is obtained by performing matrix transformation on the edge feature matrix of the fusion node using the target model weight parameters of the target business risk prediction model. Obtain the target business node corresponding to the cross-party anchor encrypted data, and query the neighbor node features of the target business node from the transformation feature matrix based on the privacy node adjacency matrix, and aggregate the neighbor node features to obtain the neighbor context features. Feature extraction is performed on the target business node to obtain the target business node features, and the target business node features and neighbor context features are fused to obtain the target fusion features of each target participant; The target fusion features of each target participant are fused to obtain cross-party global features; Business risk scores are calculated based on cross-party global features to obtain business risk assessment data for the target business object.
[0163] Specific limitations regarding the business risk analysis device for multi-party secure computation can be found in the limitations of the business risk analysis method for multi-party secure computation mentioned above, and will not be repeated here. Each module in the aforementioned business risk analysis device for multi-party secure computation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0164] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements a multi-party secure computation business risk analysis method or the server-side functions or steps of a multi-party secure computation business risk analysis method.
[0165] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements a multi-party secure computation business risk analysis method or the client-side functions or steps of a multi-party secure computation business risk analysis method.
[0166] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: In response to a business risk analysis request for a target business object, obtain the participant business graph of at least two target participants; Obtain cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, which is used to locate the anchor point node belonging to the same target business object in the business graph of different participants; The cross-anchor point identification data is encrypted to obtain encrypted cross-anchor point data. Based on the cross-party anchor point encrypted data, the anchor point associated business sub-graphs of each target participant are extracted from the business graphs of each participant. By using a pre-trained target business risk prediction model, multi-party business risk assessments are conducted on the business subgraphs associated with each anchor point to obtain business risk assessment data for the target business object.
[0167] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: In response to a business risk analysis request for a target business object, obtain the participant business graph of at least two target participants; Obtain cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, which is used to locate the anchor point node belonging to the same target business object in the business graph of different participants; The cross-anchor point identification data is encrypted to obtain encrypted cross-anchor point data. Based on the cross-party anchor point encrypted data, the anchor point associated business sub-graphs of each target participant are extracted from the business graphs of each participant. By using a pre-trained target business risk prediction model, multi-party business risk assessments are conducted on the business subgraphs associated with each anchor point to obtain business risk assessment data for the target business object.
[0168] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0169] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0170] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0171] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.
[0172] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A business risk analysis method for multi-party secure computation, characterized in that, include: In response to a business risk analysis request for a target business object, obtain the participant business graph of at least two target participants; Obtain cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, and is used to locate anchor point nodes belonging to the same target business object in the business graphs of different participants; The cross-anchor point identification data is encrypted to obtain encrypted cross-anchor point data; Based on the cross-party anchor point encrypted data, the anchor point associated business sub-graphs of each target participant are extracted from the business graphs of each participant. The pre-trained target business risk prediction model is used to conduct multi-party business risk assessments on the business subgraphs associated with each anchor point, thereby obtaining the business risk assessment data of the target business object.
2. The method as described in claim 1, characterized in that, Before obtaining the business risk assessment data of the target business object by performing multi-party business risk assessment on each anchor-related business subgraph using a pre-trained target business risk prediction model, the method further includes: Graph feature privacy processing is performed on the business subgraphs associated with each anchor point to obtain the target subgraph features; The target subgraph features are used to predict business risks through a preset original business risk prediction model, thereby obtaining the predicted business risk assessment data of the target business object in each of the target participants. Obtain the actual business risk assessment data of the target business object, and calculate the loss based on the predicted business risk assessment data and the actual business risk assessment data to obtain the target loss value of each target participant; The original business risk prediction model is back-trained by multiple parties based on the target loss values of each target participant to obtain the target business risk prediction model.
3. The method as described in claim 2, characterized in that, The anchor point associated service subgraph includes subgraph service nodes and node connection edges between the subgraph service nodes; the target subgraph features include a fused node edge feature matrix and a privacy-preserving node adjacency matrix; The step of performing graph feature privacy processing on the anchor-related business subgraphs of each of the target participants to obtain target subgraph features includes: Feature extraction is performed on the subgraph service nodes to obtain a node feature matrix, and feature extraction is performed on the node connecting edges to obtain a node connecting edge feature matrix; The node feature matrix and the node connection edge feature matrix are fused to obtain the fused node edge feature matrix; An initial node adjacency matrix is constructed based on the subgraph service nodes and the node connecting edges; wherein, the initial node adjacency matrix is used to characterize the edge connection structure relationship between the subgraph service nodes; Differential privacy processing is performed on the initial node adjacency matrix to obtain the privacy-enhanced node adjacency matrix.
4. The method as described in claim 3, characterized in that, The step of performing differential privacy processing on the initial node adjacency matrix to obtain the privacy-enhanced node adjacency matrix includes: Obtain the matrix elements of the initial node adjacency matrix; The element flipping probability or element retention probability of the matrix element is obtained according to the preset element random response logic; The matrix elements are flipped based on the element flipping probability to obtain flipped matrix elements, and the matrix elements are kept based on the element keeping probability to obtain kept matrix elements. The privacy-preserving node adjacency matrix is determined based on the flipped matrix elements and the preserved matrix elements.
5. The method as described in claim 2, characterized in that, The process of performing multi-party back-training on the original business risk prediction model based on the target loss values of each of the target participants to obtain the target business risk prediction model includes: Based on the target loss value of each target participant, the inverse gradient calculation is performed on the original model weight parameters of the original business risk prediction model to obtain the model weight gradient; The model weight gradient is encrypted to obtain the encrypted model weight gradient of each of the target participants; The global model gradient of all target participants is obtained by weighted calculation based on the encryption model weight gradient of each target participant. The original business risk prediction model is trained with weight parameters based on the global model gradient to obtain the target business risk prediction model.
6. The method as described in claim 3, characterized in that, The step of performing multi-party business risk assessment on the anchor-related business subgraph using a pre-trained target business risk prediction model to obtain business risk assessment data for the target business object includes: The transformed feature matrix is obtained by performing matrix transformation on the edge feature matrix of the fusion node using the target model weight parameters of the target business risk prediction model. Obtain the target business node corresponding to the cross-party anchor encrypted data, and query the neighbor node features of the target business node from the transformation feature matrix based on the privacy node adjacency matrix, and aggregate the neighbor node features to obtain the neighbor context features; Feature extraction is performed on the target service node to obtain the target service node features, and the target service node features and the neighbor context features are fused to obtain the target fusion features of each target participant; The target fusion features of each of the target participants are fused to obtain cross-party global features; The cross-party global features are used to score business risks, thereby obtaining business risk assessment data for the target business object.
7. The method as described in claim 6, characterized in that, The step of extracting the anchor-related business sub-graph of each target participant from the business graph of each participant based on the cross-party anchor encrypted data includes: In the node index table preset by each of the target participants, the cross-party anchor point encrypted data is matched with node query data; wherein, the node index table is used to store the mapping relationship between the graph business nodes of each participant's business graph and the node encryption identifier of the graph business nodes; If the matching node query data is used to indicate that the cross-party anchor encrypted data is successfully matched with the graph service node, then the target service node corresponding to the cross-party anchor encrypted data is obtained; Using the target business node as the central node, a multi-hop topology traversal is performed on the business graphs of each participant to obtain the anchor point associated business subgraph of each target participant.
8. A business risk analysis device for multi-party secure computation, characterized in that, include: The participant business graph acquisition module is used to obtain the participant business graphs of at least two target participants in response to a business risk analysis request for a target business object. The cross-party anchor point identification data module is used to obtain the cross-party anchor point identification data of the target business object; wherein, the cross-party anchor point identification data is the identity identifier shared by the target business object among the target participants, and is used to locate the anchor point node belonging to the same target business object in the business graph of different participants; An encryption processing module is used to encrypt the cross-anchor point identification data to obtain encrypted cross-anchor point data. Anchor point associated business subgraph extraction module is used to extract the anchor point associated business subgraph of each target participant from the business graph of each participant based on the cross-party anchor point encrypted data; The multi-party business risk assessment module is used to perform multi-party business risk assessment on each anchor point-related business subgraph using a pre-trained target business risk prediction model, and obtain the business risk assessment data of the target business object.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the business risk analysis method for multi-party secure computation as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the business risk analysis method for multi-party secure computation as described in any one of claims 1 to 7.
