An intelligent risk control system based on blockchain

Abstract

The present invention discloses an intelligent risk control system based on blockchain, which relates to the field of financial technology and includes: a data integration module that integrates multiple data sources in real time and uses a hash algorithm to encrypt and process transaction data. A graph neural network model construction module models the dependency relationship between nodes through graph convolution operations to identify potential fraud risks between participants; a fraud propagation chain identification module identifies core nodes and indirectly affected nodes; a node feature update module dynamically adjusts the risk level of nodes; a real-time anti-fraud detection and response module obtains transaction data in real time, evaluates the risk of each transaction node, and triggers an early warning and response mechanism when a potential fraud risk is identified; a fraud early warning module evaluates the risk level of a transaction node in real time, and triggers a response mechanism through a smart contract when potential fraud is identified. By combining blockchain with graph neural networks, the ability to identify, predict, and prevent fraud risks has been significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and in particular to an intelligent risk control system based on blockchain. Background Art

[0002] With the digitization and networking of global financial transactions, financial fraud is on the rise in areas such as cross-border payments, supply chain finance, and trading platforms. Financial institutions, businesses, and individuals face severe credit risks and financial losses. Therefore, effectively identifying and preventing financial fraud has become a major technical challenge in today's financial risk management field.

[0003] Traditional risk control systems typically rely on rule engines and historical data analysis, identifying potential risks by setting a fixed set of risk control rules. However, this approach often struggles to address complex and dynamically changing fraudulent activities, particularly in cross-border payments and multi-node supply chain environments. Fraudulent activity can spread through multiple links and participants, forming fraud networks, and fraud patterns constantly evolve and escalate. Relying solely on traditional risk control methods cannot identify complex fraud risks in a timely and accurate manner.

[0004] However, existing risk control systems that integrate graph neural networks and blockchain technology often suffer from low data processing efficiency, difficulty identifying the chain of fraud transmission, and insufficient real-time warning and response mechanisms. Therefore, we propose a blockchain-based intelligent risk control system. Summary of the Invention

[0005] This invention aims to address the problem that fraud, in existing multi-node, multi-party scenarios such as supply chains and cross-border payments, often propagates through multiple links, forming complex fraud networks. Existing technologies often struggle to effectively identify these hidden fraud chains and fail to promptly identify potential risk nodes and transmission paths. This paper proposes a blockchain-based intelligent risk control system.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A blockchain-based intelligent risk control system, comprising:

[0008] Data integration module, used to integrate multiple data sources in real time through API interfaces or big data integration tools;

[0009] A graph neural network model construction module is used to convert the integrated data source into graph structured data, wherein: each node represents a participant, and the transaction relationship or upstream and downstream supply chain relationship between the nodes serves as an edge; each node carries feature data related to it; the graph neural network model models the relationship between nodes through graph convolution operations to obtain a graph neural network model, thereby performing deep learning and analysis on the dependency relationship between nodes and their adjacent nodes in the graph; the graph convolution operation includes a multi-level convolution process, wherein: in the first layer of convolution, each node updates its own features by aggregating information of its directly adjacent nodes; in the second and higher layers of convolution, the node expands the information propagation range layer by layer by aggregating information of the adjacent nodes of its adjacent nodes, captures deeper dependencies, and helps the model identify potential indirect influences; the multi-level convolution process helps the graph neural network gradually build a deep dependency network and form global information aggregation between nodes, supports remote association analysis of nodes, and identifies fraudulent behavior spread through indirect relationships;

[0010] The fraud propagation chain identification module is used to identify core nodes related to fraud in the graph and trace the fraud propagation path. It iteratively identifies directly and indirectly affected nodes through graph convolution operations. Based on each node's influence value and position in the propagation chain, it updates its risk level and outputs the complete fraud propagation path, including core nodes, indirect nodes, their risk levels, and influence.

[0011] A node feature updating module, configured to update the features of each node based on the fraudulent behavior and propagation path identified by the fraudulent behavior propagation chain identification module;

[0012] A real-time anti-fraud detection and response module is used to acquire and process transaction data in real time using streaming data processing technology, assess the fraud risk of each transaction node, and trigger an early warning and response mechanism when potential fraud risks are identified. The processing of transaction data is recorded and audited using blockchain technology;

[0013] The fraud warning module is used to evaluate the risk level of each transaction node in real time based on the results of the node feature update module and combined with a machine learning algorithm. When potential fraud risks are identified, measures are taken through smart contracts, which are executed through blockchain technology.

[0014] Preferably, the graph neural network model building module updates the features of each node in the following manner:

[0015] In the graph neural network model, the features of each node are iteratively calculated through graph convolution operations. The specific update process is as follows:

[0016]

[0017] in, is the feature representation of node v in the k+1th round, is the weight matrix, is the set of adjacent nodes of node v, is the inter-node normalization coefficient, is an activation function; the node features are dynamically updated in combination with the information of adjacent nodes to capture the dependencies between nodes.

[0018] Preferably, the graph neural network model is used to identify the fraudulent behavior of the core nodes in the direct fraud chain, further including: identifying the core nodes related to the fraudulent behavior in the graph through the graph neural network model; determining the core nodes as the implementation nodes of the fraud through feature analysis of the core nodes; tracing the fraudulent behavior of the core nodes based on the graph neural network model, and finding their connection relationship with other nodes in the graph; tracing the nodes at each level through graph convolution operation, and propagating the fraudulent behavior of the core nodes to the direct or indirect nodes associated with them; using the graph neural network model to identify indirect nodes, that is, nodes affected by intermediary nodes; iteratively updating the influence value of each node through graph convolution calculation, and finally identifying all affected indirect nodes; for each node, calculating its influence in the fraud propagation chain; nodes with greater influence are considered to be core fraud nodes, and nodes with less influence are considered to be indirectly affected nodes; after calculating the influence, using the influence value of each node as an indicator of the node characteristics, and updating its risk level according to its position in the fraud chain; based on the calculation of the graph neural network, outputting the complete fraud propagation path, including core nodes, indirect nodes, and the risk level and influence of each node.

[0019] Preferably, the fraud propagation chain identification module is used to trace the propagation path of fraud in the supply chain. The back-propagation algorithm is used to trace the propagation path starting from the core node. The graph neural network model is used to identify potentially affected nodes through feature propagation between nodes. For each node, the influence between it and the core node is calculated, and based on the size of the influence, it is determined whether it is part of the fraud chain. The influence of each node is calculated using the following formula:

[0020]

[0021] in, For nodes To Node The propagation coefficient between For nodes Features at layer I.

[0022] Preferably, the node feature update module updates the features of each node based on the fraudulent behavior and propagation path identified by the fraudulent behavior propagation chain identification module, specifically including: propagating the information of adjacent nodes to the target node through iterative calculation of the graph neural network model, gradually updating the feature data of the target node to reflect the impact of the fraudulent behavior; dynamically adjusting the risk level of the node based on the results of the node feature update, where the node with greater influence is considered to be the core fraud node, and the node with less influence is considered to be the indirectly affected node; the node feature update process updates the features of the node and its adjacent nodes based on the graph convolution operation;

[0023] The dynamic adjustment of the node's risk level is performed through multiple rounds of iterative calculations of the graph neural network, gradually updating the node features. After each round of iteration, the node's risk level is adjusted according to its updated features;

[0024] For each node, its risk score is calculated based on its distance from the fraudulent node or the influence spread in the graph neural network model, and the risk level is dynamically updated;

[0025] The formula for updating the node's risk level is as follows:

[0026]

[0027] in, is the risk level of node v in round k+1, is the influence of node v, is the update function of the risk score.

[0028] Preferably, the node feature update module further expands the propagation range of the node features through iterative calculations of the graph convolutional network, specifically including: in the process of updating the node features, gradually transmitting the feature information of the adjacent nodes to the target node, and expanding to more distant nodes through multiple rounds of iterations to ensure that the impact of the fraudulent behavior can be spread throughout the supply chain network. After each round of iteration, the node features are updated, and the node features ultimately reflect the potential risk of the node in the fraud chain.

[0029] Preferably, the real-time anti-fraud detection and response module realizes real-time monitoring and response to potential fraudulent behaviors by: real-time monitoring of transaction data, obtaining transaction information through streaming data processing technology, and triggering the anti-fraud response mechanism when high-risk transactions are detected; real-time evaluation of the fraud risk of each transaction node through streaming computing, and triggering automatic account freezing and transaction suspension measures when necessary.

[0030] Preferably, the real-time anti-fraud detection and response module evaluates the fraud risk of each transaction node by combining a machine learning algorithm. The specific evaluation method is as follows: the characteristic data of each node is trained using a machine learning algorithm, and the transaction node is evaluated for fraud risk in real time using the trained model. When the risk value of the node exceeds a preset threshold, the corresponding anti-fraud measures are triggered.

[0031] Preferably, the fraud warning module automatically executes anti-fraud measures through smart contracts, specifically including: based on the risk level of the node, automatically executing anti-fraud operations through smart contracts; when the node risk level exceeds the threshold, the smart contract automatically executes measures such as freezing the account, suspending transactions, and notifying relevant parties; all anti-fraud measures are executed through smart contracts.

[0032] Preferably, the multiple data sources include transaction data, supply chain data, and external data sources, and the data sources include transaction records, payment history, credit reports, and account change data of each node; each transaction data is encrypted using the hash algorithm SHA-256 to generate a unique hash value for each transaction; and the hash value is stored through the blockchain.

[0033] The present invention has the following beneficial effects:

[0034] In the present invention, by capturing the complex relationships between nodes and the propagation paths of fraudulent behavior through a graph structure, the system can identify cross-node fraudulent behavior that cannot be discovered by single-node fraud detection, providing more comprehensive risk assessment and prevention. At the same time, through real-time node feature updates and risk level adjustments, the system can dynamically assess and respond to risks to ensure the timely identification and suppression of fraudulent behavior. In addition, the use of smart contract technology to automatically execute anti-fraud measures such as freezing accounts and suspending transactions significantly shortens the time for manual intervention and improves response speed and efficiency. The application of blockchain technology ensures the immutability and transparency of data, reduces the risk of data tampering or omission, and enhances the auditability and traceability of the system. Using streaming data processing technology to monitor and analyze transaction data in real time, it is possible to quickly respond to and handle fraudulent behavior and provide real-time warnings.

[0035] This invention not only solves the problem of fraud detection at a single node but also deeply analyzes the relationships between nodes, ensuring comprehensive and in-depth risk identification. By automating anti-fraud measures and real-time monitoring, the time required for manual intervention is reduced, work efficiency is improved, and the possibility of human error is reduced. Furthermore, through streaming data processing technology and feedback loops, the system can continuously learn and adjust model parameters, improving the accuracy of fraud identification and continuously optimizing risk control strategies. The entire system provides a highly automated, real-time, efficient, and reliable solution for fraud identification and prevention in the financial sector, especially in supply chain management. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a system architecture diagram of the blockchain-based intelligent risk control system proposed in this invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] like Figure 1 As shown, the present invention proposes a blockchain-based intelligent risk control system, including:

[0039] Data integration module, used to integrate multiple data sources in real time through API interfaces or big data integration tools;

[0040] The graph neural network model building module is used to convert the integrated data source into graph-structured data, in which: each node represents a participant, and the transaction relationship or upstream and downstream supply chain relationship between nodes serves as an edge; each node carries feature data related to it; the graph neural network model models the relationship between nodes through graph convolution operations to obtain a graph neural network model, thereby performing deep learning and analysis on the dependency relationship between nodes and their adjacent nodes in the graph; the graph convolution operation includes a multi-level convolution process, in which: in the first layer of convolution, each node updates its own features by aggregating information from its directly adjacent nodes; in the second and higher layers of convolution, the node aggregating information from the adjacent nodes of its adjacent nodes, gradually expanding the range of information dissemination, capturing deeper dependencies, and helping the model identify potential indirect influences; the multi-level convolution process helps the graph neural network gradually build a deep dependency network and form a global information aggregation between nodes, supporting remote association analysis of nodes and identifying fraudulent behavior spread through indirect relationships;

[0041] The fraud propagation chain identification module is used to identify core nodes related to fraud in the graph and trace the fraud propagation path. It iteratively identifies directly and indirectly affected nodes through graph convolution operations. Based on each node's influence value and position in the propagation chain, it updates its risk level and outputs the complete fraud propagation path, including core nodes, indirect nodes, their risk levels, and influence.

[0042] A node feature update module is used to update the features of each node based on the fraudulent behavior and propagation path identified by the fraud propagation chain identification module;

[0043] A real-time anti-fraud detection and response module is used to acquire and process transaction data in real time through streaming data processing technology, assess the fraud risk of each transaction node, and trigger early warning and response mechanisms when potential fraud risks are identified. The transaction data processing process is recorded and audited through blockchain technology;

[0044] The fraud warning module is used to conduct real-time risk assessment of each transaction node based on the results of the node feature update module and combined with machine learning algorithms. When potential fraud risks are identified, measures are taken through smart contracts, which are executed through blockchain technology.

[0045] The solution focuses on identifying and preventing fraud in the financial sector, particularly in supply chain management. Traditional anti-fraud technologies focus on detecting fraud at a single node or transaction. However, the interconnectedness between supply chain nodes or multiple transaction nodes is often overlooked. When fraud occurs at one node, it often directly or indirectly impacts other related nodes.

[0046] 1. Data Integration Module:

[0047] Data source integration: Transaction data: Real-time acquisition of transaction amount, time, identities of both parties, transaction devices, geographic location, and other detailed information from transaction platforms (such as banks, payment gateways, e-commerce platforms, etc.). Specifically including:

[0048] Obtain transaction amount, transaction time, transaction type (such as transfer, payment, refund, etc.), transaction status (such as success, failure, pending, etc.), account information of transaction participants, IP address or geographic location information of transaction devices, etc. from bank transaction data.

[0049] Obtain transaction payment information from the payment gateway, such as payment method (such as credit card, e-wallet, etc.), payment status, payment amount, payment time, payment channel information, etc.

[0050] Obtain order information from the e-commerce platform, such as order amount, order time, order status (such as paid, unpaid, refunded, etc.), information of the buyer and seller, product information, etc.

[0051] Supply chain data: Obtain data such as suppliers, customers, and logistics information from the supply chain management system to ensure complete records of each node (such as suppliers, customers, etc.). Specifically including:

[0052] Obtain supplier information from the supply chain system, such as supplier name, address, contact information, credit rating, delivery history, payment history, etc.

[0053] Obtain customer information, such as customer name, address, contact information, credit rating, purchase history, payment history, etc.

[0054] Obtain logistics information such as transportation method, transportation time, transportation status, transportation company, cargo quantity, cargo value, etc.

[0055] External data sources:

[0056] Financial services data: Obtain credit ratings, payment history and other related data from financial services institutions. Specifically including:

[0057] Obtain data such as a company's or individual's credit rating, credit report, credit history, credit risk, etc. from credit rating agencies.

[0058] Obtain payment history from payment platforms or banks, such as payment amount, payment frequency, payment channel, payment status, etc.

[0059] Government cloud business administration information service data: obtains enterprise registration information, tax information, etc. from government departments. Specifically including:

[0060] Obtain company registration information from the government’s industrial and commercial registration system, such as company name, registered address, registered capital, legal representative, shareholder information, business scope, etc.

[0061] Obtain the company's tax records, tax risk information, tax credit rating, etc. from the tax department.

[0062] Legal information service data: Obtain legal risk information of enterprises or individuals from legal databases, such as litigation, bankruptcy, etc. Specifically including:

[0063] Obtain legal risk information from the legal database regarding whether the enterprise or individual is involved in litigation, legal disputes, bankruptcy, liquidation, compulsory execution, etc.

[0064] Data preprocessing:

[0065] Denoising and supplementation: Use the Kalman filter algorithm to denoise the original data to eliminate the impact of noise on data accuracy:

[0066] For missing data, linear interpolation or the average value of historical data is used to fill in the missing data to ensure data integrity.

[0067] standardization:

[0068] Standardize data from different data sources to ensure dimensional consistency:

[0069] For numerical data (such as transaction amount, frequency, etc.), use Z-score standardization to make the data distributed on the same scale:

[0070] For categorical data (such as transaction type, party type, etc.), use one-hot encoding or other appropriate encoding methods to make the data suitable for model training.

[0071] Ensure data integrity and immutability:

[0072] Use the SHA-256 algorithm to generate a unique hash value for each transaction data:

[0073] Convert transaction data into strings and ensure the data is in the same order (for example, sorted by field name).

[0074] Hash the sorted data string to generate a unique and irreversible hash value.

[0075] Blockchain storage:

[0076] The generated hash value and related transaction data are recorded in the blockchain to ensure data security and transparency:

[0077] Block generation: Packing transaction data into new blocks. Each block contains multiple transaction data and their hash values.

[0078] Hash link: connect the hash value of the new block with the hash value of the previous block to form a chain structure.

[0079] Verification and consensus: Verify and confirm the legitimacy of the block through consensus mechanisms (such as Proof of Work (PoW) and Proof of Stake (PoS)).

[0080] Blockchain update: Verified blocks are added to the blockchain to ensure the immutability and transparency of transaction data.

[0081] 2. Graph Neural Network Model Building Module:

[0082] Graph structure construction: Convert the preprocessed data into a graph structure and build the relationship between nodes:

[0083] Node: Each node represents a transaction participant, such as a supplier, customer, financial institution, etc. Node characteristics include transaction records (such as transaction amount, time, frequency, etc.), credit score, account changes and other data. Specific node characteristics include:

[0084] Transaction amount: Indicates the amount of the transaction.

[0085] Transaction time: indicates the time when the transaction occurs.

[0086] Transaction frequency: indicates the frequency of transactions.

[0087] Credit Score: The credit rating of the party obtained from an external data source.

[0088] Account changes: Indicates changes in the account balances of participants.

[0089] Rule-based relationship building:

[0090] Association rule setting: Determine the relationship between nodes through preset association rules. These rules may include:

[0091] Transaction frequency: If two nodes have a high transaction frequency within a specific time period, they are considered to have a stronger connection.

[0092] Transaction amount: Relationships between nodes with large and frequent transaction amounts may have higher weights.

[0093] Participant relationship: If a node has a common trading partner or upstream and downstream supply chain relationship with other nodes, connections can be established through this "indirect" relationship.

[0094] Rule example: If node A and node B have more than 10 transactions with an amount greater than 5,000 in the past 30 days, the system will consider that there is a strong trading relationship between A and B.

[0095] If the transaction objects of node A and node B are similar (for example, they participate in the same transaction or fund flow), they will be considered as associated nodes.

[0096] Edge weight assignment

[0097] For each pair of constructed edges, a weight needs to be assigned to it, and the weight value represents the strength of the association. This can usually be assigned based on the following method:

[0098] Transaction amount weighting: The weight of the edge can be weighted according to the transaction amount between the two nodes. The larger the amount, the higher the associated weight.

[0099] Transaction frequency weighting: Edges with high transaction frequency can be assigned higher weights, indicating that the interactions between the two nodes are more frequent.

[0100] Transaction type weighting: If the transaction type is high-risk (such as capital flow), higher weights can be given to the edges between related nodes.

[0101] Relationship strength weighting: Set weights based on the strength of the relationship between nodes, such as long-term partners with higher weights.

[0102] Graph model initialization steps: Node and edge initialization:

[0103] When initially establishing the graph model, the characteristics of each node will include: the legitimacy of the node's behavior: for example, whether the node has a record of violations (credit assessment, historical fraud records, etc.).

[0104] Node transaction frequency: the frequency of the node's transaction behavior and the intensity of its interaction with other nodes.

[0105] Node credit score: such as the node's credit score, which is used to determine whether the node behavior is normal.

[0106] The attributes of each edge will include: Transaction type: used to mark the transaction category between nodes.

[0107] Association strength: Based on the weight distribution mentioned above, it reflects the closeness of the relationship between nodes.

[0108] Initial graph structure construction: The initial graph model should be completed by building basic node and edge properties. The graph model at this point can be considered a sparse graph because not every pair of nodes has a direct connection. Graph structure representation: Build a graph ,in:

[0109] V is a set of nodes, each node Representing a participant (such as a user, merchant, etc.), the node carries its characteristic data (such as transaction history, behavioral data, etc.).

[0110] It is a set of edges, which represents the relationship between nodes, usually transaction relationships or upstream and downstream relationships in the supply chain.

[0111] Steps for modeling multi-level relationships between nodes:

[0112] Introduction of multi-level relationships:

[0113] Multi-level relationship modeling: Beyond direct transaction relationships, we further analyze and model indirect connections between nodes. For example, historical transaction chains can be constructed by analyzing historical transaction records. For example, the relationship between nodes A and B extends beyond the current transaction and can also be traced back to other historically connected nodes (such as C and D).

[0114] Supply chain relationships: Node A may be the supplier of a product, B may be its customer, and C may be B's supplier. By tracing these upstream and downstream supply chain relationships, a richer graph model can be formed.

[0115] Enhanced graph model recognition capabilities: Indirect relationships: By introducing indirect relationships between nodes, graph models can capture nodes that appear to have no direct transaction relationship but are potentially risky. For example, cross-node indirect relationships: If node A is directly connected to node B, and node B also has a strong transaction relationship with node C, then nodes A and C may be indirectly affected.

[0116] Deep optimization of graph models: By modeling multi-level and multi-dimensional node relationships, graph neural networks can capture more complex behavioral patterns from a deep learning perspective and identify potential fraudulent behaviors or abnormal patterns.

[0117] Use graph neural network (GNN) for node feature learning and updating:

[0118] Graph convolution operation: The first layer of convolution updates the node features using the following formula:

[0119]

[0120] in, is the feature representation of node v in the k+1th round, is the weight matrix, is the set of adjacent nodes of node v, is the inter-node normalization coefficient, is the activation function; (such as ReLU).

[0121] Second and higher convolution layers (recursive information aggregation):

[0122] For higher levels of convolution, nodes obtain information from their neighbors through their neighbors. The convolution of the layer is done by Aggregate the features of the layer:

[0123]

[0124] This multi-layer convolutional mechanism helps capture deeper dependencies, thereby supporting the identification of fraud propagation paths.

[0125] Iterative calculation: Through multiple rounds of iteration, node features gradually reflect the relationship between nodes and potential fraud. In each iteration, the node obtains information from its adjacent nodes and updates its own feature representation, forming a dynamically updated graph model that can reflect new nodes or relationships in real time.

[0126] Model output: The final output of the graph neural network is the feature of each node These features contain a deep learning representation of the relationships between nodes. This representation will be used in the subsequent fraud propagation chain identification module.

[0127] Use supervised learning with historical data to train a model to identify fraudulent activity:

[0128] Training dataset: Create a sample dataset containing known fraudulent and non-fraudulent transactions.

[0129] Each sample in the dataset includes transaction feature data (such as transaction amount, time, identity of participants, etc.) and a label (whether it is fraudulent). Specifically, it includes:

[0130] Extract features from historical transaction data, such as transaction amount, transaction time, transaction frequency, transaction participant identity, transaction device information, geographic location, etc.

[0131] Based on the history of known fraudulent transactions, fraudulent transactions and non-fraudulent transaction samples are marked.

[0132] Optimization goal: Through the back-propagation algorithm, adjust the model parameters to minimize the deviation between the model output and the true label, and maximize the model's accuracy in identifying fraudulent behavior.

[0133] Model parameter adjustment: Update the weight matrix through the gradient descent algorithm.

[0134] Determine risk points and related factors based on business categories, evaluate them according to risk control ratios, and form rating results:

[0135] Model construction: Determine risk points and related factors based on business categories, and make assessments based on risk control ratios.

[0136] By integrating relevant data interfaces such as financial services, government cloud business administration information services, and legal information services, key information data that may affect individual or corporate financial risk assessments is stored in the database for rating. Specifically, it includes:

[0137] Financial services data interface: obtain credit ratings, payment history, etc. from financial services institutions.

[0138] Government cloud business administration information service data interface: obtain enterprise registration information, tax information, etc. from government departments.

[0139] Legal information service data interface: obtain legal risk information of enterprises or individuals, such as litigation, bankruptcy, etc. from the legal database.

[0140] Rating results: Displays the rating results of the initial business review, including the rating of the enterprise / business.

[0141] Statistical portrait: Automatically store and count customer information, form an analysis and statistics page, and visualize user portraits.

[0142] Rating methods: including linear scoring and grade mapping - matrix scoring method:

[0143] Linear scoring: Calculate the risk score by taking a weighted average based on the weight of each risk point.

[0144] Matrix scoring: Matrix mapping of each risk point according to importance and impact to determine the rating level.

[0145] 3. Fraud Propagation Chain Identification Module: In financial risk control, a fraud propagation chain refers to a malicious account or node influencing and controlling other accounts through multiple transactions, forming a behavioral pattern or risk transmission chain. This chain may involve the interaction of multiple accounts, propagating fraud through different paths.

[0146] For example, suppose account A (the account initiating the fraud) conducts a series of transactions with accounts B, C, and D, which may also have transactions with other accounts, causing the fraud to spread between accounts. In this case, accounts A, B, C, and D form a fraud transmission chain.

[0147] Using the node features output by the graph neural network model, we identify core nodes associated with fraudulent activity and trace the path through which the fraudulent activity spreads. This module uses depth-first search (DFS) or breadth-first search (BFS) to traverse the graph structure, calculate the path, and identify influence.

[0148] Core node identification: Node fraud probability: The fraud probability of each node output by the graph neural network model , we can determine whether a node is a core fraud node. We set a threshold T, and nodes exceeding this threshold are considered core fraud nodes.

[0149] The specific calculation process is as follows:

[0150] Get the node fraud probability output by the model .

[0151] Set the threshold T, if >T, then the node Considered a core fraud node.

[0152] Traverse all nodes and compare their fraud probability With the threshold T, all core fraud nodes are identified , and add it to the core fraud node list.

[0153] The fraud propagation path tracing starts from the core node and uses graph convolution operations and graph traversal algorithms to trace the fraud propagation path in the graph and identify the affected nodes.

[0154] Propagation path calculation: For each core node , using graph convolution to calculate the influence of its neighboring nodes. The influence is calculated based on the propagation coefficient between nodes , that is, node and its adjacent nodes The influence between them.

[0155] The influence calculation formula is as follows:

[0156]

[0157] in, For nodes To Node The propagation coefficient between For nodes Features at layer I.

[0158] Propagation path identification: Through the depth-first search (DFS) or breadth-first search (BFS) algorithm, from the core node Starting from the core node, traverse the graph structure and identify all directly or indirectly affected nodes. The specific steps are as follows: Start by traversing the graph structure using the DFS or BFS algorithm.

[0159] For each traversed node , calculate its relationship with the core node The influence between them.

[0160] If the node If the influence of a node exceeds a certain threshold, the node is considered to be an affected node.

[0161] The node Add to the fraud propagation chain and continue traversing.

[0162] 2.3 Indirect Node Identification and Influence Update

[0163] The influence of each node is iteratively updated through the graph neural network model, and all affected indirect nodes are eventually identified.

[0164] Influence update: The influence of each node is updated through graph convolution operations. The update process is as follows:

[0165] in, For nodes The features at layer I+1 represent the nodes Updated influence.

[0166] Influence size judgment: Nodes with greater influence are considered core fraud nodes, while nodes with less influence are considered indirectly affected nodes.

[0167] Output: Based on the graph neural network calculation, the complete fraud propagation path is output, including:

[0168] List of core fraud nodes.

[0169] List of indirectly affected nodes.

[0170] The risk level and influence of each node in the fraud propagation chain.

[0171] 4. Node feature update module: Automatically updates the attribute information of each node based on real-time data changes in the blockchain network:

[0172] Feature Update: Through graph convolution, information from neighboring nodes is propagated to the target node, updating its feature representation. The specific steps include: Obtaining feature data from neighboring nodes. Updating the target node's feature representation through graph convolution, incorporating information from neighboring nodes. Dynamically adjusting the node's risk level based on node characteristics and the influence of fraudulent activity:

[0173] Risk level adjustment: Calculate the risk score of the node using the following formula:

[0174]

[0175] in, is the risk level of node v in round k+1, is the influence of node v, is the update function for the risk score. It can be a linear or nonlinear function.

[0176] The node's risk level is dynamically adjusted based on the risk score. A higher risk score indicates the node's key position in the fraud chain. Specific adjustment methods include:

[0177] Increase risk level threshold: When a node's risk score exceeds the preset threshold, its risk level is increased.

[0178] Weight adjustment: Based on the node’s fraud risk score, adjust its weight in the fraud propagation path.

[0179] 5. Real-time anti-fraud detection and response module:

[0180] Use streaming data processing technologies (such as Apache Kafka and Apache Flink) for real-time data monitoring and processing to ensure fast response:

[0181] Real-time data processing: Transaction data is obtained from the trading platform in real time and quickly processed through stream processing engines such as Kafka or Flink.

[0182] Perform pre-processing on transaction data in real time, such as denoising and standardization, to make it suitable for model input.

[0183] Real-time risk assessment of each transaction node is performed through a trained machine learning model: features such as transaction amount, transaction time, identities of both parties to the transaction, device information, geographic location, etc. are extracted from real-time transaction data.

[0184] Use the trained model to predict the risk score of the transaction. The specific steps include:

[0185] Get real-time transaction data.

[0186] Extract transaction features.

[0187] Input the feature data into the trained model to predict the risk score.

[0188] When high-risk transactions are identified, anti-fraud measures are automatically triggered:

[0189] Response mechanism triggering: Set a risk score threshold (e.g. 90). When the risk score of a transaction node is higher than the threshold, the following measures are executed:

[0190] Freeze Account: Automatically freeze high-risk accounts associated with the transaction.

[0191] Suspend transaction: Suspend transaction activities related to this node.

[0192] Notify relevant parties: Inform relevant regulatory agencies or responsible parties through smart contracts or other notification mechanisms.

[0193] 6. Fraud warning module:

[0194] Combined with machine learning algorithms, the risk level of each transaction node is assessed in real time: features are extracted from transaction data. The trained model is used to evaluate the feature data of each node and dynamically adjust the node's risk level. The specific steps include:

[0195] Get transaction data.

[0196] Extract transaction features.

[0197] Input feature data into the trained model to predict the risk score and dynamically adjust the risk level.

[0198] Automatically execute anti-fraud measures through smart contracts: Write smart contracts on the blockchain and predefine various anti-fraud actions, such as freezing accounts and suspending transactions. When the risk level exceeds the threshold, the smart contract automatically triggers the corresponding anti-fraud measures: Trigger condition: Check whether the node's risk score exceeds the threshold. If it does, the smart contract automatically triggers the corresponding anti-fraud measures:

[0199] Freeze Account: Freeze high-risk accounts associated with the transaction.

[0200] Suspend transaction: Suspend transaction activities related to this node.

[0201] Send notifications: Inform relevant regulatory agencies or responsible parties through smart contracts or other notification mechanisms.

[0202] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A blockchain-based intelligent risk control system, characterized by: include: Data integration module, used to integrate multiple data sources in real time through API interfaces or big data integration tools; A graph neural network model construction module is used to convert the integrated data source into graph structured data, wherein: each node represents a participant, and the transaction relationship or upstream and downstream supply chain relationship between the nodes serves as an edge; each node carries feature data related to it; the graph neural network model models the relationship between nodes through graph convolution operations to obtain a graph neural network model, thereby performing deep learning and analysis on the dependency relationship between nodes and their adjacent nodes in the graph; the graph convolution operation includes a multi-level convolution process, wherein: in the first layer of convolution, each node updates its own features by aggregating information of its directly adjacent nodes; in the second and higher layers of convolution, the node expands the information propagation range layer by layer by aggregating information of the adjacent nodes of its adjacent nodes, captures deeper dependencies, and helps the model identify potential indirect influences; the multi-level convolution process helps the graph neural network gradually build a deep dependency network and form global information aggregation between nodes, supports remote association analysis of nodes, and identifies fraudulent behavior spread through indirect relationships; The fraud propagation chain identification module is used to identify core nodes related to fraud in the graph and trace the fraud propagation path. It iteratively identifies directly and indirectly affected nodes through graph convolution operations. Based on each node's influence value and position in the propagation chain, it updates its risk level and outputs the complete fraud propagation path, including core nodes, indirect nodes, their risk levels, and influence. A node feature updating module, configured to update the features of each node based on the fraudulent behavior and propagation path identified by the fraudulent behavior propagation chain identification module; A real-time anti-fraud detection and response module is used to acquire and process transaction data in real time using streaming data processing technology, assess the fraud risk of each transaction node, and trigger an early warning and response mechanism when potential fraud risks are identified. The processing of transaction data is recorded and audited using blockchain technology; The fraud warning module is used to evaluate the risk level of each transaction node in real time based on the results of the node feature update module and combined with a machine learning algorithm. When potential fraud risks are identified, measures are taken through smart contracts, which are executed through blockchain technology.

2. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The graph neural network model building module updates the features of each node in the following way: In the graph neural network model, the features of each node are iteratively calculated through graph convolution operations. The specific update process is as follows: in, is the feature representation of node v in the k+1th round, is the weight matrix, is the set of adjacent nodes of node v, is the inter-node normalization coefficient, is an activation function; the node features are dynamically updated in combination with the information of adjacent nodes to capture the dependencies between nodes.

3. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The graph neural network model is used to identify the fraudulent behavior of the core nodes in the direct fraud chain, further including: identifying the core nodes related to the fraudulent behavior in the graph through the graph neural network model; determining that the core nodes are the implementation nodes of the fraud through feature analysis of the core nodes; tracing the fraudulent behavior of the core nodes based on the graph neural network model, and finding their connection relationship with other nodes in the graph; tracing back the nodes at each level through graph convolution operations, and propagating the fraudulent behavior of the core nodes to the direct or indirect nodes associated with them; using the graph neural network model to identify indirect nodes, that is, nodes affected by intermediary nodes; iteratively updating the influence value of each node through graph convolution calculation, and finally identifying all affected indirect nodes; for each node, calculating its influence in the fraud propagation chain; nodes with greater influence are considered to be core fraud nodes, and nodes with less influence are considered to be indirectly affected nodes; after calculating the influence, using the influence value of each node as an indicator of the node characteristics, and updating its risk level according to its position in the fraud chain; based on the calculation of the graph neural network, outputting the complete fraud propagation path, including core nodes, indirect nodes, and the risk level and influence of each node.

4. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The fraud propagation chain identification module is used to trace the propagation path of fraud in the supply chain. It uses the back-propagation algorithm to trace the propagation path starting from the core node. Using the graph neural network model, it identifies potentially affected nodes by propagating features between nodes. For each node, its influence with the core node is calculated. Based on the size of the influence, it is determined whether it is part of the fraud chain. The influence of each node is calculated using the following formula: in, For nodes To Node The propagation coefficient between For nodes Features at layer I.

5. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The node feature update module updates the features of each node based on the fraudulent behavior and propagation path identified by the fraud propagation chain identification module. Specifically, the module propagates the information of adjacent nodes to the target node through iterative calculations of the graph neural network model, gradually updating the feature data of the target node to reflect the impact of the fraudulent behavior; based on the results of the node feature update, the risk level of the node is dynamically adjusted, with nodes with greater influence being considered core fraud nodes and nodes with less influence being considered indirectly affected nodes; the node feature update process updates the features of the node and its adjacent nodes based on graph convolution operations; The dynamic adjustment of the node's risk level is performed through multiple rounds of iterative calculations of the graph neural network, gradually updating the node features. After each round of iteration, the node's risk level is adjusted according to its updated features; For each node, its risk score is calculated based on its distance from the fraudulent node or the influence spread in the graph neural network model, and the risk level is dynamically updated; The formula for updating the node's risk level is as follows: in, is the risk level of node v in round k+1, is the influence of node v, is the update function of the risk score.

6. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The node feature update module further expands the propagation range of node features through iterative calculations of the graph convolutional network. Specifically, during the node feature update process, the feature information of adjacent nodes is gradually transmitted to the target node, and expanded to more distant nodes through multiple rounds of iterations to ensure that the impact of fraudulent behavior can be spread throughout the entire supply chain network. After each round of iteration, the node features are updated, and the node features ultimately reflect the potential risk of the node in the fraud chain.

7. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The real-time anti-fraud detection and response module monitors and responds to potential fraudulent activities in real time by: monitoring transaction data in real time, acquiring transaction information through streaming data processing technology, and triggering the anti-fraud response mechanism when high-risk transactions are detected; assessing the fraud risk of each transaction node in real time through streaming computing, and triggering automatic account freezing and transaction suspension measures when necessary.

8. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The real-time anti-fraud detection and response module evaluates the fraud risk of each transaction node by combining a machine learning algorithm. The specific evaluation method is as follows: the feature data of each node is trained using a machine learning algorithm, and the transaction node is evaluated for fraud risk in real time using the trained model. When the risk value of the node exceeds a preset threshold, the corresponding anti-fraud measures are triggered.

9. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The fraud warning module automatically executes anti-fraud measures through smart contracts, specifically including: automatically executing anti-fraud operations based on the node's risk level through smart contracts. When the node risk level exceeds the threshold, the smart contract automatically executes measures such as freezing accounts, suspending transactions, and notifying relevant parties. All anti-fraud measures are executed through smart contracts.

10. The blockchain-based intelligent risk control system according to claim 1, characterized in that: The multiple data sources include transaction data, supply chain data, and external data sources, and the data sources include transaction records, payment history, credit reports, and account change data for each node; Each transaction data is encrypted using the hash algorithm SHA-256 to generate a unique hash value for each transaction; the hash value is stored through the blockchain.

Images