Digital financial fraud identification method and system based on alliance chain and deep learning

Abstract

The invention discloses a digital financial fraud identification method and system based on an alliance chain and deep learning, and the method provides a safe, transparent and non-tampering financial transaction and data sharing platform based on the alliance chain, guarantees the authenticity and traceability of all transaction records, reduces the risk of data counterfeiting and tampering, and improves the safety and reliability of the system. Meanwhile, through strong data processing and feature extraction capabilities of deep learning, the system can analyze mass and unstructured data from different industries and scenes, and capture subtle modes and abnormal signs hidden behind various transaction behaviors, so that digital financial fraudulent behaviors in two financial business modes can be accurately identified, and the system can be widely applied to the field of digital financial fraudulent behaviors. And the generalization ability and robustness of the financial fraud identification algorithm are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of financial fraud identification, and in particular, to a digital financial fraud identification method and system, an electronic device, and a computer-readable storage medium based on a consortium blockchain and deep learning. Background Art

[0002] Existing financial business models can be divided into two main types. The first is digital financial business fully built on the blockchain. Such businesses make full use of the core features of blockchain technology, such as decentralization, immutability, and transparency. Under this model, financial services are directly designed and operated relying on the blockchain network, aiming to provide a more secure, efficient, and transparent service experience. However, due to the anonymity and cross-border characteristics of blockchain technology, some lawbreakers may use these features to carry out illegal activities. For example, they may use complex transaction structures to conceal the source or destination of funds for money laundering purposes; issue unregulated cryptocurrencies to raise funds; or set up Ponzi schemes to attract investors to invest, ultimately causing a large number of investors to suffer losses; and steal sensitive information such as users' private keys through phishing websites. Such behaviors seriously violate laws and regulations and cause great harm to the public interest. The second is the combination of traditional financial business models and blockchain technology, that is, the so-called "off-chain" financial activities plus an "on-chain" data sharing mechanism. Under this model, actual financial transactions are still completed offline according to current rules, but all parties involved can achieve efficient and secure financial data exchange through the blockchain platform. Through smart contracts and permission management modules, the data sharing platform can precisely control the access levels of different participants to specific data, ensuring that only authorized entities can read or modify information under preset conditions. This fine-grained permission control system not only enhances data security and privacy protection but also promotes trust and cooperation among financial institutions, reduces operating costs, and improves transaction efficiency. In addition, since all operations are recorded on the blockchain, this provides a clear traceability path for auditing and compliance checks. Since the data uploaded to the blockchain in the two digital financial business models is different, the data features that can be extracted are also different, resulting in the inability of existing digital financial fraud identification algorithms based on blockchain data to accurately identify fraud behaviors in traditional financial business models.

[0003] In addition, current digital financial fraud identification methods usually use machine learning identification algorithms to implement, which require a large amount of labeled data to train the identification model. From the perspective of data security, data between different application platforms will not be shared, resulting in the need to retrain the model when the identification model is applied to different platform environments, and the generalization ability and robustness of the model are poor. Summary of the Invention

[0004] The present invention provides a method and system for identifying digital financial fraud based on a consortium blockchain and deep learning, an electronic device, and a computer-readable storage medium, which can accurately identify digital financial fraud behaviors in two financial business models and improve the generalization ability and robustness of the financial fraud identification algorithm.

[0005] According to one aspect of the present invention, there is provided a method for identifying digital financial fraud based on a consortium blockchain and deep learning, including the following:

[0006] Receiving transaction data, where the transaction data is on-chain transaction data or off-chain transaction data;

[0007] Writing the transaction data into the consortium blockchain and reaching a consensus among all nodes;

[0008] Extracting corresponding financial business features from the transaction data according to the type of the transaction data, and inputting the extracted financial business features into a pre-trained deep learning model to identify whether the transaction behavior belongs to a financial fraud behavior, and identifying whether the transaction behavior belongs to a fraud behavior in traditional financial business or a fraud behavior in blockchain-based digital financial business.

[0009] Furthermore, the consortium blockchain creates a governance role, a contract administrator role, and a user role based on smart contracts. The governance role plays the role of a top-level manager in the consortium blockchain, used to establish and maintain the governance rules and order on the chain and be responsible for the election and operation of the governance committee, and make decisions on major matters through this committee. The contract administrator role is used to define and manage the users who can deploy new smart contracts, and which users or entities can call specific interfaces of the deployed contracts. The user role participates in the business on the premise of complying with the established rules. After the transaction data is uploaded to the chain, the consortium blockchain conducts a preliminary identification of the financial fraud of the user's transaction behavior.

[0010] Furthermore, the process of the consortium blockchain conducting a preliminary identification of the financial fraud of the user includes the following:

[0011] After sharing the transaction data on the blockchain, the governance role elects a governance committee to manage the permissions for deploying contracts. The contract administrator role controls the user access permissions for the deployed contracts. Initially, all user roles are set to white list permissions, and at this time, the initial value of the warning mechanism is 0. By collecting and analyzing the long-term behavior patterns of users, a normal behavior baseline is established. For critical operations, users are required to provide additional authentication information. The user addresses with critical operation behaviors deviating from the norm at this time are recorded, and the value of the warning mechanism is increased. Then, the addresses of suspicious users are sent to the governance role, and the behaviors of suspicious users and their warning mechanism values are continuously tracked in subsequent transactions. When the warning mechanism value of a suspicious user reaches the threshold, the governance role freezes the account of the suspicious user, and the contract administrator role sets the account to black list permissions. Then, the governance role conducts a new proposal to elect the governance committee, continuously updates the contract and iterates, realizing the initial financial fraud identification process.

[0012] Further, if the transaction data is on-chain transaction data, first construct a transaction information graph of the consortium blockchain. The transaction information graph is a graph network with a topological logical structure constructed based on the transaction data. Different graph networks are used to represent the association relationships between different blockchain entities. At the same time, the corresponding labeled data and off-chain behavior characteristics can be introduced into the corresponding transaction information graph as node attributes or edge attributes. Then, extract the financial business characteristics from the transaction information graph. Among them, the extracted financial business characteristics include graph statistical characteristics and graph embedding characteristics. If the transaction data is off-chain transaction data, the financial business characteristics include time series characteristics, text characteristics, and image characteristics.

[0013] Further, the transaction information graph includes an address graph, a transaction graph, an address-transaction graph, a user graph, and a transaction sub-graph. The transaction graph and the address graph provide the corresponding transaction information and address information for the address-transaction graph. The user graph extracts the transaction pattern based on the address-transaction graph, and the transaction sub-graph extracts the entity relationship based on the address-transaction graph.

[0014] Further, the address-transaction graph is a combination of the address graph and the transaction graph, used to describe the correspondence between transactions and addresses. Directed edges are used to connect input addresses, transactions, and output addresses, and the weight value of the directed edge represents the transaction amount.

[0015] Further, the user graph is represented as G = {U, M, E}, used to describe the flow of digital currency among clustered users in the blockchain. U represents the user set, each user is a user abstracted by clustering, M represents the address set, and E represents the transaction set. The weight value of each transaction edge represents the transaction amount.

[0016] In addition, the present invention also provides a digital financial fraud identification system based on the consortium blockchain and deep learning, including:

[0017] A transaction data receiving module, configured to receive transaction data, where the transaction data is on-chain transaction data or off-chain transaction data;

[0018] A transaction data writing module, configured to write the transaction data into the consortium blockchain and reach a consensus among all nodes;

[0019] A financial fraud identification module, configured to extract corresponding financial business features from the transaction data according to the type of the transaction data, input the extracted financial business features into a pre-trained deep learning model, to identify whether the transaction behavior belongs to a financial fraud behavior, and identify that the transaction behavior belongs to a fraud behavior in traditional financial business or a fraud behavior in blockchain-based digital financial business.

[0020] In addition, the present invention also provides an electronic device, including a processor and a memory, where a computer program is stored in the memory, and the processor is configured to execute the steps of the method as described above by calling the computer program stored in the memory.

[0021] In addition, the present invention also provides a computer-readable storage medium, configured to store a computer program for digital financial fraud identification based on a consortium blockchain and deep learning, and the computer program executes the steps of the method as described above when running on a computer.

[0022] The present invention has the following beneficial effects:

[0023] The digital financial fraud identification method based on a consortium blockchain and deep learning of the present invention provides a secure, transparent and tamper-proof financial transaction and data sharing platform based on the consortium blockchain, ensures the authenticity and traceability of all transaction records, reduces the risk of data forgery and tampering, and at the same time, through the powerful data processing and feature extraction capabilities of deep learning, the system can analyze massive, unstructured data from different industries and scenarios, capture the subtle patterns and abnormal signs hidden behind various transaction behaviors, so as to accurately identify digital financial fraud behaviors in two financial business models, and improve the generalization ability and robustness of the financial fraud identification algorithm.

[0024] In addition, the digital financial fraud identification system based on a consortium blockchain and deep learning of the present invention also has the above advantages.

[0025] In addition to the purposes, features and advantages described above, the present invention has other purposes, features and advantages. The following will refer to the drawings to further elaborate on the present invention in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings forming a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0027] Figure 1 is a digital financial fraud identification method based on consortium blockchain and deep learning in a preferred embodiment of the present application;

[0028] Figure 2 is a schematic diagram of the principle of the preliminary warning mechanism for fraud identification in a preferred embodiment of the present application;

[0029] Figure 3 is a schematic diagram of the UTXO transaction model in a preferred embodiment of the present application;

[0030] Figure 4 is a schematic diagram of the account transaction model in a preferred embodiment of the present application;

[0031] Figure 5 is a schematic diagram of the address map in a preferred embodiment of the present application;

[0032] Figure 6 is a schematic diagram of the transaction graph in a preferred embodiment of the present application;

[0033] Figure 7 is a schematic diagram of the address transaction graph in a preferred embodiment of the present application;

[0034] Figure 8 is a schematic diagram of the user graph in a preferred embodiment of the present application;

[0035] Figure 9 is a schematic diagram of the transaction sub-graph in a preferred embodiment of the present application;

[0036] Figure 10 is a schematic diagram of the system architecture of a digital financial fraud identification platform constructed according to the digital financial fraud identification method based on consortium blockchain and deep learning in a preferred embodiment of the present application;

[0037] Figure 11 is Figure 10 a schematic diagram of the business process of the digital financial fraud identification platform in;

[0038] Figure 12 is Figure 10 a schematic diagram of the algorithm of the digital financial fraud identification platform in;

[0039] Figure 13 is a schematic diagram of the module structure of a digital financial fraud identification system based on consortium blockchain and deep learning in another embodiment of the present application. Detailed implementation manners

[0040] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] Reference Figure 1 , a preferred embodiment of the present application provides a digital financial fraud recognition method based on a consortium blockchain and deep learning, including the following:

[0042] Step S1: Receive transaction data, where the transaction data is on-chain transaction data or off-chain transaction data;

[0043] Step S2: Write the transaction data into the consortium blockchain and reach a consensus among all nodes;

[0044] Step S3: Extract the corresponding financial business features from the transaction data according to the type of the transaction data, and input the extracted financial business features into a pre-trained deep learning model to identify whether the transaction behavior belongs to a financial fraud behavior, and identify whether the transaction behavior belongs to a fraud behavior in traditional financial business or a fraud behavior in digital financial business based on blockchain.

[0045] It can be understood that the digital financial fraud recognition method based on a consortium blockchain and deep learning in this embodiment provides a secure, transparent and tamper-proof financial transaction and data sharing platform based on the consortium blockchain, ensuring the authenticity and traceability of all transaction records, reducing the risk of data forgery and tampering. At the same time, through the powerful data processing and feature extraction capabilities of deep learning, the system can analyze massive, unstructured data from different industries and scenarios, capture the subtle patterns and abnormal signs hidden behind various transaction behaviors, so as to accurately identify digital financial fraud behaviors in two financial business models, and improve the generalization ability and robustness of the financial fraud recognition algorithm.

[0046] It can be understood that in step S1, transaction data from different sources enters the recognition system of the present invention through an API interface or a data transmission protocol (such as a P2P network). The types of transaction data include on-chain transaction data or off-chain transaction data. Among them, off-chain transaction data is mainly traditional financial business data, which covers a series of transaction information related to financial services, such as credit card transaction data, insurance application and claim details, personal or corporate credit activities, and company internal financial statements, etc.; on-chain transaction data refers to the blockchain transaction details data, contracts deployed on the chain, and log data, etc. that occur in financial business built on blockchain itself and are certified and uploaded to the chain.

[0047] It can be understood that in step S2, after the data layer of the consortium blockchain receives the uploaded transaction data, it will initiate a data writing request to the blockchain network. The consortium blockchain controls the writing process of the transaction data through smart contracts to ensure that the data is stored on the blockchain according to predefined rules and verifies the validity and format of the data. After the transaction data is written to the blockchain, it forms an immutable record, and a hash calculation is performed on the written data to generate a unique hash value, ensuring the uniqueness and anti-tampering property of the transaction data. Each block contains multiple transaction records, ensuring the transparency and security of the data. Then, all nodes of the consortium blockchain synchronize the latest block data to ensure that all participating nodes have the same blockchain state, thus guaranteeing the reliability and consistency of the system, and reaching a consensus through consensus algorithms (such as PBFT, Raft, etc.) to confirm the validity of the transaction. The consensus mechanism ensures the consistent recognition of the data by all nodes and synchronizes the data to all nodes.

[0048] In addition, the smart contract layer of the consortium blockchain also ensures that only authorized parties can perform specific operations through the permission control mechanism. The smart contract automatically executes the transaction logic to ensure that the transaction proceeds according to the predefined rules and calculates the Gas fees required for the transaction to provide resources for the execution of the smart contract.

[0049] It can be understood that under the consortium blockchain architecture, each financial institution joins the network as a node to jointly maintain a distributed ledger, ensuring that all transaction records are transparent and immutable, enhancing the trust foundation among all parties; secondly, using smart contracts to automatically execute predefined rules and conditions simplifies the approval process for data sharing and improves operational efficiency; at the same time, the consortium blockchain supports various privacy protection mechanisms, such as zero-knowledge proofs, homomorphic encryption, etc., which can achieve data verification and calculation without revealing sensitive information, protecting user privacy and business secrets; in addition, based on the permission control module, the consortium blockchain can finely manage the access rights of different roles to specific data sets to prevent unauthorized operations.

[0050] Optionally, in the consortium blockchain of the present invention, governance roles, contract administrator roles, and user roles are created based on smart contracts. The governance role plays the role of the top manager in the consortium blockchain. Its main task is to establish and maintain the governance rules and order on the chain. The governance role is responsible for the election and operation of the governance committee, and major matters are decided through this committee, such as the freezing or thawing of accounts, etc.; the main responsibility of the contract administrator role is to define and manage who can deploy new smart contracts, and which users or entities can call specific interfaces of the deployed contracts. The contract administrator role sets access permissions according to business requirements, and uses the whitelist or blacklist mode to restrict or allow certain operations. Once a smart contract is deployed, the contract administrator role can choose to designate other accounts as subsequent managers, or default to itself to serve as this role. If the contract administrator role fails to properly perform its duties, the governance committee has the right to intervene and vote on whether to replace the contract administrator to ensure the security and efficiency of the smart contract layer; and the primary responsibility of the user role is to participate in the business while abiding by the established rules, such as initiating transactions or other forms of interactions. Whether the user role can participate in specific business activities depends on the permissions set for it by the contract administrator role. By default, all users can call the interfaces of smart contracts without special permission restrictions. When it comes to sensitive or critical operations, the user role should follow the principle of least privilege and only request the necessary permissions to complete the required tasks. Based on the above three roles, the present invention sets up a preliminary fraud detection early warning mechanism through smart contracts to preliminarily identify financial fraud in user behavior. Among them, as Figure 2 shown, the process of preliminarily identifying financial fraud in users includes the following contents:

[0051] After sharing the transaction data on the chain, the governance role elects a governance committee to manage the permissions for deploying contracts. The contract administrator role controls the user access permissions for the deployed contracts. Initially, the white list permissions are set for all user roles by default. At this time, the initial value of the early warning mechanism is 0; by collecting and analyzing the long-term behavior patterns of users, a normal behavior baseline is established, and for critical operations, users are required to provide additional identity verification information, and the user addresses with critical operation behaviors deviating from the normal at this time are recorded, and the early warning mechanism value is increased; then the addresses of suspicious users are sent to the governance role and the behaviors of suspicious users and their early warning mechanism values are continuously followed up in subsequent transactions. When the early warning mechanism value of a suspicious user reaches the threshold, the governance role freezes the account of the suspicious user, and the contract administrator role sets the account to blacklist permissions; then, the governance role conducts a new proposal to elect a governance committee, continuously updates the contract and iterates to achieve the preliminary financial fraud identification process. In addition, due to a certain error tolerance probability in the preliminary financial fraud identification process, for the suspicious accounts that are not frozen, further identification needs to be carried out through subsequent financial business characteristics and deep learning models.

[0052] It can be understood that the consortium chain of the present invention creates a governance role, a contract administrator role, and a user role through a smart contract, and for the first time creates a preliminary fraud identification early warning mechanism based on these three roles, which can preliminarily identify financial fraud behaviors and improve the identification efficiency.

[0053] In addition, in the step S3, for the transaction data in two different financial business scenarios, different financial business characteristics need to be extracted. The deep learning model can identify whether a transaction behavior belongs to a financial fraud behavior based on the extracted financial business characteristics, and identify whether the transaction behavior belongs to a fraud behavior in traditional financial business or a fraud behavior in blockchain-based digital financial business, so as to improve the identification accuracy of financial fraud.

[0054] Specifically, if the transaction data is on-chain transaction data, that is, for the blockchain-based digital financial business model, first construct a transaction information graph of the consortium chain, and then extract financial business characteristics from the transaction information graph. Among them, the extracted financial business characteristics include graph statistical characteristics and graph embedding characteristics. Among them, the transaction information graph is a graph network with a topological logical structure constructed based on transaction data. Different graph networks are used to represent the association relationships between different blockchain entities. At the same time, the corresponding labeled data and off-chain behavior characteristics can be introduced into the corresponding transaction information graph as node attributes or edge attributes. Currently, the transaction information graph is constructed based on two mainstream transaction models. One is the UTXO transaction model represented by Bitcoin, and the other is the account transaction model represented by Ethereum.

[0055] For example, as Figure 3 shown, the UTXO transaction model, that is, the unspent transaction output model. In Bitcoin transactions, an account can control multiple addresses. Bitcoin transactions consist of several transfer accounts and several receiving accounts. This model is similar to the process of commodity change. If the balance of the transfer accounts (such as m1, m2, m3) is higher than the balance of the receiving accounts (such as m4, m5, m6), an automatically generated change address is used to store the unreceived balance and enter the next transaction. In Figure 3 it, the transfer accounts transfer a total of 12 bitcoins, and the receiving accounts receive a total of 11 bitcoins. After deducting a handling fee of 0.2 bitcoins, 0.8 bitcoins are deposited into the change address m7 for the next transaction. Since the Bitcoin transaction handling fee is usually a non-integer non-negative decimal, the transfer balance is not equal to the receiving balance, resulting in a change address. In Bitcoin transactions, the completion of a transaction requires the transaction input party to confirm and sign with a private key. When the private key is not leaked, all input addresses can be regarded as the same entity. In this system, the change address m7 and the transfer accounts (m1, m2, m3) are regarded as the same entity.

[0056] As Figure 4As shown, the account transaction model is based on the Ethereum platform. Ethereum is a smart contract platform that can develop smart contracts for various transactions. A smart contract is a program stored on the blockchain, and each blockchain node must execute this contract. Once the contract trigger conditions are met, the transaction is automatically executed. This account transaction model is similar to the process of keeping accounts in a bank. The accounts on Ethereum are mainly divided into two categories: external accounts and contract accounts. An external account is an account created by a user based on a private key and mainly records the account balance. A contract account is an account based on a smart contract, created by an external account or a contract, and automatically assigned an address at the time of creation, enabling the contract code to be called and executed between external accounts or other contract accounts. The transaction fee on Ethereum is measured by the total amount of gas consumed by the execution of the transaction. There is a pre-specified total amount of gas for Ethereum operation instructions, and the transaction fee for each transaction is determined by the product of the gas price given by the user and the amount of gas consumed by the transaction. In Figure 4 it, m1 to m5 are external accounts, and SC1 and SC2 are contract accounts, with a total of 6 transactions. Each transaction is a one-to-one transfer that can occur between any external account and contract account. For example, transaction Tx1 records the transaction between external account m1 and contract account SC1, transaction Tx3 records the process of contract account SC2 calling contract account SC1, and transaction Tx5 records the transaction between external accounts m2 and m4.

[0057] The main application scenario of the present invention is account transactions, so an account transaction model is used to construct a transaction information graph. Among them, the transaction information graph of the present invention includes an address graph, a transaction graph, an address-transaction graph, a user graph, and a transaction sub-graph. The transaction graph and the address graph provide corresponding transaction information and address information for the address-transaction graph. The user graph extracts the transaction pattern based on the address-transaction graph, and the transaction sub-graph extracts the entity relationship based on the address-transaction graph.

[0058] It can be understood that the commonly used transaction information graph includes an address graph and a transaction graph. The address graph is a tool used to depict the interaction of digital currencies between different addresses. As a basic model for behavior recognition, it shows the fund flow path in an intuitive and easy-to-construct way. However, despite its simple logic and convenient construction, the address graph has significant resource consumption problems, especially a large storage requirement. This makes the effectiveness and efficiency of the address graph affected when dealing with large-scale transactions, which is not conducive to complex behavior analysis and recognition. Therefore, although the address graph has certain advantages in small-scale or preliminary analysis, its limitations become particularly prominent when faced with massive transaction data. For example, as Figure 5As shown, the address graph G = {M, E} is used to represent the interaction pattern of digital currency between addresses. M = {m1, m2, …, m12} is the vertex set of graph G, and E = {e1, e2, …, e12} is the edge set of graph G. Each vertex represents an account address of the blockchain, each edge represents a transaction between addresses, and the weight value of the edge represents the transaction amount. Among them, the edges in set E are all directed, indicating the direction of the transaction. Figure 5 There are a total of 5 transactions. For example, in transaction Tx1, a total of 7 eth is transferred from accounts {m1, m2, m3} to account {m4}, and in transaction Tx4, 1 eth is transferred from account {m4} to account {m9}, and 2 eth is transferred from account {m8} to account {m9}.

[0059] The transaction graph is a graphical representation method used to depict the flow state of digital currency between different transactions over time. It uses the hash value of the transaction as a node, which not only simplifies the expression of transaction logic but also effectively reduces the required resource occupancy. In this way, the transaction graph can display the transaction process in a more concise and efficient form, making data processing and analysis easier and applicable to scenarios that need to process a large amount of transaction information. This method optimizes the use of storage and computing resources and provides a more refined solution for tracking and analyzing digital currency transactions. For example, as Figure 6 shown, the transaction graph G = {T, E} describes the process of digital currency flowing between transactions over time. Each vertex Txi (i = 1, 2, 3, 4, 5) represents a transaction and is identified by a hash value. Each directed edge ei (i = 1, 2, 3, 4, 5) represents the amount flowing between transactions. There are no multiple edges in the transaction graph, that is, any two transactions are one-way edges and are unique. Taking transaction Tx4 as the starting point, the input address m4 comes from transaction Tx1 and inputs 1 eth, the input address m8 comes from transaction Tx3 and inputs 2 eth, and the output address m9 comes from transaction Tx5 and outputs 1.5 eth.

[0060] The address-transaction graph of the present invention is a combination of the address graph and the transaction graph, used to describe the correspondence between transactions and addresses. The directed edge is used to connect the input address, transaction, and output address, and the weight value of the directed edge represents the transaction amount. Among them, the address-transaction graph can be expressed as: G = {M, T, E}. As Figure 7 shown, in transaction Tx1, address m1 transfers out 2 eth, and address m4 transfers in 7 eth.

[0061] It can be understood that the present invention constructs an address transaction graph by combining an address graph and a transaction graph. It synthesizes the graphical representation of transaction details and address data, covering not only individual transaction information but also incorporating the address details of the participating transactions. Therefore, it can provide a more detailed and comprehensive view of transactions. Through the address transaction graph, it is possible to more clearly understand how digital currency flows between different addresses and the correlations between these transactions, providing users with richer contextual information to analyze transaction activities.

[0062] In addition, the present invention also extracts transaction patterns from the address transaction graph to construct a user graph, which is represented as G = {U, M, E} and is used to describe the flow of digital currency among clustered users in the blockchain. U represents the set of users, where each user is an abstracted user through clustering. M represents the set of addresses, and E represents the set of transactions. The weight value of each transaction edge represents the transaction amount. For example, as Figure 8 shown, this user graph mainly contains 6 users, namely U1{m1, m2, m3}, U2{m4, m5}, U3{m9, m10, m11}, U4{m7, m8}, {m6}, and {m12}.

[0063] It can be understood that the present invention constructs a user graph by extracting transaction patterns from the address transaction graph, which can well display the flow status of digital currency among blockchain users. It forms an abstraction of the behavior of user groups based on the address graph through entity clustering, thus enabling more clearly identifying and analyzing the behavior patterns between users. This not only enhances the understanding of user behavior logic but also makes complex behavior relationships easier to analyze.

[0064] In addition, the present invention also extracts entity relationships from the address transaction graph to construct a transaction subgraph. The transaction subgraph is a partial view extracted from the complete transaction graph and can be used to identify blockchain digital financial fraud behaviors and extract corresponding characteristic patterns. For example, sub-topological structures that appear multiple times in the transaction graph are often related to specific fraud behaviors. For instance, a triangular structure is often considered a money laundering behavior. For example, as Figure 9 shown, the central node of the transaction subgraph is m9, and the central node can iterate to any ordinary node.

[0065] It can be understood that the present invention simplifies and abstracts the transactions of the address graph to obtain a transaction subgraph, thereby intuitively observing the relationships between fraud behavior nodes and victims, collaborators, and neighbor nodes.

[0066] In addition, if the transaction data is off-chain transaction data, the financial business features include time series features, text features, and image features. Specifically, in traditional financial business, time series data such as stock prices and transaction records contain rich dynamic information. Deep learning algorithms use LSTM or GRU operators in recurrent neural networks (RNNs) to extract the temporal features of this data. Specifically, first, the data is preprocessed, including steps such as missing value filling, normalization, and creation of time windows. Then, the data at each time step is input into the LSTM or GRU operator, and through non-linear transformation, a hidden state vector containing current and historical information is generated as a local feature representation. As multiple sequences are input into the operator, these features accumulate to form a global feature representation. When used, local or global features will be extracted according to specific circumstances.

[0067] For traditional financial business involving a large amount of text information, such as news reports, company announcements, and customer reviews, deep learning can provide powerful text feature extraction tools. Among them, the extraction of text features uses the Encoder operator as the core component, and the input text data is processed by stacking multiple layers of encoders. Specifically, the Encoder operator takes the text related to financial transactions as input and uses its internal self-attention mechanism to capture the long-term dependencies and context information between words, thereby generating a word vector representation rich in semantics. In addition, for the need of fraud recognition, the parameters of the Encoder operator can be further optimized through fine-tuning to obtain more accurate text feature extraction capabilities. In addition, by combining traditional methods such as TF-IDF and Word2Vec, key features in financial texts can be extracted more comprehensively.

[0068] In addition, in some financial scenarios, image data also plays an important role, such as check scans, signature verification, and bill review. The convolution operator is the main tool for image feature extraction. It gradually abstracts the spatial hierarchical features of the image through a series of convolution calculations, pooling layers, and activation functions. Specifically, the convolution operator performs local operations on the input image through a sliding window (i.e., the convolution kernel), capturing spatial patterns. The convolution kernel is a small matrix, such as 3x3 or 5x5 in size. It multiplies and sums the pixel values in the corresponding area of the image element by element to obtain a value in the output feature map. Combined with a non-linear activation function (such as ReLU), the network can learn complex feature patterns. The convolution operator reduces the model parameters through parameter sharing, improving the computational efficiency. To enhance the feature representation and reduce overfitting, the convolution operator uses a pooling layer (such as max pooling or average pooling) to reduce the data dimension and retain important information. Therefore, a well-designed convolution operator can effectively transform from the original image to abstract features, supporting the extraction of various types of financial image data.

[0069] In addition, for fraud in blockchain-based digital financial services, the extracted financial service features include graph statistical features and graph embedding features. Among them, the graph statistical features include transaction node features, address node features, contract features, and subgraph features, and the graph embedding features include node embedding features and whole graph embedding features. It can be understood that by extracting the graph statistical features and graph embedding features of the transaction information graph, the present invention can characterize the common features of fraud behavior, which is beneficial to improving the generalization of the deep learning model.

[0070] The model layer of the system includes multiple pre-trained deep learning models, such as graph neural network models, convolutional neural network models, recurrent neural network models, Transformer models, GPT models, BERT models, etc. After the output of the feature layer is input into the model layer, for blockchain money laundering behavior, the model layer can calculate graph statistical features (such as degree centrality, clustering coefficient) to identify abnormally active or isolated accounts, use graph neural network (GNN) to learn low-dimensional embedding representations, capture complex relationship patterns, introduce supervision signals and binary classification loss functions to optimize model parameters, and use adversarial training to simulate new money laundering methods. The model layer can monitor suspicious transactions in real time, reveal hidden money laundering network structures, provide support for regulatory agencies, and enhance the transparency and security of the financial system. Additionally, for illegal token issuance behavior, the model layer can also collect text information such as project white papers and code libraries, use the BERT model to extract semantic features, identify false propaganda, conduct static analysis based on smart contract code, check for malicious logic, construct a token issuance network graph, evaluate the association strength through the graph embedding features output by the feature layer, and can also combine time series data (such as price fluctuations, trading volume changes) to enrich the feature space, and adopt a hybrid loss function to optimize classification accuracy and regression error, so as to effectively distinguish between legal and illegal token issuances, protect the interests of investors, and combat illegal financial activities. Additionally, for Ponzi schemes, the model layer can also construct a fund flow graph, analyze the time distribution, frequency, and scale, identify abnormal circular paths, use graph embedding technology to map the roles and influences of participants, combine text features such as project descriptions, mine risk signals with NLP, use a multi-task learning framework to optimize classification and regression tasks simultaneously, predict development trends, and introduce historical case transfer learning to improve the adaptability to new Ponzi schemes, so as to provide early warnings and prevent the spread of the scheme, ensure the stability of the financial market, and protect investors from being defrauded by Ponzi schemes. Additionally, for phishing behavior, the model layer can also extract time series features (such as trading frequency, amount size) from blockchain transaction records to identify abnormal patterns, use GNN to construct a user social network graph, calculate graph embedding features to evaluate interaction patterns, apply convolutional neural network (CNN) to analyze phishing website screenshots or URL links, identify forged interfaces, analyze the content of phishing emails, extract text features to judge inducement language, and adopt a multi-modal fusion strategy to integrate different data sources to improve detection accuracy, so as to monitor user behavior in real time, give timely warnings to prevent falling into phishing traps, and protect asset security.

[0071] In addition, for fraud in traditional financial services, the extracted financial service features include time series features, text features, and image features. The model layer can detect abnormal activities by analyzing the time series features and behavior patterns of transaction data, use recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to capture the temporal relationships between transactions, and use autoencoders to identify abnormal transactions based on reconstruction errors. By combining with random forests to screen important features, it can evaluate the risk of each transaction in real time, respond quickly, and minimize losses, significantly improving the efficiency and accuracy of credit card fraud detection. Additionally, the model layer can also utilize natural language processing (NLP) techniques and graph neural networks (GNNs) to parse claim documents and construct user association graphs, understand semantic content through pre-trained language models such as BERT to identify suspicious expressions, use GNNs to detect implicit gang fraud behaviors, and combine convolutional neural networks (CNNs) with RNNs to process mixed image and text data to comprehensively evaluate the authenticity of claim materials. After model optimization, it is deployed in the insurance company's system to assist in efficient and accurate case review, effectively combat insurance fraud, and maintain a healthy industry ecosystem. Moreover, for financial fraud, the model layer can also integrate information from multiple channels, such as personal identity information and bank statements, to form a comprehensive customer profile, apply the Transformer architecture to process large-scale text data, capture nuances in complex contexts, introduce OCR technology and NLP to parse the text content in scanned documents, verify their consistency and authenticity, use cross-validation and AUC values to evaluate the performance during model training to ensure generalization ability, and the system monitors new loan applications in real time and issues early warnings in a timely manner to help financial institutions reduce credit risks, prevent financial loan fraud, and ensure the security of funds. Furthermore, for financial statement fraud, the model layer uses RNNs to model the changing patterns of financial indicators over time, combines attention mechanisms to highlight key information, and a multi-modal learning framework integrates various forms of data such as text, tables, and charts to provide a comprehensive risk assessment perspective, analyzes the text content of management discussions to check whether it matches the financial figures, the model focuses on the false negative rate in the confusion matrix to prevent missed reports, and after deployment, it assists auditors in preliminary screening, continuously learns new cases, improves recognition accuracy, and provides reliable decision-making support for investors. Among them, the recognition principles of the deep learning algorithms in the present invention for various financial fraud behaviors belong to the prior art and will not be elaborated here.

[0072] It can be understood that a digital financial fraud recognition platform can be constructed according to the digital financial fraud recognition method based on the consortium chain and deep learning of the present invention. The system architecture of this platform is as Figure 10As shown. The system architecture of the digital financial fraud identification platform includes data layer, network layer, consensus layer, smart contract layer, feature layer, model layer and application layer. The data layer is used to collect off-chain transaction data, on-chain transaction data and marked fraud behavior data of financial institutions, core enterprises, small and medium-sized enterprises and supporting enterprises in the supply chain. The network layer is used to broadcast the data of the data layer across the entire network using alliance chain technology to achieve data sharing. The consensus layer is a key mechanism to ensure that all nodes on the alliance chain reach a consensus on the order and status of digital financial transactions. The smart contract layer is used to automatically execute, verify or enforce contract terms without intermediaries, update in real time, and prevent tampering. The feature layer is used to extract features of various financial data using deep learning technology, including graph statistical features, graph embedding features, time series features, text features, image features, etc. The model layer uses the features extracted by the feature layer to train the deep learning model to improve the generalization of the model. The application layer identifies digital financial fraud behaviors based on the trained model, such as fraud behaviors in blockchain-based financial business models, including blockchain money laundering prevention, illegal token issuance detection, Ponzi scheme warning and phishing behaviors, etc.; and fraud behaviors in traditional financial business models, including credit card fraud identification, insurance fraud identification, financial loan fraud identification and financial statement fraud identification. In addition, the platform's business processes are as follows: Figure 11 As shown in the figure, it shows the entire business process from data reading to fraud identification, including the following:

[0073] 1) The system receives data from different sources, including off-chain transaction data and on-chain transaction data. These data enter the system through API interfaces or data transmission protocols (such as P2P networks).

[0074] 2). After receiving the data, the data layer initiates a data write request to the blockchain network. This process ensures the security and integrity of the data.

[0075] 3). The data writing process is controlled by smart contracts to ensure that the data is stored on the blockchain according to predefined rules. Smart contracts verify the validity and format of the data.

[0076] 4). The data is written into the blockchain to form an unalterable record. Each block contains multiple transaction records to ensure the transparency and security of the data.

[0077] 5). Perform hash calculation on the written data to generate a unique hash value. This step ensures the uniqueness and tamper-proofness of the data.

[0078] 6). All nodes synchronize the latest block data and maintain consistency. This step ensures that all participating nodes have the same blockchain status, thereby ensuring the reliability and consistency of the system.

[0079] 7). Achieve consensus through consensus algorithms (such as PBFT, Raft, etc.), confirm the validity of transactions. The consensus mechanism ensures that all nodes recognize the consistency of data and synchronize the data to all nodes.

[0080] 8). Through the permission control mechanism, ensure that only authorized parties can perform specific operations. The smart contract automatically executes the transaction logic, ensures that the transaction proceeds according to the predetermined rules, and calculates the Gas fees required for the transaction to provide resources for the execution of the smart contract.

[0081] 9). Extract key statistical features from the transaction information graph for subsequent analysis. These features include graph statistical features, graph embedding features, image features, time series features, text features, etc., for more in-depth analysis.

[0082] 10). Use deep learning algorithms (such as graph neural network GNN, convolutional neural network CNN, and recurrent neural network RNN) to train the extracted features to identify potential fraud patterns. GNN captures the complex associations between nodes by constructing a transaction network graph and identifies hidden gang behaviors and abnormal paths; CNN can capture spatial structure features and is suitable for image and sequence data; RNN can process time series data and capture long-term dependencies.

[0083] 11). The final model identifies whether it belongs to blockchain-based financial business fraud or credit card fraud, insurance fraud, financial loan fraud, or financial statement fraud in traditional financial business, etc., and returns the identification result. The system feeds back the identification result to the application layer for decision-makers to refer to, thus achieving efficient fraud behavior identification.

[0084] In addition, the above business process can be described by the following algorithm. Such as Figure 12As shown, the system input is the on-chain financial transaction data (OnChainData) and off-chain financial transaction data (OffChainData) in the data layer, corresponding to two types of financial services respectively, and the system output is the fraud recognition result. In the algorithm, flagDB is used to control the writing of off-chain data to the database, with an initial value of False, and by default, it is the on-chain transaction data of the digital financial service based on the blockchain itself; and flagLedger is used to control whether to read on-chain financial data or off-chain financial data, with an initial value of OnChainData, that is, by default, read on-chain financial data. First, the type of financial service is judged. If it is an on-chain transaction, flagDB remains the default value; if it is an off-chain transaction, the off-chain financial data is collected. When the collected off-chain transaction data reaches a certain amount, flagDB becomes True. Then, the collected off-chain financial data is written into the database, the storage address of the written data and the calculated hash value are obtained, and the relevant data is uploaded to the blockchain ledger, and at the same time, flagDB is changed to False. Then, the financial data is processed according to the value of the flagLedger variable. If the value of flagLedger is OnChainData, the financial data at this time is the on-chain financial data read; if the value of flagLedger is OffChainData, it means that the financial data at this time is off-chain financial data, and it is necessary to further judge whether the data is obtained correctly. Specifically, first, the hash value and storage address corresponding to the off-chain financial data are read from the blockchain ledger, and the financial data is read from the off-chain database according to the address and the hash value is calculated. Then, the two hash values are compared. If they are equal, it means that the data is obtained correctly, otherwise an error is returned. Then, the corresponding graph features (graph embedding features and graph statistical features), time series features, text features, and image features are extracted from the obtained financial data. Then, the algorithm trains the deep learning algorithm model based on the extracted features and identifies the result. This algorithm uses a while loop to control the execution of the program, and the algorithm stops only when an error is sent.

[0085] In addition, as Figure 13 shown, another embodiment of the present invention further provides a digital financial fraud recognition system based on a consortium blockchain and deep learning, preferably adopting the digital financial fraud recognition method described above, including:

[0086] A transaction data receiving module, configured to receive transaction data, where the transaction data is on-chain transaction data or off-chain transaction data;

[0087] A transaction data writing module, configured to write the transaction data into the consortium blockchain and reach a consensus among all nodes;

[0088] A financial fraud identification module is used to extract corresponding financial business features from transaction data according to the type of transaction data, and input the extracted financial business features into a pre-trained deep learning model to identify whether a transaction behavior belongs to a financial fraud behavior, and identify whether the transaction behavior belongs to a fraud behavior in traditional financial business or a fraud behavior in blockchain-based digital financial business.

[0089] It can be understood that the digital financial fraud identification system based on the consortium blockchain and deep learning in this embodiment provides a secure, transparent and tamper-proof financial transaction and data sharing platform based on the consortium blockchain, ensuring the authenticity and traceability of all transaction records, reducing the risk of data forgery and tampering. At the same time, through the powerful data processing and feature extraction capabilities of deep learning, the system can analyze massive, unstructured data from different industries and scenarios, capture the subtle patterns and abnormal signs hidden behind various transaction behaviors, so as to accurately identify digital financial fraud behaviors in two financial business models, and improve the generalization ability and robustness of the financial fraud identification algorithm.

[0090] In addition, another embodiment of the present invention further provides an electronic device, including a processor and a memory. A computer program is stored in the memory, and the processor is used to execute the steps of the method as described above by calling the computer program stored in the memory.

[0091] In addition, another embodiment of the present invention further provides a computer-readable storage medium for storing a computer program for digital financial fraud identification based on the consortium blockchain and deep learning. When the computer program runs on a computer, it executes the steps of the method as described above.

[0092] The forms of computer-readable storage media generally include: floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, any other optical media, punch cards, paper tapes, any other physical media with a pattern of holes, random access memories (RAMs), programmable read-only memories (PROMs), erasable programmable read-only memories (EPROMs), flash erasable programmable read-only memories (FLASH-EPROMs), any other memory chips or cartridges, or any other media readable by a computer. Instructions can further be transmitted or received by a transmission medium. The term "transmission medium" can include any tangible or intangible medium that can be used to store, encode, or carry instructions for execution by a machine, and includes digital or analog communication signals or the intangible medium that facilitates the communication of the above instructions. Transmission media include coaxial cables, copper wires, and optical fibers, which include the wires of a bus used to transmit a computer data signal.

[0093] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0094] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0095] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction means that implements the functions specified in one or more of the flow Figure 1 acts or a plurality of acts and / or boxes Figure 1 specified in one box or a plurality of boxes.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the flow Figure 1 acts or a plurality of acts and / or boxes Figure 1 specified in one box or a plurality of boxes.

[0097] Although the preferred embodiments of the present application have been described, additional changes and modifications can be made by those skilled in the art once they learn the basic inventive concept. Therefore, the appended claims are intended to be construed to cover the preferred embodiments as well as all changes and modifications that fall within the scope of the present application.

[0098] It is apparent that those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

[0099] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A digital financial fraud identification method based on alliance chain and deep learning, characterized in that: Includes the following: Receive transaction data, which can be on-chain transaction data or off-chain transaction data; Write transaction data into the consortium chain and reach consensus among all nodes; The corresponding financial business features are extracted from the transaction data according to the type of transaction data, and the extracted financial business features are input into the pre-trained deep learning model to identify whether the transaction behavior constitutes financial fraud, and to identify whether the transaction behavior constitutes fraud in traditional financial business or fraud in blockchain-based digital financial business.

2. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 1 is characterized in that: Based on smart contracts, the alliance chain creates governance roles, contract administrator roles and user roles. The governance role plays the role of top-level managers in the alliance chain, and is used to establish and maintain the governance rules and order on the chain and is responsible for the election and operation of the governance committee, and makes decisions on major issues through the committee. The contract administrator role is used to define and manage users who can deploy new smart contracts, and which users or entities can call specific interfaces of deployed contracts. User roles participate in the business under the premise of complying with established rules. After the transaction data is uploaded to the chain, the alliance chain conducts preliminary identification of financial fraud on the user's transaction behavior.

3. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 2 is characterized in that: The process of the alliance chain's preliminary identification of financial fraud by users includes the following: After the transaction data is shared on the chain, the governance role elects a governance committee to manage the permissions for deploying contracts. The contract administrator role controls the user access permissions for the deployed contracts. Initially, all user roles are set to whitelist permissions. At this time, the initial value of the early warning mechanism is 0. By collecting and analyzing the long-term behavior patterns of users, a normal behavior baseline is established, and for key operations, users are required to provide additional identity authentication information. The user addresses that deviate from the normal key operation behaviors are recorded at this time, and the early warning mechanism value is increased; The address of the suspicious user is then sent to the governance role, and the suspicious user behavior and its early warning mechanism value are continuously followed up in subsequent transactions. When the early warning mechanism value of the suspicious user reaches the threshold, the governance role freezes the account of the suspicious user, and the contract administrator role sets the account to blacklist permissions; then, the governance role re-proposes to elect a governance committee, continuously updates the contract and iterates, and realizes the initial financial fraud identification process.

4. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 1 is characterized in that: If the transaction data is on-chain transaction data, first build a transaction information graph of the alliance chain. The transaction information graph is a graph network with a topological logical structure based on the transaction data. Different graph networks are used to characterize the relationship between different blockchain entities. At the same time, the corresponding labeled data and off-chain behavior characteristics can be introduced into the corresponding transaction information graph as node attributes or edge attributes, and then the financial business features are extracted from the transaction information graph. The extracted financial business features include graph statistical features and graph embedding features. If the transaction data is off-chain transaction data, the financial business features include time series features, text features and image features.

5. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 4 is characterized in that: The transaction information graph includes the address graph, transaction graph, address transaction graph, user graph and transaction sub-graph. The transaction graph and address graph provide corresponding transaction information and address information for the address transaction graph. The user graph extracts transaction patterns based on the address transaction graph, and the transaction sub-graph extracts entity relationships based on the address transaction graph.

6. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 5 is characterized in that: The address transaction graph is a combination of the address graph and the transaction graph. It is used to describe the correspondence between transactions and addresses. Directed edges are used to connect input addresses, transactions, and output addresses. The weight value of the directed edge represents the transaction amount.

7. The digital financial fraud identification method based on alliance chain and deep learning as claimed in claim 5 is characterized in that: The user graph is represented by G={U,M,E}, which is used to describe the flow of digital currency between clustered users in the blockchain. U represents the user set, each user is a user abstracted by clustering, M represents the address set, E represents the transaction set, and the weight value of each transaction edge represents the transaction amount.

8. A digital financial fraud identification system based on alliance chain and deep learning, characterized in that: include: A transaction data receiving module is used to receive transaction data, where the transaction data is on-chain transaction data or off-chain transaction data; The transaction data writing module is used to write transaction data into the alliance chain and reach consensus among all nodes; The financial fraud identification module is used to extract corresponding financial business features from the transaction data according to the type of transaction data, and input the extracted financial business features into the pre-trained deep learning model to identify whether the transaction behavior is financial fraud, and identify whether the transaction behavior is fraud in traditional financial business or fraud in blockchain-based digital financial business.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the processor executes the steps of the method according to any one of claims 1 to 7 by calling the computer program stored in the memory.

10. A computer-readable storage medium for storing a computer program for digital financial fraud identification based on alliance chain and deep learning, characterized in that: When the computer program is run on a computer, the steps of the method according to any one of claims 1 to 7 are executed.

Images