Transaction risk assessment method and device, program product and electronic equipment

Abstract

The invention discloses a transaction risk assessment method and device, a program product and electronic equipment, and relates to the technical field of federal learning, and the method comprises the steps: collecting the transaction information of a financial transaction initiated by a user after the authorization information of the user is received; transmitting the transaction information of the financial transaction to a local target model of the target website; performing feature extraction on the transaction information of the financial transaction through a local target model to obtain L transaction features; determining a risk score of the financial transaction based on the L transaction features; and when the risk score is greater than or equal to a preset score, generating a risk report corresponding to the financial transaction, the risk report at least comprising the risk score, the risk type and the risk transaction link of the financial transaction. The technical problem that in the prior art, due to poisoning attack in the modeling process of a federated model, the accuracy of transaction risk detection through a local model obtained through training based on the federated learning technology is low is solved.

Description

Technical Field

[0001] This application relates to the field of federal learning technology, and more specifically, to a method, apparatus, program product, and electronic device for assessing transaction risk. Background Technology

[0002] With the accelerating pace of globalization, financial institutions are opening more and more branches in different regions. As business scenarios become increasingly diverse, the local data of a single branch is often insufficient to meet the data usage needs in certain scenarios due to its limited volume. In such cases, it is necessary to combine data from other branches for joint use. Against this backdrop, federated learning, as a distributed machine learning technology, has become an ideal solution for financial institutions to achieve cross-regional data cooperation.

[0003] Federated learning technology allows multiple devices to collaboratively train models without leaking local data. While ensuring privacy, it also boasts high computational efficiency and good scalability, making it widely used in many data-sensitive fields.

[0004] In practical applications, participants in federated learning face interference from third parties who attempt to manipulate model training results to achieve their attack objectives. This is particularly true in horizontal federated learning, where the similarity of participants in data dimensions makes poisoning attacks launched by malicious devices more covert, threatening the quality of the trained model and the security of the entire federated learning ecosystem. This results in the technical problem of low accuracy in detecting transaction risks in financial transactions by local models trained based on federated learning technology.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a method, apparatus, program product, and electronic device for assessing transaction risk, in order to at least solve the technical problem in the prior art where the accuracy of local models trained based on federated learning technology for detecting transaction risk is low due to poisoning attacks during the modeling process of federated models.

[0007] According to one aspect of this application, a method for assessing transaction risk is provided, comprising: after receiving authorization information from a user, collecting transaction information of a financial transaction initiated by the user; transmitting the transaction information of the financial transaction to a local target model of a target branch, wherein the model parameters corresponding to the local target model and the global target model are the same, the global target model is a federated model trained on transaction information of historical transactions in P financial branches based on federated learning technology, where P is a positive integer, and the target branch is one of the P financial branches; extracting features from the transaction information of the financial transaction through the local target model to obtain L transaction features, where L is a positive integer; determining a risk score for the financial transaction based on the L transaction features; and generating a risk report corresponding to the financial transaction if the risk score is greater than or equal to a preset score, wherein the risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0008] Optionally, the training steps of the global target model include: iteratively training the local initial model of each financial outlet based on the transaction information of historical transactions in each financial outlet to obtain a local training model; transmitting the gradients of the P local training models corresponding to the P financial outlets to the central server in a private sharing manner; aggregating the gradients of the P local training models through the central server to obtain the global gradient; updating the model parameters of the global initial model based on the global gradient to obtain the global training model; broadcasting the private share of the model parameters of the global training model to the P financial outlets; updating the model parameters of the local training model of each financial outlet based on the private share of the model parameters of the global training model to obtain a new local training model; and determining the global target model based on the P new local training models corresponding to the P financial outlets.

[0009] Optionally, based on the P new local training models corresponding to the P financial outlets, a global target model is determined, including: transmitting the gradients of the P new local training models corresponding to the P financial outlets to a central server in a private sharing manner; aggregating the gradients of the P new local training models through the central server to obtain a new global gradient; iteratively updating the model parameters of the global training model based on the new global gradient until the number of training iterations of the global training model reaches a preset maximum number of iterations / the global training model converges; and using the global training model obtained from the last update as the global target model.

[0010] Optionally, before iteratively training the local initial model for each financial outlet, the method for assessing transaction risk further includes: performing an initialization operation through a central server to obtain a global initial model, wherein the central server includes at least a first device, a second device, and a third device, the first and second devices are used to interact with P financial outlets in a data encryption manner, and the third device is used to generate the multiplication triples used for data encryption; broadcasting the private share of the model parameters of the global initial model to the P financial outlets; and establishing a local initial model for each financial outlet based on the private share of the model parameters of the global initial model.

[0011] Optionally, the gradients of the P locally trained models are aggregated through a central server to obtain the global gradient, including: determining the weight of each gradient in the gradients of the P locally trained models through the central server; and aggregating the gradients of the P locally trained models based on the weight of each gradient using matrix-vector multiplication to obtain the global gradient.

[0012] Optionally, the weight of each gradient in the gradients of the P locally trained models is determined through a central server, including: controlling a first device and a second device in the central server to each generate a random vector; encrypting the gradients of the P locally trained models based on their respective generated random vectors using the first and second devices to obtain P encrypted gradient vectors, and transmitting the P encrypted gradient vectors to a third device in the central server; determining the median of the P encrypted gradient vectors in each vector dimension using the third device to obtain the median gradient vector corresponding to the P encrypted gradient vectors, and transmitting the median gradient vector to the first and second devices in a private sharing manner; determining the Pearson correlation coefficient between the median gradient vector and each encrypted gradient vector using the first and second devices to obtain P Pearson correlation coefficients corresponding to the P encrypted gradient vectors; taking the derivative of the P Pearson correlation coefficients using the third device to obtain P derivative values ​​corresponding to the P Pearson correlation coefficients; and determining the weight of the gradient corresponding to each derivative value based on the P derivative values.

[0013] Optionally, the data transmitted via private sharing includes at least: a private share and a verification share, wherein the private share is used to reconstruct the private value of the transmitted data, and the verification share is used to verify the integrity of the private value.

[0014] According to another aspect of this application, a transaction risk assessment device is also provided, comprising: an information collection unit, used to collect transaction information of a financial transaction initiated by a user after receiving authorization information from a user; an information transmission unit, used to transmit the transaction information of the financial transaction to a local target model of a target network, wherein the model parameters corresponding to the local target model and the global target model are the same, the global target model is a federated model trained on the transaction information of historical transactions in P financial networks based on federated learning technology, where P is a positive integer, and the target network is one of the P financial networks; a feature extraction unit, used to extract features from the transaction information of the financial transaction through the local target model to obtain L transaction features, where L is a positive integer; a risk score determination unit, used to determine the risk score of the financial transaction based on the L transaction features; and a risk report generation unit, used to generate a risk report corresponding to the financial transaction if the risk score is greater than or equal to a preset score, wherein the risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0015] According to another aspect of this application, a computer program product is also provided, which stores a computer program, wherein, when the computer program is running, the computer program product is controlled to perform the transaction risk assessment method of any of the above.

[0016] According to another aspect of this application, an electronic device is also provided, wherein the electronic device includes one or more processors and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the transaction risk assessment method of any of the above.

[0017] In this application, after receiving the user's authorization information, the transaction information of the financial transaction initiated by the user is first collected. Then, the transaction information is transmitted to the local target model of the target branch. The local target model has the same model parameters as the global target model. The global target model is a federated model trained using federated learning technology on the transaction information of historical transactions in P financial branches, where P is a positive integer, and the target branch is one of the P financial branches. Next, the local target model extracts features from the transaction information of the financial transaction, obtaining L transaction features, where L is a positive integer. Subsequently, based on the L transaction features, the risk score of the financial transaction is determined. If the risk score is greater than or equal to a preset score, the application generates a risk report corresponding to the financial transaction. The risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0018] As described above, this application employs federated learning to break down data isolation between financial institutions in different regions without directly sharing their local data (i.e., transaction information). Furthermore, by training data from multiple financial institutions using federated learning technology, this application aims to improve the accuracy of financial institutions' assessment of transaction risks. In addition, during the federated training process, this application utilizes private sharing technology and pre-defined encryption methods to maintain data privacy while enhancing robustness against poisoning attacks. This solves the technical problem in existing technologies where poisoning attacks during the modeling process of federated models lead to low accuracy in detecting transaction risks using local models trained using federated learning technology. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for assessing transaction risk, according to an embodiment of this application.

[0021] Figure 2 This is a flowchart of an optional method for assessing transaction risk according to an embodiment of this application;

[0022] Figure 3 This is a flowchart of an optional federated learning-based method for resisting poisoning attacks according to an embodiment of this application;

[0023] Figure 4 This is a schematic diagram of an optional transaction risk assessment device according to an embodiment of this application;

[0024] Figure 5 This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] It should also be noted that all relevant information (including but not limited to transaction information of collected financial transactions) and data (including but not limited to data used for display and analysis) involved in this application are information and data authorized by the user or fully authorized by all parties. For example, this system has an interface with relevant users or institutions. Before obtaining relevant information, it needs to send an acquisition request to the aforementioned user or institution through the interface, and obtain the relevant information after receiving consent from the aforementioned user or institution.

[0028] Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of relevant information and data involved in this application all comply with the relevant laws, regulations, and standards of the relevant regions, and necessary confidentiality measures have been taken. This application does not violate public order and good morals. In addition, this application provides a corresponding operation entry point for users to choose to agree to or refuse authorization. If the user chooses to refuse authorization, the corresponding expert decision-making process will be initiated.

[0029] The present invention will now be described in detail with reference to various embodiments.

[0030] Example 1

[0031] According to an embodiment of this application, an embodiment of a method for assessing transaction risk is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0032] The methods and embodiments provided in this application can be executed on mobile terminals, computer terminals, or similar computing devices. Figure 1 This is a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for assessing transaction risk, according to an embodiment of this application. Figure 1As shown, the computer terminal 10 (or mobile device) may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0033] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0034] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the transaction risk assessment method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned transaction risk assessment method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0035] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0036] The display can be configured as a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0037] Under the aforementioned operating environment, this application provides a transaction risk assessment system (hereinafter referred to as the assessment system) for implementing the transaction risk assessment method described in this application. Figure 2 This is a flowchart of an optional transaction risk assessment method according to an embodiment of this application, such as... Figure 2 As shown, the method includes the following steps:

[0038] Step S201: After receiving the user's authorization information, collect the transaction information of the financial transaction initiated by the user.

[0039] Optionally, authorization information refers to the formal permission granted by a user to a financial institution to access and process its transaction data.

[0040] Optionally, transaction information refers to the details of a financial transaction, such as the transaction amount, transaction time, transaction location, and the identity information of the parties involved.

[0041] Optionally, step S201 ensures that the transaction information collection process complies with the data protection agreements of different financial outlets through the authorization information confirmation mechanism, thereby protecting the security of transaction information and improving the compliance of the assessment system.

[0042] Step S202: Transmit the transaction information of the financial transaction to the local target model of the target network. The local target model has the same model parameters as the global target model. The global target model is a federated model trained on the transaction information of historical transactions in P financial networks based on federated learning technology. P is a positive integer and the target network is one of the P financial networks.

[0043] Optionally, the local target model refers to the local model trained locally by each financial outlet that can be used for user credit scoring, transaction risk detection, or transaction link tracing. Here, it specifically refers to the local model used for transaction risk detection.

[0044] Optionally, the model parameters of the local target model are determined by the model parameters of the global target model. In the evaluation system, the model parameters of the global target model can be constructed based on the local models of each financial outlet using federated learning technology, without obtaining the original local data of each of the P financial outlets.

[0045] Optionally, step S202 utilizes federated learning technology, with the help of a central processor, to update model parameters in a private sharing manner while the local data of each financial branch is in a data isolation state, thereby avoiding the risk of local data leakage and improving the data security of the local data of each financial branch.

[0046] Step S203: Extract features from the transaction information of financial transactions using the local target model to obtain L transaction features, where L is a positive integer.

[0047] Optionally, the evaluation system can simplify the complex raw transaction information into a feature vector expression by using a local target model to extract features from the transaction information of financial transactions. This helps the local model to quickly calculate the transaction risk score, avoids the computational pressure caused by directly processing a large amount of raw local data, and improves the speed of transaction risk assessment.

[0048] Step S204: Determine the risk score of the financial transaction based on L transaction characteristics.

[0049] Optionally, risk scoring refers to the quantitative value of the transaction risk obtained by analyzing the transaction characteristics through a local model of the financial outlet where the user who initiated the financial transaction is located.

[0050] Optionally, by generating risk scores, the assessment system can further implement different risk management measures based on different levels of risk, thereby achieving refined and automated risk classification management.

[0051] Step S205: If the risk score is greater than or equal to the preset score, generate a risk report corresponding to the financial transaction. The risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0052] Optionally, if the risk score is greater than or equal to the preset score, the assessment system generates a risk report corresponding to the financial transaction, providing financial institutions with a real-time risk warning mechanism, which helps to prevent potential financial risk behaviors in a timely manner and improves the security of users' financial accounts.

[0053] As described above, this application employs federated learning to break down data isolation between financial institutions in different regions without directly sharing their local data (i.e., transaction information). Furthermore, by training data from multiple financial institutions using federated learning technology, this application aims to improve the accuracy of financial institutions' assessment of transaction risks. In addition, during the federated training process, this application utilizes private sharing technology and pre-defined encryption methods to maintain data privacy while enhancing robustness against poisoning attacks. This solves the technical problem in existing technologies where poisoning attacks during the modeling process of federated models lead to low accuracy in detecting transaction risks using local models trained using federated learning technology.

[0054] In one optional embodiment, to train the global target model, the evaluation system first iteratively trains the local initial model of each financial outlet based on the transaction information of historical transactions in each financial outlet, obtaining a local training model. Then, the evaluation system transmits the gradients of the P local training models corresponding to the P financial outlets to the central server in a private sharing manner. Next, the evaluation system aggregates the gradients of the P local training models through the central server to obtain the global gradient. Subsequently, the evaluation system updates the model parameters of the global initial model based on the global gradient, obtaining the global training model. Then, the evaluation system broadcasts the private share of the model parameters of the global training model to the P financial outlets. Next, the evaluation system updates the model parameters of the local training model of each financial outlet based on the private share of the model parameters of the global training model, obtaining a new local training model. Finally, the evaluation system determines the global target model based on the P new local training models corresponding to the P financial outlets.

[0055] Optionally, historical transactions refer to financial transactions that occurred at a financial outlet within a historical time period.

[0056] Optionally, the model parameters of the local initial model are determined by the model parameters of the global initial model obtained by the central processing unit performing initialization operations.

[0057] Optionally, the evaluation system iteratively trains the local initial model based on local data, using the stochastic gradient descent algorithm during the iterative training process. This enables the model to quickly learn and adapt to the specific transaction behavior patterns of each financial outlet, thereby improving the training efficiency of the locally trained model.

[0058] Optionally, gradient refers to the direction and magnitude of the adjustment of model parameters.

[0059] Optionally, the evaluation system transmits each gradient from the gradients of the P locally trained models to a first device and a second device in a private sharing format to the central server. Then, the first device and the second device each obtain a private share of each gradient. The private share of the gradient refers to the encrypted and segmented gradient information. Through the data transmission format of the private share, even if the gradient information is intercepted by an attacker, the attacker cannot directly obtain the local data and the original gradient value, thereby enhancing the security of the gradient information during transmission.

[0060] Optionally, the aggregation operation refers to the process by which the central server merges the gradient information uploaded by all financial outlets to determine the new update direction of the model parameters. The preferred aggregation protocol used in this process is a robust aggregation protocol based on the median and Pearson correlation coefficient.

[0061] Optionally, the evaluation system uses the global gradient obtained through aggregation operations to integrate the data characteristics and transaction patterns of different financial outlets, thereby enhancing the generalization ability of the global target model. This enables the final trained global target model to more accurately identify and predict cross-outlet transaction risks.

[0062] Optionally, the evaluation system ensures the security of model parameters by broadcasting a private share of the model parameters of the globally trained model to P financial outlets, thereby preventing unauthorized third-party parsing of the model parameters during transmission among federation members.

[0063] Optionally, the evaluation system updates the model parameters of the local training model of each financial outlet based on the private share of the model parameters of the global training model, so that the local model can be iteratively improved in sync with the global model. Since the construction process of the global model integrates the data characteristics and transaction patterns of different financial outlets, using the local model obtained in sync with the global model to conduct transaction risk assessment improves the accuracy of transaction risk assessment.

[0064] In summary, through the above steps, the evaluation system can not only enhance the privacy protection and data security of the model, but also promote the performance improvement of the local model and the ability to identify risky transactions across branches without directly sharing the local transaction information of each financial branch.

[0065] In one optional embodiment, in the process of determining the global target model based on P new local training models corresponding to P financial outlets, the evaluation system first transmits the gradients of the P new local training models corresponding to the P financial outlets to the central server in a private sharing manner; then, the evaluation system aggregates the gradients of the P new local training models through the central server to obtain a new global gradient; then, the evaluation system iteratively updates the model parameters of the global training model based on the new global gradient until the number of training iterations of the global training model reaches a preset maximum number of iterations / the global training model converges; finally, the evaluation system uses the global training model obtained from the last update as the global target model.

[0066] Optionally, the central server refers to the server that plays a coordinating and aggregating role in the federated learning architecture. It is responsible for receiving and aggregating gradients, as well as updating and broadcasting model parameters. The central server in the evaluation system adopts a three-tier architecture, including service providers SP0 and SP1 and auxiliary computing servers CS (i.e., the first device, the second device, and the third device).

[0067] Optionally, the evaluation system transmits the gradients of P new locally trained models in a private sharing manner. That is, the private share of each gradient in the generated P new locally trained models is transmitted to service providers SP0 and SP1 respectively. This data transmission method ensures the security of gradient information during transmission and reduces the risk of data leakage of the original gradient information.

[0068] Optionally, the evaluation system aggregates the gradients of P new locally trained models through a central server to obtain a new global gradient. This gradient aggregation enables the model to learn from data from multiple financial outlets, improving the representativeness and generalization ability of the global model, making the model more robust and adaptable to a wider range of trading scenarios and risk types.

[0069] Optionally, by setting a preset maximum number of iterations, the evaluation system can avoid overfitting of the global model due to excessive training time. At the same time, by performing dynamic convergence judgment of the model, the evaluation system can terminate training when the global model reaches the training objective, thereby saving the time cost of training resources for the model.

[0070] In an optional embodiment, before iteratively training the local initial model for each financial outlet, the method for assessing transaction risk further includes: the assessment system first performs an initialization operation through a central server to obtain a global initial model, wherein the central server includes at least a first device, a second device, and a third device, the first and second devices are used to interact with P financial outlets in a data encryption manner, and the third device is used to generate the multiplication triples used for data encryption; then, the assessment system broadcasts the private share of the model parameters of the global initial model to the P financial outlets; then, the assessment system establishes a local initial model for each financial outlet based on the private share of the model parameters of the global initial model.

[0071] Optionally, the first device and the second device, namely service providers SP0 and SP1, are used to receive the private share of the gradient sent by the user, and to perform poisoning attack detection and gradient aggregation on the private share.

[0072] Optionally, a third device, namely the auxiliary computing server CS, is used to assist service providers SP0 and SP1 in performing data calculations.

[0073] Alternatively, a multiplication triple refers to a data structure used for secure multi-party computation, specifically for performing multiplication operations in an encrypted state.

[0074] Optionally, the specific steps of the initialization operation performed by the central server include: First, the auxiliary computing server CS generates the multiplication triples required for private shared computing and sends them to two service providers SP0 and SP1; the two service providers SP0 and SP1 randomly initialize the private share of the model parameters of the global initial model and send it to P local devices corresponding to P financial outlets; each local device establishes a local initial model based on the private share of the received model parameters.

[0075] Optionally, the distributed collaboration and data encryption functions of the first, second, and third devices in the central server improve the security of the model parameters of the global initial model during the distribution phase. The use of data encryption operations and multiplication triples avoids unauthorized third-party access to and use of interactive data, enhances the privacy protection of the data to be transmitted during federated training, and avoids the risk of privacy leakage caused by transmitting interactive data in plaintext.

[0076] Optionally, the evaluation system initializes and builds a local model for each financial outlet by using a private share of the model parameters of the global initial model, enabling each financial outlet to train its local model in parallel, thereby accelerating the entire federated learning process and improving the overall model training efficiency.

[0077] In one optional embodiment, during the aggregation operation of the gradients of P locally trained models, the evaluation system first determines the weight of each gradient in the gradients of the P locally trained models through a central server; then, the evaluation system uses matrix-vector multiplication to aggregate the gradients of the P locally trained models based on the weight of each gradient to obtain the global gradient.

[0078] Optionally, when performing gradient aggregation, the evaluation system uses matrix-vector multiplication, which reduces communication rounds and computational overhead, improving the computational efficiency of global gradients while reducing the computational resources used.

[0079] Optionally, the evaluation system controls the central server to perform weighted aggregation of gradients. The evaluation system can filter out the gradients of local training models that come from high-quality local data and thus exhibit more stable learning behavior. This effectively suppresses poisoning attacks during the aggregation process. The weighting mechanism ensures that benign gradient information dominates the model update, which helps to reduce the negative impact of abnormal local data or poisoning attacks on model training, thereby improving the accuracy of the trained global target model.

[0080] In one optional embodiment, in determining the weight of each gradient among the gradients of P locally trained models, the evaluation system first controls a first device and a second device in the central server to each generate a random vector. Then, the evaluation system, through the first and second devices, encrypts the gradients of the P locally trained models based on their respective generated random vectors, obtaining P encrypted gradient vectors, and transmits these P encrypted gradient vectors to a third device in the central server. Next, the evaluation system, through the third device, determines the median of each of the P encrypted gradient vectors in each vector dimension, obtaining the median gradient vector corresponding to the P encrypted gradient vectors, and transmits the median gradient vector to the first and second devices in a private sharing manner. Subsequently, the evaluation system, through the first and second devices, determines the Pearson correlation coefficient between the median gradient vector and each encrypted gradient vector, obtaining P Pearson correlation coefficients corresponding to the P encrypted gradient vectors. Then, the evaluation system, through the third device, calculates the derivative of the P Pearson correlation coefficients, obtaining P derivative values ​​corresponding to the P Pearson correlation coefficients. Finally, the evaluation system determines the weight of the gradient corresponding to each of the P derivative values.

[0081] Optionally, the evaluation system encrypts the gradient by using random vectors independently generated by the first and second devices, thus obfuscating the true value of the gradient. Even if the gradient is intercepted during transmission, it is difficult for attackers to infer the original gradient information from the encrypted gradient vector. The use of encrypted gradient vectors improves the security of gradient information in the federated learning process.

[0082] Optionally, the evaluation system uses a third device to determine the median of each of the P encrypted gradient vectors in each vector dimension, thereby obtaining the median gradient vector corresponding to the P encrypted gradient vectors. The median gradient vector is a median gradient vector containing obfuscated medians. Subsequently, the correlation coefficient is calculated using the median gradient vector containing obfuscated medians, further reducing the risk of leakage of the original gradient information.

[0083] Optionally, the evaluation system uses the first and second devices to determine the private share of the median gradient vector based on a private sharing method. This ensures that the value of the median gradient vector is not directly exposed to either the first or second device, but is instead stored in a distributed manner through private sharing, further enhancing the security of the original gradient information.

[0084] Optionally, the Pearson correlation coefficient is a statistical indicator that measures the degree of linear correlation between two variables. Its value ranges from -1 to 1. The larger the absolute value of the Pearson correlation coefficient, the stronger the linear relationship.

[0085] In one optional embodiment, the data transmitted in the evaluation system via private sharing includes at least: a private share and a verification share, wherein the private share is used to reconstruct the private value of the transmitted data, and the verification share is used to perform integrity verification on the private value.

[0086] Optionally, for a private value (i.e., data that needs to be transmitted in the form of private sharing), its general private sharing form (i.e., including only the private share) is represented as a, while a verifiable private sharing form (i.e., including both the private share and the verification share) is represented as [a].

[0087] Based on the verifiable private sharing methods provided in the above embodiments, the following formula must be satisfied:

[0088] ;

[0089] ;

[0090] in, , as well as To verify the share, and For private use only.

[0091] Optionally, both service providers SP0 and SP1 hold [the following]. SP0 and SP1 respectively hold , , as well as .

[0092] Optionally, 'a' is generated by the auxiliary computing server CS and is unknown to the two service providers SP0 and SP1. During the data encryption computation performed by the central server based on the privacy protocol, if one service provider attempts to tamper with its own private share, it must simultaneously modify... and Otherwise, the above equation would not hold.

[0093] However, since both service providers SP0 and SP1 knew... The value of , therefore, It cannot be tampered with by a single service provider, and since neither service provider SP0 nor SP1 knows the value of 'a', it cannot be modified correctly. By verifying whether the above formula holds true, the evaluation system can detect whether the private share has been tampered with. Therefore, using verifiable private shares instead of general private shares to complete gradient aggregation operations can further prevent malicious service providers from attacking the gradient aggregation protocol.

[0094] As described above, this application employs federated learning to break down data isolation between financial institutions in different regions without directly sharing their local data (i.e., transaction information). Furthermore, by training data from multiple financial institutions using federated learning technology, this application aims to improve the accuracy of financial institutions' assessment of transaction risks. In addition, during the federated training process, this application utilizes private sharing technology and pre-defined encryption methods to maintain data privacy while enhancing robustness against poisoning attacks. This solves the technical problem in existing technologies where poisoning attacks during the modeling process of federated models lead to low accuracy in detecting transaction risks using local models trained using federated learning technology.

[0095] In one alternative embodiment, Figure 3 This is a flowchart of an optional federated learning-based anti-poisoning attack method according to an embodiment of this application, such as... Figure 3 As shown, the method includes:

[0096] (1) Preprocessing stage: The preprocessing stage is executed only once in the entire federated learning process. First, the auxiliary computing server CS generates the multiplication triples required for private shared computing and sends them to the two service providers SP0 and SP1. Then, the two service providers SP0 and SP1 randomly initialize the private share of the model parameters of the global model and send it to the local devices corresponding to all financial outlets. The local devices of each financial outlet reconstruct the model parameters based on the private share of the received model parameters and establish the initial local model based on the model parameters.

[0097] (2) Local training phase: After all financial outlets receive the private share of the model parameters of the global model, the local devices of the financial outlets train the model locally based on their respective local user transaction data to obtain their respective local training models.

[0098] Unlike traditional machine learning methods, local devices at financial outlets only need to upload the model parameters (such as gradients) of the locally trained model to the central server. After receiving the model parameters of the locally trained model uploaded by all the local devices at financial outlets, the central server will aggregate and update the model parameters (such as gradients) of the global model to obtain a new global training model, and continuously cycle through this training process.

[0099] (3) Gradient aggregation stage:

[0100] After local training is completed on the local equipment at the financial outlet, the gradient information in floating-point representation is updated using a fixed-point representation method to obtain a corresponding integer. Then, in order to prevent faulty equipment from affecting the training process, this method uses a robust aggregation protocol based on the median and Pearson correlation coefficient to perform gradient aggregation. That is, the median gradient corresponding to all gradients is used as the benchmark, and the weight of the gradient in the aggregation is determined according to the Pearson correlation coefficient between the gradient and the benchmark.

[0101] Optionally, due to the benign training model among the P locally trained models (i.e. Figure 3 Since the majority of users are honest (in the context of a well-trained model), the median gradient vector is closer to the gradient vector of a well-trained model. In other words, the correlation between the gradient vector of a well-trained model and the median gradient vector is stronger, whereas that of a poisoned model (i.e., a poorly trained model) is much stronger. Figure 3 The gradient vector of a malicious user (in the context of a median user) will have a lower correlation with the median gradient.

[0102] Based on this, the evaluation system calculates gradient weights using a logarithmic function, encouraging higher divergence in values ​​near the tail. When the correlation coefficient, used to characterize correlation, is less than a certain threshold, the gradient weight is clipped to 0. Since the poisoning gradient sent by the poisoning training model corresponds to a lower correlation coefficient, the weight of the poisoning gradient sent by the poisoning training model will be clipped to 0, thus avoiding the impact of the poisoning gradient sent by the poisoning training model on the federated training process.

[0103] Optionally, the protocols used in the gradient aggregation stage include, but are not limited to: gradient median privacy computation protocol, linear correlation coefficient privacy computation protocol, and gradient aggregation privacy computation protocol.

[0104] Optionally, the gradient median privacy computation protocol is described below:

[0105] (1) Two service providers, SP0 and SP1, each generate a random vector;

[0106] (2) Based on the generated random vector, the two service providers SP0 and SP1 respectively calculate the encrypted gradient vector and send the encrypted gradient vector to the auxiliary computing server CS;

[0107] (3) After the auxiliary computing server CS receives the encrypted gradient vector, it calculates the median of each vector dimension and generates a median gradient vector containing the confused median based on all the calculated medians.

[0108] (4) The auxiliary computing server CS generates a private share based on the median gradient vector containing the confused median, and sends it to the two service providers SP0 and SP1 respectively;

[0109] (5) The two service providers SP0 and SP1 perform a privacy calculation protocol for the linear correlation coefficient based on the privacy share of the received median gradient vector.

[0110] Optionally, after the two service providers SP0 and SP1 obtain the private shares of the gradient and the intermediate gradient with median information, they calculate the Pearson correlation coefficient between the gradient and the intermediate gradient with median information using a linear correlation coefficient privacy computation protocol. Then, after completing the private sharing multiplication calculation, they send the calculated private shares to the auxiliary computation server CS. After the auxiliary computation server CS reconstructs the private shares, it can recover the correlation coefficient from the intermediate data. Next, the gradient aggregation privacy computation protocol is executed, that is, after the auxiliary computation server CS calculates the correlation coefficient, it calculates the gradient update coefficient (i.e., weight), which is used for subsequent gradient aggregation calculation. Finally, the two service providers SP0 and SP1 send the aggregated gradient to the local devices of all financial outlets. The local devices of each financial outlet reconstruct and update their local models for a new round of iterative training.

[0111] In summary, the above-mentioned federated learning-based anti-poisoning attack method has the following advantages:

[0112] (1) By introducing a new system architecture and designing a privacy computing protocol for poisoning detection and gradient aggregation in horizontal federated learning, complex computation steps are reduced and execution efficiency is improved;

[0113] (2) By constructing a three-layer central server, a high-efficiency privacy computing protocol based on private sharing was designed, which achieved the goal of efficiently completing the anti-poisoning attack algorithm and the horizontal federated learning aggregation algorithm without exposing the original gradient;

[0114] (3) By controlling most of the computing operations to be performed on the central server, the computing burden of edge devices with limited resources is reduced. Furthermore, the protocols in the above method are mainly conducted between two service providers, SP0 and SP1. In addition, the communication overhead of local devices is lower than that of CS and SP. Therefore, the increase in the total number of devices caused by adopting a three-tier architecture will not affect the communication cost of a single local device.

[0115] Example 2

[0116] This application embodiment can also provide a transaction risk assessment device. It should be noted that the transaction risk assessment device of this application embodiment can be used to execute the transaction risk assessment method provided in this application embodiment. The transaction risk assessment device provided in this application embodiment will be described below.

[0117] According to an embodiment of this application, an apparatus for implementing the above-described method for assessing transaction risks is also provided. Figure 4This is a schematic diagram of an optional transaction risk assessment device according to an embodiment of this application, such as... Figure 4 As shown, the device includes: an information acquisition unit 401, an information transmission unit 402, a feature extraction unit 403, a risk scoring determination unit 404, and a risk report generation unit 405.

[0118] Optionally, the information collection unit 401 is used to collect transaction information of the financial transaction initiated by the user after receiving the user's authorization information; the information transmission unit 402 is used to transmit the transaction information of the financial transaction to the local target model of the target network, wherein the model parameters corresponding to the local target model and the global target model are the same, and the global target model is a federated model trained on the transaction information of historical transactions in P financial networks based on federated learning technology, where P is a positive integer and the target network is one of the P financial networks; the feature extraction unit 403 is used to extract features from the transaction information of the financial transaction through the local target model to obtain L transaction features, where L is a positive integer; the risk score determination unit 404 is used to determine the risk score of the financial transaction based on the L transaction features; and the risk report generation unit 405 is used to generate a risk report corresponding to the financial transaction if the risk score is greater than or equal to a preset score, wherein the risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0119] In one optional embodiment, the transaction risk assessment device further includes: a first training unit, a first transmission unit, a first aggregation unit, a first update unit, a second transmission unit, a second update unit, and a first determination unit.

[0120] Optionally, the first training unit is used to iteratively train the local initial model of each financial outlet based on the transaction information of historical transactions in each financial outlet to obtain a local training model; the first transmission unit is used to transmit the gradients of the P local training models corresponding to the P financial outlets to the central server in a private sharing manner; the first aggregation unit is used to aggregate the gradients of the P local training models through the central server to obtain the global gradient; the first update unit is used to update the model parameters of the global initial model based on the global gradient to obtain the global training model; the second transmission unit is used to broadcast the private share of the model parameters of the global training model to the P financial outlets; the second update unit is used to update the model parameters of the local training model of each financial outlet based on the private share of the model parameters of the global training model to obtain a new local training model; and the first determination unit is used to determine the global target model based on the P new local training models corresponding to the P financial outlets.

[0121] In an optional embodiment, the transaction risk assessment device further includes: a third transmission unit, a second aggregation unit, a third update unit, and a second determination unit.

[0122] Optionally, the third transmission unit is used to transmit the gradients of the P new local training models corresponding to the P financial outlets to the central server in a private sharing manner; the second aggregation unit is used to perform an aggregation operation on the gradients of the P new local training models through the central server to obtain a new global gradient; the third update unit is used to iteratively update the model parameters of the global training model based on the new global gradient until the number of training iterations of the global training model reaches a preset maximum number of iterations / the global training model converges; and the second determination unit is used to take the global training model obtained from the last update as the global target model.

[0123] In an optional embodiment, the transaction risk assessment device further includes: a first initialization unit, a fourth transmission unit, and a second initialization unit.

[0124] Optionally, the first initialization unit is used to perform an initialization operation through a central server to obtain a global initial model, wherein the central server includes at least a first device, a second device, and a third device. The first and second devices are used to interact with P financial outlets in a data encryption manner, and the third device is used to generate the multiplication triples used for data encryption. The fourth transmission unit is used to broadcast the private share of the model parameters of the global initial model to the P financial outlets. The second initialization unit is used to establish a local initial model for each financial outlet based on the private share of the model parameters of the global initial model.

[0125] In one alternative embodiment, the first aggregation unit includes: a first determining subunit and an aggregation subunit.

[0126] Optionally, the first determining subunit is used to determine the weight of each gradient in the gradients of the P locally trained models through a central server; the aggregation subunit is used to aggregate the gradients of the P locally trained models based on the weight of each gradient using matrix-vector multiplication to obtain the global gradient.

[0127] In one optional embodiment, the first determining subunit includes: a random vector generation module, a gradient encryption module, a median gradient determination module, a correlation coefficient determination module, a derivative calculation module, and a weight determination module.

[0128] Optionally, the random vector generation module controls the first and second devices in the central server to each generate a random vector; the gradient encryption module encrypts the gradients of P locally trained models based on the random vectors generated by the first and second devices, obtaining P encrypted gradient vectors, and transmits the P encrypted gradient vectors to the third device in the central server; the median gradient determination module determines the median of the P encrypted gradient vectors in each vector dimension through the third device, obtaining the median gradient vector corresponding to the P encrypted gradient vectors, and transmits the median gradient vector to the first and second devices in a private sharing manner; the correlation coefficient determination module determines the Pearson correlation coefficient between the median gradient vector and each encrypted gradient vector through the first and second devices, obtaining P Pearson correlation coefficients corresponding to the P encrypted gradient vectors; the derivative module calculates the derivative of the P Pearson correlation coefficients through the third device, obtaining P derivative values ​​corresponding to the P Pearson correlation coefficients; and the weight determination module determines the weight of the gradient corresponding to each derivative value among the P derivative values.

[0129] In one optional embodiment, the data transmitted in the transaction risk assessment device via private sharing includes at least: a private share and a verification share, wherein the private share is used to reconstruct the private value of the transmitted data, and the verification share is used to verify the integrity of the private value.

[0130] As described above, this application employs federated learning to break down data isolation between financial institutions in different regions without directly sharing their local data (i.e., transaction information). Furthermore, by training data from multiple financial institutions using federated learning technology, this application aims to improve the accuracy of financial institutions' assessment of transaction risks. In addition, during the federated training process, this application utilizes private sharing technology and pre-defined encryption methods to maintain data privacy while enhancing robustness against poisoning attacks. This solves the technical problem in existing technologies where poisoning attacks during the modeling process of federated models lead to low accuracy in detecting transaction risks using local models trained using federated learning technology.

[0131] It should be noted that the information acquisition unit 401, information transmission unit 402, feature extraction unit 403, risk scoring determination unit 404, and risk report generation unit 405 mentioned above correspond to steps S201 to S205 in the method embodiment. The instances and application scenarios implemented by the above units and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above modules can also be part of the device and run in the computer terminal 10 provided in the embodiment.

[0132] Example 3

[0133] Embodiments of this application can also provide an electronic device. Figure 5 This is a structural block diagram of an electronic device according to an embodiment of this application, such as... Figure 5 As shown, the electronic device includes: one or more ( Figure 5 (Only one is shown) processor 502, memory 504, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0134] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and devices in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the above-mentioned method for assessing transaction risks.

[0135] The memory may include high-speed random access memory (RAM), and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0136] The processor can access information and applications stored in memory via a transmission device to execute the following steps: After receiving authorization information from the user, it collects transaction information of the financial transaction initiated by the user; it transmits the transaction information of the financial transaction to the local target model of the target network, wherein the model parameters of the local target model and the global target model are the same, and the global target model is a federated model trained on the transaction information of historical transactions in P financial networks based on federated learning technology, where P is a positive integer and the target network is one of the P financial networks; it extracts features from the transaction information of the financial transaction through the local target model to obtain L transaction features, where L is a positive integer; it determines the risk score of the financial transaction based on the L transaction features; if the risk score is greater than or equal to a preset score, it generates a risk report corresponding to the financial transaction, wherein the risk report includes at least the risk score, risk type, and risk transaction link of the financial transaction.

[0137] This application provides a scheme for assessing transaction risk. It employs federated learning, breaking down data isolation between financial institutions in different regions without directly sharing their local data (i.e., transaction information). Furthermore, by training data from multiple financial institutions using federated learning, it improves the accuracy of financial institutions' assessments of user transaction risks. Additionally, during the federated model training process, this application utilizes private sharing technology and pre-defined encryption methods to maintain data privacy while enhancing robustness against poisoning attacks. This solves the technical problem in existing technologies where poisoning attacks during the federated model modeling process lead to low accuracy in detecting transaction risks using local models trained using federated learning.

[0138] Those skilled in the art will understand that Figure 5 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones, tablets, PDAs, mobile internet devices, PADs, and other terminal devices. Figure 5 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 5 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 5 The different configurations shown.

[0139] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0140] Example 4

[0141] Embodiments of this application may also provide a storage medium.

[0142] Optionally, in this embodiment of the application, the storage medium can be used to store the program code executed by the transaction risk assessment method provided in the above method embodiment.

[0143] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0144] This application also provides a computer program product, which, when executed on a data processing device, is suitable for performing steps of a method for assessing transaction risks.

[0145] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0146] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0147] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0148] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0149] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0150] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0151] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method of assessing transaction risk, characterized by, Comprise: After receiving the authorization information of the user, collect transaction information of a financial transaction initiated by the user; Transmit the transaction information of the financial transaction to a local target model of a target site, wherein the local target model has the same model parameters as a global target model, the global target model is a federal model obtained by training transaction information of historical transactions in P financial sites based on federal learning technology, P is a positive integer, and the target site is one of the P financial sites; Extract features from the transaction information of the financial transaction through the local target model to obtain L transaction features, wherein L is a positive integer; Determine a risk score of the financial transaction based on the L transaction features; If the risk score is greater than or equal to a preset score, generate a risk report corresponding to the financial transaction, wherein the risk report at least includes the risk score, risk type and risk transaction link of the financial transaction.

2. The method of claim 1, wherein, The training steps of the global target model include: Train a local initial model of each financial site based on transaction information of historical transactions in each financial site to obtain a local training model; In the form of private sharing, transmit gradients of P local training models corresponding to P financial sites to a central server; Aggregate the gradients of the P local training models through the central server to obtain a global gradient; Update model parameters of a global initial model based on the global gradient to obtain a global training model; Broadcast private shares of the model parameters of the global training model to the P financial sites; Update model parameters of the local training model of each financial site based on the private shares of the model parameters of the global training model to obtain a new local training model; Determine the global target model based on the P new local training models corresponding to the P financial sites.

3. The method of claim 2, wherein, Determine the global target model based on the P new local training models corresponding to the P financial sites, including: In the form of private sharing, transmit gradients of P new local training models corresponding to P financial sites to the central server; Aggregate the gradients of the P new local training models through the central server to obtain a new global gradient; Update the model parameters of the global training model based on the new global gradient until the training times of the global training model reach a preset maximum number / the global training model converges; Take the global training model obtained by the last update as the global target model.

4. The method of claim 2, wherein, Before iteratively training the local initial model of each financial site, the transaction risk assessment method further includes: Performing an initialization operation through the central server to obtain a global initial model, wherein the central server at least includes a first device, a second device, and a third device, the first device and the second device are used at least for data interaction with the P financial sites in a data encryption mode, and the third device is used to generate a multiplicative triple used for the data encryption; Broadcasting a private share of model parameters of the global initial model to the P financial sites; Based on the private share of model parameters of the global initial model, establishing a local initial model of each financial site.

5. The method of claim 2, wherein, Through the central server, performing an aggregation operation on the gradients of the P local training models to obtain a global gradient, including: Through the central server, determining the weight of each gradient in the gradients of the P local training models; Based on the weight of each gradient, performing aggregation on the gradients of the P local training models by using a matrix-vector multiplication operation to obtain the global gradient.

6. The method of claim 5, wherein, Through the central server, determining the weight of each gradient in the gradients of the P local training models, including: Controlling the first device and the second device in the central server to each generate a random vector; Through the first device and the second device, encrypting the gradients of the P local training models based on the respective generated random vectors to obtain P encrypted gradient vectors, and transmitting the P encrypted gradient vectors to the third device in the central server; Through the third device, determining the median of the P encrypted gradient vectors in each vector dimension to obtain a median gradient vector corresponding to the P encrypted gradient vectors, and transmitting the median gradient vector to the first device and the second device in the form of private sharing; Through the first device and the second device, determining the Pearson correlation coefficient between the median gradient vector and each encrypted gradient vector to obtain P Pearson correlation coefficients corresponding to the P encrypted gradient vectors; Through the third device, deriving the P Pearson correlation coefficients to obtain P derivative values corresponding to the P Pearson correlation coefficients; Based on each derivative value in the P derivative values, determining the weight of the gradient corresponding to the derivative value.

7. The method of claim 2, wherein the step of determining the risk of the transaction is performed by a risk assessment engine. The data transmitted in the form of private sharing at least includes a private share and a verification share, wherein the private share is used to reconstruct a private value of the transmitted data, and the verification share is used to perform integrity verification on the private value.

8. A device for assessing transaction risk, characterized in that, Including: An information collection unit configured to collect transaction information of a financial transaction initiated by a user after receiving authorization information of the user; An information transmission unit configured to transmit the transaction information of the financial transaction to a local target model of a target site, wherein the local target model has the same model parameters as a global target model, the global target model is a federated model obtained by training transaction information of historical transactions in P financial sites based on a federated learning technique, P is a positive integer, and the target site is one of the P financial sites. The feature extraction unit is configured to perform feature extraction on transaction information of the financial transaction by using the local target model to obtain L transaction features, where L is a positive integer. The risk score determination unit is configured to determine a risk score of the financial transaction based on the L transaction features. The risk report generation unit is configured to generate a risk report corresponding to the financial transaction in a case where the risk score is greater than or equal to a preset score, where the risk report at least includes the risk score, a risk type, and a risk transaction link of the financial transaction.

9. A computer program product, characterised in that, The computer program product includes a computer program, wherein the computer program controls the computer program product to execute the transaction risk evaluation method in any one of claims 1 to 7 when the computer program is running.

10. An electronic device, comprising: The computer program product includes one or more processors and a memory, and the memory is configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the transaction risk evaluation method in any one of claims 1 to 7.

Images