Fraud detection method, device and system
By employing privacy-preserving intersection and encrypted computation techniques, a scalable privacy-preserving neural network is constructed, solving the privacy protection problem when different institutions share financial big data, and enabling data sharing and fraud detection while protecting user privacy.
Patent Information
- Application Number
- CN202511289834.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-10
- Publication Date
- 2026-01-23
AI Technical Summary
When sharing financial big data among different institutions to achieve fraud detection, how can data sharing be carried out while protecting privacy and avoiding problems such as identity theft, identity exposure, leakage of transaction privacy, and leakage of trade secrets?
By using privacy-preserving intersection techniques to identify common users across different datasets, and leveraging encrypted computation and semi-honest servers for data sharing, a scalable privacy-preserving neural network is constructed. This ensures user privacy while achieving data matching and alignment. Computation is performed using arithmetic secret sharing and additive homomorphic encryption techniques, and complex computations are delegated to high-performance semi-honest servers.
Data sharing is achieved while protecting user privacy, which improves the performance of fraud detection models, ensures data quality and consistency, and enables accurate fraud detection.
Smart Images
Figure CN121389149A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, in particular to a fraud detection method, a fraud detection device, a fraud detection system, a machine readable storage medium and a computer program product. BACKGROUND
[0002] With the continuous development of artificial intelligence technology, deep neural networks have made significant progress in various practical applications under the condition of providing a large amount of training data. However, data silos and privacy leaks have become serious problems, especially the inability of different parties / organizations to share data.
[0003] Taking the fraud detection scenario of a user as an example, suppose two companies both have a batch of financial big data of users with different users, and the two companies hope to jointly and safely establish a better fraud detection method for detecting whether a user is fraudulent. However, the financial big data usually contains a variety of private information, such as user personal information and transaction information, and leaking these information may cause problems such as identity theft, identity exposure, transaction privacy leakage and business secret leakage. Since the financial big data is jointly held by different companies, financial institutions or third-party data providers, how to realize data sharing under the premise of protecting privacy and then realize fraud detection becomes a problem to be solved. SUMMARY
[0004] The purpose of the embodiments of the present application is to provide a fraud detection method, device and system to solve the problem of how to realize data sharing under the premise of protecting privacy and then realize fraud detection.
[0005] In order to achieve the above-mentioned purpose, the fraud detection method provided by the embodiments of the present application comprises: controlling a terminal to collect a first data set of a first institution and a second data set of a second institution; the first data set comprises first feature data of a plurality of users; the second data set comprises second feature data of a plurality of users; controlling the terminal to perform privacy intersection based on the first data set and the second data set, and determining the first feature data and the second feature data of the common users between the first data set and the second data set; controlling the terminal to perform encrypted calculation based on the first feature data and the second feature data of the common users by using model parameters of a first hidden layer, to obtain a first output result; controlling a semi-honest server to perform calculation based on the first output result by using model parameters of a second hidden layer, to obtain a second output result; controlling the terminal to perform encrypted calculation based on the second output result by using model parameters of a third hidden layer, to obtain a fraud detection result; The first hidden layer, the second hidden layer, and the third hidden layer together constitute a hidden layer model structure of the fraud detection model.
[0006] Optionally, the control of the terminal based on the first feature data and the second feature data of the common user, using the model parameters of the first hidden layer to perform encrypted calculation to obtain the first output result, includes: The control of the terminal based on the first feature data and the second feature data of the common user, using the model parameters of the first hidden layer to perform arithmetic secret sharing and homomorphic encryption, to obtain the first output result.
[0007] Optionally, the control of the terminal based on the first feature data and the second feature data of the common user, using the model parameters of the first hidden layer to perform homomorphic encryption, includes: The control of the terminal receiving the public key sent by the semi-honest server; The control of the terminal using the public key and a first random number to encrypt the first feature data to obtain first encrypted data, and using the public key and a second random number to encrypt the second feature data to obtain second encrypted data; So that the semi-honest server performs homomorphic addition calculation based on the first encrypted data and the second encrypted data.
[0008] Optionally, when the terminal controls the arithmetic secret sharing based on the first feature data and the second feature data of the common user, decimal calculation is performed through fixed-point representation method.
[0009] Optionally, the fraud detection model is obtained by the following steps: Control the terminal to collect a first sample data set of a first institution and a second sample data set of a second institution; the first sample data set includes first sample feature data of a plurality of users; the second sample data set includes second sample feature data of a plurality of users; Control the terminal to perform private intersection based on the first sample data set and the second sample data set to determine the first sample feature data and the second sample feature data of the common user between the first sample data set and the second sample data set; Control the terminal and the semi-honest server to cooperatively adjust the model parameters of the first hidden layer, the second hidden layer, and the third hidden layer based on the first feature data and the second feature data of the common user using the stochastic gradient Langevin dynamics method to obtain the trained fraud detection model.
[0010] Optionally, before the control of the terminal to determine the first feature data and the second feature data of the common users between the first data set and the second data set based on the first data set and the second data set, the method further comprises: controlling the terminal to perform at least one of data preprocessing operations of data cleaning, format conversion and feature extraction based on the first data set and the second data set.
[0011] In another aspect, the embodiments of the present application further provide a fraud detection device, comprising: a first control module configured to control a terminal to collect a first data set of a first institution and a second data set of a second institution; the first data set comprises first feature data of a plurality of users; and the second data set comprises second feature data of the plurality of users; a second control module configured to control the terminal to determine the first feature data and the second feature data of the common users between the first data set and the second data set based on the first data set and the second data set; a third control module configured to control the terminal to perform encrypted calculation based on the first feature data and the second feature data of the common users by using model parameters of a first hidden layer to obtain a first output result; a fourth control module configured to control a semi-honest server to perform calculation based on the first output result by using model parameters of a second hidden layer to obtain a second output result; a fifth control module configured to control the terminal to perform encrypted calculation based on the second output result by using model parameters of a third hidden layer to obtain a fraud detection result; wherein the first hidden layer, the second hidden layer and the third hidden layer jointly constitute a hidden layer model structure of a fraud detection model.
[0012] In another aspect, the embodiments of the present application further provide a fraud detection system, comprising a terminal and a semi-honest server; the terminal is configured to collect a first data set of a first institution and a second data set of a second institution; the first data set comprises first feature data of a plurality of users; and the second data set comprises second feature data of the plurality of users; to determine the first feature data and the second feature data of the common users between the first data set and the second data set based on the first data set and the second data set; to perform encrypted calculation based on the first feature data and the second feature data of the common users by using model parameters of a first hidden layer to obtain a first output result; and to perform calculation based on a second output result by using model parameters of a third hidden layer to obtain a fraud detection result; The semi-honest server is configured to calculate, based on the first output result, a second output result by using model parameters of a second hidden layer. The first hidden layer, the second hidden layer, and the third hidden layer together constitute a hidden layer model structure of the fraud detection model.
[0013] In another aspect, the present application also provides a machine readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the fraud detection method.
[0014] In another aspect, the present application also provides a computer program product comprising a computer program, the computer program being executed by a processor to implement the fraud detection method.
[0015] Through the above technical solution, the present application embodiment finds out the common users between different data sets by private intersection, ensures that the user data can be accurately matched and aligned while protecting the user privacy. The present application embodiment also uses the encryption technology to perform the calculation related to the private data of the user by the terminal through the first hidden layer and the third hidden layer, and the rest of the calculation is delegated to the semi-honest server with high computing power to perform the calculation through the second hidden layer. Thus, the present application embodiment realizes the data sharing and further realizes the fraud detection under the premise of protecting the privacy.
[0016] Other features and advantages of the present application embodiment will be described in detail in the following specific implementation part. BRIEF DESCRIPTION OF DRAWINGS
[0017] The accompanying drawings are used to provide a further understanding of the present application embodiment, and constitute a part of the specification, and are used to explain the present application embodiment together with the following specific implementation, but do not constitute a limitation to the present application embodiment. In the drawings: Figure 1 is one of the flowcharts of the fraud detection method provided by the present application; Figure 2 is the second flowchart of the fraud detection method provided by the present application; Figure 3 is the structural schematic diagram of the fraud detection device provided by the present application; Figure 4 is the structural schematic diagram of the fraud detection system provided by the present application. DETAILED DESCRIPTION
[0018] The specific implementation of the present application embodiment will be described in detail below in combination with the drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present application embodiment, and is not used to limit the present application embodiment.
[0019] Method embodiment Please refer to Figure 1 The embodiment of the present application provides a fraud detection method, comprising: Step 100, control the terminal to collect a first data set of a first institution and a second data set of a second institution; the first data set comprises first feature data of a plurality of users; the second data set comprises second feature data of the plurality of users.
[0020] The first institution and the second institution can be any one of a company, a financial institution or a third-party data provider. The first feature data in the first data set and the second feature data in the second data set can be personal information and transaction messages of the user. In the embodiment of the present application, the first feature data and the second feature data are set to distinguish the user feature data in the first institution and the second institution. Taking the first institution as A and the second institution as B for example. Please refer to Figure 2 For example, A has feature data (denoted as XA, represented by yellow dots, for example, which can be user personal information) and labels (y, represented by yellow squares) of some users, and B has feature data (denoted as XB, represented by green dots, for example, which can be user transaction information) of the same batch of users. In the embodiment of the present application, taking the user fraud detection scenario as an example, it is assumed that two companies both have a batch of users with different user features, and they hope to jointly and safely establish a better fraud detection method. The embodiment of the present application first controls the terminal to collect a first data set of a first institution A and a second data set of a second institution B.
[0021] Step 200, control the terminal to perform privacy intersection based on the first data set and the second data set, and determine the first feature data and the second feature data of the common users between the first data set and the second data set.
[0022] Although the first institution A can build a deep neural network for fraud detection using its own feature data, the performance of the model can be improved by adding the feature data of the second institution B. However, the two companies cannot share each other's data because it violates relevant regulations to leak user privacy data. It is a challenge for both parties to cooperate to build an extensible privacy-protected neural network without leaking private data. In the embodiment of the present application, only the case where two data holders (i.e. the first institution A and the second institution B) have the same sample set is considered, one of which (A) has part of the features and labels of a batch of users, and the other (B) has the remaining part of the features of the same batch of users. The proposed algorithm can be naturally extended to more than two parties. The embodiment of the present application uses the private set intersection (PSI) technology to find the cross-users (i.e. common users, intersection users) between the data sets of the two institutions. By using the private set intersection technology, the cross-users between different data sets are found. This step ensures that the user feature data can be accurately matched and aligned while protecting user privacy.
[0023] In some embodiments, before the controlling the terminal to perform the private set intersection based on the first data set and the second data set, and determining the first feature data and the second feature data of the common users between the first data set and the second data set, the method further includes: controlling the terminal to perform at least one of data preprocessing operations including data cleaning, format conversion and feature extraction based on the first data set and the second data set.
[0024] Before the controlling the terminal to perform the private set intersection based on the first data set and the second data set, and determining the first feature data and the second feature data of the common users between the first data set and the second data set, the method further includes: controlling the terminal to perform data preprocessing based on the first data set and the second data set to ensure the quality and consistency of the data. For example, the embodiment of the present application can control the terminal to perform at least one of data preprocessing operations including data cleaning, format conversion and feature extraction based on the first data set and the second data set. In one embodiment, in order to maximize the quality and consistency of the data, the terminal is controlled to perform data cleaning, format conversion and feature extraction based on the first data set and the second data set. The data cleaning can use missing value processing and outlier processing. The missing value processing can use direct deletion method or use statistical quantities such as mean and median to fill in. The format conversion can use data standardization techniques, such as using the min-max normalization method to scale the user's feature data to the [0, 1] interval. The feature extraction can use feature dimension reduction methods, such as using principal component analysis to extract features from the user's feature data.
[0025] The collected user feature data is preprocessed, including at least one of data cleaning, format conversion and feature extraction, and samples between users are aligned to ensure the quality and consistency of the data.
[0026] Step 300, controlling the terminal to perform encrypted calculation based on the first feature data and the second feature data of the common user by using model parameters of a first hidden layer, to obtain a first output result.
[0027] Step 400, controlling the semi-honest server to perform calculation based on the first output result by using model parameters of a second hidden layer, to obtain a second output result.
[0028] Step 500, controlling the terminal to perform encrypted calculation based on the second output result by using model parameters of a third hidden layer, to obtain a fraud detection result.
[0029] The first hidden layer, the second hidden layer and the third hidden layer jointly constitute a hidden layer model structure of the fraud detection model.
[0030] The hidden layer model structure of the fraud detection model in the embodiment of the application can adopt various deep neural network model structures. For example, the hidden layer model structure of the fraud detection model can utilize existing deep learning platforms such as TensorFlow and PyTorch, and various deep neural network models can be easily implemented. In the embodiment of the application, the hidden layer model structure of the fraud detection model can learn complex functions by combining multiple nonlinear modules, so as to convert low-level original input into high-level abstract representation. Mathematically, the forward process of the hidden layer model structure of the fraud detection model can be defined as a representation function f, which maps the input X to the output y, i.e. y = f(X, θ), where θ is the model parameter of the hidden layer model structure. Assuming that the hidden layer model structure of the fraud detection model has L layers, then f is composed of L sub-functions f l|l∈[1,L], which are connected in a chain. That is, f(X) = f L(...f2 (f1 (X, θ0)), θ1)..., θ L-1). The hidden layer model structure of the fraud detection model can be defined as a hierarchical representation function. The computational graph of the hidden layer model structure of the fraud detection model is decoupled into two types in the embodiments of the present application. That is, the computation related to the private data of the user is performed by the terminal using encryption techniques (such as secret sharing, homomorphic encryption, garbled circuits, etc.), and the rest of the computation is delegated to a semi-honest server with high computing power. The private data here refers to the input and output of the fraud detection model, which is equivalent to the feature data of the user collected by the terminal and the label of the feature data. By letting the semi-honest server perform a large number of computations in the hidden layer model of the fraud detection model in plaintext form, such as nonlinear activation functions and max pooling, these operations are expensive for encryption techniques, so the semi-honest server can avoid many complex nonlinear operations. In addition, the semi-honest protocol of the semi-honest server is also the key to forming a scalable and privacy-preserving deep neural network (Scalable and Privacy-preserving Deep Neural Network). Specifically, the model parameters (θ) of the hidden layer model structure of the fraud detection model are divided into three parts in the embodiments of the present application: 1 is the computation (θ A and θ B ) related to the private features of the two data holders (the feature data of the user) on the terminal, 2 is the rest of the computation (θ S ) related to the hidden layer on the semi-honest server, and 3 is the computation (θ Y ) related to the private label on the data holder with the label on the terminal. Please refer to Figure 2 , the first part is the computation related to the private data of the institution (the feature data of the user XA, XB), so it is performed by the data holder himself using encryption techniques such as secret sharing, homomorphic encryption, garbled circuits, etc., and the second part is delegated to a semi-honest server with rich computing resources.
[0031] Specifically, the first hidden layer is calculated by the first institution A and the second institution B in cooperation. The computation of the subsequent second hidden layer is completed by the semi-honest server. After obtaining the first output result of the first hidden layer, it is sent to the semi-honest server for related computation of the second hidden layer, that is: hL=f(hl, θ S ). This is the same as the traditional neural network. Given the lth hidden layer hl, where 1≤l≤L-1 and L is the number of hidden layers, the calculation formula of the (l+1)th hidden layer is: h l+1=fl(hl,θ1), where θ1 is the model parameter of the 1st layer and fl is the effective function of the lth layer. These are the most time-consuming calculations because there are many nonlinear operations, such as max pooling, which are not friendly to cryptography. The embodiments of the present application delegate these heavy calculations to the powerful semi-honest server to complete. For this purpose, the fraud detection model of the embodiments of the present application can be extended to large data sets. In addition, various deep neural network models can be easily implemented using existing deep learning platforms such as TensorFlow and PyTorch. After the semi-honest server completes the second hidden layer related calculation, it will send the final second output result hL to the data holder (terminal, that is, A in this example) with labels for calculating the prediction result. That is =fl(hl,θ1), where θ1 is the model parameter of the 1st layer and fl is the effective function of the lth layer. These are the most time-consuming calculations because there are many nonlinear operations, such as max pooling, which are not friendly to cryptography. The embodiments of the present application delegate these heavy calculations to the powerful semi-honest server to complete. For this purpose, the fraud detection model of the embodiments of the present application can be extended to large data sets. In addition, various deep neural network models can be easily implemented using existing deep learning platforms such as TensorFlow and PyTorch. After the semi-honest server completes the second hidden layer related calculation, it will send the final second output result hL to the data holder (terminal, that is, A in this example) with labels for calculating the prediction result. That is Y ). Where δ is designed according to different prediction tasks, for example, δ is a Softmax function for classification tasks. Therefore, the three-layer hidden layer composed of the terminal and the semi-honest server of the embodiments of the present application proposes a scalable and privacy-protected deep neural network learning framework from the common perspective of algorithm and cryptography.
[0032] The embodiments of the present application find out the common users between different data sets by private intersection finding, ensuring that the user data can be accurately matched and aligned while protecting user privacy. The embodiments of the present application also perform the calculation related to the private data of the user by the terminal through the first hidden layer and the third hidden layer using encryption technology, and the remaining calculation is delegated to the semi-honest server with high computing power to calculate through the second hidden layer. Thus, the embodiments of the present application realize data sharing and fraud detection under the premise of protecting privacy.
[0033] In other aspects of the embodiments of the present application, the control of the terminal based on the first feature data and the second feature data of the common user, and the model parameter of the first hidden layer is used for encrypted calculation to obtain the first output result, comprising: controlling the terminal to perform arithmetic secret sharing and additive homomorphic encryption based on the first feature data and the second feature data of the common user, and the model parameter of the first hidden layer to obtain the first output result.
[0034] In the embodiments of the present application, the first institution A and the second institution B hope that: (1) cooperatively calculate a common function, that is, h1=f(XA, XB; θ A , θ B ); (2) keep their respective features, that is, XA and XB are private. In mathematics, A and B respectively own part of the features (XA and XB) and part of the model parameters (θ A and θ B ), and hope to cooperatively calculate the output of the first hidden layer. That is, A and B want to calculate h1=XAθ A +XB θ B =(XA XB) (θ A θ B ),in This indicates a connection operation. The aforementioned secure computation problem can be solved using encryption techniques. In one embodiment, the terminal controls the sharing of arithmetic secrets and addition-based homomorphic encryption based on the first and second feature data of the common user, using the model parameters of the first hidden layer, to obtain the first output result.
[0035] A and B each share an arithmetic secret of their user feature data. Through this secret-sharing technique, they collaboratively compute the hidden layer outputs of a deep neural network (DNN) without revealing the original data.
[0036] First, we will introduce how to solve the aforementioned secure computation problem using arithmetic sharing. The main technique is secret sharing based on matrix addition and multiplication using fixed-point numbers.
[0037] Arithmetic secret sharing: Suppose there are two institutions (P0 and P1), P0 has a ℓ-bit secret 'a', and P1 has a ℓ-bit secret 'b'. Here, ℓ is a security parameter representing the bit length of the secret. For P0, in order to share secret 'a' (this operation is denoted as Shr(...)... P0 is uniformly and randomly selected from the integer ring. Choose an integer from ,Will Send to P1 as a shared share and retain As another shared share. Similarly, P1 can share with P0. Specifically: P1 from Randomly select integers from the middle ,reserve and will Send to P0. In this way, and The data is split into two shared shares, held by P0 and P1 respectively. Safe addition and multiplication operations can then be performed based on these shared shares, and the processing of decimal numbers and vectors is supported.
[0038] Addition and Multiplication: Suppose P0 and P1 want to use an arithmetic secret to calculate a + b. P0 locally calculates hci0 = hai0 + hbi0 mod 2. ℓ And P1 calculates hci1 = hai1 + hbi1 mod 2 locally. ℓ To reconstruct this secret (this operation is denoted as Rec( , ) ), one party sends his share to the other party, who reconstructs the plaintext by c = hci0+ hci1, which equals a + b. To compute a * b using arithmetic secret sharing, Beaver's multiplication triplets are usually needed. Specifically, to multiply two secret shared values (a and b), P0 and P1 first need to cooperate to generate a triplet hui, hvi and hwi, where u, v are uniform random values in ℓ . Then, P0 locally computes hei0= hai0- hui0and hfi0= hbi0- hvi0, and P1 locally computes hei1= hai1- hui1and hfi1= hbi1- hvi1. Next, they respectively reconstruct e and f by Rec(hei0, hei1) and Rec(hfi0, hfi1). Finally, P0 gets hci0= fhai0+ e hbi0+ hwi0. And P1 gets hci1= ef + fhai1+ ehbi1+ hwi1, where hci0+ hci1= ab.
[0039] It is to be noted that when controlling the terminal to perform arithmetic secret sharing based on the first feature data and the second feature data of the common user, decimal calculation is performed by fixed-point representation. The above protocol is only applicable to finite fields because it needs to uniformly sample in . However, in neural networks, features and model parameters are usually decimal vectors. For this purpose, the embodiment of the present application uses fixed-point representation to efficiently approximate decimal arithmetic. Simply put, at most lF bits are used to represent the decimal part of a decimal number. Specifically, suppose a and b are two decimals with at most lF bits in the decimal part, to do fixed-point multiplication, first convert them to integers, let a' = 2lFa and b' = 2lFb, then compute c = a' * b'. Finally, truncate the last lF bits of c so that it has at most lF bits representing the decimal part. When c is a secret share, this truncation technique also works. The embodiment of the present application sets lF = 16. Thus, it is easy to vectorize the addition and multiplication protocols under arithmetic sharing.
[0040] In some embodiments, controlling the terminal to perform additive homomorphic encryption based on the first and second feature data of the common user using model parameters of the first hidden layer includes: controlling the terminal to receive a public key sent by the semi-honest server; controlling the terminal to encrypt the first feature data using the public key and a first random number to obtain first encrypted data, and encrypting the second feature data using the public key and the second random number to obtain second encrypted data; so that the semi-honest server performs homomorphic addition calculation based on the first encrypted data and the second encrypted data.
[0041] A semi-honest server generates a public-private key pair and distributes the public key to a first organization A and a second organization B. A and B use the public key to encrypt their own feature data. The semi-honest server performs homomorphic addition on the encrypted features to obtain the encrypted output of the second hidden layer (the second output result). Additive homomorphic encryption is an encryption scheme that allows third parties (such as cloud providers or service providers) to perform addition operations on encrypted data while preserving the characteristics of the addition operation and the format of the encrypted data. Specifically, there is a semi-honest server with key generation capabilities and several terminals possessing private data. In this scenario, the main steps of using additive homomorphic encryption are as follows: First, key generation: The semi-honest server generates a public-private key pair (pk, sk) and distributes the public key pk to the participants (first organization A and second organization B); second, encryption: First organization A possesses the plaintext data. (Its characteristic data), using public key and random numbers right Encrypt and generate ciphertext: .in express The encrypted result (ciphertext). Secondary organization B possesses the plaintext data. (Its characteristic data), using public key and random numbers right Encrypt and generate ciphertext: .in express The encrypted result (ciphertext) is then processed. Finally, homomorphic addition is performed: the semi-honest server receives the encrypted data. and Perform homomorphic addition: According to the properties of additive homomorphic encryption, the result of adding ciphertext is equal to the encrypted form of adding plaintext, that is: .in It is the merged random number (the specific value depends on the encryption scheme, and may be generated by...). and The derivation is made, but the semi-honest server does not need to know it.
[0042] In other aspects of the embodiments of the present application, the fraud detection model is trained by the following steps: a terminal is controlled to collect a first sample data set of the first institution and a second sample data set of the second institution; the first sample data set includes first sample feature data of a plurality of users; the second sample data set includes second sample feature data of the plurality of users; the terminal is controlled to perform private set intersection based on the first sample data set and the second sample data set, to determine first sample feature data and second sample feature data of common users between the first sample data set and the second sample data set; the terminal and the semi-honest server are controlled to cooperate to perform model parameter adjustment on the first hidden layer, the second hidden layer and the third hidden layer based on the first feature data and the second feature data of the common users by using a stochastic gradient Langevin dynamics method, to obtain the trained fraud detection model.
[0043] The first sample data set can be a historical data set corresponding to users in the first data set. The second sample data set can be a historical data set corresponding to users in the second data set. In model training, the same is to find out the cross users between different sample data sets by using the private set intersection technology. This step ensures that the sample feature data of the users can be accurately matched and aligned while protecting the privacy of the users. The embodiments of the present application match the sample feature data of the cross users by using the private set intersection technology, and align these sample feature data as training samples.
[0044] The security assumption of the embodiments of the present application is that the semi-honest server gradient calculation method is based on the chain rule for back propagation, which is similar to the forward propagation process of the algorithm. Specifically, given the loss calculated by a data holder with labels, the data holder first calculates the gradient of its model parameters, and then sends the model update to the semi-honest server. Then, the semi-honest server calculates the gradient of its own model parameters layer by layer, and then sends the model update of the previous several hidden layers (i.e. the layers related to private features) to the data holder. Finally, each data holder calculates the corresponding model gradient. Both forward calculation and backward calculation need to be communicated in a decentralized manner between the first institution A, the second institution B and the server. During the training process, all private data (XA, XB, y) and model parameters (θ A , θ B and θ Y ) related to private data are saved by the data holder. Therefore, the data privacy is largely protected.
[0045] The fraud detection model of this invention can be extended to multiple parties, including data holders, to jointly compute i (1≤i≤L) hidden layers, instead of just computing the first hidden layer. They aim to collaboratively build a better fraud detection system. To effectively achieve this goal, this invention employs the SGLD (Stochastic Gradient Langevin Dynamics) method to train the proposed fraud detection model (a neural network with a specific structure, such as a deep neural network or convolutional neural network, depending on the application scenario) to generate accurate fraud detection results. Model parameter tuning using the SGLD method includes parameter initialization, iterative updates, and convergence determination. The parameter initialization step includes randomly generating parameters. θ 0. Learning rate η and number of iterations T The initialization process involves iterative updates, including random sampling of mini-batch data; calculation of stochastic gradients based on the mini-batch data; injection of Gaussian noise; and finally, updating the model parameters based on the stochastic gradients and Gaussian noise. Convergence is judged by the criterion that the parameter sequence gradually stabilizes in the high-probability region of the posterior distribution.
[0046] A fraud detection model is trained using the stochastic gradient lattice (SRL) method. This method not only introduces noise during training to protect privacy but also improves the model's generalization ability. Through continuous iteration and optimization, the fraud detection model can accurately identify fraudulent behavior. The trained model is evaluated using a validation set to ensure its performance meets requirements. Then, the model is applied to real-world scenarios to make predictions and detect potential fraudulent activities. Thus, this invention proposes a scalable and privacy-preserving deep neural network learning framework, SPNN (Scalable and Privacy-preserving Deep Neural Network), from the combined perspectives of algorithms and cryptography. Furthermore, a novel vertically federated graph neural network (VFGNN) is formed by a three-layer hidden layer consisting of a terminal and a semi-honest server to attempt to address the problem of feature / edge information separation.
[0047] Device Examples Please refer to Figure 3 On the other hand, embodiments of the present invention also provide a fraud detection device, comprising: The first control module 301 is used to control the terminal to collect a first dataset from the first institution and a second dataset from the second institution; the first dataset includes first feature data of multiple users; the second dataset includes second feature data of multiple users. a second control module 302, configured to control the terminal to determine first feature data and second feature data of a common user between the first data set and the second data set based on the first data set and the second data set by privacy intersection; a third control module 303, configured to control the terminal to perform encrypted calculation based on the first feature data and the second feature data of the common user by using model parameters of a first hidden layer, to obtain a first output result; a fourth control module 304, configured to control the semi-honest server to perform calculation based on the first output result by using model parameters of a second hidden layer, to obtain a second output result; a fifth control module 305, configured to control the terminal to perform encrypted calculation based on the second output result by using model parameters of a third hidden layer, to obtain a fraud detection result; wherein the first hidden layer, the second hidden layer and the third hidden layer jointly constitute a hidden layer model structure of a fraud detection model.
[0048] The embodiment of the application finds the common user between different data sets by privacy intersection, and ensures that the user data can be accurately matched and aligned while protecting the privacy of the user. The embodiment of the application also performs the calculation related to the private data of the user by the terminal using the encryption technology through the first hidden layer and the third hidden layer, and performs the rest of the calculation by the semi-honest server through the second hidden layer. Thus, the embodiment of the application realizes data sharing and fraud detection under the premise of protecting the privacy.
[0049] Optionally, the control of the terminal to perform encrypted calculation based on the first feature data and the second feature data of the common user by using model parameters of the first hidden layer to obtain the first output result comprises: the control of the terminal to perform arithmetic secret sharing and homomorphic encryption of addition based on the first feature data and the second feature data of the common user by using model parameters of the first hidden layer to obtain the first output result.
[0050] Optionally, the control of the terminal to perform homomorphic encryption of addition based on the first feature data and the second feature data of the common user by using model parameters of the first hidden layer comprises: the control of the terminal to receive a public key sent by the semi-honest server; the control of the terminal to encrypt the first feature data by using the public key and a first random number to obtain first encrypted data, and to encrypt the second feature data by using the public key and a second random number to obtain second encrypted data; so that the semi-honest server performs homomorphic addition calculation based on the first encrypted data and the second encrypted data.
[0051] Optionally, when the terminal is controlled to perform arithmetic secret sharing based on the first feature data and the second feature data of the common user, decimal calculation is performed through fixed-point representation.
[0052] Optionally, the fraud detection model is obtained through the following steps of training: The terminal is controlled to collect a first sample data set of the first institution and a second sample data set of the second institution; the first sample data set includes first sample feature data of a plurality of users; and the second sample data set includes second sample feature data of the plurality of users. The terminal is controlled to perform private set intersection based on the first sample data set and the second sample data set, to determine first sample feature data and second sample feature data of a common user between the first sample data set and the second sample data set. The terminal and the semi-honest server are controlled to cooperatively perform model parameter adjustment on the first hidden layer, the second hidden layer and the third hidden layer based on the first feature data and the second feature data of the common user by using a stochastic gradient Langevin dynamics method, to obtain the trained fraud detection model.
[0053] Optionally, before the terminal is controlled to perform private set intersection based on the first data set and the second data set to determine first feature data and second feature data of a common user between the first data set and the second data set, the method further includes: The terminal is controlled to perform at least one data preprocessing operation of data cleaning, format conversion and feature extraction based on the first data set and the second data set.
[0054] The fraud detection device includes a processor and a memory, and the first control module 301, the second control module 302, the third control module 303, the fourth control module 304 and the fifth control module 305 are stored in the memory as program units, and the processor executes the program units stored in the memory to realize the corresponding functions.
[0055] The processor includes a core, and the core retrieves the corresponding program units from the memory. The core can be provided with one or more than one.
[0056] The memory can include a non-permanent memory in a computer readable medium, a random access memory (RAM) and / or a non-volatile memory such as a read-only memory (ROM) or a flash memory (flash RAM), and the memory includes at least one memory chip.
[0057] On the other hand, please refer to Figure 4The present invention also provides a fraud detection system, including a terminal 10 and a semi-honest server 20; The terminal 10 is used to collect a first dataset from a first institution and a second dataset from a second institution; the first dataset includes first feature data of multiple users; the second dataset includes second feature data of multiple users; privacy intersection is performed based on the first dataset and the second dataset to determine the first feature data and second feature data of common users between the first dataset and the second dataset; based on the first feature data and second feature data of the common users, encrypted calculation is performed using the model parameters of the first hidden layer to obtain a first output result; and based on the second output result, calculation is performed using the model parameters of the third hidden layer to obtain a fraud detection result; The semi-honest server 20 is used to calculate, based on the first output result and using the model parameters of the second hidden layer, to obtain the second output result; The first hidden layer, the second hidden layer, and the third hidden layer together constitute the hidden layer model structure of the fraud detection model.
[0058] This invention utilizes privacy-preserving intersection to identify common users across different datasets, ensuring accurate matching and alignment of user data while protecting user privacy. Furthermore, calculations related to users' private data are performed by the terminal using encryption techniques through a first and third hidden layer, while remaining calculations are delegated to a high-performance semi-honest server using a second hidden layer. Thus, this invention achieves data sharing while protecting privacy, thereby enabling fraud detection.
[0059] In another aspect, the present application also provides a computer program product comprising a computer program, which can be stored on a machine-readable storage medium, and which, when executed by a processor, enables a computer to perform a fraud detection method, the method comprising: controlling a terminal to collect a first data set of a first institution and a second data set of a second institution; the first data set comprising first feature data of a plurality of users; the second data set comprising second feature data of the plurality of users; controlling the terminal to perform a privacy intersection based on the first data set and the second data set, to determine first feature data and second feature data of common users between the first data set and the second data set; controlling the terminal to perform an encrypted calculation based on the first feature data and the second feature data of the common users, using model parameters of a first hidden layer, to obtain a first output result; controlling a semi-honest server to perform a calculation based on the first output result, using model parameters of a second hidden layer, to obtain a second output result; and controlling the terminal to perform an encrypted calculation based on the second output result, using model parameters of a third hidden layer, to obtain a fraud detection result; wherein the first hidden layer, the second hidden layer, and the third hidden layer together constitute a hidden layer model structure of a fraud detection model.
[0060] In yet another aspect, the present application also provides a machine-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements a fraud detection method, the method comprising: controlling a terminal to collect a first data set of a first institution and a second data set of a second institution; the first data set comprising first feature data of a plurality of users; the second data set comprising second feature data of the plurality of users; controlling the terminal to perform a privacy intersection based on the first data set and the second data set, to determine first feature data and second feature data of common users between the first data set and the second data set; controlling the terminal to perform an encrypted calculation based on the first feature data and the second feature data of the common users, using model parameters of a first hidden layer, to obtain a first output result; controlling a semi-honest server to perform a calculation based on the first output result, using model parameters of a second hidden layer, to obtain a second output result; and controlling the terminal to perform an encrypted calculation based on the second output result, using model parameters of a third hidden layer, to obtain a fraud detection result; wherein the first hidden layer, the second hidden layer, and the third hidden layer together constitute a hidden layer model structure of a fraud detection model.
[0061] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.
[0062] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.
[0063] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A fraud detection method, characterized in that, include: The control terminal collects a first dataset from the first organization and a second dataset from the second organization; the first dataset includes first feature data of multiple users. The second dataset includes second feature data from multiple users; The terminal is controlled to perform privacy intersection based on the first dataset and the second dataset to determine the first and second feature data of common users between the first dataset and the second dataset; The terminal is controlled to perform encrypted calculations based on the first and second feature data of the common user, using the model parameters of the first hidden layer, to obtain a first output result. The semi-honest server is controlled to calculate the second output result based on the first output result and using the model parameters of the second hidden layer; The terminal is controlled to perform encrypted calculations based on the second output result and using the model parameters of the third hidden layer to obtain the fraud detection result. The first hidden layer, the second hidden layer, and the third hidden layer together constitute the hidden layer model structure of the fraud detection model.
2. The fraud detection method according to claim 1, characterized in that, The terminal controls the encryption calculation based on the first and second feature data of the common user, using the model parameters of the first hidden layer, to obtain a first output result, including: The terminal controls the sharing of arithmetic secrets and addition homomorphic encryption based on the first and second feature data of the common user, using the model parameters of the first hidden layer, to obtain the first output result.
3. The fraud detection method according to claim 2, characterized in that, Controlling the terminal to perform additive homomorphic encryption based on the first and second feature data of the common user, using the model parameters of the first hidden layer, includes: Control the terminal to receive the public key sent by the semi-honest server; The terminal is controlled to encrypt the first feature data using the public key and the first random number to obtain the first encrypted data, and to encrypt the second feature data using the public key and the second random number to obtain the second encrypted data. So that the semi-honest server can perform homomorphic addition calculations based on the first encrypted data and the second encrypted data.
4. The fraud detection method according to claim 2, characterized in that, When the terminal controls the arithmetic secret sharing based on the first and second feature data of the common user, it performs decimal calculations using fixed-point representation.
5. The fraud detection method according to claim 1, characterized in that, The fraud detection model is trained through the following steps: The control terminal collects a first sample dataset from a first institution and a second sample dataset from a second institution; the first sample dataset includes first sample feature data of multiple users; the second sample dataset includes second sample feature data of multiple users. The terminal is controlled to perform privacy intersection based on the first sample dataset and the second sample dataset to determine the first sample feature data and the second sample feature data of the common users between the first sample dataset and the second sample dataset. The terminal and the semi-honest server collaborate to adjust the model parameters of the first hidden layer, the second hidden layer, and the third hidden layer based on the first and second feature data of the common user, using the stochastic gradient Langevin dynamics method, in order to obtain the trained fraud detection model.
6. The fraud detection method according to claim 1, characterized in that, Before controlling the terminal to perform privacy intersection based on the first dataset and the second dataset to determine the first feature data and second feature data of common users between the first dataset and the second dataset, the method further includes: The terminal is controlled to perform at least one data preprocessing operation based on the first dataset and the second dataset, including data cleaning, format conversion, and feature extraction.
7. A fraud detection device, characterized in that, include: The first control module is used to control the terminal to collect a first dataset from a first organization and a second dataset from a second organization; the first dataset includes first feature data of multiple users. The second dataset includes second feature data from multiple users; The second control module is used to control the terminal to perform privacy intersection based on the first dataset and the second dataset to determine the first feature data and the second feature data of the common users between the first dataset and the second dataset. The third control module is used to control the terminal to perform encrypted calculations based on the first and second feature data of the common user, using the model parameters of the first hidden layer, to obtain the first output result. The fourth control module is used to control the semi-honest server to perform calculations based on the first output result using the model parameters of the second hidden layer to obtain the second output result; The fifth control module is used to control the terminal to perform encrypted calculations based on the second output result and the model parameters of the third hidden layer to obtain the fraud detection result; The first hidden layer, the second hidden layer, and the third hidden layer together constitute the hidden layer model structure of the fraud detection model.
8. A fraud detection system, characterized in that, Including terminals and semi-honest servers; The terminal is used to collect a first dataset from a first organization and a second dataset from a second organization; the first dataset includes first feature data of multiple users; The second dataset includes second feature data of multiple users; based on the first dataset and the second dataset, a privacy intersection is performed to determine the first feature data and second feature data of users who are common to the first dataset and the second dataset. Based on the first and second feature data of the common users, encrypted calculations are performed using the model parameters of the first hidden layer to obtain the first output result. And based on the second output, the fraud detection result is obtained by calculating using the model parameters of the third hidden layer; The semi-honest server is used to calculate the second output result based on the first output result and the model parameters of the second hidden layer. The first hidden layer, the second hidden layer, and the third hidden layer together constitute the hidden layer model structure of the fraud detection model.
9. A machine-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the fraud detection method according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the fraud detection method according to any one of claims 1 to 6.