CN113362160A - Federal learning method and device for credit card anti-fraud - Google Patents

Federal learning method and device for credit card anti-fraud Download PDF

Info

Publication number
CN113362160A
CN113362160A CN202110635863.6A CN202110635863A CN113362160A CN 113362160 A CN113362160 A CN 113362160A CN 202110635863 A CN202110635863 A CN 202110635863A CN 113362160 A CN113362160 A CN 113362160A
Authority
CN
China
Prior art keywords
neural network
convolutional neural
network model
graph
federal learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110635863.6A
Other languages
Chinese (zh)
Other versions
CN113362160B (en
Inventor
胡凯
吴佳胜
陆美霞
李姚根
徐露娟
夏旻
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202110635863.6A priority Critical patent/CN113362160B/en
Publication of CN113362160A publication Critical patent/CN113362160A/en
Application granted granted Critical
Publication of CN113362160B publication Critical patent/CN113362160B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/03—Credit; Loans; Processing thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Probability & Statistics with Applications (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • Technology Law (AREA)
  • General Business, Economics & Management (AREA)
  • Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
  • Character Discrimination (AREA)

Abstract

The invention discloses a federal learning method and a device for anti-fraud of a credit card, wherein the method comprises the following steps: building a partial graph convolutional neural network model corresponding to K federal learning participants with different fraud classes; carrying out federal learning training by using a local graph convolutional neural network model; the method comprises the following steps that an attention mechanism is adopted to improve the aggregation process of the federal learning parameters, so that each partial graph convolutional neural network model has adaptive weight to aggregate; and outputting a global graph convolution neural network model, wherein the global graph convolution neural network model is used for processing the imported user data and identifying the corresponding fraud category. Aiming at the existing problems of the existing credit card fraud assessment method and the classical federal learning algorithm, the invention provides a federal learning algorithm which is suitable for non-European space data and the personalized characteristics of participants to process financial data and carry out credit card anti-fraud judgment.

Description

Federal learning method and device for credit card anti-fraud
Technical Field
The invention relates to the technical field of credit card anti-fraud, in particular to a federal learning method and a device for credit card anti-fraud.
Background
With the rapid development of the internet, the financial technology based on the artificial intelligence technology has a profound influence on the consumption behavior of people. But the financial data often relate to privacy, and data of financial institutions such as different banks, loan institutions and the like cannot be directly shared, so that a data island is formed. However, if the artificial intelligence algorithm is to achieve higher precision, a large amount of data support is needed, and the algorithm model trained by a single data owner locally and independently cannot accurately evaluate whether the credit card is fraudulent.
Federal learning is used as a distributed machine learning/deep learning framework for protecting data privacy, and a good solution can be provided for the problems of data isolated island, serious data discretization, data isomerism, unbalanced data distribution and the like. At the present stage, machine learning and deep learning are also successful in various fields, and a foundation is laid for the Federal learning algorithm model to obtain better performance. The existing federal learning algorithm only carries out averaging processing on parameters of each local model, and firstly, personalization of each local model is not considered (for example, for an anti-fraud analysis system, user data characteristics of different financial institutions in a sample are not consistent, and the whole amount of loan is also different according to different regional economic levels), and the influence of different gravity centers of samples of characteristics caused by different environments of each client cannot be dealt with; secondly, the fact that most data in the actual environment are in non-Euclidean space, such as the association between users, financial knowledge maps and the like, is not considered, because the evaluation standards of the non-Euclidean space data are inconsistent, the data structure is irregular, and a high-performance model is difficult to obtain by the combined training of the data. Such a problem causes a drawback that learning is not accurate enough.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a federal learning method for anti-fraud of a credit card, and provides a federal learning algorithm which is suitable for non-European space data and personalized characteristics of participants to process financial data and judge anti-fraud of the credit card aiming at the prior problems of the prior credit card fraud assessment method and the prior classic federal learning algorithm.
In order to achieve the purpose, the invention adopts the following technical scheme:
in a first aspect, an embodiment of the present invention provides a federal learning method for anti-fraud of a credit card, where the federal learning method includes the following steps:
s1, building partial graph convolutional neural network models corresponding to K federal learning participants with different fraud categories; the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises a plurality of key characteristic information including user information, loan amount, deposit amount and credit investigation data, and the edge set between the nodes is ei,j＝(vi,vj) E belongs to E; a represents an adjacency matrix and defines the interconnection relationship among nodes; the fraud category comprises three types of card theft fraud, virtual application fraud and non-fraud;
s2, carrying out federal learning training by using a local graph convolutional neural network model; the method comprises the following steps that an attention mechanism is adopted to improve the aggregation process of the federal learning parameters, so that each partial graph convolutional neural network model has adaptive weight to aggregate;
and S3, outputting a global graph convolution neural network model, wherein the global graph convolution neural network model is used for processing the imported user data and identifying the corresponding fraud category.
Optionally, the adjacency matrix includes a region adjacency matrix for expressing a region distance or a relationship and an attribute adjacency matrix for expressing whether the user-related information is similar.
Optionally, in step S1, the process of building the partial graph convolutional neural network model corresponding to the K federate learning participants with different fraud classes includes the following steps:
s11, carrying out normalization preprocessing on the node data by adopting the following formula to obtain data to be processed:
Figure BDA0003105675530000021
wherein xiFor the characteristic raw data of the respective nodes,
Figure BDA0003105675530000022
is the characteristic mean, mu is the variance
S12, building an embedding layer, and embedding each node in the graph by a graph embedding method according to the following formula:
Figure BDA0003105675530000023
wherein N is the number of graph nodes, the superscript 0 represents the 0 th layer, namely the input layer, h represents the characteristic vector of each node, and omega represents the trainable weight matrix of the corresponding layer;
s13, building a graph convolution layer, aggregating the adjacent node characteristics according to the following formula, and updating the characteristic vector of the node:
Figure BDA0003105675530000024
where the superscript l denotes the number of layers,
Figure BDA0003105675530000025
adding an identity matrix to the original adjacency matrix to obtain a new adjacency matrix, so as to contain the self node, when λ is 1, the complete self node is contained,
Figure BDA0003105675530000026
obtaining a degree matrix according to a new adjacent matrix, wherein sigma represents an activation function;
s14, adding an attention mechanism module; in aggregation, different nodes are assigned different weights according to the following formula:
Figure BDA0003105675530000027
wherein
Figure BDA0003105675530000028
A feature vector representing the i-th node in the l-th network,
Figure BDA0003105675530000029
to represent
Figure BDA00031056755300000210
Of a neighboring node, WlRepresenting a feature vector dimension transformation matrix, att () representing an attention coefficient calculation function, i.e. calculating a correlation coefficient; transformed features of two neighboring nodes
Figure BDA00031056755300000211
Matrix transverse splicing and trainable parameters
Figure BDA00031056755300000212
And performing dot product operation to form a single-layer perceptron of which the hidden layer only has one neuron, inputting the characteristics of the spliced nodes, and outputting the similarity between the two nodes. The symbol represents the dot product operation, and the symbol | represents the matrix transverse splicing;
s15, normalizing the attention coefficient using the softmax function:
Figure BDA00031056755300000213
wherein
Figure BDA00031056755300000214
The serial number of the neighbor node of the i node;
s16, distributing attention weight, calculating and averaging the attention coefficients for multiple times by using a formula (4) for multiple times, accumulating the coefficients, averaging to obtain a final attention coefficient, adding the final attention coefficient into a graph convolution network, and modifying a feature vector updating formula to obtain:
Figure BDA0003105675530000031
where num _ att represents the number of attention coefficient calculations.
Alternatively, in step S14, the attention mechanism only works within a first-order neighbor node, i.e., only node pairs with directly connected edges are considered.
Optionally, in step S2, the process of performing federal learning training using the partial graph convolutional neural network model includes the following steps:
s21, initializing model parameters of the global graph convolutional neural network model;
s22, randomly selecting the federal learning participants according to the following formula:
Figure BDA0003105675530000032
wherein num _ fed is the number of the participants in the federal study, the federal study has K participants, the proportion of the participants participating in the calculation in each round is C, and the symbol
Figure BDA0003105675530000033
Is rounded down, max () is taken to be the maximum;
s23, downloading the to-be-learned parameters initialized by the global graph convolutional neural network model by the local graph convolutional neural network model;
s24, according to the downloaded parameters to be learned, the local graph convolution neural network model starts to train:
s241, setting a loss function, and updating model parameters of the local graph convolution neural network model by a batch stochastic gradient descent method according to the following formula:
Figure BDA0003105675530000034
wherein W represents the parameter to be learned in each partial graph convolutional neural network, eta represents the learning rate,
Figure BDA00031056755300000311
the loss-expressing function is used to calculate the difference between the predicted value of the neural network output of the graph and the true label,
Figure BDA0003105675530000035
it is shown that the partial derivative is taken,
Figure BDA0003105675530000036
the output result of the last layer of the graph convolution neural network is obtained;
s242, carrying out federal attention mechanism calculation on all the partial graph convolutional neural network models; and performing attention mechanism calculation on the kth local graph convolutional neural network model, uploading the calculated attention weight coefficient to the global graph convolutional neural network model and aggregating the calculated attention weight coefficient with other local graph convolutional neural network models:
Figure BDA0003105675530000037
wherein
Figure BDA0003105675530000038
An attention weight coefficient representing the kth local graph convolutional neural network model of the l layer, an
Figure BDA0003105675530000039
att () represents an attention mechanism calculation function, calculates a correlation between the local graph convolutional neural network model and the global graph convolutional neural network model,
Figure BDA00031056755300000310
trainable parameters, w, representing the kth local graph convolution neural network model at the l-th layerlTrainable parameters representing a first layer global graph convolutional neural network model;
s25, updating model parameters of the global graph convolutional neural network model, uploading the calculated attention weight coefficients of each local graph convolutional neural network model and the calculated parameters of the local graph convolutional neural network model to the global graph convolutional neural network model for aggregation:
Figure BDA0003105675530000041
wherein
Figure BDA0003105675530000042
Representing the attention weight coefficient assigned to the kth participant model at time t,
Figure BDA0003105675530000043
and the l-th layer parameters of the global graph convolutional neural network model after the t +1 time aggregation are represented.
In a second aspect, an embodiment of the present invention provides a federal learning apparatus for credit card fraud prevention, where the apparatus includes:
the local model building module is used for building local graph convolutional neural network models corresponding to K federal learning participants with different fraud categories; the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises a plurality of key characteristic information including user information, loan amount, deposit amount and credit investigation data, and the edge set between the nodes is ei,j＝(vi,vj) E belongs to E; a represents an adjacency matrix and defines the interconnection relationship among nodes; the fraud category comprises three types of card theft fraud, virtual application fraud and non-fraud;
the federal learning training module is used for performing federal learning training by using a local graph convolutional neural network model; the method comprises the following steps that an attention mechanism is adopted to improve the aggregation process of the federal learning parameters, so that each partial graph convolutional neural network model has adaptive weight to aggregate;
and the global graph convolutional neural network model is used for processing the imported user data and identifying corresponding fraud categories.
The invention has the beneficial effects that:
the invention takes the individuation of each participant into consideration, introduces the attention mechanism, and in deep learning, the attention mechanism can emphasize and highlight the characteristics of the participant sample, thereby taking the individuation problem of the participants into consideration. According to the method, the data of the non-European space can be considered, the relevance among the user data is fully utilized, the training precision of the federal learning model can be well improved, and the financial fraud assessment accuracy is further improved.
Drawings
Fig. 1 is a schematic diagram of a federated learning framework based on a graph convolutional neural network according to an embodiment of the present invention.
Fig. 2 is a schematic structural diagram of a graph convolution neural network according to an embodiment of the present invention.
Fig. 3 is a schematic diagram illustrating an attention mechanism of the neural network according to an embodiment of the present invention.
FIG. 4 is a schematic diagram of a neural network aggregation with attention mechanism according to an embodiment of the present invention.
Fig. 5 is a flowchart of a federal learning method for credit card fraud prevention according to an embodiment of the present invention.
Detailed Description
The present invention will now be described in further detail with reference to the accompanying drawings.
It should be noted that the terms "upper", "lower", "left", "right", "front", "back", etc. used in the present invention are for clarity of description only, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not limited by the technical contents of the essential changes.
Example one
Fig. 5 is a flowchart of a federal learning method for credit card fraud prevention according to an embodiment of the present invention. The present embodiment is applicable to the case where the credit card fraud category is identified by a device such as a server, and the method may be performed by a federal learning apparatus for credit card fraud prevention, which may be implemented in software and/or hardware, and may be integrated in an electronic device, such as an integrated server device.
The federal learning method is mainly used for identifying two most common credit card fraud types, wherein the type 1 is card theft fraud: act of surrendering or embezzlement of the lost credit card for transaction; type 2 is virtual application fraud: the applicant uses false information to apply credit cards and avoids the audit of the card issuing organization. Therefore, the objective of the federal learning method proposed in this embodiment is to improve the accuracy of credit card fraud three classification tasks (type 1, type 2, non-fraud) with limited samples and protection of private data of the parties.
Assuming that the federal study has K participants, the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises 20 types of key characteristic information such as user information, loan amount, deposit amount, credit investigation data and the like, and the edge set between the nodes is ei,j＝(vi,vj) E, for example, the user name of the node i is called zhang san, the nodes connected with the node i by the edge have certain correlation, such as related companies, the same contact way, and the like, a represents an adjacency matrix, defines the interconnection relationship between the nodes, and is divided into two types of adjacency matrices according to different connection relationships, as shown below.
Region adjacency matrix
Figure BDA0003105675530000051
Wherein GD is a regionEnglish acronym for Distance geographic Distance. According to whether the living addresses provided by the user nodes i and j are adjacent, if the linear distance of the living location is less than M kilometers, the living addresses are judged to be adjacent, and because different users with similar distances are close to each other, important consideration is needed. Or if the prior knowledge can know that the relatives exist among the users, the users are also judged to be in the adjacent relationship.
Figure BDA0003105675530000052
Attribute adjacency matrix
Figure BDA0003105675530000053
Where FD is an English acronym for Feature Distance. According to whether the user-related information provided by the user nodes i and j is similar, for example, the contact telephone of different users is the same, and the working units are the same.
Figure BDA0003105675530000054
In the training process, the proportion of participants participating in calculation in each round is C, the number of times of completely training respective local data of each participant in each round is Epoch, and the minimum batch data volume B is used for updating the participant model.
Referring to fig. 5, step 1: and building a local graph convolution neural network (GCN) model of the federal learning participator. The innovation point of the step 1 is that the graph convolution neural network is set as a local model of federal learning, irregular data can be excellently processed according to the characteristic that the graph convolution network can, and therefore the problem that various non-European space data cannot be well utilized by a conventional deep learning network in real life is solved, and the specific steps are shown in steps 1.2-1.4.
Step 1.1: and carrying out normalization pretreatment on the node data to obtain data to be processed. Each node viCorrespond to respective features xiHowever, the range difference of the original values may be very largeIn many cases, the objective function may not work properly, and also to accelerate the convergence speed of the gradient descent. The invention normalizes the original data by mean variance normalization.
Figure BDA0003105675530000061
Wherein xiFor the characteristic raw data of the respective nodes,
Figure BDA0003105675530000062
μ is the variance, which is the feature mean.
Step 1.2: building an embedding layer, and embedding and representing each node in the graph by a graph embedding method, as shown in formula (2):
Figure BDA0003105675530000063
wherein N is the number of nodes of a graph, the superscript 0 represents the 0 th layer, namely the input layer, h represents the feature vector of each node, and omega represents the trainable weight matrix of the corresponding layer.
Step 1.3: and (4) building a graph convolution layer, as shown in a formula (3), aggregating the characteristics of adjacent nodes, and updating the characteristic vectors of the nodes to achieve the purpose of extracting the characteristics.
Figure BDA0003105675530000064
Where the superscript l denotes the number of layers,
Figure BDA0003105675530000065
adding an identity matrix to the original adjacency matrix to obtain a new adjacency matrix, so as to contain the self node, when λ is 1, the complete self node is contained,
Figure BDA0003105675530000066
is based on a new adjacency matrix degree matrix, and sigma represents the ReLU activation functionAnd (4) counting. Fig. 2 is a schematic structural diagram of a graph convolution neural network according to an embodiment of the present invention.
Step 1.4: and an attention mechanism module is added, and the adjacent nodes with large influence are focused. Different nodes are assigned different weights when aggregated, as shown in equation (4).
Figure BDA0003105675530000067
Wherein
Figure BDA0003105675530000068
A feature vector representing the i-th node in the l-th network,
Figure BDA0003105675530000069
to represent
Figure BDA00031056755300000610
Of a neighboring node, WlRepresenting a feature vector dimension transformation matrix, att () representing an attention coefficient calculation function, namely calculating a correlation coefficient, firstly transforming the features of two neighboring nodes
Figure BDA00031056755300000611
Matrix transverse splicing and trainable parameters
Figure BDA00031056755300000612
And performing dot product operation to form a single-layer perceptron of which the hidden layer only has one neuron, inputting the characteristics of the spliced nodes, and outputting the similarity between the two nodes. The symbol represents the dot product operation, and the symbol | | | represents the matrix horizontal concatenation. To reduce the amount of computation, the attention mechanism only works within a first order neighbor node, i.e., only node pairs with directly connected edges are considered.
Step 1.4.1: and normalizing the attention coefficient to facilitate the application of the attention coefficient to each node, as shown in formula (5). Normalization was performed using the softmax function.
Figure BDA00031056755300000613
Wherein
Figure BDA00031056755300000614
The sequence number of the neighbor node of the inode. The softmax function is well known in the art and will not be described in detail herein. Fig. 3 is a schematic diagram illustrating an attention mechanism of the neural network according to an embodiment of the present invention.
Step 1.4.2: and (3) allocating attention weight, using formula (4) for multiple times, namely performing multiple attention coefficient calculation averaging, accumulating the coefficients, averaging, adding the accumulated coefficients into a graph convolution network, and modifying the feature vector updating formula in formula (3) as shown in formula (6).
Figure BDA0003105675530000071
The equation (6) is different from the equation (3) in that an adjacency matrix is used when neighbor node features are aggregated in the equation (3), the same weight coefficient is distributed but cannot be changed, and the equation (6) distributes different weight coefficients for different neighbor nodes, so that the correlation among the nodes is better considered. FIG. 4 is a schematic diagram of a neural network aggregation with attention mechanism according to an embodiment of the present invention.
Step 2: federal learning training is performed using a graph convolutional neural network (GCN). Fig. 1 is a schematic diagram of a federated learning framework based on a graph convolutional neural network according to an embodiment of the present invention. The innovation point of the step 2 is that: unlike the traditional federal learning average aggregation algorithm, which is to average and aggregate the local models of the federal learning participants in the global model with the same weight, this approach is not favorable for the individualization of the model, i.e. each local model cannot be adapted to the respective field. The patent improves the aggregation process of the federal learning parameters through an attention mechanism, so that each local model has a weight which is relatively suitable for the local model to aggregate. Thereby reducing the influence of data noise and increasing the degree of individuation. The specific steps are step 2.4
Step 2.1: initializing global graph convolutional neural network GCN _ G model parameters
Step 2.2: the federal learning participants are randomly selected as shown in equation (7).
Figure BDA0003105675530000072
Wherein num _ fed is the number of the participants in the federal study, the federal study has K participants, the proportion of the participants participating in the calculation in each round is C, and the symbol
Figure BDA0003105675530000073
Is rounded down, if C · K is 2.99, then
Figure BDA0003105675530000074
max () is taken to be maximum, max (2,1) ═ 2.
Step 2.3: and (4) downloading the parameters to be learned after the global model is initialized by the local model (the global model is consistent with the graph convolution neural network structure built in the step 1).
Step 2.4: local graph convolutional neural network GCN _ L model training
Step 2.4.1: a loss function is set and the parameters are updated by a batch stochastic gradient descent method, as shown in equation (8) below.
Figure BDA0003105675530000075
Wherein W represents the parameter to be learned in each partial graph convolutional neural network, eta represents the learning rate,
Figure BDA0003105675530000078
the loss-expressing function is used to calculate the difference between the predicted value y' of the graph neural network output and the true label y,
Figure BDA0003105675530000076
indicating partial derivative. At this time
Figure BDA0003105675530000077
Output results for the last layer of the graph convolution neural network
Step 2.4.2: federal attention mechanism calculations. Calculating GCN _ L for kth local graph convolution neural network modelkAnd performing attention mechanism operation, and calculating attention weight coefficients so as to establish a personalized model for uploading to the global model and aggregating with other local models. As shown in equation (9).
Figure BDA0003105675530000081
Wherein
Figure BDA0003105675530000082
Represents the attention weight coefficient of the kth local graph convolution network of the l layer, and
Figure BDA0003105675530000083
att () represents an attention mechanism calculation function, calculates the correlation between local GCN _ L and global GCN _ G,
Figure BDA0003105675530000084
denotes the kth local GCN _ L of the L-th layerkTrainable parameters of, w1Trainable parameters representing the global GCN _ G of the l-th layer. As with equation (4), the notation · represents the dot product operation, and the notation | | | represents the matrix horizontal concatenation.
Step 2.5: updating global GCN _ G model parameters, and calculating each local GCN _ L obtained in the step 2.4.2kAttention weight coefficient in conjunction with local GCN _ L calculated in step 2.4.1kModel parameters are uploaded into the GCN _ G global model for aggregation. As shown in equation (10).
Figure BDA0003105675530000085
Wherein
Figure BDA0003105675530000086
Representing the attention weight coefficient assigned to the kth participant model at time t,
Figure BDA0003105675530000087
and the l-th layer parameters of the global model after aggregation at the t +1 moment are represented.
The invention still adopts a frame of a classical horizontal federal learning model, the innovative content of the model is mainly divided into two modules, the first module is used for improving the capability of the model for processing non-European data, the relevance between the data can be mined by utilizing a graph convolution neural network, a participator adopts the graph convolution neural network as a local model to carry out data modeling, and an attention mechanism module is used in the graph convolution neural network so as to reduce the influence of data noise. The second module provides an improved attention mechanism algorithm for relieving the problems that the average aggregation algorithm causes the model to lack individuation and the irregular data has noise, and provides a proper attention weight for each participant model, and the weight is applied to the parameters of each layer so as to improve the individuation degree of the model parameters, cause the participant models to be more suitable for respective fields and further reduce the noise influence caused by the irregular data structure to a certain degree.
Example two
The embodiment of the invention provides a federal learning device for anti-fraud of a credit card, which comprises a local model building module, a federal learning training module and a global graph convolution neural network model.
The local model building module is used for building local graph convolutional neural network models corresponding to K federal learning participants with different fraud categories; the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises user information and loan amountA plurality of key characteristic information including deposit amount and credit investigation data, and the edge set between nodes is ei,j＝(vi,vj) E belongs to E; a represents an adjacency matrix and defines the interconnection relationship among nodes; the fraud categories include card theft fraud, virtual application fraud and no fraud.
The federal learning training module is used for performing federal learning training by using a local graph convolutional neural network model; and improving the aggregation process of the federal learning parameters by adopting an attention mechanism, so that each partial graph convolutional neural network model has a weight adapted to the partial graph convolutional neural network model for aggregation.
And the global graph convolutional neural network model is used for processing the imported user data and identifying corresponding fraud categories.
Through the federal learning device in the second embodiment of the invention, the transmission object is determined by establishing the data inclusion relation of the whole application, so that the aim of identifying the credit card fraud category is achieved. The federal learning device provided by the embodiment of the invention can execute the federal learning method for anti-fraud of the credit card provided by any embodiment of the invention, and has corresponding functional modules and beneficial effects of the execution method.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.

Claims (6)

1. A federal learning method for credit card fraud prevention, comprising the steps of:
s1, building partial graph convolutional neural network models corresponding to K federal learning participants with different fraud categories; the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises a plurality of key characteristic information including user information, loan amount, deposit amount and credit investigation data, and the edge set between the nodes is ei，j＝(vi，vj) E belongs to E; a represents an adjacency matrix and defines the interconnection relationship among nodes; the fraud category comprises three types of card theft fraud, virtual application fraud and non-fraud;
s2, carrying out federal learning training by using a local graph convolutional neural network model; the method comprises the following steps that an attention mechanism is adopted to improve the aggregation process of the federal learning parameters, so that each partial graph convolutional neural network model has adaptive weight to aggregate;
and S3, outputting a global graph convolution neural network model, wherein the global graph convolution neural network model is used for processing the imported user data and identifying the corresponding fraud category.
2. The federal learning method for credit card fraud prevention as claimed in claim 1, wherein the adjacency matrix includes both a regional adjacency matrix for expressing regional distance or relationship and an attribute adjacency matrix for expressing whether the user-related information is similar.
3. The federal learning method for credit card fraud prevention as claimed in claim l, wherein in step S1, the process of building partial graph convolutional neural network models corresponding to K federal learning participants with different fraud classes includes the following steps:
s11, carrying out normalization preprocessing on the node data by adopting the following formula to obtain data to be processed:
Figure FDA0003105675520000011
wherein xiFor the characteristic raw data of the respective nodes,
Figure FDA0003105675520000012
is the characteristic mean, mu is the variance
S12, building an embedding layer, and embedding each node in the graph by a graph embedding method according to the following formula:
Figure FDA0003105675520000013
wherein N is the number of graph nodes, the superscript 0 represents the 0 th layer, namely the input layer, h represents the characteristic vector of each node, and omega represents the trainable weight matrix of the corresponding layer;
s13, building a graph convolution layer, aggregating the adjacent node characteristics according to the following formula, and updating the characteristic vector of the node:
Figure FDA0003105675520000014
where the superscript l denotes the number of layers,
Figure FDA0003105675520000015
adding an identity matrix to the original adjacency matrix to obtain a new adjacency matrix, so as to contain the self node, when λ is 1, the complete self node is contained,
Figure FDA0003105675520000016
obtaining a degree matrix according to a new adjacent matrix, wherein sigma represents an activation function;
s14, adding an attention mechanism module; in aggregation, different nodes are assigned different weights according to the following formula:
Figure FDA0003105675520000017
wherein
Figure FDA0003105675520000021
A feature vector representing the i-th node in the l-th network,
Figure FDA0003105675520000022
to represent
Figure FDA0003105675520000023
Of a neighboring node, WlRepresenting a feature vector dimension transformation matrix, att () representing an attention coefficient calculation function, i.e. calculating a correlation coefficient; transformed features of two neighboring nodes
Figure FDA0003105675520000024
Matrix transverse splicing and trainable parameters
Figure FDA0003105675520000025
And performing dot product operation to form a single-layer perceptron of which the hidden layer only has one neuron, inputting the characteristics of the spliced nodes, and outputting the similarity between the two nodes. The symbol represents the dot product operation, and the symbol | represents the matrix transverse splicing;
s15, normalizing the attention coefficient using the softmax function:
Figure FDA0003105675520000026
wherein
Figure FDA0003105675520000027
The serial number of the neighbor node of the i node;
s16, distributing attention weight, calculating and averaging the attention coefficients for multiple times by using a formula (4) for multiple times, accumulating the coefficients, averaging to obtain a final attention coefficient, adding the final attention coefficient into a graph convolution network, and modifying a feature vector updating formula to obtain:
Figure FDA0003105675520000028
where num _ att represents the number of attention coefficient calculations.
4. A federal learning method for credit card fraud prevention as claimed in claim 3, wherein in step S14, the attention mechanism is applied only in a first-order neighbor node, i.e., only the node pairs with directly connected edges are considered.
5. The federal learning method for credit card fraud prevention as claimed in claim 1, wherein the process of using the partial graph convolutional neural network model for federal learning training in step S2 includes the steps of:
s21, initializing model parameters of the global graph convolutional neural network model;
s22, randomly selecting the federal learning participants according to the following formula:
Figure FDA0003105675520000029
wherein num _ fed is the number of the participants in the federal study, the federal study has K participants, the proportion of the participants participating in the calculation in each round is C, and the symbol
Figure FDA00031056755200000210
Is rounded down, max () is taken to be the maximum;
s23, downloading the to-be-learned parameters initialized by the global graph convolutional neural network model by the local graph convolutional neural network model;
s24, according to the downloaded parameters to be learned, the local graph convolution neural network model starts to train:
s241, setting a loss function, and updating model parameters of the local graph convolution neural network model by a batch stochastic gradient descent method according to the following formula:
Figure FDA00031056755200000211
wherein W represents the parameter to be learned in each partial graph convolutional neural network, eta represents the learning rate,
Figure FDA00031056755200000214
the loss-expressing function is used to calculate the difference between the predicted value of the neural network output of the graph and the true label,
Figure FDA00031056755200000212
it is shown that the partial derivative is taken,
Figure FDA00031056755200000213
the output result of the last layer of the graph convolution neural network is obtained;
s242, carrying out federal attention mechanism calculation on all the partial graph convolutional neural network models; and performing attention mechanism calculation on the kth local graph convolutional neural network model, uploading the calculated attention weight coefficient to the global graph convolutional neural network model and aggregating the calculated attention weight coefficient with other local graph convolutional neural network models:
Figure FDA0003105675520000031
wherein
Figure FDA0003105675520000032
An attention weight coefficient representing the kth local graph convolutional neural network model of the l layer, an
Figure FDA0003105675520000033
att () represents an attention mechanism calculation function, calculates a correlation between the local graph convolutional neural network model and the global graph convolutional neural network model,
Figure FDA0003105675520000034
trainable parameters, w, representing the kth local graph convolution neural network model at the l-th layerlTrainable parameters representing a first layer global graph convolutional neural network model;
s25, updating model parameters of the global graph convolutional neural network model, uploading the calculated attention weight coefficients of each local graph convolutional neural network model and the calculated parameters of the local graph convolutional neural network model to the global graph convolutional neural network model for aggregation:
Figure FDA0003105675520000035
wherein
Figure FDA0003105675520000036
Representing the attention weight coefficient assigned to the kth participant model at time t,
Figure FDA0003105675520000037
and the l-th layer parameters of the global graph convolutional neural network model after the t +1 time aggregation are represented.
6. A federal learning device for credit card fraud prevention, the federal learning device comprising:
the local model building module is used for building local graph convolutional neural network models corresponding to K federal learning participants with different fraud categories; the local undirected graph structure data owned by each participant is Gi(V, E, A) (i ∈ K), where the set of nodes in the graph structure is Vi∈V，viThe feature on a node is xiBelongs to X, each node comprises a plurality of key characteristic information including user information, loan amount, deposit amount and credit investigation data, and the edge set between the nodes is ei,j＝(vi,vj) E belongs to E; a represents an adjacency matrix and defines the interconnection relationship among nodes; the fraud categories include stolen card fraud, virtual application fraud and no fraudThree types are adopted;
the federal learning training module is used for performing federal learning training by using a local graph convolutional neural network model; the method comprises the following steps that an attention mechanism is adopted to improve the aggregation process of the federal learning parameters, so that each partial graph convolutional neural network model has adaptive weight to aggregate;
and the global graph convolutional neural network model is used for processing the imported user data and identifying corresponding fraud categories.
CN202110635863.6A 2021-06-08 2021-06-08 Federal learning method and device for credit card anti-fraud Active CN113362160B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110635863.6A CN113362160B (en) 2021-06-08 2021-06-08 Federal learning method and device for credit card anti-fraud

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110635863.6A CN113362160B (en) 2021-06-08 2021-06-08 Federal learning method and device for credit card anti-fraud

Publications (2)

Publication Number Publication Date
CN113362160A true CN113362160A (en) 2021-09-07
CN113362160B CN113362160B (en) 2023-08-22

Family

ID=77533029

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110635863.6A Active CN113362160B (en) 2021-06-08 2021-06-08 Federal learning method and device for credit card anti-fraud

Country Status (1)

Country Link
CN (1) CN113362160B (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743677A (en) * 2021-09-16 2021-12-03 成都数融科技有限公司 Personal credit evaluation model training method and evaluation method based on federal learning
CN114117926A (en) * 2021-12-01 2022-03-01 南京富尔登科技发展有限公司 Robot cooperative control algorithm based on federal learning
CN114676849A (en) * 2022-03-24 2022-06-28 支付宝(杭州)信息技术有限公司 Method and system for updating model parameters based on federal learning
CN114693317A (en) * 2022-04-08 2022-07-01 重庆邮电大学 Telecommunication fraud security federation detection method fusing homogeneous graph and bipartite graph
CN114781553A (en) * 2022-06-20 2022-07-22 浙江大学滨江研究院 Unsupervised patent clustering method based on parallel multi-graph convolution neural network
CN115577858A (en) * 2022-11-21 2023-01-06 山东能源数智云科技有限公司 Block chain-based carbon emission prediction method and device and electronic equipment
CN116049769A (en) * 2023-04-03 2023-05-02 湖南大学 Discrete object data relevance prediction method and system and storage medium
CN116229560A (en) * 2022-09-08 2023-06-06 广东省泰维思信息科技有限公司 Abnormal behavior recognition method and system based on human body posture
CN116501978A (en) * 2023-06-28 2023-07-28 杭州金智塔科技有限公司 Recommendation model generation method and device based on privacy protection machine forgetting algorithm
CN116502709A (en) * 2023-06-26 2023-07-28 浙江大学滨江研究院 Heterogeneous federal learning method and device
CN116527824A (en) * 2023-07-03 2023-08-01 北京数牍科技有限公司 Method, device and equipment for training graph convolution neural network

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020029585A1 (en) * 2018-08-10 2020-02-13 深圳前海微众银行股份有限公司 Neural network federation modeling method and device employing transfer learning, and storage medium
CN111325619A (en) * 2018-12-15 2020-06-23 深圳先进技术研究院 Credit card fraud detection model updating method and device based on joint learning
CN112199702A (en) * 2020-10-16 2021-01-08 鹏城实验室 Privacy protection method, storage medium and system based on federal learning
CN112418520A (en) * 2020-11-22 2021-02-26 同济大学 Credit card transaction risk prediction method based on federal learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020029585A1 (en) * 2018-08-10 2020-02-13 深圳前海微众银行股份有限公司 Neural network federation modeling method and device employing transfer learning, and storage medium
CN111325619A (en) * 2018-12-15 2020-06-23 深圳先进技术研究院 Credit card fraud detection model updating method and device based on joint learning
CN112199702A (en) * 2020-10-16 2021-01-08 鹏城实验室 Privacy protection method, storage medium and system based on federal learning
CN112418520A (en) * 2020-11-22 2021-02-26 同济大学 Credit card transaction risk prediction method based on federal learning

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743677A (en) * 2021-09-16 2021-12-03 成都数融科技有限公司 Personal credit evaluation model training method and evaluation method based on federal learning
CN113743677B (en) * 2021-09-16 2023-06-30 成都数融科技有限公司 Personal credit evaluation model training method and evaluation method based on federal learning
CN114117926A (en) * 2021-12-01 2022-03-01 南京富尔登科技发展有限公司 Robot cooperative control algorithm based on federal learning
CN114117926B (en) * 2021-12-01 2024-05-14 南京富尔登科技发展有限公司 Robot cooperative control algorithm based on federal learning
CN114676849A (en) * 2022-03-24 2022-06-28 支付宝(杭州)信息技术有限公司 Method and system for updating model parameters based on federal learning
CN114693317A (en) * 2022-04-08 2022-07-01 重庆邮电大学 Telecommunication fraud security federation detection method fusing homogeneous graph and bipartite graph
CN114781553A (en) * 2022-06-20 2022-07-22 浙江大学滨江研究院 Unsupervised patent clustering method based on parallel multi-graph convolution neural network
CN116229560B (en) * 2022-09-08 2024-03-19 广东省泰维思信息科技有限公司 Abnormal behavior recognition method and system based on human body posture
CN116229560A (en) * 2022-09-08 2023-06-06 广东省泰维思信息科技有限公司 Abnormal behavior recognition method and system based on human body posture
CN115577858A (en) * 2022-11-21 2023-01-06 山东能源数智云科技有限公司 Block chain-based carbon emission prediction method and device and electronic equipment
CN116049769A (en) * 2023-04-03 2023-05-02 湖南大学 Discrete object data relevance prediction method and system and storage medium
CN116502709A (en) * 2023-06-26 2023-07-28 浙江大学滨江研究院 Heterogeneous federal learning method and device
CN116501978A (en) * 2023-06-28 2023-07-28 杭州金智塔科技有限公司 Recommendation model generation method and device based on privacy protection machine forgetting algorithm
CN116527824A (en) * 2023-07-03 2023-08-01 北京数牍科技有限公司 Method, device and equipment for training graph convolution neural network
CN116527824B (en) * 2023-07-03 2023-08-25 北京数牍科技有限公司 Method, device and equipment for training graph convolution neural network

Also Published As

Publication number Publication date
CN113362160B (en) 2023-08-22

Similar Documents

Publication Publication Date Title
CN113362160B (en) Federal learning method and device for credit card anti-fraud
CN109255506B (en) Internet financial user loan overdue prediction method based on big data
CN112418520B (en) Credit card transaction risk prediction method based on federal learning
CN108985929B (en) Training method, business data classification processing method and device, and electronic equipment
CN110188198A (en) A kind of anti-fraud method and device of knowledge based map
KR102009309B1 (en) Management automation system for financial products and management automation method using the same
CN108734338A (en) Credit risk forecast method and device based on LSTM models
CN106897918A (en) A kind of hybrid machine learning credit scoring model construction method
CN106960358A (en) A kind of financial fraud behavior based on rural area electronic commerce big data deep learning quantifies detecting system
CN110138595A (en) Time link prediction technique, device, equipment and the medium of dynamic weighting network
CN106056444A (en) Data processing method and device
CN111260462A (en) Transaction fraud detection method based on heterogeneous relation network attention mechanism
Sayjadah et al. Credit card default prediction using machine learning techniques
CN109739844B (en) Data classification method based on attenuation weight
CN109767225B (en) Network payment fraud detection method based on self-learning sliding time window
CN108446291A (en) The real-time methods of marking and points-scoring system of user credit
CN113011895B (en) Associated account sample screening method, device and equipment and computer storage medium
CN112581265A (en) Internet financial client application fraud detection method based on AdaBoost
CN110889759A (en) Credit data determination method, device and storage medium
Zhang et al. Surface and high-altitude combined rainfall forecasting using convolutional neural network
CN114255121A (en) Credit risk prediction model training method and credit risk prediction method
CN113409157B (en) Cross-social network user alignment method and device
CN115510948A (en) Block chain fishing detection method based on robust graph classification
Adedoyin et al. Evaluating Case-Based Reasoning Knowledge Discovery in Fraud Detection.
CN110942391A (en) Method for determining main activity label of multi-activity label user in block chain

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant