Cross-domain multi-dimensional federal learning method and system for anti-fraud model training

Abstract

The invention discloses a cross-domain multi-dimensional federated learning method and system for anti-fraud model training. The cross-domain multi-dimensional federal learning method for anti-fraud model training comprises the following steps: establishing to-be-trained global models of different fraud types; acquiring undertaking information of each regional dimension service system; building a local training model of the to-be-trained global model of each fraud type; setting a regional dimension service system associated with the local training model; setting a client dimension system associated with the local training model; clustering is carried out according to a local training model associated with a client dimension system, so that clustering clusters are obtained, and each clustering cluster is coordinated through a coordinator; and each client dimension system in the clustering cluster is used for training the local training model associated with the client dimension system. According to the method, different regions are trained through different types of to-be-trained global models, so that the purpose of reasonable allocation according to data characteristics of different regions is achieved.

Description

Technical Field

[0001] The present application relates to the field of model training technology, and specifically to a cross-domain multi-dimensional federated learning method for anti-fraud model training and a cross-domain multi-dimensional federated learning system for anti-fraud model training. Background Art

[0002] In the existing technology, there are many types of fraud, and each type of fraud can now be identified through neural networks. However, the main fraud methods that may exist in different provinces are inconsistent, and there may not be enough data to train some methods. However, based on anti-fraud needs, all types of threats need to be able to be identified and reported. Therefore, it is ultimately necessary to integrate the models of different provinces, aggregate them nationwide to obtain the final model and distribute it. Summary of the invention

[0003] The purpose of the present invention is to provide a cross-domain multi-dimensional federated learning method for anti-fraud model training to solve at least one of the above-mentioned technical problems.

[0004] One aspect of the present invention provides a cross-domain multi-dimensional federated learning method for anti-fraud model training, the cross-domain multi-dimensional federated learning method for anti-fraud model training comprising: The central service system establishes multiple global models to be trained for different fraud types; The central service system obtains the undertaking information of each regional dimension service system; According to the acceptance information of each regional dimension service system, a local training model of the global model to be trained for each fraud type is built, wherein a global model to be trained for one fraud type can build at least two or more local training models; The central service system sets up the regional dimension service system associated with each local training model; Each regional dimension service system sets up the client dimension system associated with each local training model; Clustering is performed according to the local training models associated with the client dimension systems in the same regional dimension service system, so as to obtain cluster clusters. The client dimension systems in each cluster cluster coordinate through a coordinator when training their respective local training models. The clusters corresponding to each coordinator are managed through the coordinator, so that each client dimension system in the cluster can train its associated local training model.

[0005] Optionally, the cross-domain multi-dimensional federated learning method for anti-fraud model training further includes: Each client dimension system in each cluster trains its associated local training model and obtains the client dimension gradient information corresponding to each client dimension system; The coordinator of each cluster obtains the dimension gradient information of each client in the cluster; The coordinator of each cluster generates the cluster gradient information corresponding to the cluster according to the dimension gradient information of each client under the cluster; The regional dimension service system generates regional gradient information based on the gradient information of each cluster cluster obtained; The central service system obtains the gradient information of each region and obtains the gradient information of the global model through the gradient information of each region.

[0006] Optionally, the fraud types include contract fraud, bill fraud, loan fraud, credit card fraud, letter of credit fraud, financial certificate fraud, fund-raising fraud, securities fraud, insurance fraud, and swindling; The global models to be trained include contract fraud global models to be trained, bill fraud global models to be trained, loan fraud global models to be trained, credit card fraud global models to be trained, letter of credit fraud global models to be trained, financial voucher fraud global models to be trained, fundraising fraud global models to be trained, securities fraud global models to be trained, insurance fraud global models to be trained, and swindling global models to be trained.

[0007] Optionally, the cross-domain multi-dimensional federated learning method for anti-fraud model training includes: Acquire a region type library, wherein the region type library includes at least one preset region name and a preset fraud type corresponding to each preset region name; Get the preset fraud type corresponding to the preset region name that is the same as the name of the region dimension service system; According to the preset fraud type, the global model to be trained that the regional dimension service system needs to train is obtained as the acceptance information.

[0008] Optionally, the method of building a local training model for the global model to be trained for each fraud type according to the acceptance information of each regional dimension service system includes: Perform the following operations for each fraud type's global model to be trained: Obtain the number of regional dimension service systems that need to train the global model to be trained for the fraud type; A local training model is established for each regional dimension service system that needs to train the global model to be trained for the fraud type, and one local training model corresponds to one regional dimension service system.

[0009] Optionally, each regional dimension service system sets a client dimension system associated with each local training model, including: Obtain a client dimension type library, wherein the client dimension type library includes at least one preset client name and a preset client fraud type corresponding to each preset client name; Get the preset client fraud type corresponding to the preset client name that is the same as the name of the client dimension system; According to the fraud type of the preset client, the local training model that needs to be trained for each preset client is obtained and the preset client is associated with the local training model that needs to be trained.

[0010] Optionally, clustering according to each local training model associated with each client dimension system in the same regional dimension service system to obtain a cluster cluster includes: Obtain the local training model that needs to be trained for each client dimension system; Multiple client-dimensional systems with the same local training models that need to be trained are grouped into a cluster.

[0011] Optionally, when a client dimension system is associated with a plurality of local training models, clustering the local training models associated with the client dimension systems in the same regional dimension service system to obtain cluster clusters further includes: Setting a training priority for each associated local training model to be trained by a client dimension system having multiple associated local training models; The local training model with the highest training priority is obtained as a basis for forming clusters.

[0012] The present application also provides a cross-domain multi-dimensional federated learning system for anti-fraud model training, wherein the cross-domain multi-dimensional federated learning system for anti-fraud model training includes a central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems, wherein the central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems cooperate to implement the cross-domain multi-dimensional federated learning method for anti-fraud model training as described above.

[0013] The present application also provides a cross-domain multi-dimensional federated learning system for anti-fraud model training, wherein the cross-domain multi-dimensional federated learning system for anti-fraud model training includes a central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems, wherein the central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems cooperate to implement the cross-domain multi-dimensional federated learning method for anti-fraud model training as described above.

[0014] Beneficial effects:

[0015] The cross-domain multi-dimensional federated learning method for anti-fraud model training of the present application gives different types of global models to be trained to different regions for training, thereby achieving reasonable allocation according to the data characteristics and data volume of different regions. In addition, the training speed can be accelerated by setting the coordinator through clustering clusters. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a structural diagram of a cross-domain multi-dimensional federated learning method for anti-fraud model training according to an embodiment of the present application; Figure 2 It is a system schematic diagram of a cross-domain multi-dimensional federated learning system for anti-fraud model training in one embodiment of the present application; Figure 3 It is a regional aggregation diagram of a cross-domain multi-dimensional federated learning method for anti-fraud model training in one embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the embodiment of this application will be described in more detail below in conjunction with the drawings in the embodiment of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described embodiments are part of the embodiments of this application, not all of them. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be construed as limitations on this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The embodiments of this application are described in detail below in conjunction with the drawings.

[0018] Figure 1 It is a structural diagram of a cross-domain multi-dimensional federated learning method for anti-fraud model training in one embodiment of the present application.

[0019] like Figure 1 The cross-domain multi-dimensional federated learning method for anti-fraud model training shown includes: The central service system establishes multiple global models to be trained for different fraud types; The central service system obtains the undertaking information of each regional dimension service system; According to the acceptance information of each regional dimension service system, a local training model of the global model to be trained for each fraud type is built, wherein a global model to be trained for one fraud type can build at least two or more local training models; The central service system sets up the regional dimension service system associated with each local training model; Each regional dimension service system sets up the client dimension system associated with each local training model; Clustering is performed according to the local training models associated with the client dimension systems in the same regional dimension service system, so as to obtain cluster clusters. The client dimension systems in each cluster cluster coordinate through a coordinator when training their respective local training models. The clusters corresponding to each coordinator are managed through the coordinator, so that each client dimension system in the cluster can train its associated local training model.

[0020] The cross-domain multi-dimensional federated learning method for anti-fraud model training of the present application gives different types of global models to be trained to different regions for training, thereby achieving reasonable allocation according to the data characteristics and data volume of different regions. In addition, the training speed can be accelerated by setting the coordinator through clustering clusters.

[0021] In this embodiment, the cross-domain multi-dimensional federated learning method for anti-fraud model training further includes: Each client dimension system in each cluster trains its associated local training model and obtains the client dimension gradient information corresponding to each client dimension system; The coordinator of each cluster obtains the dimension gradient information of each client in the cluster; The coordinator of each cluster generates the cluster gradient information corresponding to the cluster according to the dimension gradient information of each client under the cluster; The regional dimension service system generates regional gradient information based on the gradient information of each cluster cluster obtained; The central service system obtains the gradient information of each region and obtains the gradient information of the global model through the gradient information of each region.

[0022] For example, suppose there is a regional dimension service system A1 in regional dimension A; there are m regional coordinators under regional dimension service system A1, and each coordinator is represented as .

[0023] Taking the i-th coordinator as an example, there are n client dimension systems in the domain, represented as ,express The j-th client dimension system within; Client Dimension System On-premises data Train on the local model to obtain the client dimension gradient information updated by the local model , and Upload to , Coordinator Receive client dimension gradient information of all client dimension systems Afterwards, for each Calculate the sum of the Euclidean distances to the rest of the gradients , select the one with the smallest score d , as the aggregated cluster gradient information Sent to the regional dimension service system A1 (krum robust aggregation).

[0024] The regional dimension service system A1 receives the gradient data sent by all coordinators Then, the average value of all gradient data is calculated. It is used as the regional gradient information for training within the region and sent to the central service system.

[0025] The above operations are performed in the remaining areas, and finally the central service system obtains regional gradient information from different areas. .

[0026] Assume that the central service system CS has obtained regional gradient information from different regions .

[0027] See also Figure 3 , the central server CS adopts The mapping function maps all gradients into the subspace sphere and performs feature clustering in the space, which is expressed as F: . Then, CS will vector The vector feature matrix M is obtained by vertical splicing by row, and the mean of each column in M ​​is calculated to finally obtain the aggregated vector As the gradient information of the global model, Distributed to each regional dimension service system, the regional dimension service system then sends it down to the client dimension system layer by layer, and the client dimension system uses Update the local model.

[0028] M vector matrix concatenation example: Suppose there are three vectors , then the final vector matrix M = .

[0029] In this embodiment, the fraud types include contract fraud, bill fraud, loan fraud, credit card fraud, letter of credit fraud, financial certificate fraud, fund-raising fraud, securities fraud, insurance fraud, and swindling; The global models to be trained include contract fraud global models to be trained, bill fraud global models to be trained, loan fraud global models to be trained, credit card fraud global models to be trained, letter of credit fraud global models to be trained, financial voucher fraud global models to be trained, fundraising fraud global models to be trained, securities fraud global models to be trained, insurance fraud global models to be trained, and swindling global models to be trained.

[0030] In this embodiment, the cross-domain multi-dimensional federated learning method for anti-fraud model training includes: Acquire a region type library, wherein the region type library includes at least one preset region name and a preset fraud type corresponding to each preset region name; Get the preset fraud type corresponding to the preset region name that is the same as the name of the region dimension service system; According to the preset fraud type, the global model to be trained that the regional dimension service system needs to train is obtained as the acceptance information.

[0031] In this embodiment, the local training model for building the global model to be trained for each fraud type according to the acceptance information of each regional dimension service system includes: Perform the following operations for each fraud type's global model to be trained: Obtain the number of regional dimension service systems that need to train the global model to be trained for the fraud type; A local training model is established for each regional dimension service system that needs to train the global model to be trained for the fraud type, and one local training model corresponds to one regional dimension service system.

[0032] Optionally, each regional dimension service system sets a client dimension system associated with each local training model, including: Obtain a client dimension type library, wherein the client dimension type library includes at least one preset client name and a preset client fraud type corresponding to each preset client name, for example, a preset client name is H, and the corresponding preset client fraud type is financial credential fraud; Obtain the preset client fraud type corresponding to the preset client name that is the same as the name of the client dimension system, obtain the client dimension system H that has the same name as the preset client name H, and the preset client fraud type corresponding to it is financial certificate fraud; According to the fraud type of the preset client, the local training model that needs to be trained for each preset client is obtained and the preset client is associated with the local training model that needs to be trained.

[0033] In this embodiment, clustering is performed according to each local training model associated with each client dimension system in the same regional dimension service system to obtain a cluster cluster, including: Obtaining a local training model required for training each client dimension system. It is understandable that the local training model required for training the client dimension system can be set in a preset manner; Multiple client dimension systems with the same local training models that need to be trained are grouped into a cluster. For example, there are three client dimension systems E, F, and J. The local training models that need to be trained for E and F are both global models to be trained for insurance fraud. In other words, the data owned by E and F are suitable for training the global model to be trained for insurance fraud. At this time, E and F are a cluster, and J is a cluster.

[0034] In this embodiment, when there are multiple local training models associated with a client dimension system, clustering the local training models associated with the client dimension systems in the same regional dimension service system to obtain cluster clusters further includes: Setting a training priority for each associated local training model to be trained by a client dimension system having multiple associated local training models; The local training model with the highest training priority is obtained as a basis for forming clusters.

[0035] For example, in addition to the insurance fraud global model to be trained, the above-mentioned E also has a swindling global model to be trained. At this time, the priority can be set for the two models. The priority can be set manually or according to the amount of training or training data.

[0036] See also Figure 3 The present application also provides a cross-domain multi-dimensional federated learning system for anti-fraud model training, wherein the cross-domain multi-dimensional federated learning system for anti-fraud model training includes a central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems, wherein the central service system, multiple regional dimension service systems, multiple coordinators, and multiple client dimension systems cooperate to implement the cross-domain multi-dimensional federated learning method for anti-fraud model training as described above.

[0037] It should be noted that the above explanations of the method embodiment are also applicable to the system of this embodiment and will not be repeated here.

[0038] Although the present invention has been described in detail above with general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements may be made thereto based on the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.

Claims

1. A cross-domain multi-dimensional federated learning method for anti-fraud model training, characterized in that: The cross-domain multi-dimensional federated learning method for anti-fraud model training includes: The central service system establishes multiple global models to be trained for different fraud types; The central service system obtains the undertaking information of each regional dimension service system; According to the acceptance information of each regional dimension service system, a local training model of the global model to be trained for each fraud type is built, wherein a global model to be trained for one fraud type can build at least two or more local training models; The central service system sets up the regional dimension service system associated with each local training model; Each regional dimension service system sets up the client dimension system associated with each local training model; Clustering is performed according to the local training models associated with the client dimension systems in the same regional dimension service system, so as to obtain cluster clusters. The client dimension systems in each cluster cluster coordinate through a coordinator when training their respective local training models. The clusters corresponding to each coordinator are managed through the coordinator, so that each client dimension system in the cluster can train its associated local training model.

2. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 1, characterized in that: The cross-domain multi-dimensional federated learning method for anti-fraud model training further includes: Each client dimension system in each cluster trains its associated local training model and obtains the client dimension gradient information corresponding to each client dimension system; The coordinator of each cluster obtains the dimension gradient information of each client in the cluster; The coordinator of each cluster generates the cluster gradient information corresponding to the cluster according to the dimension gradient information of each client under the cluster; The regional dimension service system generates regional gradient information based on the gradient information of each cluster cluster obtained; The central service system obtains the gradient information of each region and obtains the gradient information of the global model through the gradient information of each region.

3. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 2, characterized in that: The types of fraud include contract fraud, bill fraud, loan fraud, credit card fraud, letter of credit fraud, financial certificate fraud, fund-raising fraud, securities fraud, insurance fraud, and swindling; The global models to be trained include contract fraud global models to be trained, bill fraud global models to be trained, loan fraud global models to be trained, credit card fraud global models to be trained, letter of credit fraud global models to be trained, financial voucher fraud global models to be trained, fundraising fraud global models to be trained, securities fraud global models to be trained, insurance fraud global models to be trained, and swindling global models to be trained.

4. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 3, characterized in that: The cross-domain multi-dimensional federated learning method for anti-fraud model training includes: Acquire a region type library, wherein the region type library includes at least one preset region name and a preset fraud type corresponding to each preset region name; Get the preset fraud type corresponding to the preset region name that is the same as the name of the region dimension service system; According to the preset fraud type, the global model to be trained that the regional dimension service system needs to train is obtained as the acceptance information.

5. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 4, characterized in that: The local training model for building the global model to be trained for each fraud type according to the acceptance information of each regional dimension service system includes: Perform the following operations for each fraud type's global model to be trained: Obtain the number of regional dimension service systems that need to train the global model to be trained for the fraud type; A local training model is established for each regional dimension service system that needs to train the global model to be trained for the fraud type, and one local training model corresponds to one regional dimension service system.

6. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 5, characterized in that: The client dimension system associated with each local training model set by each regional dimension service system includes: Obtain a client dimension type library, wherein the client dimension type library includes at least one preset client name and a preset client fraud type corresponding to each preset client name; Get the preset client fraud type corresponding to the preset client name that is the same as the name of the client dimension system; According to the fraud type of the preset client, the local training model that needs to be trained for each preset client is obtained and the preset client is associated with the local training model that needs to be trained.

7. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 6, characterized in that: The clustering according to each local training model associated with each client dimension system in the same regional dimension service system to obtain a cluster cluster includes: Obtain the local training model that needs to be trained for each client dimension system; Multiple client-dimensional systems with the same local training models that need to be trained are grouped into a cluster.

8. The cross-domain multi-dimensional federated learning method for anti-fraud model training according to claim 7, characterized in that: When there are multiple local training models associated with a client dimension system, clustering the local training models associated with the client dimension systems in the same regional dimension service system to obtain cluster clusters further includes: Setting a training priority for each associated local training model to be trained by a client dimension system having multiple associated local training models; The local training model with the highest training priority is obtained as a basis for forming clusters.

9. A cross-domain multi-dimensional federated learning system for anti-fraud model training, characterized in that: The cross-domain multi-dimensional federated learning system for anti-fraud model training includes a central service system, multiple regional dimension service systems, multiple coordinators and multiple client dimension systems, wherein the central service system, multiple regional dimension service systems, multiple coordinators and multiple client dimension systems cooperate to implement the cross-domain multi-dimensional federated learning method for anti-fraud model training as described in any one of claims 1 to 8.

Images