CN114580661A

CN114580661A - Data processing method and device based on federal learning and computer equipment

Info

Publication number: CN114580661A
Application number: CN202210181539.6A
Authority: CN
Inventors: 郭清宇; 蓝利君; 李超; 周义朋
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-02-25
Filing date: 2022-02-25
Publication date: 2022-06-03

Abstract

The application relates to a data processing method, a data processing device, a computer device, a storage medium and a computer program product based on federal learning. The method can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic, auxiliary driving and the like, for example, the method is applied to each parameter server in a federal learning architecture and comprises the following steps: receiving gradient data obtained by training participants in the corresponding participant cluster by using local data; summarizing the gradient data of the participant cluster to obtain gradient summarized data; acquiring a topological structure of a parameter server; exchanging gradient summary data with adjacent parameter servers based on the parameter server topology; and updating parameters of the combined model according to gradient summarized data of all parameter servers obtained by exchanging. The method improves the stability of the system, and provides a foundation for improving the data exchange efficiency between the parameter servers because the topological structure of the parameter servers is constructed based on the response time between the parameter servers and is not fixed and unchanged.

Description

Data processing method and device based on federal learning and computer equipment

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a data processing method and apparatus, a computer device, a storage medium, and a computer program product based on federal learning.

Background

Federal Learning (Federal Learning) is an emerging artificial intelligence basic technology, is a distributed machine Learning training mode based on privacy protection, and aims to train a high-quality prediction model while protecting the data privacy of a client when training data are dispersed on a large number of unreliable clients with high network delay.

The framework of federated learning is shown in fig. 1, and is a federated learning structure with a single-parameter server and multiple participants. As the name implies, the parameter server only has a single node and performs model parameter exchange with a plurality of participants respectively. The specific process is shown in FIG. 1. The single-parameter server and the multi-party structure are adopted, the parameter server is easy to become a performance bottleneck in the whole training process, meanwhile, the robustness of the whole training process is low due to the fact that the parameter server is a single point, and if the parameter server breaks down or has network problems, the whole federal learning training process has problems.

Disclosure of Invention

In view of the above, it is necessary to provide a data processing method, an apparatus, a computer device, a computer readable storage medium, and a computer program product based on federal learning, which can improve stability.

In a first aspect, the present application provides a data processing method based on federal learning. The method comprises the following steps:

receiving gradient data obtained by training participants in the corresponding participant cluster by using local data;

summarizing the gradient data of the participants of the participant cluster to obtain gradient summarized data;

acquiring a parameter server topological structure, wherein the parameter server topological structure is constructed based on response time among parameter servers when a federal learning model is initialized;

exchanging the gradient summary data with an adjoining parameter server based on the parameter server topology;

and updating parameters of the combined model according to gradient summarized data of all parameter servers obtained by exchanging.

In a second aspect, the application further provides a data processing device based on federal learning. Each parameter server applied to the federal learning architecture, the device comprises:

the receiving module is used for receiving gradient data obtained by training the participants in the corresponding participant cluster by using local data;

the summarizing module is used for summarizing the gradient data of the participants in the participant cluster to obtain gradient summarized data;

the structure acquisition module is used for acquiring a parameter server topological structure, wherein the parameter server topological structure is constructed based on the response time among the parameter servers when a federal learning model is initialized;

an exchange module for exchanging the gradient summary data with an adjacent parameter server based on the parameter server topology;

and the updating module is used for updating the parameters of the combined model according to the gradient summarized data of all the parameter servers obtained by exchanging.

In a third aspect, the present application also provides a computer device. The computer device comprises a memory storing a computer program and a processor implementing the following steps when executing the computer program:

In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of:

In a fifth aspect, the present application further provides a computer program product. The computer program product comprising a computer program which when executed by a processor performs the steps of:

According to the data processing method, device, computer equipment, storage medium and computer program product based on the federal learning, each parameter server corresponds to a participant cluster and receives gradient data of training of the corresponding parameter cluster, the data processing method is a multi-participant and multi-parameter server framework, each parameter server and the corresponding participant cluster perform data exchange, and even if a parameter server fails, the federal learning training cannot be influenced, and the system stability is improved. Meanwhile, each parameter server constructs a parameter server topological structure according to the response time among the parameter servers, and after each parameter server obtains the gradient summarized data of the corresponding participant cluster, the gradient summarized data are exchanged based on the parameter server topological structure. The topological structure of the parameter server is constructed based on the response time between the parameter servers and is not fixed and unchangeable, so that a foundation is provided for improving the data exchange efficiency between the parameter servers.

Drawings

FIG. 1 is a diagram of a single parameter server, multi-party federated learning architecture in one embodiment;

FIG. 2 is a diagram of a federated learning architecture for a multi-parameter server and multi-parameter party, according to an embodiment;

FIG. 3 is a diagram illustrating a direct connection topology of a multi-parameter server in one embodiment;

FIG. 4 is a diagram illustrating a star topology of a multi-parameter server in one embodiment;

FIG. 5 is a diagram of an application environment of a federated learning-based data processing method in one embodiment;

FIG. 6 is a flow diagram illustrating a federated learning-based data processing method in one embodiment;

FIG. 7 is a block diagram illustrating a cluster of participants corresponding to parameter servers in an embodiment;

FIG. 8 is a diagram of a parameter server topology in one embodiment;

FIG. 9 is a diagram illustrating response times between parameter servers according to one embodiment;

FIG. 10 is a diagram illustrating a direct connection type parameter server topology according to an embodiment;

FIG. 11 is a diagram illustrating a topology of a direct connection type parameter server in another embodiment;

FIG. 12 is a relational diagram that illustrates the topology of the parameter servers in one embodiment;

FIG. 13 is a diagram illustrating a data processing method based on federated learning in another embodiment;

FIG. 14 is a comparison of experimental results of classification of data sets in MINIST images in one embodiment;

FIG. 15 shows comparative results of Synthetic experiments in one example;

FIG. 16 is a comparison of complexity and robustness of different topologies in one embodiment;

FIG. 17 is a block diagram of a data processing apparatus based on federated learning in one embodiment;

FIG. 18 is a diagram illustrating an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

With the research and progress of artificial intelligence technology, the artificial intelligence technology is developed and applied in a plurality of fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, unmanned aerial vehicles, robots, smart medical care, smart customer service, and the like.

The scheme provided by the embodiment of the application relates to the technologies of artificial intelligence, such as federal learning and the like. Federal Learning (Federal Learning) is an emerging artificial intelligence basic technology, is a distributed machine Learning training mode based on privacy protection, and aims to train a high-quality prediction model while protecting the data privacy of a client when training data are dispersed on a large number of unreliable clients with high network delay. On the premise of guaranteeing information safety during big data exchange, protecting terminal data and personal data privacy and guaranteeing legal compliance, efficient machine learning can be developed among multiple parties or multiple computing nodes.

In the application scenario of multi-party participation in federal training, data are distributed on different clients (participants) of the same organization, and a parameter server is used as an intermediate server for model parameter exchange. Aiming at the problem that the federal learning structure of a single-parameter server and multiple-party is unstable, the federal learning structure of the multiple-parameter server and the multiple-party is provided.

The multi-parameter server and the multi-party federal learning structure are provided with a plurality of parameter servers, the parameter servers respectively exchange parameters with the corresponding party groups, and the specific training process is shown in fig. 2:

step 21: and the participator client (such as a mobile phone of a user, a plurality of banks, namely a product user) downloads the shared prediction model from the corresponding parameter server.

Step 22: the participant client performs training iterations on the model using local data.

Step 23: and the participator client updates the gradient obtained after the model training, encrypts and uploads the gradient to the corresponding parameter server.

Step 24: all the parameter server sides wait for collecting gradient updates from all the participants, then perform gradient exchange and integration with the neighbor server, and then update the shared model of the own side.

And repeating the processes of the steps 21 to 24, and continuously downloading the model updated each time by the server side to the local client side by the participant for updating iteration until the model is converged and stopping updating.

In the federal learning structure of the multi-parameter servers and the multi-party, the structures among the multi-parameter servers are mainly direct connection type and star type structures, the direct connection type structure among the parameter servers is shown in figure 3, and the star type structure among the parameter servers is shown in figure 4.

The data processing method based on federal learning provided by the embodiment of the application can be applied to a multi-parameter server and a multi-party federal learning structure as shown in fig. 5. The participants communicate with a parameter server over a network. Each parameter service receives gradient data obtained by training participants of corresponding participant clusters by using local data; summarizing the gradient data of the participant cluster to obtain gradient summarized data; acquiring a parameter server topological structure constructed based on response time among parameter servers; exchanging gradient summarized data with an adjacent parameter server based on the topology structure of the parameter server; and updating parameters of the combined model according to gradient summarized data of all parameter servers obtained by exchanging. Participant 102 may be a user terminal, which includes but is not limited to a mobile phone, a computer, an intelligent voice interaction device, an intelligent appliance, a vehicle-mounted terminal, an aircraft, and the like. The parameter server 104 may be implemented as a stand-alone server or a server cluster composed of a plurality of servers. The embodiment of the invention can be applied to various scenes including but not limited to cloud technology, artificial intelligence, intelligent traffic, driving assistance and the like.

In one embodiment, as shown in fig. 6, a data processing method based on federal learning is provided, which is described by taking the method as an example of applying the method to one participating server in fig. 1, and includes the following steps 602 to 612, it is understood that when a multi-participating server performs federal learning, each participating server performs the following steps 602 to 610, respectively:

step 602, gradient data obtained by training the participants in the corresponding participant cluster by using the local data is received.

The federated learning is that participants train together to obtain a global model, each participant trains the model based on own local data, and then exchanges and summarizes through a parameter server to obtain the global model, and in the process, user data is always local and is not sent outside, so that the requirements of data safety and privacy protection are met. The parties to federal learning generally include data parties, algorithm parties, coordination parties, calculation parties, result parties, task promoters, and the like. Taking the example of federal learning applied to wind-controlled federated modeling between different banks, the participants include banks, users and credit agencies.

The participant cluster is a cluster of some participants, each participant cluster corresponds to one parameter server, and gradient data obtained by local data training is reported to the parameter server.

When the federal learning is initialized, the participant cluster corresponding to each parameter server can be initialized. In one embodiment, the parameter servers may be divided into regions, a participant cluster corresponding to each parameter server is determined, and participants in the same or similar regions are constructed as the participant cluster of the local parameter server. In one embodiment, each server may be used as a central node, and response times of the servers and the participants are clustered to form a plurality of participant clusters (corresponding to the number of servers). The whole clustering algorithm minimizes the average response time of the participants and the corresponding servers which fall into the same cluster. In one mode, the parameter server may send a gradient update request to the participants, and select the participants to construct the corresponding participant cluster according to the response time of the participants. In one mode, the parameter server may send a test request to the participants, and select the participants to construct the corresponding participant cluster according to the response time of the participants.

Through initialization, a cluster of participants corresponding to each parameter server is constructed, as shown in fig. 7. In federal learning, a parameter server issues a training program to a participant, and the participant calculates the descending gradient and loss by using local data to obtain gradient data of a round of training, encrypts and uploads the gradient data to the corresponding parameter server. Therefore, for the parameter server, it receives gradient data obtained by one iterative training of each participant in the corresponding participant cluster. As shown in fig. 7, the parameter server 1 receives gradient data obtained by each participant in the participant cluster 1 during the current iterative training.

It can be understood that, in order to ensure the confidentiality of data in the training process, the participant encrypts the gradient data and sends the gradient data to the corresponding parameter server.

And step 604, summarizing the gradient data of the participant cluster to obtain gradient summarized data.

Specifically, each parameter server summarizes the received gradient data of the participant cluster to obtain gradient summarized data. Wherein, the summary formula is as follows:

step 606, obtaining a parameter server topological structure, wherein the parameter server topological structure is constructed based on the response time between the parameter servers when initializing the federal learning model.

The data processing method based on the federal learning is suitable for the application scene of the federal learning of a multi-party multi-parameter server. And each parameter server receives gradient data of one participant cluster for summarizing. After the summary, gradient summary data needs to be exchanged between the parameter servers.

Usually, the parameter servers are in a direct connection topology as shown in fig. 3, or in a star topology as shown in fig. 4. The main defects of the topology structure of the direct connection type multi-parameter server are that the model convergence speed is low, each parameter server needs to interact with parameters for multiple times to enable the model to converge, and the system stability is poor due to single-point faults. The topological structure of the multi-parameter server with the star structure has high model convergence speed, but the poor system stability caused by single-point failure of the central node still exists.

In the embodiment, the parameter server topological structure is constructed based on the response time among the parameter servers, so that the topological structure among the parameter servers is not fixed and invariable, but is flexibly constructed based on the response time among the parameter servers, and the influence of the response delay length among the parameter servers on the gradient summarized data exchange efficiency among the parameter servers in the federal learning process can be truly reflected.

In the federal learning model initialization process, each parameter server can send a request to other parameter servers to obtain the response time of each other parameter server to the request. Each server obtains the response time with other servers, as shown in fig. 7 for C parameter servers.

In one approach, after the response times between all parameter servers are obtained, a direct connection type parameter server topology is constructed with the shortest exchange time as the target.

The direct connection type parameter server topology is shown in fig. 3, and all the parameter servers are connected in sequence. The direct connection type parameter server topological structure is constructed by utilizing response time among parameter servers, and is different from the existing fixed direct connection type parameter server topological structure. By shortening the data exchange time among the parameter servers, the data exchange efficiency among the parameter servers is improved, and the efficiency of federal learning is further improved.

In one approach, the parameter server topology is constructed with the goal of minimum exchange time and maximum constraints after the response times between all parameter servers are obtained.

The maximum degree constraint means that the edges between the parameter servers meet the maximum number requirement. Due to the fact that maximum degree constraint is set, each parameter server and the plurality of parameter servers have a connection relation, and the problem of single point failure easily existing in direct connection type connection is solved. The parameter server topological structure constructed by taking the shortest exchange time and the maximum degree constraint as the targets not only considers the efficiency of data exchange, but also considers the stability of data, and avoids the problem of poor system stability caused by single-point failure.

Step 608, gradient summary data is exchanged with the neighboring parameter servers based on the parameter server topology.

Specifically, with the parameter server topology, each parameter server exchanges gradient summary data with an adjacent parameter server having a connection relationship. Taking the topology structure of the parameter server shown in fig. 8 as an example, the parameter server a exchanges the gradient summary data to the parameter server B, the parameter server C, and the parameter server E, and correspondingly, through round exchange, the parameter server a can obtain the gradient summary data of the parameter server B, the parameter server E, and the parameter server C. Meanwhile, in the first round of exchange, the gradients of the parameter servers adjacent to other parameter servers, for example, the gradients of the parameter server a, the parameter server C and the parameter server E, which are obtained by the parameter server B, are summarized.

After the gradient summarized data of the adjacent parameter servers are obtained by exchanging of each parameter server, the gradient summarized data of the adjacent parameter servers obtained by exchanging are subjected to gradient summarization with the gradient summarized data of the parameter servers, and the gradient summarized data comprise the gradient summarized data of the parameter servers and the gradient summarized data of the adjacent parameter servers. For example, after the first round of exchange, the parameter server a performs gradient summarization on gradient summarized data of itself, the parameter server B, the parameter server E, and the parameter server C.

If the mutual gradient summary data cannot be obtained in the first round of exchange between the parameter servers which are not directly connected, the round of exchange is continued. For example, if there is no direct connection between the parameter server B and the parameter server D, the parameter server C obtains the gradient summary data of the parameter server D in the first round of exchange, and through the second round of exchange, the parameter server C exchanges the gradient summary data including the parameter server D to the parameter server B. It can be understood that through multiple rounds of exchange and gradient summarization, gradient summary data of all the parameter servers are obtained among the parameter servers.

It can be understood that, in order to ensure the confidentiality of data in the training process, gradient summarized data is encrypted between the parameter servers and then exchanged with the adjacent parameter servers, that is, gradient summarized data after homomorphic encryption is exchanged between the parameter servers.

And step 610, updating parameters of the combined model according to the gradient summarized data of all the parameter servers obtained by exchanging.

Specifically, through multiple rounds of exchange and gradient summarization, gradient summary data of all parameter servers are obtained among the parameter servers, parameters of the combined model are updated, and one iterative training of federal learning is completed.

And the parameter server continuously transmits the updated combined model to the corresponding parameter server cluster according to the updated combined model, each participant continues training by using local data, and each parameter server continuously updates the combined model by repeating the steps 602 to 610 until the training is finished.

According to the data processing method based on the federal learning, each parameter server corresponds to a participant cluster, gradient data corresponding to training of the parameter cluster are received, the data processing method is a multi-participant and multi-parameter server framework, each parameter server and the corresponding participant cluster perform data exchange, and even if a parameter server fails, the federal learning training cannot be influenced, and the system stability is improved. Meanwhile, each parameter server constructs a parameter server topological structure according to the response time among the parameter servers, and after each parameter server obtains the gradient summarized data of the corresponding participant cluster, the gradient summarized data are exchanged based on the parameter server topological structure. The topological structure of the parameter server is constructed based on the response time between the parameter servers and is not fixed and unchangeable, so that a foundation is provided for improving the data exchange efficiency between the parameter servers.

In another embodiment, receiving gradient data trained by participants in a corresponding cluster of participants using local data comprises: and sending a gradient updating request to the corresponding participant cluster, and receiving gradient data obtained by training the first N participants who respond to the gradient updating request most quickly in the participant cluster by using local data.

Specifically, each participant of federal learning has different response time due to different regions, network environments and the like, and the participant with slow response time in the same federal seriously affects the whole training time. Aiming at the problem, firstly clustering the participants based on the response time of the participants to the parameter server to construct the participant cluster of the parameter server. Specifically, each parameter party with a short response time to the parameter server is clustered as a participant of the parameter server, so that all participants in federal learning are clustered into different clusters, and the participants falling into the same cluster have a short response time.

The factors that generally affect the response speed of the participants are the distance between the participants and the parameter server and the processing capacity of the participants themselves, and if the participants and the parameter server are in the same area, the requests of the parameter server can be responded to faster, so that a cluster of participants selected by one parameter server based on the response time is generally composed of participants with closer geographical positions. Each parameter server is used as an intermediate server for model parameter exchange, receives gradient data obtained by training the participants based on local data, and does not affect the training result even if the participants are selected by a plurality of parameter servers, so that the cluster division of the participants can be performed only by response time in the embodiment.

Further, in each training iteration of the federal learning, the server only selects the first N participants in the same cluster to perform the joint model iteration, so that the tiredness of the participants with slow response to the participants with fast response is greatly reduced.

Specifically, the construction of the participant cluster may be performed during an initialization process of the federated learning modeling, where the initialization includes the initialization of the participant cluster. And taking each parameter server as a center, and respectively sending a request to all participants by each parameter server to acquire the response time of each participant for responding to the request. The response time is taken as a cluster, and the aim of the cluster is to minimize the average response time of the participants and the corresponding servers which fall into the same cluster. Specifically, according to the response time, a plurality of fastest and nearer participant servers are selected to construct a corresponding participant cluster. Thereby forming a cluster of participants for each server.

Based on the participant clusters, in each iteration training, each parameter server sends a gradient updating request to the corresponding participant cluster, and receives gradient data obtained by the first N participants who respond to the gradient updating request most quickly in the participant clusters through local data training.

Specifically, each parameter server sends a gradient update request to each participant, and receives only gradient data sent by the first N participants with the fastest responses. For example, N is 50, the number of the parameter servers is 5, 1000 participants are distributed in each region, each parameter server corresponds to one participant cluster, and there are 5 participant clusters in total. Each parameter server selects only the gradient data of the top 50 first responding participants.

It is foreseeable that, because the parameter server selects the gradient data of the first N participants in the participant cluster with the fastest response according to the response time, the participant with the fastest response will not be tired by the participant with the slow response, thereby reducing the total waiting time of the parameter server and improving the efficiency of the joint modeling.

In another embodiment, a method for constructing a topology of parameter servers based on response times between the parameter servers includes: acquiring response time between parameter servers; and constructing a direct-connection type parameter server topological structure by taking the shortest exchange time as a target based on the response time.

The direct connection type parameter server topology structure is shown in fig. 4, and all participating servers are connected in sequence. Different from the traditional direct connection type parameter server topology structure, the direct connection type parameter server topology structure in the embodiment considers the exchange time based on the response time between servers. The data exchange efficiency is improved by the constructed direct connection type parameter server topological structure by enabling the exchange time to be shortest.

Specifically, based on the response time, with the shortest exchange time as the target, a parameter server topology is constructed, including: respectively taking each parameter server as a starting point, iterating the step of selecting the parameter server which has the shortest response time with the starting point and is in an unconnected state as a next starting point according to the response time among the parameter servers until all the parameter servers are connected to obtain a plurality of candidate parameter server topological structures; and taking the candidate parameter server with the shortest total response time as the final parameter server topological structure.

Wherein the parameter server topology can be constructed at initialization of federated learning joint modeling. Each parameter server sends a request to other parameter servers to obtain the response time of each parameter server for responding to the request of the parameter server. In one embodiment, the response time between parameter servers is shown in FIG. 9.

On the basis, the following steps are executed:

s10, each parameter server is used as a starting point, and the parameter server having the shortest response time to the starting point is selected as the next starting point. Taking fig. 9 as an example, the selection server a, the parameter server B, the parameter server C, the parameter server D, and the parameter server E are selected as starting points, and the parameter server having the shortest response time to the starting point is selected as the next starting point. Taking server a as the starting point, parameter server B with the shortest response time to parameter server a is selected as the next node.

And S11, repeating the step S10 until all the parameter servers are connected, and obtaining a plurality of candidate parameter server topological structures.

Specifically, according to the response time between the parameter servers, the parameter server which has the shortest response time with the starting point and is in an unconnected state is selected as the next starting point. For example, taking the parameter server a as a starting point, the second node is the parameter server B, the third node is the parameter server C, the fourth node is the parameter server D, and when the fifth node is determined, the response time of the parameter server B being the parameter server E is the same, but the parameter server B is in a connected state, the parameter server E is selected as the fifth node. The resulting parameter server topology is constructed starting from parameter server a as shown in fig. 10. The topology of the parameter server constructed with the parameter server E as the starting point is shown in fig. 11.

And respectively taking each parameter server as a starting point, so that a plurality of candidate parameter server topological structures can be constructed and obtained.

And S12, taking the candidate parameter server with the shortest total response time as the final parameter server topological structure.

Wherein the total response time is the sum of the response times between the parameter servers of the respective connections. Taking the parameter server topology shown in fig. 10 as an example, the total response time is 10, and taking the parameter server topology shown in fig. 11 as an example, the total response time is 12. And if the total response time is shortest, the time for data exchange between the parameter servers is also shortest based on the topological structure of the parameter servers. By comparison, if the total response time of the direct connection type parameter server topology starting from a is shortest, the structure shown in fig. 10 is used as the final parameter server topology. The total response time is shortest, so that the time spent on exchanging the gradient data by using the direct-connection type parameter server topological structure is shortest, and the data exchange efficiency can be improved.

The direct connection type parameter server needs to interact for multiple times to enable the model to be converged, and meanwhile, the system stability is poor due to single-point faults. On the basis, another construction mode of the parameter server topology is provided.

In another embodiment, the method for constructing the topology of the parameter server based on the response time between the parameter servers comprises the following steps: acquiring response time between parameter servers; and constructing a parameter server topological structure by taking the shortest exchange time and the maximum degree constraint as targets based on the response time.

The maximum degree constraint refers to the maximum number of edges of each node in the parameter server topology structure. By setting the maximum number to be at least 2, namely, the number of edges of each node to be at least 2, the interaction relationship among the parameter servers can be various, so that single-point failure is avoided, and efficiency is not influenced by too much. And the exchange efficiency and the stability of the system are integrated through the shortest exchange time and the maximum degree constraint.

Specifically, based on response time, with the shortest exchange time and the maximum degree constraint as targets, a parameter server topology is constructed, which includes:

s121: and selecting a parameter server which has the shortest response time with the starting point and is in an unconnected state as a next starting point according to the response time between the parameter servers by taking any parameter server as the starting point.

As shown in fig. 12, based on the response times between all parameter servers, a parameter server a having the shortest response time to the parameter server B is selected as the next starting point from an arbitrary parameter server, for example, from the parameter server B. And if the nodes with the same response time are encountered, randomly selecting one of the nodes as a next starting point.

And S122, iterating the step S121 until all the parameter servers are connected to obtain a parameter server communication structure.

Taking fig. 12 as an example, taking the parameter server B as a starting point, the parameter server a with the shortest response time with the parameter server B is selected as a second node, the parameter server C is selected as a third node, the parameter server D is selected as a fourth node, and the parameter server E is selected as a fifth node, and the obtained parameter server communication structure is as shown in fig. 12.

And S123, sequentially adding the response relations which are not added into the parameter server communication structure based on maximum constraint according to the sequence of the response time of the parameter server from small to large, and obtaining the final parameter server topological structure.

Specifically, the response relationship refers to a connection relationship between servers for indicating response time, and may be an edge in a topology structure. The response relations (edges) which are not added into the parameter server communication structure can be sequentially added into the parameter server communication structure according to the sequence of the response time of the parameter server from small to large. In the iteration process, if the degree of any node on two sides of the edge to be added is equal to the maximum degree number, the edge is abandoned to be added. And finally forming a network topology structure of the server nodes until the edges which can be added are all added into the parameter server communication structure. As shown in fig. 12, the maximum degree number is set to 3, and after sorting according to the response time, edges (B- > C, B- > E, a-E) are sequentially added to the graph, so as to form a final parameter server topology.

In the embodiment, the response time between the servers is used for constructing the parameter server topological structure, and the influence of the problems of node response delay, hardware region and the like on federal learning is considered. The parameter servers are not in a sequential single-point connection relationship and have multi-directionality, and even if a single-point fault occurs in one node, data can still be exchanged with other parameter servers through other connection relationships. The method takes into account transfer efficiency and system stability.

The data processing method based on the federal learning can be applied to joint modeling of any plurality of institutions on the premise of not exposing respective private data, such as wind control joint modeling among different banks or credit institutions, case diagnosis joint modeling among different medical heterogeneous institutions, commodity recommendation joint modeling among different e-commerce platforms and the like. In the federated model training process based on federated learning, a decentralization algorithm of self-adaptive adjustment can be realized, the overall federated learning efficiency and the system robustness are improved on the premise of not losing the federated modeling effect, and the product use experience is improved.

Specifically, in an embodiment, the data processing method based on federal learning, as shown in fig. 13, includes the following steps:

s131: federal learning is initialized. In this step, parameter server topology initialization, participant cluster initialization, and joint model parameter initialization.

Specifically, the parameter server topology initialization includes:

1) and recording the response time among all the servers.

2) And iterating the steps of selecting the parameter server which has the shortest response time with the starting point and is in an unconnected state as the next starting point according to the response time between the parameter servers by taking any parameter server as the starting point until all the parameter servers are connected, so as to obtain the parameter server communication structure.

3) And sequentially adding the response relations which are not added into the parameter server communicating structure according to the sequence of the response time of the parameter servers from small to large. In the iteration process, if the degree of any one node on two sides of the edge (response relation) to be added is equal to the maximum degree number, the edge is abandoned. And finally forming a network topology structure of the server nodes until the edges which can be added are all added into the parameter server communication structure.

The initialization of the participant cluster specifically comprises the following steps: each parameter server sends a request to each participant, and each server is taken as a central node, and the response time of the server and the participant is clustered to form a plurality of participant clusters (corresponding to the number of the servers). The whole clustering algorithm enables the average response time of the participators and the corresponding servers falling into the same cluster to be shortest, and the participator cluster corresponding to each parameter server is constructed.

And initializing the federal model parameters to ensure that the model parameters on each server are initialized to be the same set.

S132, downloading the combined model from the server side by the client side of the participant (such as the mobile phone of the user, a plurality of banks, namely a product user), wherein the combined model is initialized by the server cluster in the training starting stage (the combined models of all the servers are the same).

And S133, the participant client performs training iteration on the model by using the local data.

And S134, with each parameter server as a central node, clustering the first n participants with the fastest response by each participant, and uploading the summary gradient to the corresponding server. The size of n is set by a server, and the maximum number of participants participating in each iteration is controlled by the Federal model through the size of n.

S135, each server in the server cluster waits to collect the gradient updates reported from the corresponding participant. And after the server receives the gradient updates of the first n participants, aggregating all the collected gradients, then performing gradient exchange with an adjacent server, and after the exchange is finished, updating the combined model by each server side. Then the next iteration is entered (each server invites all corresponding participants to cluster again for joint model download and local model update).

In each round of training iteration, the server requests gradient report of n participants in the corresponding participant cluster, and after receiving the request, the participant cluster returns the gradient of the first n participants with the fastest response. Because the participants receiving the requests have similar response time and only return the first n fastest participants, the participants with fast response cannot be tired by the participants with slow response, thereby reducing the total waiting time of the server and improving the joint modeling efficiency.

And each server waits for the gradient reported by the corresponding participant cluster, and then performs gradient summarization, wherein a specific formula is shown as a formula 1.

After gradient summarization is carried out on each server, own gradient is exchanged with the node of the leading population according to the topological structure, and after the exchange, gradient summarization and parameter updating are carried out again to complete the updating iteration of the model.

And repeating the steps from S132 to S135, and continuously downloading the model updated each time by the server side to the local client side of the participant until the model converges, and stopping updating.

The data processing method based on the federal learning can be a universal decentralization method of self-adaptive adjustment, can effectively improve the training efficiency of the model within a limited training time budget under the scenes of federal learning or distributed training and the like, and meanwhile guarantees the accuracy rate of the model and the stability of network topology. From the application angle, the scheme is not only limited to be applied to the financial anti-fraud combined modeling scene under the participation of multiple banks or mutual-fund enterprises, but also can be applied to various scenes of multi-client combined modeling under the privacy protection constraint. From the model perspective, the model architecture used by the server in the combined modeling can be changed flexibly according to the actual application scene.

The self-adaptive adjustment decentralized federal training method can generate an optimal topological structure of the parameter server and a mode of clustering and selecting participants through response time in a self-adaptive mode under a limited model training budget, improves the model training efficiency and the stability of a federal learning system, and achieves the maximum model correctness in limited training time. Fig. 14 shows experimental comparison results for the MINIST image classification dataset under 1000 participants, 10 servers, 3 different topologies, and fig. 15 shows experimental comparison results for Synthetic under 1000 participants, 10 servers, 3 different topologies. The time delay result shows that the self-adaptive topological structure has excellent convergence speed compared with the star-type and direct-connection-type models, and the model effect (accuracy) is not attenuated. Fig. 16 shows the complexity and robustness of different topologies, and the result also proves that the method is advantageous in enhancing the stability and robustness of the system. The effectiveness of the scheme is further verified.

It should be understood that, although the steps in the flowcharts related to the embodiments as described above are sequentially displayed as indicated by arrows, the steps are not necessarily performed sequentially as indicated by the arrows. The steps are not limited to being performed in the exact order illustrated and, unless explicitly stated herein, may be performed in other orders. Moreover, at least a part of the steps in the flowcharts related to the embodiments described above may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the execution order of the steps or stages is not necessarily sequential, but may be rotated or alternated with other steps or at least a part of the steps or stages in other steps.

Based on the same inventive concept, the embodiment of the application also provides a data processing device based on the federal learning, which is used for realizing the related data processing method based on the federal learning. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the method, so that the specific limitations in one or more federate learning-based data processing device embodiments provided below may refer to the limitations on the federate learning-based data processing method in the foregoing, and are not described herein again.

In one embodiment, as shown in fig. 17, there is provided a data processing apparatus based on federal learning, applied to parameter servers in a federal learning architecture, including:

a receiving module 1702, configured to receive gradient data obtained by training participants in the corresponding participant cluster by using the local data.

A summarizing module 1704, configured to summarize the gradient data of the participants in the participant cluster to obtain gradient summarized data.

A structure obtaining module 1706, configured to obtain a parameter server topology structure, where the parameter server topology structure is constructed based on response time between parameter servers when the federal learning model is initialized.

An exchanging module 1708, configured to exchange the gradient summary data with an adjacent parameter server based on the topology of the parameter server.

An updating module 1710, configured to update the parameters of the combination model according to the gradient summarized data of all the parameter servers obtained through the exchanging.

In another embodiment, the receiving module is configured to send a gradient update request to a corresponding participant cluster, and receive gradient data obtained by training, using local data, the first N participants in the participant cluster that respond to the gradient update request most quickly.

In another embodiment, a structure acquisition module includes:

and the time acquisition module is used for acquiring the response time between the parameter servers.

And the building module is used for building a direct-connection type parameter server topological structure by taking the shortest exchange time as a target based on the response time.

In another embodiment, the construction module is configured to iterate the step of selecting, according to response times among the parameter servers, a parameter server that has the shortest response time with the start point and is in an unconnected state as a next start point until all the parameter servers are connected, with each parameter server as a start point, and obtain a plurality of candidate parameter server topology structures; and taking the candidate parameter server with the shortest total response time as a final parameter server topological structure.

In another embodiment, the method comprises constructing a parameter server topology based on the response times, with a minimum exchange time and a maximum degree constraint as targets, the number of maximum degrees being at least 2.

In another embodiment, the construction module is configured to iterate the step of selecting, according to response time between the parameter servers, a parameter server that has the shortest response time with the starting point and is in an unconnected state as a next starting point, with an arbitrary parameter server as the starting point, until all the parameter servers are connected, so as to obtain a parameter server connection structure; and sequentially adding the response relations which are not added into the parameter server communication structure based on maximum degree constraint according to the sequence of the response time of the parameter server from small to large, so as to obtain the final parameter server topological structure.

The various modules in the above-described federally learned data processing apparatus can be implemented in whole or in part by software, hardware, and combinations thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in fig. 18. The computer device includes a processor, a memory, an Input/Output interface (I/O for short), and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store federally learned training data. The input/output interface of the computer device is used for exchanging information between the processor and an external device. The communication interface of the computer device is used for connecting and communicating with an external terminal through a network. The computer program is executed by a processor to implement a federated learning-based data processing method.

Those skilled in the art will appreciate that the architecture shown in fig. 18 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the data processing method based on federal learning of the above embodiments when executing the computer program.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the federated learning-based data processing method of the above-described embodiments.

In one embodiment, a computer program product is provided, which comprises a computer program that, when executed by a processor, implements the steps of the federated learning-based data processing method of the above-described embodiments.

It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data need to comply with the relevant laws and regulations and standards of the relevant country and region.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high-density embedded nonvolatile Memory, resistive Random Access Memory (ReRAM), Magnetic Random Access Memory (MRAM), Ferroelectric Random Access Memory (FRAM), Phase Change Memory (PCM), graphene Memory, and the like. Volatile Memory can include Random Access Memory (RAM), external cache Memory, and the like. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), for example. The databases referred to in various embodiments provided herein may include at least one of relational and non-relational databases. The non-relational database may include, but is not limited to, a block chain based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing based data processing logic devices, etc., without limitation.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A data processing method based on federal learning is characterized in that the method is applied to each parameter server in a federal learning framework, and the method comprises the following steps:

summarizing the gradient data of the participant cluster to obtain gradient summarized data;

2. The method of claim 1, wherein receiving gradient data trained by participants in a corresponding cluster of participants using local data comprises:

sending a gradient updating request to a corresponding participant cluster, and receiving gradient data obtained by training the first N participants which respond to the gradient updating request most quickly in the participant cluster by using local data.

3. The method of claim 1, wherein constructing a topology of parameter servers based on response times between the parameter servers comprises:

acquiring response time between parameter servers;

and constructing a direct-connection type parameter server topological structure by taking the shortest exchange time as a target based on the response time.

4. The method of claim 3, wherein the constructing a parameter server topology based on the response time and targeting a shortest exchange time comprises:

respectively taking each parameter server as a starting point, iterating the step of selecting the parameter server which has the shortest response time with the starting point and is in an unconnected state as a next starting point according to the response time among the parameter servers until all the parameter servers are connected to obtain a plurality of candidate parameter server topological structures;

and taking the candidate parameter server with the shortest total response time as a final parameter server topological structure.

5. The method of claim 1, wherein the means for constructing a topology of parameter servers based on response times between the parameter servers comprises:

acquiring response time between parameter servers;

and constructing a parameter server topological structure by taking the shortest exchange time and the maximum degree constraint as targets based on the response time, wherein the number of the maximum degrees is at least 2.

6. The method of claim 5, wherein constructing a parameter server topology based on the response times with a minimum exchange time and a maximum degree constraint as targets comprises:

taking any parameter server as a starting point, iterating the step of selecting the parameter server which has the shortest response time with the starting point and is in an unconnected state as a next starting point according to the response time between the parameter servers until all the parameter servers are connected to obtain a parameter server communication structure;

and sequentially adding the response relations which are not added into the parameter server communication structure based on maximum degree constraint according to the sequence of the response time of the parameter server from small to large, so as to obtain the final parameter server topological structure.

7. A data processing apparatus based on federal learning, which is applied to each parameter server in the federal learning architecture, and the apparatus comprises:

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the method of any of claims 1 to 6.

9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that the computer program realizes the steps of the method of any one of claims 1 to 6 when executed by a processor.