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

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, device and computer equipment based on federated learning

technical field

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

Background technique

Federated Learning (Federated Learning) is an emerging artificial intelligence basic technology, which is a distributed machine learning training method based on privacy protection. On the end, it can train a high-quality prediction model while protecting the privacy of client data.

A federated learning architecture is shown in Figure 1, which is a single-parameter server, multi-participant federated learning structure. As the name suggests, the parameter server has only a single node, which exchanges model parameters with multiple participants respectively. The specific process is shown in Figure 1. Single-parameter server and multi-participant structure, the parameter server can easily become a performance bottleneck in the entire training process. At the same time, because the parameter server is a single point, the robustness of the entire training process will be low. If the parameter server fails or If there is a network problem, then the entire federated learning training process will have problems.

SUMMARY OF THE INVENTION

Based on this, it is necessary to provide a federated learning-based data processing method, apparatus, computer equipment, computer-readable storage medium and computer program product that can improve the stability for the above technical problems.

In a first aspect, the present application provides a data processing method based on federated learning. The method includes:

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

Aggregating the gradient data of the participants of the participant cluster to obtain the gradient summary data;

obtaining a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized;

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

The parameters of the joint model are updated according to the gradient summary data of all parameter servers obtained by the exchange.

In a second aspect, the present application also provides a data processing apparatus based on federated learning. Applied to each parameter server in the federated learning architecture, the device includes:

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

an aggregation module, configured to aggregate the gradient data of the participants of the participant cluster to obtain the gradient aggregated data;

a structure acquisition module, configured to acquire a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized;

an exchange module, configured to exchange the gradient summary data with an adjacent parameter server based on the parameter server topology;

The update module is used to summarize the data according to the gradients of all parameter servers obtained by the exchange, and update the parameters of the joint model.

In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

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

Aggregating the gradient data of the participants of the participant cluster to obtain the gradient summary data;

obtaining a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized;

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

The parameters of the joint model are updated according to the gradient summary data of all parameter servers obtained by the exchange.

In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium has a computer program stored thereon, and when the computer program is executed by the processor, the following steps are implemented:

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

Aggregating the gradient data of the participants of the participant cluster to obtain the gradient summary data;

obtaining a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized;

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

The parameters of the joint model are updated according to the gradient summary data of all parameter servers obtained by the exchange.

In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

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

Aggregating the gradient data of the participants of the participant cluster to obtain the gradient summary data;

obtaining a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized;

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

The parameters of the joint model are updated according to the gradient summary data of all parameter servers obtained by the exchange.

The above-mentioned data processing method, device, computer equipment, storage medium and computer program product based on federated learning, each parameter server corresponds to a participant cluster and receives the gradient data trained by the corresponding parameter cluster, which is a multi-participant, multi-parameter The architecture of the server, each parameter server exchanges data with the corresponding participant cluster, even if a parameter server fails, it will not affect the training of federated learning and improve the system stability. At the same time, each parameter server constructs a parameter server topology structure according to the response time between the parameter servers. After each parameter server obtains the gradient summary data of the corresponding participant cluster, it exchanges the gradient summary data based on the parameter server topology structure. Since the parameter server topology is constructed based on the response time between parameter servers, it is not fixed, which provides a basis for improving the data exchange efficiency between parameter servers.

Description of drawings

1 is a schematic diagram of a federated learning architecture of a single-parameter server and multiple participants in one embodiment;

2 is a schematic diagram of a federated learning architecture of a multi-parameter server and a multi-parameter party in one embodiment;

3 is a schematic diagram of a direct-connected topology structure of a multi-parameter server in one embodiment;

4 is a schematic diagram of a star topology of a multi-parameter server in one embodiment;

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

6 is a schematic flowchart of a data processing method based on federated learning in one embodiment;

7 is a schematic diagram of the framework of a participant cluster corresponding to each parameter server in one embodiment;

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

9 is a schematic diagram of the relationship between the response times between parameter servers in one embodiment;

10 is a schematic diagram of a topology structure of a direct-connected parameter server in one embodiment;

11 is a schematic diagram of a topology structure of a direct-connected parameter server in another embodiment;

12 is a schematic diagram of the relationship of parameter server topology construction in one embodiment;

13 is a schematic diagram of a data processing method based on federated learning in another embodiment;

Fig. 14 is the experimental comparison result in MINIST image classification data set in one embodiment;

Fig. 15 is the experimental contrast result in Synthetic in one embodiment;

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

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

Figure 18 is a diagram of the internal structure of a computer device in one embodiment.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

With the research and progress of artificial intelligence technology, artificial intelligence technology has been researched and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important value.

The solutions provided in the embodiments of this application involve technologies such as federated learning of artificial intelligence. Federated Learning (Federated Learning) is an emerging artificial intelligence basic technology, which is a distributed machine learning training method based on privacy protection. At the same time, a high-quality prediction model can be trained while protecting the privacy of client data. Under the premise of ensuring information security during big data exchange, protecting the privacy of terminal data and personal data, and ensuring legal compliance, efficient machine learning can be carried out between multiple participants or multiple computing nodes.

For the application scenario where multiple participants participate in federated training, the data is distributed in different clients (participants) of the same organization, and the parameter server is used as an intermediate server for model parameter exchange. Aiming at the instability of the federated learning structure of single-parameter server and multi-participant, a federated learning structure of multi-parameter server and multi-parameter party is proposed.

Among them, the multi-parameter server and multi-participant federated learning structure has multiple parameter servers, and multiple parameter servers exchange parameters with their corresponding participant groups. The specific training process is shown in Figure 2:

Step 21: Participating client clients (eg, user's mobile phone, multiple banks, ie product users) download the shared prediction model from their corresponding parameter servers.

Step 22: The participant client uses local data to train the model and iterate.

Step 23: Participating clients update the gradients obtained after model training, encrypt and upload them to their corresponding parameter servers.

Step 24: All parameter servers wait to collect gradient updates from all participants, then exchange and integrate gradients with neighbor servers, and then update the shared model on their own side.

Repeating the process of steps 21 to 24 above, the participants continuously download the model updated each time on the server side to the local client for update iteration until the model converges and stop updating.

In the multi-parameter server and multi-participant federated learning structure, the structures between multi-parameter servers are mainly direct-connected and star-shaped structures. The direct-connected structure between parameter servers is shown in Figure 3. The structure is shown in Figure 4.

The data processing method based on federated learning provided by the embodiments of the present application can be applied to the federated learning structure of multi-parameter server and multi-participant as shown in FIG. 5 . The participants communicate with the parameter server over the network. Each parameter service receives the gradient data obtained by the participants of the corresponding participant cluster using local data training; summarizes the gradient data of the participant cluster to obtain the gradient summary data; obtains the parameter server topology constructed based on the response time between the parameter servers Structure; based on the parameter server topology, exchange gradient summary data with adjacent parameter servers; update the parameters of the joint model according to the gradient summary data of all parameter servers obtained from the exchange. The participant 102 may be a user terminal, and the user terminal includes but is not limited to a mobile phone, a computer, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, and the like. The parameter server 104 can be implemented by an independent server or a server cluster composed of multiple servers. The embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, and the like.

In one embodiment, as shown in FIG. 6 , a data processing method based on federated learning is provided. Taking the method applied to a participating server in FIG. 1 as an example, the method includes the following steps 602 to 612 , which can be It is understood that when the multi-participant server performs federated learning, each participant server performs the following steps 602 to 610 respectively:

Step 602: Receive gradient data obtained by participants in the corresponding participant cluster using local data training.

Among them, federated learning is a global model that is jointly trained by the participants. Each participant trains the model based on their own local data, and then exchanges and summarizes the global model through the parameter server. In this process, the user data is always local, not external. Send, meet data security and privacy protection requirements. The participants of federated learning generally include the data party, the algorithm party, the coordinator, the calculation party, the result party, and the task initiator. Take the application of federated learning to the joint modeling of risk control among different banks as an example, the participants include banks, users and credit institutions.

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

Among them, the participant cluster corresponding to each parameter server can be initialized when the federated learning is initialized. In one embodiment, regions can be divided to determine the participant cluster corresponding to each parameter server, and the participants in the same or similar regions can be constructed as the local parameter server participant cluster. In one embodiment, each server can be used as a central node, and the response time of the server and the participants can be used as clusters to form a cluster of multiple participants (corresponding to the number of servers). The whole clustering algorithm makes the average response time of the participants who fall into the same cluster and the corresponding servers the shortest. In one way, the parameter server may send a gradient update request to the participants, and select the participants to construct a corresponding participant cluster according to the response time of the participants. In one method, the parameter server may send a test request to the participant, and select the participant to construct the corresponding participant cluster according to the response time of the participant.

Through initialization, the participant cluster corresponding to each parameter server is constructed, as shown in Figure 7. Among them, in federated learning, the parameter server sends the training program to the participants, and the participants use the local data to calculate the gradient and loss of the descent, obtain the gradient data of one round of training, encrypt and upload it to the corresponding parameter server. Therefore, for the parameter server, what it receives is the gradient data obtained by one iteration of training of each participant in the corresponding participant cluster. As shown in FIG. 7 , the parameter server 1 receives the gradient data obtained by each participant in the participant cluster 1 during this iterative training.

It is understandable that, in order to ensure the confidentiality of the data during the training process, the participants encrypt the gradient data and send it to the corresponding parameter server.

Step 604: Summarize the gradient data of the participant clusters to obtain the gradient summary data.

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

Step 606: Obtain a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when the federated learning model is initialized.

The data processing method based on federated learning of the present application is suitable for the application scenario of federated learning of multi-participant and multi-parameter servers. Each parameter server receives gradient data for a cluster of participants and aggregates them. After the aggregation, the gradient aggregation data needs to be exchanged between the parameter servers.

Usually, the parameter servers are in a direct connection topology as shown in Figure 3, or a star topology as shown in Figure 4. The main disadvantage of the directly connected multi-parameter server topology is that the model convergence speed is slow, each parameter server needs to exchange parameters for many times to make the model converge, and the system stability is also poor due to a single point of failure. In the star-structured multi-parameter server topology, the model converges quickly, but there is still a single point of failure in the central node, which leads to poor system stability.

In this embodiment, the parameter server topology is constructed based on the response time between the parameter servers, so that the topology between the parameter servers is not fixed, but is flexibly constructed based on the response time between the parameter servers, which can truly reflect the process of federated learning. , the effect of short response time between parameter servers on the exchange efficiency of gradient summary data between parameter servers.

Wherein, during the initialization process of the federated learning model, each parameter server may send a request to other parameter servers to obtain the response time of the other parameter servers to the request. As shown in Figure 7, there are C parameter servers, each server obtains the response time with other servers.

In one way, after obtaining the response time between all parameter servers, a direct-connected parameter server topology is constructed with the shortest exchange time as the goal.

The topological structure of the parameter server of the direct connection type is shown in Figure 3, and each parameter server is connected in sequence. Aiming at the shortest exchange time, the directly-connected parameter server topology is constructed by utilizing the response time between parameter servers, which is different from the existing fixed direct-connected parameter server topology. By making the data exchange time between parameter servers the shortest, the data exchange efficiency between parameter servers is improved, thereby improving the efficiency of federated learning.

In one way, after obtaining the response time between all parameter servers, the parameter server topology is constructed with the shortest exchange time and the maximum degree constraint as the goal.

The maximum degree constraint means that the edges between parameter servers meet the maximum number of requirements. Due to the setting of the maximum degree constraint, each parameter server has a connection relationship with multiple parameter servers, and the single point of failure problem that is easy to exist in the direct connection type is avoided. Aiming at the shortest exchange time and maximum degree constraints, the constructed parameter server topology not only considers the efficiency of data exchange, but also the stability of data, so as to avoid the problem of poor system stability caused by single point of failure.

Step 608 , exchange gradient summary data with adjacent parameter servers based on the parameter server topology.

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

After each parameter server exchanges and obtains the gradient summary data of the adjacent parameter servers, the gradient summary data of the adjacent parameter servers obtained by the exchange is again carried out with its own gradient summary data. At this time, the gradient summary data includes itself and the adjacent parameters. Gradient summary data for the server. For example, after the first round of exchange, parameter server A performs gradient summary on the gradient summary data of itself, parameter server B, parameter server E, and parameter server C.

If the parameter servers without direct connection cannot obtain the mutual gradient summary data in the first round of exchange, the round of exchange is continued. For example, if there is no direct connection between parameter server B and parameter server D, parameter server C obtains the gradient summary data of parameter server D in the first round of exchange. Through the second round of exchange, parameter server C will include parameter server D. The gradient summary data is exchanged to parameter server B. It can be understood that, through multiple rounds of exchange and gradient aggregation, each parameter server obtains the gradient aggregation data of all parameter servers.

It is understandable that, in order to ensure the confidentiality of the data during the training process, after the gradient summary data is encrypted between the parameter servers, it is exchanged with the adjacent parameter servers, that is, the exchange between the parameter servers is after homomorphic encryption. The gradient summary data of .

Step 610: Update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through the exchange.

Specifically, through multiple rounds of exchange and gradient aggregation, each parameter server obtains the gradient aggregation data of all parameter servers, updates the parameters of the joint model, and completes an iterative training of federated learning.

The parameter server continues to distribute the updated joint model to the corresponding parameter server cluster according to the updated joint model, and each participant uses the local data to continue training, and each parameter server continuously updates the joint model by repeating steps 602 to 610, until Training is over.

In the above-mentioned data processing method based on federated learning, each parameter server corresponds to the participant cluster and receives the gradient data trained by the corresponding parameter cluster, which is a multi-participant and multi-parameter server architecture. For data exchange with the cluster cluster, even if there is a failure of the parameter server, it will not affect the training of federated learning, which improves the stability of the system. At the same time, each parameter server constructs a parameter server topology structure according to the response time between the parameter servers. After each parameter server obtains the gradient summary data of the corresponding participant cluster, it exchanges the gradient summary data based on the parameter server topology structure. Since the parameter server topology is constructed based on the response time between parameter servers, it is not fixed, which provides a basis for improving the data exchange efficiency between parameter servers.

In another embodiment, receiving the gradient data obtained by the participants in the corresponding participant cluster using local data training includes: sending a gradient update request to the corresponding participant cluster, and receiving the fastest response in the participant cluster The first N participants of the gradient update request use the gradient data obtained by local data training.

Specifically, each participant of federated learning has different response times due to different regions, network environments, etc., and the participants with slow response time in the same federation will seriously drag down the overall training time. In order to solve this problem, the participant cluster of the parameter server is constructed by clustering the participants based on the response time of the participants to the parameter server. Specifically, each parameter party with a short response time to the parameter server is clustered into a cluster, and the participants of the parameter server are clustered, so that all the participants of the federated learning are clustered into different clusters, so that those belonging to the same cluster are clustered. The parties have similar response times.

The factors that usually affect the response speed of the participant are the distance between the participant and the parameter server and the processing capability of the participant. If the participant and the parameter server are in the same area, they can respond to the request of the parameter server faster. Therefore, usually A parameter server selects a cluster of participants based on response time consisting of geographically close participants. As an intermediate server for model parameter exchange, each parameter server receives the gradient data obtained by the participants through training based on local data. Even if a participant is selected by multiple parameter servers, the training results will not be affected. Therefore, this implementation In this example, the partitioning of the participant clusters can be performed only by the response time.

Further, in each training iteration of federated learning, the server only selects the top N participants in the same cluster for joint model iteration, thus greatly reducing the drag of slow-responding participants on fast-responding participants.

Specifically, the construction of the participant cluster can be in the initialization process of the federated learning modeling, and the initialization includes the initialization of the participant cluster. Taking each parameter server as the center, each parameter server sends a request to all participants respectively, and obtains the response time of each participant to respond to the request. Taking response time as clustering, the goal of clustering is to make the average response time of participants and corresponding servers in the same cluster to be the shortest. Specifically, according to the response time, the fastest and closest multiple participant servers are selected to construct corresponding participant clusters. Thus, a cluster of participants corresponding to each server is formed.

Based on participant clustering, in each round of iterative training, each parameter server sends a gradient update request to the corresponding participant cluster, and receives the top N participants in the participant cluster that respond to the gradient update request fastest using local data. Gradient data obtained from training.

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

It is foreseeable that since the parameter server selects the gradient data of the top N participants with the fastest response in the participant cluster according to the response time, the participants with fast responses will not be dragged down by the participants with slow responses, thereby reducing the number of participants. The total waiting time of the parameter server is reduced, and the joint modeling efficiency is improved.

In another embodiment, the method of constructing a parameter server topology structure based on the response time between the parameter servers includes: obtaining the response time between the parameter servers; Topology.

Among them, the topological structure of the parameter server of the direct connection type is shown in Figure 4, and the participating servers are connected in sequence. Different from the traditional direct-connected parameter server topology, the direct-connected parameter server topology in this embodiment considers the exchange time based on the response time between servers. By making the exchange time shortest, the constructed direct-connected parameter server topology improves the data exchange efficiency.

Specifically, based on the response time, with the shortest exchange time as the goal, construct the parameter server topology structure, including: taking each parameter server as the starting point, iteratively selects the response time with the starting point with the shortest response time and is in the disconnected state according to the response time between the parameter servers. The parameter server is used as the next starting point until all parameter servers are connected, and multiple candidate parameter server topology structures are obtained; the candidate parameter server with the shortest total response time is used as the final parameter server topology structure.

Among them, the parameter server topology can be constructed during the initialization of federated learning joint modeling. Each parameter server sends a request to other parameter services, and obtains the response time of each parameter server in response to its own request. In one embodiment, the response time between the parameter servers is shown in FIG. 9 .

On this basis, the following steps were performed:

S10, each parameter server is taken as the starting point, and the parameter server with the shortest response time from the starting point is selected as the next starting point. Taking FIG. 9 as an example, the selection server A, parameter server B, parameter server C, parameter server D and parameter server E are respectively selected as the starting point, and the parameter server with the shortest response time from the starting point is selected as the next starting point. Taking server A as the starting point as an example, select parameter server B with the shortest response time from parameter server A as the next node.

S11, the above step S10 is repeated until all parameter servers are connected, and a plurality of candidate parameter server topology structures are obtained.

Specifically, according to the response time between the parameter servers, the parameter server with the shortest response time to the starting point and in an unconnected state is selected as the next starting point. For example, taking parameter server A as the starting point, the second node is parameter server B, the third node is parameter server C, the fourth node is parameter server D, and when the fifth node judges, parameter server B If the response time of parameter server E is the same, but parameter server B is connected, parameter server E is selected as the fifth node. Then, the topology structure of the parameter server constructed with the parameter server A as the starting point is shown in Fig. 10 . Figure 11 shows the topology of the parameter server constructed with the parameter server E as the starting point.

Taking each parameter server as a starting point, a plurality of candidate parameter server topology structures can be constructed and obtained.

S12, the candidate parameter server with the shortest total response time is used as the final parameter server topology structure.

Among them, the total response time is the sum of the response times between the connected parameter servers. 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. If the total response time is the shortest, based on the parameter server topology, the data exchange time between the parameter servers is also the shortest. After comparison, the total response time of the direct-connected parameter server topology with A as the starting point is the shortest, and the structure shown in Figure 10 is used as the final parameter server topology. It can be compared that, due to the shortest total response time, when using the direct-connected parameter server topology to exchange gradient data, the time spent is also the shortest, which can improve data exchange efficiency.

Directly connected parameter servers require multiple interactions to make the model converge, and also lead to poor system stability due to a single point of failure. On this basis, another construction method of parameter server topology is provided.

In another embodiment, the method of constructing a parameter server topology structure based on the response time between the parameter servers includes: obtaining the response time between the parameter servers; Topology.

The maximum degree constraint refers to the maximum number of edges of each node in the parameter server topology. By setting the maximum number of degrees to at least 2, that is, the number of edges of each node is at least 2, it is possible to make the interaction relationship between parameter servers various, avoiding single point of failure, but not too much and affecting efficiency . Through the shortest exchange time and maximum degree constraints, the exchange efficiency and the stability of the system are synthesized.

Specifically, based on response time, with the goal of shortest exchange time and maximum degree constraints, a parameter server topology is constructed, including:

S121: Taking any parameter server as a starting point, and according to the response time between the parameter servers, select a parameter server with the shortest response time from the starting point and in an unconnected state as the next starting point.

As shown in FIG. 12 , based on the response time among all parameter servers, starting from any parameter server, such as parameter server B, the parameter server A with the shortest response time with parameter server B is selected as the next starting point. If nodes with the same response time are encountered, one of them is randomly selected as the next starting point.

S122, the above step S121 is iterated until all parameter servers are connected, and a parameter server connection structure is obtained.

Taking Figure 12 as an example, with parameter server B as the starting point, parameter server A with the shortest response time with parameter server B is selected as the second node, parameter server C is selected as the third node, and parameter server D is selected as the fourth node. select the parameter server E as the fifth node, and the obtained connection structure of the parameter server is shown in Figure 12.

S123: Sort the response time of the parameter server from small to large, and sequentially add the response relationship not added to the parameter server connection structure into the parameter server connection structure based on the maximum degree constraint, to obtain the final parameter server topology structure.

Specifically, the response relationship refers to a connection relationship between servers for expressing response time, which may be an edge in a topology structure. The response time of the parameter server can be sorted from small to large, and the response relationships (edges) that are not added to the parameter server connection structure can be added to the parameter server connection structure in turn. In the iterative process, if the degree of any node on both sides of the edge to be added is equal to the maximum number of degrees, the edge will be discarded. Until all the joinable edges are added to the parameter server connection structure, the network topology of the server node is finally formed. As shown in Figure 12, the maximum number of degrees is set to 3. After sorting according to the response time, edges (B->C, B->E, A-E) are added to the graph in turn to form the final parameter server topology.

In this embodiment, by using the response time between servers to construct a parameter server topology structure, the impact on federated learning caused by problems such as short node response time and hardware region is considered. The parameter server topology is constructed based on the maximum degree constraint. Each parameter server is not connected to a single point in sequence, and has multi-directionality. Even if a node has a single point of failure, it can still exchange data with other parameter servers through other connection relationships. The method takes into account the transfer efficiency as well as the stability of the system.

The data processing method based on federated learning in this application can be applied to the joint modeling of any number of institutions without exposing their own private data, such as joint modeling of risk control between different banks or credit institutions, and between different medical heterogeneous Joint modeling of case diagnosis, joint modeling of product recommendation between different e-commerce platforms, etc. In the joint model training process based on federated learning, a decentralized algorithm for self-adaptive adjustment can be implemented, which can improve the overall efficiency and system robustness of federated learning without losing the effect of joint modeling, and improve the product experience.

Specifically, in one embodiment, the data processing method based on federated learning, as shown in FIG. 13 , includes the following steps:

S131: Federated learning initialization. In this step, the parameter server topology is initialized, the participant cluster is initialized, and the joint model parameters are initialized.

Specifically, the parameter server topology is initialized, including:

1), record the response time between all servers.

2) Take any parameter server as the starting point, iteratively select the parameter server with the shortest response time and the disconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected, and the parameter server is obtained connected structure.

3) Rank the response time of the parameter server from small to large, and sequentially add the response relationship that is not added to the parameter server connection structure to the parameter server connection structure. In the iterative process, if the degree of any node on both sides of the edge to be added (response relationship) is equal to the maximum number of degrees, the edge will be discarded. Until all the joinable edges are added to the parameter server connection structure, the network topology of the server node is finally formed.

Among them, the initialization of the participant cluster specifically includes: each parameter server sends a request to each participant, with each server as the central node, and the response time of the server and the participant as a cluster to form a plurality of participant clusters (with the number of servers). The whole clustering algorithm makes the average response time of the participants who fall into the same cluster and the corresponding servers the shortest, and constructs the participant clusters corresponding to each parameter server.

Among them, the initialization of federated model parameters ensures that the model parameters on each server are initialized to the same set.

S132, the participant client (eg, the user's mobile phone, multiple banks, ie product users) downloads the joint model from the server, and the joint model is initialized by the server cluster at the beginning of training (all servers have the same joint model).

S133 , the participant client uses local data to perform training iteration on the model.

S134 , with each parameter server as the central node, each participant cluster selects the top n participants with the fastest response, and upload the aggregated gradient to the corresponding server. The size of n is set by the server, and the maximum number of participants participating in each round of iteration is controlled by the size of n in federated model training.

S135, each server in the server cluster waits to collect the gradient updates reported by the corresponding participants. When the server receives the gradient updates of the first n participants, it aggregates all the collected gradients, and then exchanges gradients with neighboring servers. After the exchange, each server updates the joint model. Then enter the next round of iteration (each server again invites all corresponding participants to cluster for joint model download and local model update).

Among them, in each round of training iteration, the server requests the gradient report of n participants in the corresponding participant cluster. After the participant cluster receives the request, it returns the gradients of the top n participants with the fastest response. Since the participants receiving the request have similar response times and only return the fastest top n, the participants who respond quickly will not be dragged down by the participants who respond slowly, thereby reducing the total server waiting time and improving joint modeling. efficiency.

Each server waits for the gradient reported by the corresponding participant cluster, and then summarizes the gradient. The specific formula is shown in Equation 1.

After each server summarizes the gradient, it exchanges its own gradient with the neighbor node according to the topology structure. After the exchange, the gradient summary and parameter update are performed again to complete the update iteration of this model.

The above steps from S132 to S135 are repeated, and the model updated each time on the server side is continuously downloaded to the local client of the participant, until the model converges, and the update is stopped.

The data processing method based on federated learning in this application can be a general self-adaptive and decentralized method. In scenarios such as federated learning or distributed training, it can effectively improve the training of the model within a limited training time budget. efficiency, while ensuring the accuracy of the model and the stability of the network topology. From the application point of view, this scheme is not limited to the joint modeling scenario of financial anti-fraud involving the participation of multiple banks or mutual financial enterprises, and can be applied to various scenarios of joint modeling of multiple clients under the constraints of privacy protection. From a model perspective, the model architecture used by the server in joint modeling can be flexibly changed according to the actual application scenario.

The self-adaptive decentralized federation training method proposed in this scheme can automatically generate the optimal parameter server topology and response time clustering method to select participants under a limited model training budget to improve federation Learn the model training efficiency and training system stability of the system, and achieve maximum model correctness within a limited training time. Figure 14 shows the experimental comparison results on the MINIST image classification dataset under 1000 participants, 10 servers, and 3 different topologies. Under the topology structure, the results are compared in the Synthetic experiment. The delay results show that the adaptive topology has better convergence speed than the star and direct-connected models, and the model effect (correct rate) is not attenuated. Figure 16 shows the complexity and robustness of different topologies, and the results also demonstrate the advantages of this method in enhancing system stability and robustness. The effectiveness of this scheme is further verified.

It should be understood that, although the steps in the flowcharts involved in the above embodiments are sequentially displayed according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order, and the steps may be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but may be performed at different times The execution order of these steps or phases is not necessarily sequential, but may be performed alternately or alternately with other steps or at least a part of the steps or phases in the other steps.

Based on the same inventive concept, an embodiment of the present application also provides a federated learning-based data processing apparatus for implementing the above-mentioned federated learning-based data processing method. The implementation solution for solving the problem provided by the device is similar to the implementation solution described in the above method, so the specific limitations in one or more embodiments of the federated learning-based data processing device provided below can refer to the above for the federated learning-based data processing device. The limitations of the learned data processing method will not be repeated here.

In one embodiment, as shown in FIG. 17, a data processing apparatus based on federated learning is provided, which is applied to each parameter server in the federated learning architecture, including:

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

The summarizing module 1704 is configured to summarize the gradient data of the participants of the participant cluster to obtain the gradient summary data.

The structure obtaining module 1706 is configured to obtain the parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when the federated learning model is initialized.

The exchange module 1708 is configured to exchange the gradient summary data with adjacent parameter servers based on the parameter server topology.

The updating module 1710 is configured to update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through the exchange.

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

In another embodiment, the structure acquisition module includes:

The time acquisition module is used to acquire the response time between parameter servers.

The building module is configured to build a direct-connected parameter server topology with the shortest exchange time as the goal based on the response time.

In another embodiment, a building module is configured to take each parameter server as a starting point, and iteratively select the parameter server with the shortest response time and the disconnected state from the starting point according to the response time between the parameter servers as the next parameter server. until all parameter servers are connected, and multiple candidate parameter server topologies are obtained; the candidate parameter server with the shortest total response time is used as the final parameter server topology.

In another embodiment, a building module is configured to build a parameter server topology based on the response time, targeting a minimum exchange time and a maximum degree constraint, the number of maximum degrees being at least two.

In another embodiment, a building module is configured to take any parameter server as a starting point, and iteratively select a parameter server with the shortest response time and a disconnected state from the starting point according to the response time between the parameter servers as the next starting point until all parameter servers are connected, and the parameter server connection structure is obtained; according to the response time of the parameter server, the response time is sorted from small to large, and the response relationship that is not added to the parameter server connection structure is sequentially added to the parameter server connection based on the maximum degree constraint. structure to get the final parameter server topology.

Each module in the above-mentioned federated learning-based data processing apparatus may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

In one embodiment, a computer device is provided, and the computer device 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 (Input/Output, I/O for short) and a communication interface. Wherein, the processor, the memory and the input/output interface are connected through the system bus, and the communication interface is connected to the system bus through the input/output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The nonvolatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the execution of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store training data for federated learning. The input/output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. The computer program implements a federated learning-based data processing method when executed by the processor.

Those skilled in the art can understand that the structure shown in FIG. 18 is only a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied. Include more or fewer components than shown in the figures, or combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, where a computer program is stored in the memory, and when the processor executes the computer program, the processor implements the steps of the federated learning-based data processing method of the foregoing embodiments.

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

In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the federated learning-based data processing methods of the foregoing 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.) involved in this application are all It is the information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage In the medium, when the computer program is executed, it may include the processes of the above-mentioned method embodiments. Wherein, any reference to a memory, a database or other media used in the various embodiments provided in this application may include at least one of a non-volatile memory and a volatile memory. Non-volatile memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetic variable memory (Magnetoresistive Random Memory) Access Memory, MRAM), Ferroelectric Random Access Memory (FRAM), Phase Change Memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory may include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration and not limitation, the RAM may be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM). The databases involved in the various embodiments provided in this application may include at least one of relational databases and non-relational databases. The non-relational database may include a blockchain-based distributed database, etc., but is not limited thereto. The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., and are not limited to this.

The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features It is considered to be the range described in this specification.

The above-mentioned embodiments only represent several embodiments of the present application, and the descriptions thereof are relatively specific and detailed, but should not be construed as a limitation on the scope of the patent of the present application. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the scope of protection of the present application should be determined by the appended claims.

Claims (10)

1. A data processing method based on federated learning, wherein the method is applied to each parameter server in the federated learning architecture, and the method comprises: Receive the gradient data obtained by the participants in the corresponding participant cluster using the local data training; summarizing the gradient data of the participant cluster to obtain the gradient summary data; obtaining a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized; exchanging the gradient summary data with an adjacent parameter server based on the parameter server topology; The parameters of the joint model are updated according to the gradient summary data of all parameter servers obtained by the exchange. 2. The method according to claim 1, wherein receiving the gradient data obtained by the participants in the corresponding participant cluster using local data training, comprising: Send a gradient update request to the corresponding participant cluster, and receive gradient data obtained by training with local data from the top N participants in the participant cluster that respond to the gradient update request fastest. 3. The method according to claim 1, wherein the method for constructing a parameter server topology structure based on the response time between the parameter servers comprises: Get the response time between parameter servers; Based on the response time, a direct-connected parameter server topology is constructed with the shortest exchange time as the goal. 4. The method according to claim 3, wherein, based on the response time, the parameter server topology structure is constructed with the shortest exchange time as a target, comprising: Taking each parameter server as the starting point, iteratively selects the parameter server with 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, until all the parameter servers are connected, obtaining: Multiple candidate parameter server topologies; The candidate parameter server with the shortest total response time is used as the final parameter server topology. 5. The method according to claim 1, wherein the method for constructing a parameter server topology structure based on the response time between the parameter servers comprises: Get the response time between parameter servers; Based on the response time, a parameter server topology is constructed with the goal of a shortest exchange time and a maximum degree constraint, the number of which is at least two. 6. The method according to claim 5, wherein, based on the response time, with the shortest exchange time and the maximum degree constraint as the goal, constructing a parameter server topology structure, comprising: Taking any parameter server as the starting point, iteratively selects the parameter server with the shortest response time and the disconnected state from the starting point as the next starting point according to the response time between the parameter servers, until all parameter servers are connected, and the parameters are obtained. Server connectivity structure; The response time of the parameter server is sorted from small to large, and the response relationships not added to the parameter server connection structure are sequentially added to the parameter server connection structure based on the maximum degree constraint to obtain the final parameter server topology structure. 7. A data processing device based on federated learning, characterized in that it is applied to each parameter server in a federated learning architecture, the device comprising: The receiving module is used to receive the gradient data obtained by the participants in the corresponding participant cluster using the local data training; an aggregation module, configured to aggregate the gradient data of the participants of the participant cluster to obtain the gradient aggregated data; a structure acquisition module, configured to acquire a parameter server topology, wherein the parameter server topology is constructed based on the response time between the parameter servers when the federated learning model is initialized; an exchange module, configured to exchange the gradient summary data with an adjacent parameter server based on the parameter server topology; The update module is used to summarize the data according to the gradients of all parameter servers obtained by the exchange, and update the parameters of the joint model. 8. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the method according to any one of claims 1 to 6 when the processor executes the computer program. step. 9. A computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented. 10. A computer program product comprising a computer program, characterized in that the computer program implements the steps of the method according to any one of claims 1 to 6 when the computer program is executed by a processor.

Images