Data processing method, device and computer equipment based on federated learning

Abstract

The present application relates to a data processing method, device, computer equipment, storage medium and computer program product based on federated learning. The method can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, assisted driving, etc., such as being applied to each parameter server in a federated learning architecture, including: receiving gradient data obtained by participants in the corresponding participant cluster using local data training; summarizing the gradient data of the participant cluster to obtain gradient summary data; obtaining the parameter server topology; based on the parameter server topology, exchanging gradient summary data with adjacent parameter servers; updating the parameters of the joint model based on the gradient summary data of all parameter servers obtained by the exchange. The method improves system stability, and because the parameter server topology is constructed based on the response time between parameter servers and is not fixed, it provides a basis for improving the data exchange efficiency between parameter servers.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data processing method, apparatus, computer equipment, storage medium and computer program product based on federated learning. Background Art

[0002] Federated Learning is an emerging basic technology of artificial intelligence. It is a distributed machine learning training method based on privacy protection. Its goal is to protect the privacy of client data while training a high-quality prediction model when the training data is scattered on a large number of unreliable clients with high network latency.

[0003] A federated learning architecture is as follows Figure 1 As shown in Figure 1, it is a federated learning structure with a single parameter server and multiple participants. As the name implies, the parameter server has only a single node, which exchanges model parameters with multiple participants. The specific process is as follows: Figure 1 As shown in the figure, with a single parameter server and multiple participants, 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 is low. If the parameter server fails or there is a network problem, the entire federated learning training process will have problems. Summary of the invention

[0004] Based on this, it is necessary to provide a data processing method, apparatus, computer equipment, computer-readable storage medium and computer program product based on federated learning that can improve stability in response to the above technical problems.

[0005] In a first aspect, the present application provides a data processing method based on federated learning. The method comprises:

[0006] Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training;

[0007] Aggregating the gradient data of the participants in the participant cluster to obtain gradient summary data;

[0008] Acquire a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when initializing the federated learning model;

[0009] Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers;

[0010] Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

[0011] In a second aspect, the present application also provides a data processing device based on federated learning. Applicable to each parameter server in a federated learning architecture, the device includes:

[0012] A receiving module, used for receiving gradient data obtained by a participant in a corresponding participant cluster using local data training;

[0013] A summarizing module, used for summarizing the gradient data of the participants in the participant cluster to obtain gradient summary data;

[0014] A structure acquisition module, used to acquire a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when initializing the federated learning model;

[0015] An exchange module, configured to exchange the gradient summary data with an adjacent parameter server based on the parameter server topology;

[0016] The update module is used to update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through exchange.

[0017] In a third aspect, the present application further 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:

[0018] Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training;

[0019] Aggregating the gradient data of the participants in the participant cluster to obtain gradient summary data;

[0020] Acquire a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when initializing the federated learning model;

[0021] Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers;

[0022] Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

[0023] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0024] Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training;

[0025] Aggregating the gradient data of the participants in the participant cluster to obtain gradient summary data;

[0026] Acquire a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when initializing the federated learning model;

[0027] Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers;

[0028] Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

[0029] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0030] Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training;

[0031] Aggregating the gradient data of the participants in the participant cluster to obtain gradient summary data;

[0032] Acquire a parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between the parameter servers when initializing the federated learning model;

[0033] Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers;

[0034] Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

[0035] 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, receives the gradient data of the corresponding parameter cluster training, is a multi-participant, multi-parameter server architecture, 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 improves the stability of the system. At the same time, each parameter server builds a parameter server topology structure according to the response time between parameter servers. After obtaining the gradient summary data of the corresponding participant cluster, each parameter server exchanges the gradient summary data based on the parameter server topology structure. Since the parameter server topology structure is built based on the response time between parameter servers, it is not fixed, which provides a basis for improving the efficiency of data exchange between parameter servers. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1A schematic diagram of a federated learning architecture with a single parameter server and multiple participants in one embodiment;

[0037] Figure 2 A schematic diagram of a federated learning architecture of a multi-parameter server and a multi-parameter party in one embodiment;

[0038] Figure 3 Schematic diagram of a direct-connection topology structure of a multi-parameter server in one embodiment;

[0039] Figure 4 A schematic diagram of a star topology structure of a multi-parameter server in one embodiment;

[0040] Figure 5 An application environment diagram of a data processing method based on federated learning in one embodiment;

[0041] Figure 6 is a flow chart of a data processing method based on federated learning in one embodiment;

[0042] Figure 7 Schematic diagram of the architecture of the participant cluster corresponding to each parameter server in one embodiment;

[0043] Figure 8 A schematic diagram of a parameter server topology structure in one embodiment;

[0044] Fig. 9 A schematic diagram of the relationship between response times of parameter servers in one embodiment;

[0045] Fig.10 Schematic diagram of the topological structure of a direct-connection parameter server in one embodiment;

[0046] Fig.11 Schematic diagram of the topological structure of a directly connected parameter server in another embodiment;

[0047] Fig.12 A schematic diagram of the relationship between the parameter server topology structure in one embodiment;

[0048] Fig.13 A schematic diagram of a data processing method based on federated learning in another embodiment;

[0049] Fig.14 The following is an experimental comparison result on the MINIST image classification dataset in one embodiment;

[0050] Fig.15 The comparison results of the Synthetic experiment in one embodiment are shown;

[0051] Fig.16 The figure is a comparison result of the complexity and robustness of different topological structures in one embodiment;

[0052] Fig.17 is a structural block diagram of a data processing device based on federated learning in one embodiment;

[0053] Fig.18 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0056] The solution provided in the embodiments of the present application involves technologies such as federated learning of artificial intelligence. Federated learning is an emerging basic technology of artificial intelligence. It is a distributed machine learning training method based on privacy protection. Its goal is to protect the privacy of client data while training a high-quality prediction model when the training data is scattered on a large number of unreliable clients with high network latency. It can ensure information security during the exchange of big data, protect the privacy of terminal data and personal data, and ensure legality and compliance, and carry out efficient machine learning between multiple participants or multiple computing nodes.

[0057] In the application scenario where multiple parties 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. In order to solve the problem of the instability of the federated learning structure with a single parameter server and multiple participants, a federated learning structure with multiple parameter servers and multiple parameter parties is proposed.

[0058] Among them, the federated learning structure with multiple parameter servers and multiple participants has multiple parameter servers, and multiple parameter servers exchange parameters with their corresponding participant groups. The specific training process is as follows: Figure 2 As shown:

[0059] Step 21: The participating clients (e.g., the user's mobile phone, multiple banks, i.e., product users) download the shared prediction model from their respective corresponding parameter servers.

[0060] Step 22: The participating client uses local data to iterate the training of the model.

[0061] Step 23: The participating clients update the gradients obtained after model training and upload them to their corresponding parameter servers in encrypted form.

[0062] Step 24: All parameter servers wait to collect gradient updates from all participants, then exchange and integrate gradients with neighboring servers, and then update their own shared models.

[0063] Repeat the above steps 21 to 24, and the participants will continuously download the updated model from the server to the local client for update iteration until the model converges and stops updating.

[0064] In the federated learning structure with multiple parameter servers and multiple participants, the structures between multiple parameter servers are mainly direct connection and star structure. The direct connection structure between parameter servers is as follows: Figure 3 As shown in Figure 2, the star structure between parameter servers is as follows: Figure 4 shown.

[0065] The data processing method based on federated learning provided in the embodiment of the present application can be applied to Figure 5 In the federated learning structure of multiple parameter servers and multiple participants shown in the figure. The participants communicate with the parameter server through the network. Each parameter service receives the gradient data obtained by the participants of the corresponding participant cluster using local data training; the gradient data of the participant cluster is summarized to obtain gradient summary data; the parameter server topology structure constructed based on the response time between parameter servers is obtained; based on the parameter server topology structure, the gradient summary data is exchanged with the adjacent parameter servers; according to the gradient summary data of all parameter servers obtained by the exchange, the parameters of the joint model are updated. Among them, the participant 102 can be a user terminal, and the user terminal includes but is not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The parameter server 104 can be implemented with 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, smart transportation, assisted driving, etc.

[0066] In one embodiment, Figure 6 As shown in the figure, a data processing method based on federated learning is provided. Figure 1 Take one of the participating servers in the example as an example, including the following steps 602 to 612. It can be understood that when multiple participating servers perform federated learning, each participating server respectively performs the following steps 602 to 610:

[0067] Step 602: receiving gradient data obtained by participants in the corresponding participant cluster using local data for training.

[0068] Federated learning is a global model obtained by joint training by the participants. Each participant trains the model based on its own local data, and then exchanges and aggregates it through the parameter server to obtain the global model. In this process, user data is always local and not sent to the outside, meeting the requirements of data security and privacy protection. The participants of federated learning generally include data parties, algorithm parties, coordination parties, calculation parties, result parties, task initiators, etc. Taking the application of federated learning to joint risk control modeling among different banks as an example, the participants include banks, users and credit institutions.

[0069] A participant cluster is a cluster of some participants. Each participant cluster corresponds to a parameter server and reports gradient data obtained by training with local data to the parameter server.

[0070] Among them, when the federated learning is initialized, the participant clusters corresponding to each parameter server can be initialized. In one implementation, the participant clusters corresponding to each parameter server can be determined by region, and the participants in the same or similar regions can be constructed as the participant clusters of the local parameter server. In one implementation, each server can also be used as a central node, and the response time of the server and the participant can be clustered to form multiple participant clusters (corresponding to the number of servers). The entire clustering algorithm makes the average response time between the participants and the corresponding servers that fall into the same cluster the shortest. In one method, the parameter server can send a gradient update request to the participant, and select the participant to build the corresponding participant cluster based on the response time of the participant. In one method, the parameter server can send a test request to the participant, and select the participant to build the corresponding participant cluster based on the response time of the participant.

[0071] Through initialization, a cluster of participants corresponding to each parameter server is constructed, such as Figure 7 As shown in Figure 1. In federated learning, the parameter server sends the training program to the participants, and the participants use local data to calculate the gradient and loss of the descent, obtain the gradient data of a round of training, encrypt it and upload it to the corresponding parameter server. Therefore, for the parameter server, what it receives is the gradient data obtained by each participant in the corresponding participant cluster from one iteration of training. Figure 7 As shown, parameter server 1 receives the gradient data obtained by each participant in participant cluster 1 in this iterative training.

[0072] It is understandable that in order to ensure the confidentiality of data during training, the participants encrypt the gradient data and send it to the corresponding parameter server.

[0073] Step 604, summarizing the gradient data of the participant clusters to obtain gradient summary data.

[0074] Specifically, each parameter server aggregates the gradient data received from the participating clusters to obtain gradient aggregate data. The aggregate formula is as follows:

[0075]

[0076] 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 initializing the federated learning model.

[0077] The data processing method based on federated learning of the present application is applicable to the application scenario of federated learning of multiple participants and multiple parameter servers. Each parameter server receives the gradient data of a cluster of participants and summarizes them. After summarizing, the gradient summary data needs to be exchanged between the parameter servers.

[0078] Usually, the parameters between the servers are as follows Figure 3 Direct-connect topology as shown, or Figure 4 The star topology shown in the figure. The main disadvantage of the direct-connected multi-parameter server topology is that the model converges slowly, and each parameter server needs to exchange parameters multiple times to converge the model. It also leads to poor system stability due to single point failure. The star-structured multi-parameter server topology has a fast model convergence speed, but there is still a single point failure in the central node, which leads to poor system stability.

[0079] In this embodiment, the parameter server topology is constructed based on the response time between parameter servers, so that the topology between parameter servers is not fixed, but flexibly constructed based on the response time between parameter servers, which can truly reflect the impact of the length of the response time between parameter services on the efficiency of gradient summary data exchange between parameter servers during federated learning.

[0080] In the process of initializing the federated learning model, each parameter server can send a request to other parameter servers to obtain the response time of other parameter servers to the request. Figure 7 There are C parameter servers shown, each server obtains the response time between each server and other servers.

[0081] In one approach, 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.

[0082] The topology of the direct-connected parameter server is as follows: Figure 3As shown in the figure, each parameter server is connected in sequence. The direct-connected parameter server topology is constructed with the shortest exchange time as the goal. It is constructed using the response time between parameter servers, which is different from the existing fixed direct-connected parameter server topology. By minimizing the data exchange time between parameter servers, the data exchange efficiency between parameter servers is improved, thereby improving the efficiency of federated learning.

[0083] In one approach, after obtaining the response time between all parameter servers, the parameter server topology is constructed with the shortest exchange time and maximum degree constraint as the goal.

[0084] The maximum degree constraint means that the edges between parameter servers meet the maximum number requirement. Since the maximum degree constraint is set, each parameter server has a connection relationship with multiple parameter servers, avoiding the single point failure problem that is prone to exist in direct connection. With the shortest exchange time and maximum degree constraint as the goal, the parameter server topology structure constructed not only considers the efficiency of data exchange, but also the stability of data, avoiding the problem of poor system stability caused by single point failure.

[0085] Step 608: Exchange gradient summary data with adjacent parameter servers based on the parameter server topology.

[0086] Specifically, using the parameter server topology, each parameter server exchanges gradient summary data with adjacent parameter servers with which it has a connection relationship. Figure 8 Taking the parameter server topology shown in FIG. 1 as an example, parameter server A exchanges gradient summary data with parameter server B, parameter server C, and parameter server E. Correspondingly, through rounds of exchange, parameter server A can obtain the gradient summary data of parameter server B, parameter server E, and parameter server C. At the same time, in the first round of exchange, the gradients of the parameter servers adjacent to other parameter servers, for example, the gradient summary data of parameter server A, parameter server C, and parameter server E obtained by parameter server B.

[0087] After each parameter server obtains the gradient summary data of the adjacent parameter server through exchange, it again performs gradient summary on the gradient summary data of the adjacent parameter server obtained through exchange and its own gradient summary data. At this time, the gradient summary data includes the gradient summary data of itself and the adjacent parameter 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.

[0088] If the parameter servers that are not directly connected cannot obtain each other's gradient summary data in the first round of exchange, they will continue to exchange rounds. For example, if parameter server B and parameter server D are not directly connected, 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 exchanges the gradient summary data including parameter server D to parameter server B. It can be understood that through multiple rounds of exchange and gradient summary, each parameter server obtains the gradient summary data of all parameter servers.

[0089] It is understandable that in order to ensure the confidentiality of data during the training process, the gradient summary data is encrypted between parameter servers and then exchanged with adjacent parameter servers. In other words, the gradient summary data exchanged between parameter servers is homomorphically encrypted.

[0090] Step 610: Update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through exchange.

[0091] Specifically, through multiple rounds of exchanges and gradient aggregation, each parameter server obtains the gradient summary data of all parameter servers, updates the parameters of the joint model, and completes an iterative training of federated learning.

[0092] The parameter server continues to send the updated joint model to the corresponding parameter server cluster based on the updated joint model, and each participant continues to train using local data. Each parameter server continuously updates the joint model by repeating steps 602 to 610 until the training is completed.

[0093] In the above-mentioned data processing method based on federated learning, each parameter server corresponds to a participant cluster and receives the gradient data of the corresponding parameter cluster training. It is a multi-participant, multi-parameter server architecture. 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, thereby improving the stability of the system. At the same time, each parameter server builds a parameter server topology structure based on the response time between parameter servers. After obtaining the gradient summary data of the corresponding participant cluster, each parameter server exchanges the gradient summary data based on the parameter server topology structure. Since the parameter server topology structure is built based on the response time between parameter servers and is not fixed, it provides a basis for improving the data exchange efficiency between parameter servers.

[0094] In another embodiment, receiving gradient data obtained by participants in a corresponding participant cluster using local data training includes: sending a gradient update request to the corresponding participant cluster, and receiving gradient data obtained by first N participants in the participant cluster who respond fastest to the gradient update request using local data training.

[0095] Specifically, the response time of each participant in federated learning varies due to differences in region, network environment, etc., and participants with slow response time in the same federation will seriously drag down the overall training time. To address this problem, we first cluster the participants based on their response time to the parameter server and build a cluster of participants for the parameter server. Specifically, the parties with close response times to the parameter server are clustered into one cluster as the participant cluster of the parameter server, thereby clustering all participants in federated learning into different clusters, so that participants in the same cluster have similar response times.

[0096] The factors that usually 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. If the participants and the parameter server are in the same area, they can respond to the parameter server's request faster. Therefore, usually a participant cluster selected by a parameter server based on the response time consists of participants with close geographical locations. Each parameter server serves as an intermediate server for model parameter exchange. It receives the gradient data obtained by the participants through training based on local data. Even if a participant is selected by multiple parameter servers, it will not affect the training results. Therefore, in this embodiment, the participant clusters can be divided only by response time.

[0097] Furthermore, in each training iteration of federated learning, the server only selects the first N participants in the same cluster for joint model iteration, thereby greatly reducing the drag of slow-responding participants on fast-responding participants.

[0098] Specifically, the construction of the participant cluster can be in the initialization process of federated learning modeling, and the initialization includes the initialization of the participant cluster. With each parameter server as the center, each parameter server sends a request to all participants respectively to obtain the response time of each participant in response to the request. Taking the response time as the clustering, the goal of clustering is to make the average response time between the participants and the corresponding servers falling into the same cluster as shortest. Specifically, according to the response time, select the fastest and closest multiple participant servers to build the corresponding participant cluster. Thus, a participant cluster corresponding to each server is formed.

[0099] Based on the participant clusters, in each round of iterative training, each parameter server sends a gradient update request to the corresponding participant cluster, and receives the gradient data obtained by training with local data from the first N participants in the participant cluster who respond to the gradient update request the fastest.

[0100] Specifically, each parameter server sends a gradient update request to each participant, and only receives the gradient data sent by the first N participants who respond the fastest. 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 respond first.

[0101] It can be foreseen that since the parameter server selects the gradient data of the first N fastest responding participants in the participant cluster based on the response time, the fast responding participants will not be dragged down by the slow responding participants, thereby reducing the total waiting time of the parameter server and improving the efficiency of joint modeling.

[0102] In another embodiment, a method for constructing a parameter server topology structure based on the response time between parameter servers includes: obtaining the response time between parameter servers; based on the response time, constructing a direct-connected parameter server topology structure with the shortest exchange time as the goal.

[0103] Among them, the topology of the direct-connected parameter server is as follows: Figure 4 As shown, each participating server is connected in sequence. Different from the traditional direct-connected parameter server topology, the direct-connected parameter server topology in this embodiment takes the exchange time into consideration based on the response time between servers. By minimizing the exchange time, the constructed direct-connected parameter server topology improves the data exchange efficiency.

[0104] Specifically, based on the response time, with the shortest exchange time as the goal, a parameter server topology structure is constructed, including: taking each parameter server as the starting point, iteratively selecting a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain multiple candidate parameter server topology structures; and taking the candidate parameter server with the shortest total response time as the final parameter server topology structure.

[0105] The parameter server topology can be constructed during the 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 in response to its own request. In one embodiment, the response time between each parameter server is as follows: Fig. 9 shown.

[0106] On this basis, the following steps were performed:

[0107] S10, taking each parameter server as a starting point, select the parameter server with the shortest response time to the starting point as the next starting point. Fig. 9For example, select server A, parameter server B, parameter server C, parameter server D and parameter server E as the starting point, and select the parameter server with the shortest response time to the starting point as the next starting point. Taking server A as the starting point as an example, select parameter server B with the shortest response time to parameter server A as the next node.

[0108] S11, repeat the above step S10 until all parameter servers are connected to obtain multiple candidate parameter server topologies.

[0109] Specifically, according to the response time between parameter servers, the parameter server with the shortest response time with 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, and the fourth node is parameter server D. When judging the fifth node, parameter server B has the same response time as parameter server E, but parameter server B is in a connected state, so parameter server E is selected as the fifth node. The parameter server topology constructed with parameter server A as the starting point is as follows: Fig.10 The parameter server topology structure constructed with parameter server E as the starting point is shown in Fig.11 shown.

[0110] Taking each parameter server as a starting point, multiple candidate parameter server topologies can be constructed.

[0111] S12, taking the candidate parameter server with the shortest total response time as the final parameter server topology.

[0112] The total response time is the sum of the response times between the connected parameter servers. Fig.10 As an example, the parameter server topology shown in the figure has a total response time of 10. Fig.11 Taking the parameter server topology shown in the figure as an example, the total response time is 12. The total response time is the shortest. Based on this parameter server topology, the time for data exchange between 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. Fig.10 The structure shown is the final parameter server topology. It can be compared that, since the total response time is the shortest, the time spent on gradient data exchange using the direct-connection parameter server topology is also the shortest, which can improve data exchange efficiency.

[0113] Directly connected parameter servers require multiple interactions to converge the model, and single point failures can also lead to poor system stability. Based on this, another way to build a parameter server topology is provided.

[0114] In another embodiment, a method for constructing a parameter server topology structure based on the response time between parameter servers includes: obtaining the response time between parameter servers; and constructing the parameter server topology structure based on the response time with the shortest exchange time and the maximum degree constraint as the goal.

[0115] The maximum degree constraint refers to the maximum number of edges of each node in the parameter server topology. By setting the maximum degree to at least 2, that is, the number of edges of each node is at least 2, the interaction relationship between parameter servers can be multiple, avoiding single point failure, but not too many to affect efficiency. Through the shortest exchange time and maximum degree constraints, the exchange efficiency and system stability are combined.

[0116] Specifically, based on the response time, with the shortest exchange time and maximum degree constraint as the goal, the parameter server topology is constructed, including:

[0117] S121: Taking any parameter server as a starting point, according to the response time between the parameter servers, select a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point.

[0118] by Fig.12 As shown in the figure, based on the response time between all parameter servers, take any parameter server as the starting point, such as parameter server B, and select parameter server A with the shortest response time with parameter server B 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.

[0119] S122, iterate the above step S121 until all parameter servers are connected to obtain a parameter server connectivity structure.

[0120] by Fig.12 For example, take parameter server B as the starting point, select parameter server A with the shortest response time to parameter server B as the second node, select parameter server C as the third node, select parameter server D as the fourth node, and select parameter server E as the fifth node. The obtained parameter server connectivity structure is as follows: Fig.12 shown.

[0121] S123, sorting the parameter servers in ascending order according to their response time, and adding the response relationships that have not been added to the parameter server connectivity structure to the parameter server connectivity structure in sequence based on the maximum degree constraint, to obtain a final parameter server topology structure.

[0122] Specifically, the response relationship refers to the connection relationship between servers used to represent the response time, which can be an edge in the topology structure. The response relationships (edges) that have not been added to the parameter server connectivity structure can be sorted from small to large according to the response time of the parameter servers, and the response relationships (edges) that have not been added to the parameter server connectivity structure can be added to the parameter server connectivity structure in sequence. During the iteration process, if the degree of any node on both sides of the edge to be added is equal to the maximum degree number, then the edge is abandoned. Until all the edges that can be added are added to the parameter server connectivity structure, the network topology structure of the server nodes is finally formed. Fig.12 As shown in the figure, the maximum degree is set to 3. After sorting by response time, edges (B->C, B->E, AE) are added to the graph in sequence to form the final parameter server topology.

[0123] In this embodiment, the parameter server topology is constructed by using the response time between servers, taking into account the impact of short node response time, hardware region and other issues on federated learning. Based on the parameter server topology constructed with maximum constraints, each parameter server is not a single-point connection relationship in sequence, but has multi-directionality. Even if a node has a single-point failure, it can still exchange data with other parameter servers through other connection relationships. This method takes into account the transmission efficiency and system stability.

[0124] The data processing method based on federated learning in this application can be applied to joint modeling of any number of institutions without exposing their privacy data, such as joint risk control modeling between different banks or credit institutions, joint case diagnosis modeling between different medical heterogeneities, and joint product recommendation modeling between different e-commerce platforms. In the training process of the joint model based on federated learning, a decentralized algorithm with adaptive adjustment can be implemented to improve the overall efficiency and system robustness of federated learning without losing the joint modeling effect, thereby improving the product experience.

[0125] Specifically, in one embodiment, a data processing method based on federated learning is as follows: Fig.13 As shown, the following steps are included:

[0126] S131: Federated learning initialization. In this step, the parameter server topology is initialized, the participant clusters are initialized, and the federated model parameters are initialized.

[0127] Specifically, the parameter server topology structure is initialized, including:

[0128] 1) Record the response time between all servers.

[0129] 2) Taking any parameter server as the starting point, iterate according to the response time between parameter servers, select the parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point, until all parameter servers are connected to obtain the parameter server connectivity structure.

[0130] 3) Sort the parameter servers by their response time from small to large, and add the response relationships that have not been added to the parameter server connectivity structure to the parameter server connectivity structure. During the iteration process, if the degree of any node on both sides of the edge (response relationship) to be added is equal to the maximum degree, then the edge will be abandoned. This process will continue until all edges that can be added are added to the parameter server connectivity structure, and finally form a network topology structure of the server nodes.

[0131] The initialization of the participant cluster specifically includes: each parameter server sends a request to each participant, takes each server as the central node, clusters the response time of the server and the participant, and forms multiple participant clusters (corresponding to the number of servers). The entire clustering algorithm makes the average response time between the participants and the corresponding servers in the same cluster the shortest, and builds the participant cluster corresponding to each parameter server.

[0132] Among them, the federated model parameter initialization ensures that the model parameters on each server are initialized to the same set.

[0133] S132, the participating client (such as the user's mobile phone, multiple banks, i.e., the product user) downloads the joint model from the server side. The joint model is initialized by the server cluster at the beginning of training (the joint model of all servers is the same).

[0134] S133, the participating client uses local data to iterate the training of the model.

[0135] S134, with each parameter server as the central node, each participant cluster selects the first n participants with the fastest response and uploads the summary gradient to the corresponding server. The size of n is set by the server, and the federated model training controls the maximum number of participants participating in each round of iteration through the size of n.

[0136] S135, each server in the server cluster waits to collect the gradient updates reported by the corresponding participants. After receiving the gradient updates from the first n participants, the server aggregates all the collected gradients and then exchanges gradients with the adjacent servers. After the exchange, each server updates the joint model. Then enter the next round of iteration (each server invites all corresponding participants to cluster again to download the joint model and update the local model).

[0137] In each round of training iteration, the server requests the gradient reports of n participants in the corresponding participant cluster. After receiving the request, the participant cluster returns the gradients of the first n participants with the fastest response. Since the participants receiving the request have similar response times and only the first n fastest ones are returned, the participants with fast response will not be dragged down by the participants with slow response, thereby reducing the total waiting time of the server and improving the efficiency of joint modeling.

[0138] Each server waits for the gradient reported by the corresponding participant cluster, and then summarizes the gradient. The specific formula is shown in Formula 1.

[0139]

[0140] After each server aggregates the gradients, it exchanges its own gradients with neighboring nodes according to the topology structure. After the exchange, it aggregates the gradients and updates the parameters again to complete the update iteration of this model.

[0141] Repeat the above steps S132 to S135, and continuously download the updated model of the server to the local client of the participant until the model converges and stops updating.

[0142] The data processing method based on federated learning in this application can be a general decentralized method of adaptive adjustment. In scenarios such as federated learning or distributed training, it can effectively improve the training efficiency of the model within a limited training time budget, while ensuring the accuracy of the model and the stability of the network topology. From an application perspective, this solution is not only limited to financial anti-fraud joint modeling scenarios involving multiple banks or mutual financial companies, but can also be applied to a variety of scenarios of multi-client joint modeling under privacy protection constraints. From a model perspective, the model architecture used by the server in joint modeling can be flexibly changed according to the actual application scenario.

[0143] The adaptively adjusted decentralized federated training method proposed in this scheme can improve the model training efficiency and training system stability of the federated learning system by adaptively generating the optimal parameter server topology and response time clustering to select participants within a limited model training budget, and achieve maximum model correctness within a limited training time. Fig.14 The experimental comparison results on the MINIST image classification dataset are shown with 1000 participants, 10 servers, and 3 different topologies. Fig.15 The results of the Synthetic experiment were compared with 1,000 participants, 10 servers, and 3 different topologies. The latency results show that the adaptive topology has a better convergence speed than the star and direct-connection models, and the model effect (accuracy) is not attenuated. Fig.16The complexity and robustness of different topological structures are demonstrated, and the results also prove that this method has advantages in enhancing system stability and robustness, further verifying the effectiveness of this solution.

[0144] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0145] Based on the same inventive concept, the embodiment of the present application also provides a data processing device based on federated learning for implementing the data processing method based on federated learning involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the one or more data processing device embodiments based on federated learning provided below can refer to the limitations of the data processing method based on federated learning above, and will not be repeated here.

[0146] In one embodiment, Fig.17 As shown, a data processing device based on federated learning is provided, which is applied to each parameter server in a federated learning architecture, including:

[0147] The receiving module 1702 is used to receive gradient data obtained by a participant in the corresponding participant cluster using local data for training.

[0148] The summarizing module 1704 is used to summarize the gradient data of the participants in the participant cluster to obtain gradient summary data.

[0149] The structure acquisition module 1706 is used to obtain the parameter server topology structure, wherein the parameter server topology structure is constructed based on the response time between parameter servers when initializing the federated learning model.

[0150] The exchange module 1708 is configured to exchange the gradient summary data with an adjacent parameter server based on the parameter server topology.

[0151] The updating module 1710 is used to update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through exchange.

[0152] In another embodiment, the receiving module is used to send a gradient update request to the corresponding participant cluster, and receive gradient data obtained by training using local data from the first N participants in the participant cluster who respond to the gradient update request the fastest.

[0153] In another embodiment, the structure acquisition module includes:

[0154] The time acquisition module is used to obtain the response time between parameter servers.

[0155] A construction module is used to construct a directly connected parameter server topology structure based on the response time and with the shortest exchange time as the goal.

[0156] In another embodiment, a construction module is used to take each parameter server as a starting point, iteratively select a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain multiple candidate parameter server topology structures; and use the candidate parameter server with the shortest total response time as the final parameter server topology structure.

[0157] In another embodiment, the construction module is used to construct a parameter server topology structure based on the response time with the shortest exchange time and the maximum degree constraint as the target, and the maximum degree is at least 2.

[0158] In another embodiment, a construction module is used to take any parameter server as a starting point, iteratively select a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain a parameter server connectivity structure; sort the parameter servers in ascending order according to their response time, and add the response relationships that have not been added to the parameter server connectivity structure to the parameter server connectivity structure in turn based on a maximum degree constraint to obtain a final parameter server topology structure.

[0159] Each module in the above-mentioned data processing device based on federated learning can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0160] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.18As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, 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. Among them, the processor of the computer device is used 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 the operating system and the computer program 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 an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a data processing method based on federated learning is implemented.

[0161] Those skilled in the art will understand that Fig.18 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0162] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the data processing method based on federated learning in the above-mentioned embodiments are implemented.

[0163] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the data processing method based on federated learning in the above-mentioned embodiments are implemented.

[0164] In one embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the steps of the data processing method based on federated learning in the above-mentioned embodiments.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0166] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0167] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0168] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A data processing method based on federated learning, characterized in that: The method is applied to each parameter server in a federated learning architecture, and the method includes: Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training; Summarizing the gradient data of the participant cluster to obtain gradient summary data; Obtaining a parameter server topology structure, wherein the parameter server topology structure is a candidate parameter server topology structure with the shortest total response time determined from a plurality of directly connected candidate parameter server topologies when initializing a federated learning model; the total response time of the candidate parameter server topology structure is the sum of the response times between the connected parameter servers in the candidate parameter server topology structure; Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers; Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

2. The method according to claim 1, characterized in that Receiving gradient data obtained by participants in the corresponding participant cluster using local data training, including: A gradient update request is sent to a corresponding participant cluster, and gradient data obtained by training with local data from the first N participants in the participant cluster who respond to the gradient update request fastest.

3. The method according to claim 1, characterized in that The method of constructing a parameter server topology structure based on the response time between the parameter servers includes: 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, characterized in that: The parameter server topology structure is constructed based on the response time and with the shortest exchange time as the goal, including: Taking each parameter server as a starting point, iteratively selecting a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected, thereby obtaining a plurality of directly connected candidate parameter server topology structures; The candidate parameter server topology structure with the shortest total response time is used as the final parameter server topology structure.

5. A data processing method based on federated learning, characterized in that: The method is applied to each parameter server in a federated learning architecture, and the method includes: Receiving gradient data obtained by participants in the corresponding participant cluster using local data for training; Summarizing the gradient data of the participant cluster to obtain gradient summary data; Acquire a parameter server topology structure, wherein the parameter server topology structure is obtained by sorting the parameter servers in ascending order according to their response time when initializing the federated learning model, and adding the response relationships that are not added to the parameter server connectivity structure to the parameter server connectivity structure in sequence based on the maximum degree constraint, wherein the parameter server connectivity structure connects all provided parameter servers; the maximum degree constraint is to make the edges between the parameter servers meet the maximum number requirement; Based on the parameter server topology, exchanging the gradient summary data with adjacent parameter servers; Update the parameters of the joint model based on the gradient summary data of all parameter servers obtained through exchange.

6. The method according to claim 5, characterized in that Receiving gradient data obtained by participants in the corresponding participant cluster using local data training, including: A gradient update request is sent to a corresponding participant cluster, and gradient data obtained by training with local data from the first N participants in the participant cluster who respond to the gradient update request fastest.

7. The method according to claim 5, characterized in that The method of constructing a parameter server topology structure based on the response time between the parameter servers includes: Get the response time between parameter servers; Based on the response time, a parameter server topology is constructed with the shortest exchange time and the maximum degree constraint as the goal, and the maximum degree is at least 2.

8. The method according to claim 7, characterized in that Based on the response time, with the shortest exchange time and maximum degree constraint as the goal, a parameter server topology is constructed, including: Taking any parameter server as a starting point, iteratively selecting a parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain a parameter server connectivity structure; The parameter servers are sorted in ascending order according to their response time, and the response relationships that are not added to the parameter server connectivity structure are added to the parameter server connectivity structure in sequence based on the maximum degree constraint to obtain a final parameter server topology structure.

9. A data processing device based on federated learning, characterized in that: Applied to each parameter server in a federated learning architecture, the device comprises: A receiving module, used for receiving gradient data obtained by a participant in a corresponding participant cluster using local data training; A summarizing module, used for summarizing the gradient data of the participants in the participant cluster to obtain gradient summary data; A structure acquisition module, used to acquire a parameter server topology structure, wherein the parameter server topology structure is a candidate parameter server topology structure with the shortest total response time determined from a plurality of directly connected candidate parameter server topologies when initializing a federated learning model; the total response time of the candidate parameter server topology structure is the sum of the response times between the connected parameter servers in the candidate parameter server topology structure; 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 update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through exchange.

10. The device according to claim 9, characterized in that The receiving module is further used to send a gradient update request to the corresponding participant cluster, and receive gradient data obtained by training with local data from the first N participants in the participant cluster who respond to the gradient update request the fastest.

11. The device according to claim 9, characterized in that The structure acquisition module comprises: The time acquisition module is used to obtain the response time between parameter servers; A construction module is used to construct a directly connected parameter server topology structure based on the response time and with the shortest exchange time as the goal.

12. The device according to claim 11, characterized in that The construction module is also used to take each parameter server as the starting point, iteratively select the parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain multiple directly connected candidate parameter server topology structures; and use the candidate parameter server topology structure with the shortest total response time as the final parameter server topology structure.

13. A data processing device based on federated learning, characterized in that: The device is applied to each parameter server in a federated learning architecture, and the device includes: A receiving module, used for receiving gradient data obtained by a participant in a corresponding participant cluster using local data training; A summary module, used to summarize the gradient data of the participant cluster to obtain gradient summary data; A structure acquisition module, used to acquire a parameter server topology structure, wherein the parameter server topology structure is obtained by sorting the parameter servers from small to large according to their response time when initializing the federated learning model, and adding the response relations that are not added to the parameter server connectivity structure to the parameter server connectivity structure in sequence based on the maximum degree constraint, wherein the parameter server connectivity structure connects all provided parameter servers; 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 update the parameters of the joint model according to the gradient summary data of all parameter servers obtained through exchange.

14. The device according to claim 13, characterized in that The receiving module is further used to send a gradient update request to the corresponding participant cluster, and receive gradient data obtained by training with local data from the first N participants in the participant cluster who respond to the gradient update request the fastest.

15. The device according to claim 13, characterized in that The structure acquisition module comprises: The time acquisition module is used to obtain the response time between parameter servers; A construction module is used to construct a parameter server topology structure based on the response time, with the shortest exchange time and the maximum degree constraint as the goal, and the maximum degree is at least 2.

16. The device according to claim 15, characterized in that The construction module is also used to take any parameter server as the starting point, iteratively select the parameter server with the shortest response time to the starting point and in an unconnected state as the next starting point according to the response time between the parameter servers, until all parameter servers are connected to obtain a parameter server connectivity structure; sort the parameter servers from small to large according to their response time, and add the response relationships that have not been added to the parameter server connectivity structure to the parameter server connectivity structure in turn based on the maximum degree constraint to obtain a final parameter server topology structure.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

18. A computer-readable storage medium having a computer program stored thereon, 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 8 are implemented.

19. A computer program product comprising a computer program, 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 8 are implemented.

Images