WO2022228204A1 - 一种联邦学习方法以及装置 - Google Patents

一种联邦学习方法以及装置 Download PDF

Info

Publication number
WO2022228204A1
WO2022228204A1 PCT/CN2022/087647 CN2022087647W WO2022228204A1 WO 2022228204 A1 WO2022228204 A1 WO 2022228204A1 CN 2022087647 W CN2022087647 W CN 2022087647W WO 2022228204 A1 WO2022228204 A1 WO 2022228204A1
Authority
WO
WIPO (PCT)
Prior art keywords
server
client
federated learning
servers
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/087647
Other languages
English (en)
French (fr)
Inventor
张琦
周培晨
陈钢
陈东升
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Technologies Co Ltd
Original Assignee
Huawei Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Technologies Co Ltd filed Critical Huawei Technologies Co Ltd
Priority to EP22794675.3A priority Critical patent/EP4318338A4/en
Publication of WO2022228204A1 publication Critical patent/WO2022228204A1/zh
Priority to US18/493,136 priority patent/US20240054354A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/098—Distributed learning, e.g. federated learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions

Definitions

  • the present application relates to the field of federated learning, and in particular, to a federated learning method and device.
  • Federated learning is an organization method of artificial intelligence (AI) joint modeling. On the basis of not directly obtaining private data of devices, it completes algorithm model tasks by coordinating model training and parameter aggregation across devices, while reducing traditional centralized It has important theoretical innovation and practical implementation value.
  • cross-device federated learning also brings new challenges, such as the heterogeneous client devices of federated learning, different devices in storage, computing, communication and battery There are big differences in hardware specifications, etc., which have an impact on the performance of federated learning. Therefore, how to improve the learning efficiency of cross-device federated learning has become an urgent problem to be solved.
  • This application provides a federated learning method for synchronizing the aggregated information between each server in the process of federated learning across devices, so as to keep the data of the server synchronized during each round of iterative learning. Access to the server at all times can obtain the full amount of data, improving the overall learning efficiency.
  • the present application provides a federated learning method, which is applied to a federated learning system.
  • the federated learning system includes multiple servers, the multiple servers are connected to each other, and the multiple servers are used for iteration.
  • Federated learning the process of any round of iterative learning includes: first, the first server receives a request message sent by at least one first client, and the request message is used to request the global model stored in the first server.
  • the client includes the at least one first client; the first server sends information of the global model and training configuration parameters to the at least one first client, where the information of the global model and the training configuration parameters are used to instruct the at least one first client to use
  • the training configuration parameters are used to train the global model;
  • the first server receives first model update parameters fed back by at least one first client respectively, and the first model update parameters are parameters of the global model obtained after the at least one first client is trained;
  • the first server aggregates the first model update parameters fed back by at least one first client to obtain the first aggregated information in the current iteration;
  • the first server obtains the second aggregated information sent by the second server,
  • the second aggregation information is the information obtained by the second server in the current iteration by aggregating the received second model update parameters;
  • the first server updates the first server based on the first aggregation information and the second aggregation information on the stored global model to get the updated global model.
  • the client needs to actively send a request message to the server to request participation in federated learning, and send the model update parameters obtained after training the global model to the server. Therefore, the client accesses the server by sending a request message to the server, and the client does not need to maintain a long-term stable connection with the server to implement a cross-device federated learning system.
  • the client In addition to aggregating the model update parameters received by multiple servers, they also receive and aggregate model update parameters sent by other servers, so that each server has more data than its own aggregated information.
  • the terminal When the terminal is connected to the server, more data can be obtained, so that each round of iterative training can be performed based on more data, and the overall efficiency of federated learning can be improved.
  • the client accesses the server through a combination of requests and responses, and there is no need to maintain a long-term stable connection between the client and the server, and the client can access any server at any time. More data is needed to improve the learning efficiency of client-side federated learning.
  • the first server may receive request messages sent by multiple clients to send request messages to the server.
  • the first server can deliver the locally stored global model information and training configuration parameters to the multiple clients. Therefore, in the embodiment of the present application, the client can access the server by combining requests and responses, and the server does not need to actively address the client, which is equivalent to that the client can only access the server when there is a need, so as to achieve Federated learning across devices.
  • the first server may screen the multiple clients, and select at least one client to participate in federated learning, For example, whether to allow the client to participate in federated learning is determined according to the client's connection status or device status, so as to screen out the stable clients for federated learning and improve the overall efficiency of federated learning.
  • the number of second servers included in the foregoing multiple servers is multiple, and the first server obtains the second aggregate information sent by the second server, which may include: the first server Receive second aggregation information sent by multiple second servers respectively; the aforementioned first server updates the global model stored on the first server based on the first aggregation information and the second aggregation information, which may include: the first server The global model stored on the first server is updated based on the first aggregation information and the second aggregation information respectively sent by the plurality of second servers to obtain an updated global model.
  • the servers in the federated learning system can transmit data to each other, so that each server has the full amount of data, so that the client can obtain the full amount of data when it accesses the server at any time.
  • the accuracy of the data obtained by the client is improved, and the client obtains the full amount of data during the federated learning process, which can improve the overall efficiency of the federated learning.
  • the above-mentioned method may further include: transmitting the first aggregated information between the aforementioned multiple servers and sending the first aggregate information to each second server, so that each second server can The first aggregated information and the second aggregated information are aggregated to obtain third aggregated information.
  • the servers in the federated learning system can transmit aggregated information to each other, so that each server has more data, thereby maintaining data consistency in the federated learning system.
  • one of the multiple servers is used as the master node.
  • the master node is used to manage the multiple servers, and the first server is used as the master node.
  • Send a first trigger indication to the second server where the first trigger indication is used to instruct the second server to send the second aggregation information to the first server, so that the first server receives the second aggregation information sent by the second server.
  • the federated learning process in the federated learning system can be managed by the master node, and data is synchronized between each server under the trigger of the master node, thereby realizing data consistency in the federated learning system.
  • the master node includes a counter, and the counter is used to count the clients that send the request message, and the above-mentioned first server obtains the second aggregate in the second server under the trigger of the master node.
  • the information may include: when the value of the counter meets the first threshold, the first server sends a second trigger indication to the second server, where the second trigger indication is used to instruct the second server to perform the next round of iteration.
  • the received request messages can be counted by a counter, and when the count reaches a certain value, data synchronization between each server can be triggered, which is equivalent to limiting the customers participating in each round of federated learning iteration. To avoid the long tail effect caused by excessive client participation, each server can synchronize data in a timely manner.
  • the first threshold includes a preset value, or, the first threshold is a value determined by the number of clients accessing all servers in the federated learning system in the previous iteration process. Therefore, in the embodiment of the present application, the first threshold corresponding to the counter may be a preset value, or a value determined according to the number of clients accessing all servers in the federated learning system. Usually, the first threshold is not greater than the access The number of clients in the federated learning system, so as to avoid the long waiting time for federated learning due to waiting for client access, and improve the overall efficiency of federated learning.
  • the master node is further provided with a timer
  • the above method may further include: when the timer exceeds the second threshold, that is, when the timer times out, the first server sends the first server to each server.
  • Two trigger indications where the second trigger indication is used to trigger each server to enter the next round of iteration. Therefore, in the embodiment of the present application, the time window of each iteration can be set by a timer, and when the timer times out, the next round of iterative learning can be entered, so as to avoid the long tail effect caused by the long waiting time, even if If the client quits during federated learning, it does not affect the overall training efficiency of federated learning.
  • the second threshold is a preset value, a value determined according to the number of clients accessing the federated learning system, or an amount of data communicated between all servers and clients in the federated learning system. A sized value.
  • the time window can be determined according to the number of clients or the traffic volume, so that the time window corresponding to the timer matches the actual scene, and the time window of each iteration is set more reasonably.
  • the above method may further include: the first server receives a query message sent by a third client, where the third client is any one of multiple clients; the first server responds to the query message Send the updated global model information to the third client.
  • the client can send a query message to the server to query the latest global model. Due to the data synchronization between each server during each round of iterative learning, each server has more data than its own aggregated information, and the client can obtain more data no matter which server it accesses. data, so that the client can obtain a more accurate global model.
  • the present application provides a federated learning method, which is applied to a federated learning system.
  • the federated learning system includes multiple servers, one of the multiple servers is used as the master node, and the multiple servers are used for iteration.
  • Learning to achieve federated learning; the process of any round of iterations in iterative learning includes:
  • the master node After any one of the multiple servers receives the first request message, the master node starts the counter and the timer.
  • the counter is used to count the request messages received by the multiple servers in one iteration.
  • the request message It is used to request to obtain the global model stored in the corresponding server among the multiple servers; if the value of the counter reaches the first threshold, the master node sends a first trigger indication to each of the multiple servers, and the first trigger
  • the instruction is used to instruct multiple servers to transmit locally stored information to each other; if the value of the counter does not reach the first threshold and the value of the timer reaches the second threshold, the master node sends a second trigger instruction to each server, The second trigger indication is used to instruct each server to perform the next round of iteration.
  • one of the servers in the federated learning system is used as the master node, and the master node maintains the timer and the counter.
  • the counter counts the received request messages, and at the same time Use a timer for timing, so as to limit the number of clients participating in each iteration and the training duration of each iteration, so as to avoid meaningless waiting caused by clients falling behind, improve the overall efficiency of federated learning, and avoid long tails effect.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration. Therefore, in the embodiment of the present application, the first threshold corresponding to the counter may be a preset value, or a value determined according to the number of clients accessing all servers in the federated learning system. Usually, the first threshold is not greater than the access The number of clients in the federated learning system, so as to avoid the long waiting time for federated learning due to waiting for client access, and improve the overall efficiency of federated learning.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the federated learning in the previous iteration
  • the amount of data communicated between each server in the system and the corresponding client is related to the amount of data. Therefore, in the embodiment of the present application, the time window of each iteration can be set by a timer, and when the timer times out, the next round of iterative learning can be entered, so as to avoid the long tail effect caused by the long waiting time, even if If the client quits during federated learning, it does not affect the overall training efficiency of federated learning.
  • the present application provides a federated learning system, including multiple servers and multiple clients, the multiple servers include a first server and a second server, and both the first server and the second server are The information of the global model is stored, and multiple servers are used for iterative learning to realize federated learning. During any round of iteration in the iterative learning:
  • the first server is configured to receive the request messages respectively sent by at least one first client;
  • the first server is configured to send the information of the global model and the training configuration parameters to the at least one first client according to the request messages respectively sent by the at least one first client, and the information of the global model and the training configuration parameters are used to indicate the at least one first client.
  • the client uses the training configuration parameters to train the global model;
  • the first server is configured to receive first model update parameters fed back by at least one first client respectively, where the first model update parameters are parameters of the global model obtained after the at least one first client is trained;
  • the first server is configured to aggregate first model update parameters fed back by at least one first client respectively, to obtain first aggregated information
  • the second server is configured to receive the second model update parameters sent by the corresponding at least one second client, and aggregate the second model update parameters sent by the at least one second client to obtain second aggregation information;
  • the first server is configured to receive the second aggregation information sent by each second server;
  • the first server is configured to update the global model stored on the first server based on the first aggregation information and the second aggregation information sent by each second server to obtain an updated global model.
  • the client can actively send a request to the server to request to participate in the federated learning, and feed back the parameters of the trained global model to the server, and the server can
  • the received model update parameters are aggregated to obtain aggregated information, and the server also receives aggregated information sent by other servers, so that the server has more information than its own aggregated information. So that the client and the server do not need to maintain a long connection, the client can access the server at any time to obtain more data, which can improve the accuracy of the data obtained by the client, so that it can be based on more data.
  • Each round of iterative training improves the overall efficiency of federated learning.
  • the number of second servers included in the multiple servers is multiple; each of the multiple second servers is configured to receive at least one corresponding second client The sent second model update parameters, and aggregate the second model update parameters sent by at least one second client to obtain second aggregated information; the first server is specifically configured to receive the data sent by multiple second servers second aggregation information; the first server is specifically configured to update the global model stored on the first server based on the first aggregation information and the second aggregation information respectively sent by multiple second servers to obtain an updated global model .
  • the servers in the federated learning system can transmit data to each other, so that each server has the full amount of data, so that the client can obtain the full amount of data when it accesses the server at any time.
  • the accuracy of the data obtained by the client is improved, and the client obtains the full amount of data during the federated learning process, which can improve the overall efficiency of the federated learning.
  • the multiple servers further include a third server serving as a master node, and the master node is used to manage the multiple servers; the master node is used to send the first server to the multiple servers respectively. triggering indication; the second server is specifically configured to send the second aggregation information to the first server based on the first triggering indication.
  • the master node can trigger the mutual transmission of aggregated information between the servers, so that each server can have more comprehensive full data, so that the client can access the server at any time. Both can obtain the full amount of data, improve the accuracy of the data obtained by the client, and in the process of federated learning, the client obtains the full amount of data, which can improve the overall efficiency of federated learning.
  • the master node includes a counter, and the counter is used to count the number of request messages received by multiple servers, and the request messages are used to request to obtain the storage in the corresponding server among the multiple servers.
  • the global model; the master node is specifically used to send a first trigger instruction to each of the multiple servers when the value of the counter meets the first threshold, and the first trigger instruction is used to trigger each second server to send the first trigger instruction to the first A server sends the second aggregated information.
  • the received request messages can be counted by a counter, and when the count reaches a certain value, data synchronization between each server can be triggered, which is equivalent to limiting the customers participating in each round of federated learning iteration. To avoid the long tail effect caused by excessive client participation, each server can synchronize data in a timely manner.
  • the master node further includes a timer, and the timer starts timing from the first request message received in each round of iteration; the master node is further configured to, when the timer exceeds the second threshold, A second trigger indication is respectively sent to each of the multiple servers, where the second trigger indication is used to instruct the multiple servers to perform the next round of iteration.
  • the time window of each iteration can be set by a timer, and when the timer times out, the next round of iterative learning can be entered, so as to avoid the long tail effect caused by the long waiting time, even if If the client quits during federated learning, it does not affect the overall training efficiency of federated learning.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing each server in the federated learning system during the previous iteration, or the second threshold is related to The amount of data communicated between multiple servers and multiple clients in the last iteration is related to the size of the data. Therefore, in the embodiment of the present application, the time window can be determined according to the number of clients or the traffic volume, so that the time window corresponding to the timer matches the actual scene, and the time window of each iteration is set more reasonably.
  • the first server receives a query message sent by a third client, and the third client includes any client that accesses the federated learning system; the first server sends the query message to the third client
  • the client sends the updated global model information. Therefore, in the embodiment of the present application, the client can send a query message to the server to query the latest global model. Due to the data synchronization between each server during each round of iterative learning, each server has more data than its own aggregated information, and the client can obtain more data no matter which server it accesses. data, so that the client can obtain a more accurate global model.
  • the first server is further configured to send the first aggregation information to the second server; the second server is specifically configured to update the locally stored global information in combination with the first aggregation information and the second aggregation information model to get the updated global model. Therefore, in the embodiment of the present application, the servers in the federated learning system can transmit aggregated information to each other, so that each server has more data, thereby maintaining data consistency in the federated learning system.
  • the present application provides a server, which is applied to a federated learning system.
  • the federated learning system includes multiple servers and multiple clients.
  • the multiple servers are used for iterative federated learning.
  • Any server includes:
  • a transceiver module configured to receive a request message sent by at least one first client, where the request message is used to request the global model stored in the first server, and the multiple clients include at least one first client;
  • the transceiver module is further configured to send the information of the global model and the training configuration parameters to the at least one first client, where the information of the global model and the training configuration parameters are used to instruct the at least one first client to use the training configuration parameters to train the global model;
  • the transceiver module is further configured to receive first model update parameters fed back by at least one first client respectively, where the first model update parameters are parameters of the global model obtained after the at least one first client is trained;
  • an aggregation module configured to aggregate the first model update parameters fed back by at least one first client to obtain the first aggregation information in this round of iteration;
  • the transceiver module is further configured to obtain second aggregation information sent by the second server, where the second aggregation information is information obtained by the second server in the current iteration by aggregating the received second model update parameters;
  • the updating module is configured to update the global model stored on the first server based on the first aggregated information and the second aggregated information to obtain an updated global model.
  • the transceiver module is also used to receive the second aggregation information sent by a plurality of second servers respectively;
  • the updating module is specifically configured to update the global model stored on the first server based on the first aggregation information and the second aggregation information respectively sent by the plurality of second servers, so as to obtain an updated global model.
  • the first server is a master node in the federated learning system, and the master node is used to manage multiple servers,
  • the transceiver module is further configured to send a first trigger instruction to the second server, where the first trigger instruction is used to instruct the second server to send the second aggregation information to the first server;
  • the transceiver module is further configured to receive the second aggregation information in the second server.
  • a counter is set in the first server, and the counter is used to count the number of request messages received by multiple servers,
  • the transceiver module is specifically configured to send the first trigger indication to the second server when the value of the counter meets the first threshold.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration.
  • the master node includes a timer, and the timer starts timing after receiving the first request message in each round of iteration,
  • the transceiver module is specifically configured to receive a second trigger instruction sent by the master node when the timer exceeds the second threshold, where the second trigger instruction is used to instruct the next round of iterative learning.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the previous iteration
  • the amount of data communicated between each server and the corresponding client in the federated learning system is related to the amount of data.
  • the transceiver module is further configured to receive a query message sent by a third client, where the third client includes any one of the clients accessing the federated learning system;
  • the transceiver module is further configured to send the updated global model information to the third client for the query message.
  • the transceiver module is further configured to send the first aggregated information to the second server, so that the second server updates the locally stored global model based on the first aggregated information and the second aggregated information to Get the updated global model.
  • the present application provides a server, which is applied to a federated learning system.
  • the federated learning system includes multiple servers, one of the multiple servers serves as a master node, and the multiple servers are used for iterative learning.
  • the master node includes:
  • the startup module is used to start the counter and timer after any one of the multiple servers receives the first request message, and the counter is used for the request messages received by the multiple servers in one iteration. Counting, the request message is used to request to obtain the global model stored in the corresponding server in multiple servers;
  • a transceiver module used for if the value of the counter reaches the first threshold, the master node sends a first trigger instruction to each of the multiple servers, and the first trigger instruction is used to instruct the multiple servers to transmit local storage to each other Information;
  • the transceiver module is further configured to, if the value of the counter does not reach the first threshold and the value of the timer reaches the second threshold, the master node sends a second trigger indication to each server, and the second trigger indication is used to instruct each server proceed to the next iteration.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the federated learning in the previous iteration
  • the amount of data communicated between each server in the system and the corresponding client is related to the amount of data.
  • an embodiment of the present application provides a federated learning device, including: a processor and a memory, wherein the processor and the memory are interconnected through a line, and the processor invokes program codes in the memory to execute the first aspect or the second
  • the processing-related functions shown in any of the aspects are used in federated learning methods.
  • the federated learning device may be a chip.
  • an embodiment of the present application provides a federated learning device.
  • the federated learning device may also be referred to as a digital processing chip or a chip.
  • the chip includes a processing unit and a communication interface.
  • the processing unit obtains program instructions through the communication interface, and the program instructions are
  • the processing unit executes, and the processing unit is configured to perform processing-related functions as in any optional implementation manner of the first aspect or the second aspect.
  • an embodiment of the present application provides a computer-readable storage medium, including instructions, which, when executed on a computer, cause the computer to execute the method in any optional implementation manner of the first aspect or the second aspect. .
  • an embodiment of the present application provides a computer program product including instructions, which, when run on a computer, enables the computer to execute the method in any optional implementation manner of the first aspect or the second aspect.
  • Fig. 1 is a schematic diagram of a main frame of artificial intelligence applied by the application
  • FIG. 2 is a schematic diagram of the architecture of a federated learning system provided by an embodiment of the present application
  • FIG. 3 is a schematic structural diagram of a server according to an embodiment of the present application.
  • FIG. 4 is a schematic flowchart of a federated learning method provided by an embodiment of the present application.
  • FIG. 5 is a schematic diagram of the architecture of another federated learning system provided by an embodiment of the present application.
  • FIG. 6 is a schematic flowchart of another federated learning method provided by an embodiment of the present application.
  • FIG. 7 is a schematic flowchart of another federated learning method provided by an embodiment of the present application.
  • FIG. 8 is a schematic diagram of a polymerization mode provided in an embodiment of the present application.
  • FIG. 9 is a schematic structural diagram of a server according to an embodiment of the present application.
  • FIG. 10 is a schematic structural diagram of another server provided by an embodiment of the present application.
  • FIG. 11 is a schematic structural diagram of another server provided by an embodiment of the present application.
  • Figure 1 shows a schematic structural diagram of the main frame of artificial intelligence.
  • the above-mentioned artificial intelligence theme framework is explained in two dimensions (vertical axis).
  • the "intelligent information chain” reflects a series of processes from data acquisition to processing. For example, it can be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, data has gone through the process of "data-information-knowledge-wisdom".
  • the "IT value chain” reflects the value brought by artificial intelligence to the information technology industry from the underlying infrastructure of human intelligence, information (providing and processing technology implementation) to the industrial ecological process of the system.
  • the infrastructure provides computing power support for artificial intelligence systems, realizes communication with the outside world, and supports through the basic platform. Communicate with the outside through sensors; computing power is provided by intelligent chips, such as central processing unit (CPU), network processor (neural-network processing unit, NPU), graphics processor (English: graphics processing unit, GPU), Application specific integrated circuit (ASIC) or field programmable gate array (field programmable gate array, FPGA) and other hardware acceleration chips) are provided; the basic platform includes distributed computing framework and network related platform guarantee and support, It can include cloud storage and computing, interconnection networks, etc. For example, sensors communicate with external parties to obtain data, and these data are provided to the intelligent chips in the distributed computing system provided by the basic platform for calculation.
  • intelligent chips such as central processing unit (CPU), network processor (neural-network processing unit, NPU), graphics processor (English: graphics processing unit, GPU), Application specific integrated circuit (ASIC) or field programmable gate array (field programmable gate array, FPGA) and other hardware acceleration chips
  • CPU central processing unit
  • the data on the upper layer of the infrastructure is used to represent the data sources in the field of artificial intelligence.
  • the data involves graphics, images, voice, and text, as well as IoT data from traditional devices, including business data from existing systems and sensory data such as force, displacement, liquid level, temperature, and humidity.
  • Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making, etc.
  • machine learning and deep learning can perform symbolic and formalized intelligent information modeling, extraction, preprocessing, training, etc. on data.
  • Reasoning refers to the process of simulating human's intelligent reasoning method in a computer or intelligent system, using formalized information to carry out machine thinking and solving problems according to the reasoning control strategy, and the typical function is search and matching.
  • Decision-making refers to the process of making decisions after intelligent information is reasoned, usually providing functions such as classification, sorting, and prediction.
  • some general capabilities can be formed based on the results of data processing, such as algorithms or a general system, such as translation, text analysis, computer vision processing, speech recognition, image identification, etc.
  • Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of the overall solution of artificial intelligence, and the productization of intelligent information decision-making to achieve landing applications. Its application areas mainly include: intelligent terminals, intelligent transportation, Smart healthcare, autonomous driving, safe city, etc.
  • the embodiments of the present application can be applied to the field of federated learning, and the neural network can be trained collaboratively through the client and the server, thus involving a large number of related applications of neural networks, such as the neural network trained by the client during federated learning.
  • relevant terms and concepts of neural networks that may be involved in the embodiments of the present application are first introduced below.
  • a neural network can be composed of neural units, and a neural unit can refer to an operation unit that takes xs and intercept 1 as inputs, and the output of the operation unit can be shown in formula (1-1):
  • s 1, 2,...n, n is a natural number greater than 1
  • Ws is the weight of xs
  • b is the bias of the neural unit.
  • f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into an output signal.
  • the output signal of the activation function can be used as the input of the next convolutional layer, and the activation function can be a sigmoid function.
  • a neural network is a network formed by connecting a plurality of the above single neural units together, that is, the output of one neural unit can be the input of another neural unit.
  • the input of each neural unit can be connected with the local receptive field of the previous layer to extract the features of the local receptive field, and the local receptive field can be an area composed of several neural units.
  • a deep neural network also known as a multi-layer neural network, can be understood as a neural network with multiple intermediate layers.
  • the DNN is divided according to the position of different layers.
  • the neural network inside the DNN can be divided into three categories: input layer, intermediate layer, and output layer. Generally speaking, the first layer is the input layer, the last layer is the output layer, and the middle layers are all intermediate layers, or hidden layers.
  • the layers are fully connected, that is, any neuron in the i-th layer must be connected to any neuron in the i+1-th layer.
  • each layer can be expressed as a linear relational expression: in, is the input vector, is the output vector, is the offset vector or the bias parameter, w is the weight matrix (also called the coefficient), and ⁇ () is the activation function.
  • in is the input vector
  • the output vector is the offset vector or the bias parameter
  • w is the weight matrix (also called the coefficient)
  • ⁇ () is the activation function.
  • Each layer is just an input vector After such a simple operation to get the output vector Due to the large number of DNN layers, the coefficient W and offset vector The number is also higher.
  • the DNN Take the coefficient w as an example: Suppose that in a three-layer DNN, the linear coefficient from the 4th neuron in the second layer to the 2nd neuron in the third layer is defined as The superscript 3 represents the number of layers where the coefficient W is located, and the subscript corresponds to the output third layer index 2 and the input second layer index 4.
  • the coefficient from the kth neuron in the L-1 layer to the jth neuron in the Lth layer is defined as
  • the input layer does not have a W parameter.
  • more intermediate layers allow the network to better capture the complexities of the real world.
  • a model with more parameters is more complex and has a larger "capacity", which means that it can complete more complex learning tasks.
  • Training the deep neural network is the process of learning the weight matrix, and its ultimate goal is to obtain the weight matrix of all layers of the trained deep neural network (the weight matrix formed by the vectors W of many layers).
  • Convolutional neural network is a deep neural network with a convolutional structure.
  • the convolutional neural network contains a feature extractor composed of convolutional layers and subsampling layers, which can be regarded as a filter.
  • the convolutional layer refers to the neuron layer in the convolutional neural network that convolves the input signal.
  • a convolutional layer of a convolutional neural network a neuron can only be connected to some of its neighbors.
  • a convolutional layer usually contains several feature planes, and each feature plane can be composed of some neural units arranged in a rectangle. Neural units in the same feature plane share weights, and the shared weights here are convolution kernels. Shared weights can be understood as the way to extract image information is independent of location.
  • the convolution kernel can be initialized in the form of a matrix of random size, and the convolution kernel can obtain reasonable weights by learning during the training process of the convolutional neural network.
  • the immediate benefit of sharing weights is to reduce the connections between the layers of the convolutional neural network, while reducing the risk of overfitting.
  • the network for extracting features mentioned below in this application may include one or more layers of convolutional layers.
  • the network for extracting features may be implemented by using CNN.
  • the loss function can generally include loss functions such as mean square error, cross entropy, logarithm, and exponential.
  • mean squared error can be used as a loss function, defined as Specifically, a specific loss function can be selected according to the actual application scenario.
  • the neural network can use the error back propagation (BP) algorithm to correct the size of the parameters in the initial neural network model during the training process, so that the reconstruction error loss of the neural network model becomes smaller and smaller.
  • BP error back propagation
  • the input signal is passed forward until the output will generate error loss, and the parameters in the initial neural network model are updated by back-propagating the error loss information, so that the error loss converges.
  • the back-propagation algorithm is a back-propagation movement dominated by error loss, aiming to obtain the parameters of the optimal neural network model, such as the weight matrix.
  • the global model when the client performs model training, can be trained by using the loss function or the BP algorithm, and the trained global model can be obtained.
  • a distributed machine learning algorithm through multiple clients (clients or FL-Clients), such as mobile devices or edge servers, and the server cooperate to complete model training and algorithm update on the premise that the data is not out of the domain, and get trained
  • clients clients or FL-Clients
  • the server cooperate to complete model training and algorithm update on the premise that the data is not out of the domain, and get trained
  • the latter global model For example, the client uses the locally stored training samples to train the global model and feeds back the model parameters obtained after training to the server.
  • the server aggregates the received model parameters and updates the locally stored global model based on the aggregated information. model to get the latest global model.
  • the dataset is divided according to the user dimension, and the part of the data with the same user characteristics but different users is removed for training.
  • the client trains the model, then uploads it to the server, and the server aggregates it to obtain the final model.
  • the commonly used distributed training is the Parameter Server (PS) mode, which includes the roles of Worker (the computing node for training in the parameter server training mode) and PS (parameter server).
  • the training data is distributed to each Worker by the computing center.
  • Worker and PS communicate with PS through the network, exchange information such as models or gradients, and the PS performs aggregation and model optimization, so as to efficiently train to obtain the final model.
  • PS Parameter Server
  • each server is actually only responsible for part of the assigned parameters (multiple servers jointly maintain a global shared parameter), and each computing node is only assigned part of the data and processing tasks.
  • each server can only handle the calculation of some parameters, and requires the participants to maintain a stable and long connection with the central server.
  • Most FL-Clients are stateless, unreliable, and unaddressable, and the central server needs to aggregate all parameters and complete the aggregation calculation.
  • the PS mode training mode obviously cannot handle the challenges of cross-device terminals.
  • TFF TensorFlow Federated
  • PaddleFL provides a data-parallel-based distributed federated learning framework, which includes three roles: FL-Server, FL-Trainer, and FL-Scheduler, but FL-Client and FL-Server use the remote procedure call system (remote procedure call). calls, RPC) for cross-process communication, there is a stable and long connection between the two, and FL-Scheduler can schedule Workers to decide which Workers can participate in training before each update cycle. This is not possible in cross-device deployment scenarios.
  • RPC remote procedure call
  • this application provides a federated learning architecture, which enables clients to access servers through a looser access method, and maintains data consistency between servers, so that clients can obtain full data when they access the server at any time. , to improve the model training efficiency of the client.
  • the federated learning system provided by this application may be as shown in FIG. 2 , the system (or also referred to as a cluster) may include multiple servers, the multiple servers establish connections with each other, and each server may communicate with a Or multiple client communication.
  • the server can usually be deployed in the server, and in some scenarios, the server can also be deployed in the terminal device, which can be adjusted according to the actual application scenario, which is not limited in this application.
  • the client can be deployed in terminals, such as mobile phones, wristbands, smart TVs, etc., or can be deployed in servers, which can be adjusted according to actual application scenarios, which is not limited in this application.
  • One of the multiple servers can be used as a master node, and the master node can maintain transaction consistency in the cluster, so that each server can work together.
  • the framework may also include a scheduler (not shown in the figure) or an elastic load balance (ELB) (not shown in the figure), etc., which will not be repeated here.
  • the scheduler can be deployed on any server, and can be used to schedule communication resources in the cluster, such as time windows and channels used for transmission efficiency.
  • ELB can be deployed on any server, and can be used to route the communication between the server and the client. The ELB determines the server that the client accesses according to the load of each server. For example, the ELB can send a request message The client is connected to the server with less current load, so that each server can maintain a suitable load and avoid the situation that one or more servers are overloaded.
  • each server can deliver the model to be trained to the client that has established a connection with it, and the client can use the locally stored training samples to train the model, and use the trained model to The data is fed back to the server.
  • the server can aggregate the received data of the one or more models to obtain the aggregated data. Data, equivalent to the aggregated model.
  • the final model can be output to complete federated learning.
  • the structure of the server in the federated learning system provided by this application can refer to FIG. 3, and the server may include a training engine, a storage module, a time-limited communication module, a device management module, an encryption module, a protocol parsing module or a communication module. Wait.
  • the communication module is used for data communication between FL-Server and FL-Client, as well as dynamic networking of FL-Server cluster to achieve elastic scaling of central server resources.
  • the hypertext transfer protocol (HTTP) protocol can be used for transmission to achieve a stable short-term connection.
  • HTTP hypertext transfer protocol
  • cross-process communication can be performed through the remote procedure call system (remote procedure calls, RPC) protocol to achieve stable long-term connection.
  • Device management module responsible for device-related services such as federated learning device selection, for example, after receiving the client's request message, screening whether the client participates in federated learning.
  • Encryption and decryption module Encrypt the uploaded and downloaded model weights to reduce the possibility of obtaining original samples through model reverse attack.
  • Protocol parsing module serialize the sent communication message or deserialize the received message, etc.
  • Training Engine A module that performs federated learning calculations on data uploaded by FL-Client, including aggregators and optimizers.
  • the aggregator can be used to aggregate data, such as summing and weighted summing of data uploaded by FL-Client.
  • the optimizer can be used to calculate the gradient of the model, and optimize the model based on the calculated gradient to obtain a model with a more accurate output.
  • Storage module stores metadata required for federated learning and data such as the globally optimal model obtained from training.
  • Time-limited communication device includes a timer and a counter, and performs subsequent processing when the timer times out or the counter value reaches a threshold.
  • FIG. 4 the following is a schematic flowchart of the federated learning method provided by this application in combination with the aforementioned architecture.
  • the first server receives at least one request message sent by at least one client.
  • the first server can be any one of the multiple first servers shown in FIG. 2, and the at least one client can be a client that has established a connection with the first server, such as the at least one
  • the request message sent by the client can be routed to the first server by the ELB.
  • the client can request the first server to participate in federated learning through the request message, so as to request to use local data for model training, and to influence the global model through the local data, so as to obtain a model that carries the personalized characteristics of the client.
  • a request message can be sent to the federated learning system, and the ELB in the system will route the request message to the adapted first server to pass the request message.
  • the first server and the client cooperate to implement federated learning.
  • the client can collect personalized training data that is different from other devices. After collecting the training data, the client can actively send a request message to the federated learning system to request the use of locally stored data for federated learning. Thereby, a model adapted to the training data is obtained, and the model is synchronized to the system to complete federated learning, so that the global model in the cluster can adaptively learn various data collected by each client, and improve the performance of the global model. output precision.
  • the first server sends the global model and training configuration parameters to at least one client respectively according to the request message.
  • the first server After the first server receives the request message sent by the at least one client, if it is confirmed that the at least one client is allowed to participate in federated learning, it can send the information of the locally stored global model and the training configuration parameters to the at least one client respectively.
  • the global model is the model to be trained
  • the information of the global model may include structural parameters of the global model (such as the depth or width of the global model, etc.) or weight parameters, etc.
  • the training configuration parameters are used when training the global model. parameters, such as the learning rate, the number of epochs, or the category in the security algorithm.
  • the first server may screen the multiple clients, and select at least one Participate in federated learning.
  • the request message can carry information of the terminal, such as network connection status, power, load and other information.
  • the first server can filter out the clients with stable connection according to the terminal information to perform federated learning, and send the selected at least one
  • the client issues the model and training configuration parameters, so that the client can train the model based on the training configuration parameters.
  • the client uses the training configuration parameters to train the global model.
  • the client After the client (or the first client) receives the global model and the training configuration parameters delivered by the first server, it can use the training set to train the global model to obtain the trained global model.
  • the training set may include multiple training samples, and the training samples may be data collected by the client or data received by the client, which is not limited here.
  • the client may be an application program deployed on the user's terminal.
  • the client may collect data generated during the user's enabling of the client, and use the data as a training sample to request participation from the server.
  • Federated learning so that the collected data is applied to the model update, so that the global model in the federated learning system can be adapted to each client.
  • the client feeds back model update parameters to the first server.
  • the information of the trained model that is, the model update parameters, can be fed back to the first server.
  • the model update parameters may include weight parameters of the model.
  • the model update parameters received by server j can be expressed as: n k represents the dataset size of the kth client, That is, the weight parameters of the model obtained after client training.
  • the first server aggregates the model update parameters fed back by at least one client to obtain first aggregated information.
  • the first server can aggregate the model update parameters fed back by the at least one client to obtain first aggregated information. It is equivalent to aggregating the model fed back by the at least one client to obtain the information of the aggregated model.
  • the specific manner may include various, such as averaging, weighted fusion, etc.
  • the present application does not limit the aggregation manner, which may be adjusted according to actual application scenarios.
  • the first server obtains the second aggregation information in the second server.
  • the second server includes servers other than the first server in the federated learning system, and the number of the second server may be one or more.
  • the first server can also receive second aggregation information obtained by other servers after aggregating the received model update parameters.
  • the aggregated information in each server can maintain consistency.
  • each server can obtain the aggregated information of other servers, so that each server can maintain more data, and each client can obtain more comprehensive data when accessing the server. There is no need to maintain a long connection state with the server. Therefore, the federated learning system provided in this application can adapt to the scenario where the client is deployed on the mobile terminal.
  • the first server can receive aggregated information sent by all servers in the federated learning system except itself, that is, each server in the federated learning system can transmit aggregated information to each other, so that each The server can have full data.
  • each server in the federated learning system can transmit aggregated information to each other, so that each The server can have full data.
  • the client accesses the server at any time, it can obtain the full amount of data, thereby improving the accuracy of the data obtained by the client.
  • all servers in the federated learning system transmit aggregated information to each other as an example for illustrative description.
  • data may also be transmitted between some servers, or Some servers only receive aggregated information without sending aggregated information, etc., which can be adjusted according to actual application scenarios, which is not limited in this application.
  • the manner in which the second server or the second aggregated information is obtained is similar to the manner in which the first server acquires the first aggregated information.
  • the global model and training configuration parameters are delivered to the second client, where the global model and training configuration parameters are generally the same as the global model and training configuration parameters delivered by the first server to the first client.
  • the second client uses the locally stored training data to train the global model based on the training configuration parameters, and feeds back the parameters of the trained global model (such as weight parameters, model width or depth, etc.) to the second server.
  • the second server aggregates the received model update parameters to obtain the second aggregated information.
  • the specific process of acquiring the second aggregation information by the second server will not be repeated.
  • the master node can trigger the mutual transmission of aggregated information between various services in the cluster to achieve data synchronization in the cluster, so that all servers in the cluster can have more comprehensive data.
  • the master node may be one of the servers in the federated learning system, such as the aforementioned first server or second server.
  • At least one of a counter or a timer may be set in the master node to monitor the number of clients participating in each round of iterations or a time window, so as to effectively control the number of clients participating in federated learning.
  • the number of clients, and the duration of each iteration process, avoid ineffective waiting and reduce the effect of "falling tail".
  • the counter is used to count the request messages sent by the client received in the federated learning system, and after each round of iteration is completed, the master node can initialize the counter.
  • the master node can start the timer, and count each time a request message is received.
  • the first server can be triggered to send the model and training configuration parameters to the client.
  • the master node sends a first trigger instruction to each server to instruct each server to aggregate its own information. It is transmitted to other servers, such as instructing the second server to send the second aggregation information to the first server.
  • the master node can trigger itself to send the first aggregation information to other servers (such as all second servers), and also send a first trigger indication to the second servers to indicate each The second server sends its own second aggregation information to other servers.
  • ELB can send a notification message to the master node to notify the master node of the received master node, such as the number of request messages, reception timestamp, etc., so that the master node can Counts the number of request messages.
  • the master node sends a first trigger instruction to each server, which triggers each server (including the master node itself) in the federated learning system to transmit data to each other, thereby completing data synchronization between servers , so that each server has more comprehensive data.
  • the client accesses the server, it can obtain more comprehensive data, so that federated learning or applications can be performed based on the more comprehensive data, so as to improve the user experience of the client. user experience.
  • the first threshold includes a preset value, or is related to the number of clients accessing multiple servers in the current iteration or the previous iteration, and the like. Therefore, in the embodiment of the present application, the first threshold corresponding to the counter may be a preset value, or a value determined according to the number of clients accessing the server, and usually the first threshold is not greater than the number of clients accessing the server , so as to avoid the long waiting time for federated learning due to waiting for client access, and improve the overall efficiency of federated learning.
  • the timer can be used to set a time window in the process of federated learning, that is, only one round of federated learning is performed within the time window.
  • the master node can trigger the server to perform the next round of learning , if the model update parameter is obtained during the current learning process, the model update parameter can be discarded, and of course, the model update parameter can be retained and specified to improve the utilization of data. Therefore, in the embodiment of the present application, a timer is used to monitor the time window of each round of iterative learning, and each round of iterative learning can be performed within the time window of the timer, so that some clients may be disconnected or the training efficiency may be reduced. When it is too low, the influence of these clients on the federated learning process is reduced, and the efficiency of federated learning is improved.
  • the master node can start the timer and complete the current iteration within the time window of the timer.
  • the master node can send a second trigger instruction to each server in the federated learning system (including the master node itself), triggering each server to enter the next round of iteration.
  • the master node determines that the timer has expired, it can directly determine to enter the next round of iteration, and it does not need to send a second trigger instruction for itself. Of course, it can also generate a second trigger instruction for itself, which is not made in this application. limited.
  • the second threshold may be a preset value, or it may be related to the number of clients accessing the federated learning system in the current or previous round, or it may be related to the current or previous round of clients and the The amount of data communicated between servers is related to the size.
  • the second threshold may be positively correlated with a value determined by the number of clients accessing the federated learning system or according to the amount of data communicated between the client and the server. The greater the number of clients in the learning system, the greater the second threshold. The greater the amount of data communicated between the client and the server in the previous iteration, the greater the second threshold, etc. Adjust according to the actual application scenario. Therefore, in the embodiments of the present application, a time window that matches the actual application scenario can be set for the timer, which can prevent the time window from being too long or too short, thereby improving the overall efficiency of federated learning.
  • the master node starts the timer and the counter. If the value of the counter reaches the first threshold and the timer has not timed out, the timer can be turned off, and the next round of iteration can be entered after the aggregation of all the model update parameters fed back by the client in the current iteration process is completed. If the value of the counter does not reach the first threshold and the timer times out, it can be determined to end the current iteration and enter the next iteration, and the data in the current iteration can be discarded or retained. If the value of the counter does not reach the first threshold and the timer has not timed out, the iteration can continue.
  • the timer and the counter can work together to limit the number of clients participating in federated learning in this round of iteration, and limit the duration of this round of iteration, so as to avoid waiting for clients to access Or the long tail effect caused by the client quitting halfway.
  • the first server aggregates the first aggregation information and the second aggregation information to obtain third aggregation information.
  • the first server After the first server obtains the first aggregation information and the second aggregation information sent by other servers, it can update the locally stored global model based on the received first aggregation information and the second aggregation information, that is, steps 407 and 408 .
  • the first server can not only aggregate the model update parameters fed back by one or more clients received by itself, but also receive aggregated information obtained by other servers, and aggregate the first aggregated information calculated by itself and receive it.
  • the obtained second aggregated information is aggregated again to obtain third aggregated information with more complete data.
  • each server not only aggregates the model update parameters received by itself, but also re-aggregates the information aggregated by other servers in the federated learning system, so that each server can have
  • the data included in each server in the federation system makes the data of each server consistent, and the client can obtain the full amount of data when accessing the server at any time.
  • n k the dataset size of the kth client.
  • Each server first calculates its own and Then, the master node triggers data aggregation between servers to obtain the aggregated calculated data. and n j , each server can obtain the calculation result of the cluster dimension, that is, the full amount of data. In the next iteration, the client requests any server to get the correct full amount of data.
  • the first server can also send the first aggregation information to the second server, so that the second server can re-aggregate the second aggregation information obtained by itself and the received first aggregation information to obtain the third aggregation information.
  • Aggregate information In order to maintain the heterogeneity of data in each server, the second server can use the same aggregation method as the first server when performing aggregation, so as to finally obtain the same third aggregation information, so that each server can use the same aggregation method as the first server. can have full data in it.
  • the first server updates the stored global model.
  • the third aggregation information may include the weight parameters of the aggregated model, and the first server may use the weight parameters obtained after the final aggregation to update the locally stored global model, thereby obtaining the updated global model. Model.
  • the client can access the server by sending a request to the server, and can participate in federated learning after the server allows it, which is equivalent to the client accessing the server in a loose way, while the There is no need to maintain a long connection with the server, and it can adapt to some scenarios where clients are deployed on mobile terminals.
  • each server in addition to aggregating the model update parameters fed back by the connected clients, each server can also aggregate the information obtained by other servers after aggregation, so that each server can have the full amount of data in the federated learning system. Accessing the server at any time can obtain the full amount of data, so that the output accuracy of the final trained model is higher.
  • the present application can also count the clients accessing the server by setting a timer and/or a counter, and perform the timing during the federated learning process, and perform the federated learning within the time window of the timer. If an iterative learning process is not completed beyond the time window, the next round of learning can be directly performed, thereby avoiding the existence of some clients that affect the overall training process due to unstable connections or low training efficiency, and improving the overall learning efficiency of federated learning.
  • the client can exit at any time during the federated training process without affecting the efficiency of federated learning and avoiding binding to the client.
  • steps 401-408 may be one of the iterative processes in the federated learning process, and the federated learning process may include one or more iterations, and this application only takes one iteration as an example for illustration , which is not a limitation. Specifically, the number of iterations can be determined according to the actual application scenario.
  • the client sends a query message to the first server.
  • the client when the client (or the second client) needs to acquire the latest model, it may send a query message to the server to request to query the latest global model.
  • the client can access the federated learning system through the address of the federated learning system, and the ELB can route the query message sent by the client, so as to route the request message to the adapted server.
  • the client Sending a query message to the first server is only an exemplary illustration, and the first server here can be replaced with other servers, such as a second server or a third server, etc., which is not limited here.
  • the first server delivers the updated global model to the client.
  • the first server After receiving the query message from the client, the first server can deliver the latest local model to the client.
  • the first server updates the stored global model using the latest aggregated information, and after receiving the query message, can deliver the latest global model to the client.
  • the structure and parameters of the model can be delivered to the client, such as the number of layers of the neural network, the size of the convolution kernel, the resolution of input and output, the weight parameters in each network layer, etc., so that the client can obtain the latest model structure.
  • step 409 and step 410 in this application are optional steps.
  • each server can transmit the latest aggregated information to each other, so that each server in the federated learning system has the full amount of data, and the client can obtain the full amount of data when it accesses the server at any time. Without the need to maintain a long connection with the server, the full amount of data can be obtained, and the accuracy of the model that can be obtained by the client can be improved.
  • the federated learning system provided by this application may be shown in Figure 5, where three servers (ie, FL-Server0, FL-Server1, and FL-Server2) are used for exemplary illustration, and can also be replaced with more Server.
  • the scheduler ie, FL-Scheduler
  • the scheduler can be used to schedule communication resources or storage resources in the server cluster, so that data can be exchanged between servers.
  • the transmission control protocol (TCP) connection can be established between the servers or between the server and the scheduler, and maintain a stable long-term connection, and communicate through the TCP private protocol.
  • TCP transmission control protocol
  • IPX Internet packet exchange protocol
  • SPX sequenced Packet Exchange
  • other protocols such as Internet packet exchange protocol (internet packet exchange, IPX), sequenced packet exchange (sequenced Packet Exchange, SPX) and other protocols, which are not limited in this application.
  • the client can access the federated learning system through an IP address, and the server exposes the hypertext transfer protocol (HTTP) port to the outside to provide distributed services to the outside.
  • HTTP hypertext transfer protocol
  • the client can establish an HTTP connection with the server, that is, a short connection.
  • Server 0 can be used as the master node, other servers as slave nodes, and the master node is responsible for data synchronization within the cluster, such as triggering all nodes to aggregate, globally time or count during the federated learning process, so as to keep the data in the cluster consistent. sex.
  • the server may not be able to obtain the status of the client at any time.
  • the connection between the mobile terminal and the server may be unstable, and the mobile terminal cannot be retrieved or the connection status of the mobile terminal cannot be obtained. etc., which may lead to the interruption of communication between the server and the mobile terminal, which may affect the efficiency of federated learning.
  • FIG. 6 a schematic flowchart of another federated learning method provided by this application.
  • the timers and counters can be initialized, eg set to 0, before each iteration.
  • the counter on the master node can count.
  • the timer can be started synchronously, and a round of federated learning can be performed within the time window of the timer.
  • the first threshold may be determined according to the number of clients participating in the federated learning in the previous round. For example, if the number of clients participating in the federated learning in the previous round is less than the number of clients in the previous iteration, the set first threshold may be Decrease on the basis of a threshold value. If the number of clients participating in federated learning in the last round of federated learning is more than that in the previous iteration, it can be increased on the basis of the originally set first threshold value, so that the threshold value corresponding to the counter is increased. Adapt to clients participating in training in each iteration to maximize training efficiency.
  • FL-server 1 can deliver the global model to be trained and training configuration parameters to FL-client.
  • the global model may include various neural networks, such as DNN, CNN, Residual Network (ResNet) or a constructed network, etc., which can be adjusted according to actual application scenarios.
  • the training configuration parameters can include parameters involved in training the model, such as the learning rate, the number of epochs, or the category in the security algorithm, which are used to determine the learning rate, the number of iterations in training, or the encryption algorithm when transmitting data when the client performs model training. type, etc.
  • the model update parameters can include parameters related to the trained model, such as update The structure of the global model (such as the depth or width of the network layer, etc.), the weight parameters, or the weight parameters after reducing the sparsity, etc.
  • each FL-client can use locally stored training samples to train the received global model.
  • each training sample includes the sample and the corresponding ground truth label.
  • the sample is used as the input of the model, and then the loss function is used to measure the difference between the output of the model and the ground truth label, and through the difference Calculate the update gradient of the global model, and use the gradient to reversely update the weight parameters of the global model to obtain the updated model.
  • FL-server 1 can aggregate the model update parameters fed back by the multiple FL-clients, which is equivalent to data reduction.
  • the specific aggregation method may include summation, average value, weighted summation, etc., which may be adjusted according to actual application scenarios.
  • the data obtained by FL-server 1 aggregating the model with the new data can be expressed as: where S is the client set, represents the model update parameters fed back by each client, and n k represents the size of the model update parameters fed back by each client.
  • the master node that is, FL-server 0 triggers the server where it is located and the slave nodes (such as FL-server 1, FL-server 2, etc.) to transmit to each other between each server.
  • the slave nodes such as FL-server 1, FL-server 2, etc.
  • the aggregation information triggered by FL-server 0 from the node is used as an example for illustration.
  • FL-server 2 will send the aggregation information obtained by the aggregation of the received model update parameters to FL-server 1, and FL-server 1 will aggregate the received aggregation information and the locally stored aggregation. The information is aggregated to obtain the final aggregated information.
  • the current iteration can be terminated and the next iteration is entered, that is, the counter is initialized.
  • FL-server 0 sends a second trigger instruction to FL-server 1 and FL-server 2 to indicate entering the next round of iteration.
  • the data during the current iteration can be discarded or enabled. For example, if the data is complete, it can continue to be used, and only part of the unresponsive client data is discarded, which can be adjusted according to the actual application scenario.
  • the master node can maintain the counter and timer, count the clients that have sent the request, and when the timer reaches the threshold, the aggregation operation of the cluster dimension can be triggered, that is, each server can receive The aggregated information of other servers is aggregated again, so that each server can have the aggregated information of other servers, and each server can have the full amount of data.
  • the master node also maintains a timer.
  • the next round of iteration can be performed, which is equivalent to only one round of iterative learning within the time window of the timer to prevent the waiting time for the client's response from being too long. It affects the efficiency of federated learning and solves the long tail effect caused by clients being left behind.
  • the master node as an important node for monitoring the duration of each iteration and the number of clients participating in the iteration, can mainly be used to perform the following steps:
  • the master node After any one of the multiple servers receives the first request message, the master node starts the counter and the timer.
  • the counter is used to count the request messages received by the multiple servers in one iteration.
  • the request message It is used to request to obtain the global model stored in the corresponding server among the multiple servers; if the value of the counter reaches the first threshold, the master node sends a first trigger indication to each of the multiple servers, and the first trigger
  • the instruction is used to instruct multiple servers to transmit locally stored information to each other; if the value of the counter does not reach the first threshold and the value of the timer reaches the second threshold, the master node sends a second trigger instruction to each server, The second trigger indication is used to instruct each server to perform the next round of iteration.
  • one of the servers in the federated learning system is used as the master node, and the master node maintains the timer and the counter.
  • the counter counts the received request messages, and at the same time Use a timer for timing, so as to limit the number of clients participating in each iteration and the training duration of each iteration, so as to avoid meaningless waiting caused by clients falling behind, improve the overall efficiency of federated learning, and avoid long tails effect.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration. Therefore, in the embodiment of the present application, the first threshold corresponding to the counter may be a preset value, or a value determined according to the number of clients accessing all servers in the federated learning system. Usually, the first threshold is not greater than the access The number of clients in the federated learning system, so as to avoid the long waiting time for federated learning due to waiting for client access, and improve the overall efficiency of federated learning.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the federated learning in the previous iteration
  • the amount of data communicated between each server in the system and the corresponding client is related to the amount of data. Therefore, in the embodiment of the present application, the time window of each iteration can be set by a timer, and when the timer times out, the next round of iterative learning can be entered, so as to avoid the long tail effect caused by the long waiting time, even if If the client quits during federated learning, it does not affect the overall training efficiency of federated learning.
  • Counters and timers may be maintained by the master node, such as the FL-Server 0 shown in Figure 5 above.
  • each iteration initialize the counter and set the number of request messages from clients participating in this iteration. For example, if it is set to C, usually this value can be set to a subset of the number of clients that have communication connections with the server. That is, it usually does not exceed the number of clients that have established connections with the server, so as to avoid excessively long waiting times for clients in each round of iterations. If in the federated learning system, 100 clients have established connections with the server, but only 10 clients need to participate in federated learning, if C is set too large at this time, it may lead to too long waiting for clients to participate. , reducing the efficiency of federated learning.
  • the FL-Server receives a request message from the FL-Client.
  • the request message sent by the FL-Client can be routed by the ELB to the FL-Server, where the FL-Server can be any server in the federated learning system, such as FL-Server0, FL-Server1 or FL-Server in the aforementioned Figure 5 Any of the servers such as Server2.
  • the FL-Server can confirm whether the FL-Client is allowed to participate in federated learning. For example, random screening, first-come-first-served or other screening methods can be used to determine whether FL-Client is allowed to participate in federated learning.
  • the connection status of the FL-Client such as delay, communication duration, etc.
  • the status of the FL-Client such as low battery or high load, etc., the FL-Client may not be allowed to participate in federated learning. , to avoid FL-Client affecting the efficiency of federated learning.
  • the counter is accumulated.
  • the master node may also record the number of FL-Clients that have sent the request message, and accumulate the counter.
  • step 704. Determine whether the counter is 1, if yes, go to step 708, if not, go to step 705.
  • the value of the counter can be monitored. If the value of the counter is 1, that is, in the current iteration process, the request message sent by the first client is received, and the timing can be started at this time. , that is, step 709 is executed. At this time, only because the request message sent by only one client is received, so there is no need to perform data synchronization. After receiving the request messages sent by multiple clients, the data can be processed in the cluster dimension. Synchronize, so that the servers in the cluster maintain data consistency, that is, step 705 is executed.
  • the FL-Server processes the request and synchronizes data in the cluster dimension.
  • the model and training configuration parameters can be delivered to all or part of the multiple FL-Clients.
  • FL-Client uses locally stored training samples to train the model, and feeds back the relevant data of the trained model to FL-Server.
  • FL-Server can receive model update parameters fed back by multiple FL-Clients, and FL-Client can aggregate the model update parameters fed back by multiple FL-Clients to obtain Aggregated information (ie, first aggregated information).
  • Aggregated information ie, first aggregated information.
  • the aggregated information calculated by each FL-Server can be synchronized to realize the data synchronization of the cluster dimension.
  • each FL-Server After each FL-Server has synchronized its own aggregation information, the aggregation information obtained by its own aggregation (ie the first aggregation information) and the aggregation information sent by other FL-Servers (ie the second aggregation information) can be aggregated to obtain The final global aggregated information (that is, the third aggregated information), so that each FL-Server can have full data.
  • the manner of aggregation between FL-Servers can be specifically shown in Figure 8, wherein the federated learning system includes multiple servers, and the multiple servers are connected to each other in a ring, that is, as shown in Figure 8.
  • the shown Server 0, Server 1, Server 2, Server 3, etc. are only 4 servers as an exemplary description here, and can be replaced by more or less servers in actual application scenarios, which are not described here. limited.
  • the client includes K clients connected to the server, such as client 1 to client K shown in Figure 8, each client receives the initial weight wt sent by the connected server, and then uses the locally stored training
  • the sample trains the model corresponding to the initial weight to obtain the weight of the trained model
  • the corresponding data volume n 1 -n k , n k represents the data set size of the kth client, and then feeds the weight of the trained model and the amount of data to the connected server, for example, client 1 sends the trained model to the The weight is fed back to Server 0, and client K feeds back the weight of the model obtained after training to Server 2, etc.
  • the server j aggregates the received weights sent by the client to obtain Then the master node triggers each server to aggregate the data transmitted between the servers to obtain aggregated information.
  • the aggregated data of each server is expressed as and n j .
  • step 707 Whether the count reaches the first threshold, if yes, go to step 710, if not, go to step 702.
  • the value of the counter can be monitored in real time. If the count of the counter reaches the first threshold, it means that the number of clients participating in this round of federated learning iterative process is sufficient, and there is no need to increase the number of updates. If more clients participate in the current round of federated learning, that is, stop timing, and enter the iterative process of the next round of federated learning, that is, step 710 is executed.
  • the timer can be triggered to start timing.
  • the ELB also notifies the master node that the first request message has been received, and the master node can start the timer and The counter is used to count the request messages received by the federated learning system and time the process of this round of iteration.
  • step 710 Check whether it times out, and if so, perform step 710.
  • the state of the timer can be monitored in real time to determine whether the timer has timed out.
  • the request message or data sent by the client received within the time window T of the timer can be processed normally by the server, such as performing encryption and decryption or aggregating the received data. If C has not been reached, both the timer and the counter are reset to the initial state and enter the next round of iteration; if the counter reaches C within the time window T, the timing of the timer will be stopped, and the current iteration will be processed. After the data sent by the client, enter the next round of iteration.
  • the longest time spent in each iteration is the duration of the time window T, and during the iteration process, the communication duration of each iteration can be recorded, Thereby, the duration of the time window is dynamically adjusted, thereby eliminating the long tail effect of the client and improving the training efficiency.
  • the function of the timer may also be to trigger data synchronization between servers in the cluster, that is, the aforementioned steps 705-706, after completing steps 705-706, the next round of iteration can be entered. It can be adjusted according to actual application scenarios.
  • step 701 After the timer stops timing, it means that this round of iterative learning is ended, and after the timer and the counter are initialized, the next round of iteration can be entered, that is, step 701 is executed.
  • step 702, step 705 and step 706 can be executed by any FL-Server in the cluster, and step 701, step 703, step 704, step 707-step 710 are executed by the master node.
  • the client can access the server in a combination of request and response, and monitor the clients participating in each iteration and the iteration duration by setting counters and timers, even if the client is in The midway exit of federated learning does not affect the overall learning process of federated learning of the cluster, avoids the problem of tail drop caused by not being able to address clients, and improves the overall efficiency of federated learning.
  • data synchronization can be performed between each server, that is, each server has the full amount of data in the cluster, and the client can access the full amount of data at any time by accessing any server, thereby improving the model obtained by the client. accuracy.
  • the federated learning system can be composed of multiple servers, and the client is deployed on the user's mobile terminals, such as mobile phones, cameras or tablets.
  • the mobile terminal can collect the image through the camera, and the user can manually mark the type of the object in the photographed image in the album, such as "cat" and "dog” in the image, and the mobile terminal can use the manually marked image as the image.
  • Training samples When the number of training samples reaches a certain amount, the mobile terminal can request the federated learning system to participate in federated learning, so as to apply the data collected by itself to the classification network.
  • the mobile terminal can obtain the HTTP port exposed by the federated learning system, and send a request message to the cluster through the HTTP port.
  • the counter set on the master node counts the request message, and the master node also maintains a timer.
  • the upper limit of the counter can be preset, or determined according to the number of clients participating in the previous iteration, or determined according to the number of clients that have established connections with the cluster.
  • the upper limit of the timer may be preset, or may be determined according to the duration of the previous iteration.
  • the timer is enabled. If the request message sent by the mobile terminal is the last request message in this iteration, that is, the value of the counter reaches the preset first threshold, the mobile terminal is the last client to participate in federated learning determined in this iteration , after the mobile terminal completes the model training, it can enter the next round of iterative learning.
  • the ELB within the cluster routes the request message of the mobile terminal to the adapted server, which determines whether the mobile terminal is allowed to participate in federated learning. If the server allows the mobile terminal to participate in federated learning, the server sends the locally stored classification network to the mobile terminal, and also sends training configuration parameters, such as the learning rate used during training, and the data encryption and decryption methods. The mobile terminal uses the locally collected training samples to train the classification network according to the training configuration parameters, and obtains the trained classification network. The mobile terminal feeds back the weight parameter of the classification network after training to the server. If the time when the server receives the weight parameter is within the time window of the timer, the server aggregates the received weight parameter to obtain the first aggregation. information.
  • the server sends the locally stored classification network to the mobile terminal, and also sends training configuration parameters, such as the learning rate used during training, and the data encryption and decryption methods.
  • the mobile terminal uses the locally collected training samples to train the classification network according to the training configuration parameters, and obtains the trained
  • the aggregated information can be transmitted between servers, so that each server has full data.
  • the mobile terminal can send a query message to the server at any time to query the latest classified network, and the server can send the stored latest classified network to the mobile terminal, so that the mobile terminal can access any server at any time. to the latest full data.
  • the client when the client is deployed on the mobile terminal, the client can access the federated learning system and participate in the federated learning in a loose manner by combining requests and responses, without maintaining a stable long connection. , which is applicable to scenarios where clients are deployed on mobile terminals.
  • timers and counters are also deployed in the federated learning system. Through the cooperation of timers and counters, the low efficiency of federated learning caused by waiting for clients is avoided, the long tail effect is reduced, and the overall efficiency of federated learning is improved.
  • the upper limit of the timer or counter can be adjusted according to the information in the previous iteration or previous rounds of iterative training, and the upper limit of the timer or counter can be dynamically adjusted according to the actual scene, further improving the performance of federated learning. efficiency.
  • the present application provides a schematic structural diagram of a server.
  • the server is applied to a federated learning system.
  • the federated learning system includes multiple servers and multiple clients. Multiple servers are used for iterative federated learning. Any one of the servers includes:
  • a transceiver module 901 configured to receive a request message sent by at least one first client, where the request message is used to request the global model stored in the first server, and the multiple clients include at least one first client;
  • the transceiver module 901 is further configured to send global model information and training configuration parameters to at least one first client, where the global model information and training configuration parameters are used to instruct at least one first client to use the training configuration parameters to train the global model ;
  • the transceiver module 901 is further configured to receive first model update parameters fed back by at least one first client respectively, where the first model update parameters are parameters of the global model obtained after the at least one first client is trained;
  • an aggregation module 902 configured to aggregate the first model update parameters fed back by at least one first client to obtain the first aggregation information in this round of iteration;
  • the transceiver module 901 is further configured to acquire second aggregation information sent by the second server, where the second aggregation information is information obtained by the second server in the current iteration by aggregating the received second model update parameters;
  • the updating module 903 is configured to update the global model stored on the first server based on the first aggregation information and the second aggregation information to obtain an updated global model.
  • the transceiver module is further configured to receive second aggregated information respectively sent by multiple second servers;
  • the updating module 903 is specifically configured to update the global model stored on the first server based on the first aggregation information and the second aggregation information respectively sent by the plurality of second servers to obtain an updated global model.
  • the first server is a master node in the federated learning system, and the master node is used to manage multiple servers,
  • the transceiver module is further configured to send a first trigger instruction to the second server, where the first trigger instruction is used to instruct the second server to send the second aggregation information to the first server;
  • the transceiver module 901 is further configured to receive the second aggregation information in the second server.
  • a counter is set in the first server, and the counter is used to count the number of request messages received by multiple servers,
  • the transceiver module 901 is specifically configured to send a first trigger instruction to the second server when the value of the counter meets the first threshold.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration.
  • the master node includes a timer, and the timer starts timing after receiving the first request message in each round of iteration,
  • the transceiver module 901 is specifically configured to receive a second trigger instruction sent by the master node when the timer exceeds the second threshold, where the second trigger instruction is used to instruct the next round of iterative learning.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the previous iteration
  • the amount of data communicated between each server and the corresponding client in the federated learning system is related to the amount of data.
  • the transceiver module is further configured to receive a query message sent by a third client, where the third client includes any one of the clients accessing the federated learning system;
  • the transceiver module 901 is further configured to send the updated global model information to the third client for the query message.
  • the transceiver module is further configured to send the first aggregated information to the second server, so that the second server updates the locally stored global model based on the first aggregated information and the second aggregated information to Get the updated global model.
  • the present application provides a schematic structural diagram of another server, that is, the aforementioned master node.
  • the server is applied to the federated learning system.
  • the federated learning system includes multiple servers. One of the multiple servers serves as the master node.
  • the multiple servers are used for iterative learning to realize the federated learning.
  • the master node includes:
  • the starting module 1001 is used to start a counter and a timer after any one of the multiple servers receives the first request message, and the counter is used for the requests received by the multiple servers in one iteration.
  • the message is counted, and the request message is used to request to obtain the global model stored in the corresponding server in multiple servers;
  • the transceiver module 1002 is configured to, if the value of the counter reaches the first threshold, the master node sends a first trigger instruction to each of the multiple servers, and the first trigger instruction is used to instruct the multiple servers to transmit local data to each other. stored information;
  • the transceiver module 1002 is further configured to, if the value of the counter does not reach the first threshold and the value of the timer reaches the second threshold, the master node sends a second trigger indication to each server, and the second trigger indication is used to indicate each service end for the next iteration.
  • the first threshold is a preset value, or the first threshold is related to the number of clients accessing the federated learning system during the previous iteration.
  • the second threshold is a preset value, or the second threshold is related to the number of clients accessing the federated learning system in the previous iteration, or the second threshold is related to the federated learning in the previous iteration
  • the amount of data communicated between each server in the system and the corresponding client is related to the amount of data.
  • FIG. 11 is a schematic structural diagram of a server provided by an embodiment of the present application.
  • the server described in the embodiment corresponding to FIG. 9 may be deployed on the server 1100 to implement the functions of the server in the embodiment corresponding to FIG. 4 to FIG. 8 .
  • the server 1100 may vary greatly due to different configurations or performances, and may include one or more central processing units (CPU1122, for example, one or more processors) and memory 1132, and one or more storage applications 1142 or storage medium 1130 for data 1144 (eg, one or more mass storage devices).
  • CPU1122 central processing units
  • storage applications 1142 or storage medium 1130 for data 1144 eg, one or more mass storage devices.
  • the memory 1132 and the storage medium 1130 may be short-term storage or persistent storage.
  • the memory 1132 is a random access memory RAM, which can directly exchange data with the central processing unit 1122 for loading the data 1144 and the application program 1142 and/or the operating system 1141 for the central processing unit 1122 to run and use directly, Typically used as a temporary data storage medium for the operating system or other running programs.
  • the program stored in the storage medium 1130 may include one or more modules (not shown in FIG. 11 ), and each module may include a series of instruction operations on the server.
  • the central processing unit 1122 may be configured to communicate with the storage medium 1130 to execute a series of instruction operations in the storage medium 1130 on the server 1100 .
  • the storage medium 1130 stores program instructions and data corresponding to the method steps shown in any of the foregoing embodiments in FIG. 4 to FIG. 8 .
  • the server 1100 may also include one or more power supplies 1126, one or more wired or wireless network interfaces 1150, one or more input and output interfaces 1158, and/or, one or more operating systems 1141, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM and many more.
  • operating systems 1141 such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM and many more.
  • the central processing unit 1122 is configured to execute the method steps performed by the server shown in any of the foregoing embodiments in FIG. 4 to FIG. 8 .
  • the embodiments of the present application also provide a federated learning device, which may also be referred to as a digital processing chip or a chip.
  • the chip includes a processing unit and a communication interface.
  • the processing unit can obtain program instructions through the communication interface, and the program instructions are processed by the processing unit.
  • the processing unit is configured to execute the method steps executed by the server shown in any of the foregoing embodiments in FIG. 4 to FIG. 8 .
  • the embodiments of the present application also provide a digital processing chip.
  • the digital processing chip integrates circuits and one or more interfaces for realizing the above-mentioned central processing unit 1122 or the functions of the central processing unit 1122 .
  • the digital processing chip can perform the method steps of any one or more of the foregoing embodiments.
  • the digital processing chip does not integrate the memory, it can be connected with the external memory through the communication interface.
  • the digital processing chip implements the actions performed by the server in the above embodiment according to the program codes stored in the external memory.
  • the federated learning device when the federated learning device provided by the embodiment of the present application is a chip, the chip specifically includes: a processing unit and a communication unit, the processing unit may be, for example, a processor, and the communication unit may be, for example, an input/output interface , pins or circuits, etc.
  • the processing unit can execute the computer-executed instructions stored in the storage unit, so that the chip in the server can execute the federated learning method described in the embodiments shown in FIG. 4 to FIG. 8 .
  • the aforementioned storage unit may be a storage unit within a chip, such as a register, a cache, etc., and the storage unit may also be a storage unit located outside the chip in the wireless access device, such as a read-only storage unit.
  • Memory read-only memory, ROM
  • static storage devices that can store static information and instructions, random access memory RAM, etc.
  • the aforementioned processing unit or processor may be a central processing unit, a network processing unit (neural-network processing unit, NPU), a graphics processing unit (graphics processing unit, GPU), a digital signal processor (digital signal processor, DSP) ), application specific integrated circuit (ASIC) or field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • NPU network processing unit
  • graphics processing unit graphics processing unit
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • a general purpose processor may be a microprocessor or it may be any conventional processor or the like.
  • the processor mentioned in any one of the above may be a general-purpose central processing unit, a microprocessor, an ASIC, or one or more integrated circuits for controlling the execution of the programs of the above-mentioned methods in FIGS. 4-8 .
  • Embodiments of the present application further provide a computer-readable storage medium, where a program is stored in the computer-readable storage medium, and when the computer-readable storage medium runs on a computer, the computer executes the operations described in the foregoing embodiments shown in FIGS. 4 to 8 . steps in the method.
  • Embodiments of the present application also provide a computer program product, which, when running on a computer, causes the computer to execute the steps executed by the server in the methods described in the foregoing embodiments shown in FIGS. 4-8 .
  • the device embodiments described above are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be A physical unit, which can be located in one place or distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.
  • U disk U disk
  • mobile hard disk read-only memory ROM
  • random access memory RAM magnetic disk or optical disk, etc.
  • a computer device which can be a personal computer, server, or network device, etc.
  • the computer program product includes one or more computer instructions.
  • the computer may be a general purpose computer, special purpose computer, computer network, or other programmable device.
  • the computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be from a website site, computer, server or data.
  • the center transmits to another website site, computer, server or data center by wire (eg coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg infrared, wireless, microwave, etc.).
  • wire eg coaxial cable, optical fiber, digital subscriber line (DSL)
  • wireless eg infrared, wireless, microwave, etc.
  • the computer-readable storage medium may be any available medium that can be stored by a computer, or a data storage device such as a server, a data center, etc. that includes one or more available media integrated.
  • the usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVDs), or semiconductor media (eg, solid state disks (SSDs)), and the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Information Transfer Between Computers (AREA)

Abstract

本申请提供一种联邦学习方法以及装置，用于在跨设备的联邦学习过程中，对各个服务器之间的聚合信息进行同步，从而使每轮迭代学习过程中服务器的数据保持同步，使每个服务器上都具有更全面的数据，提高整体的学习效率。该方法包括：第一服务端接收至少一个第一客户端发送的请求消息；第一服务端向至少一个第一客户端发送全局模型和训练配置参数；第一服务端接收至少一个第一客户端分别反馈的第一模型更新参数；第一服务端对第一模型更新参数进行聚合以得到在本轮迭代中的第一聚合信息；第一服务端获取第二服务端发送的第二聚合信息；第一服务端基于第一聚合信息和第二聚合信息，更新第一服务端上存储的全局模型。

Description

一种联邦学习方法以及装置
本申请要求于2021年04月25日提交中国专利局、申请号为“202110450585.7”、申请名称为“一种联邦学习方法以及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及联邦学习领域，尤其涉及一种联邦学习方法以及装置。
背景技术
在用户隐私日益觉醒的今天，各方对于数据所有权的关注越来越多，各国对于数据安全管理也日趋严格，要解决数据孤岛的困境，仅仅靠传统的机器学习训练模式已经出现瓶颈。为了帮助公司充分利用用户的数据信息而不侵犯隐私，实现多方联合建模，以帮助企业应用提效降本，助力不同行业智慧升级，诞生了联邦学习(federated learning，FL)。联邦学习是一种人工智能(artificial intelligence，AI)联合建模的组织方式，在不直接获取设备私有数据的基础上，通过协调跨设备的模型训练和参数聚合完成算法模型任务，同时降低传统集中式机器学习的通信计算开销，具有重要理论创新和实践落地价值。
然而，在由于跨设备终端部署场景的特殊性，跨设备(cross-device)联邦学习还带来了新的挑战，如联邦学习的客户端设备异构，不同设备在存储、计算、通信和电池等硬件规格方面存在很大差异，对联邦学习性能产生影响。因此，如何提高跨设备的联邦学习的学习效率，成为亟待解决的问题。
发明内容
本申请提供一种联邦学习方法，用于在跨设备的联邦学习过程中，对各个服务端之间的聚合信息进行同步，从而使每轮迭代学习过程中服务端的数据保持同步，客户端在任何时刻接入服务端都可以获取到全量数据，提高整体的学习效率。
有鉴于此，第一方面，本申请提供一种联邦学习方法，应用于联邦学习系统，联邦学习系统中包括多个服务端，多个服务端之间互相连接，多个服务端用于进行迭代联邦学习，其中任意一轮迭代学习的过程包括：首先，第一服务端接收至少一个第一客户端发送的请求消息，请求消息用于请求第一服务端中存储的全局模型，前述的多个客户端包括该至少一个第一客户端；第一服务端向至少一个第一客户端发送全局模型的信息和训练配置参数，全局模型的信息和训练配置参数用于指示至少一个第一客户端使用训练配置参数对全局模型进行训练；第一服务端接收至少一个第一客户端分别反馈的第一模型更新参数，第一模型更新参数是至少一个第一客户端训练后得到的全局模型的参数；第一服务端对至少一个第一客户端反馈的第一模型更新参数进行聚合，以得到在本轮迭代中的第一聚合信息；第一服务端获取第二服务端发送的第二聚合信息，第二聚合信息为第二服务端在本轮迭代中，对接收到的第二模型更新参数进行聚合得到的信息；第一服务端基于第一聚合信息和第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
本申请实施方式中，在进行联邦学习的每轮迭代过程中，需要由客户端主动向服务端 发送请求消息以请求参与联邦学习，并将训练全局模型后得到的模型更新参数发送至服务端，因此客户端通过向服务端发送请求消息的方式接入服务端，客户端无需和服务端保持长期稳定的连接，以实现跨设备的联邦学习系统。多个服务端除了对自身接收到的模型更新参数进行聚合，还接收其他服务端发送的模型更新参数并进行聚合，从而使每个服务端中都具有比自身的聚合信息更多的数据，客户端在接入服务端时，即可获取到更多的数据，从而可以基于更多的数据来进行每轮迭代训练，提高联邦学习的整体效率。可以理解为，本申请中，客户端通过请求与响应结合的方式接入服务器，客户端和服务端之间无需保持长期稳定连接，客户端在任何时刻访问任意服务端都可以获取到训练模型所需的更多的数据，提高客户端侧的联邦学习的学习效率。
在一种可能的实施方式中，在第一服务端接收至少一个第一客户端分别反馈的第一模型更新参数之前，第一服务端可以接收多个客户端发送的请求消息，以向服务端请求参与联邦学习，第一服务端在接收到多个客户端发送的请求消息之后，即可将本地存储的全局模型的信息和训练配置参数下发给该多个客户端。因此，本申请实施例可以通过请求和响应结合的方式使客户端接入服务端，服务端无需主动对客户端进行寻址，相当于客户端可以仅在有需求时接入服务端，从而实现跨设备的联邦学习。
在一种可能的实施方式中，若第一服务端接收到多个客户端发送的请求消息，第一服务端可以对该多个客户端进行筛选，筛选出至少一个客户端来参与联邦学习，如根据客户端的连接状态或者设备状态等确定是否允许客户端参与联邦学习，从而筛选出状态稳定的客户端来进行联邦学习，提高联邦学习的整体效率。
在一种可能的实施方式中，前述的多个服务端中包括的第二服务端的数量为多个，第一服务端获取第二服务端发送的第二聚合信息，可以包括：第一服务端接收多个第二服务端分别发送的第二聚合信息；前述的第一服务端基于第一聚合信息和第二聚合信息，更新第一服务端上存储的全局模型，可以包括：第一服务端基于第一聚合信息和多个第二服务端分别发送的第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
本申请实施方式中，联邦学习系统内的服务端之间可以互相传输数据，使每个服务端上都具有全量数据，从而使客户端在任意时刻接入服务端，都可以获取到全量数据，提高客户端获取到的数据的准确度，且在联邦学习的过程中客户端获取到全量数据，可以提高联邦学习的整体效率。
在一种可能的实施方式中，上述方法还可以包括：前述的多个服务端之间传输第一服务端向每个第二服务端发送第一聚合信息，以使每个第二服务端对第一聚合信息和第二聚合信息进行聚合得到第三聚合信息。
因此，本申请实施方式中，联邦学习系统内的服务端之间可以互相传输聚合信息，从而使每个服务端上都具有更多数据，从而保持联邦学习系统内的数据一致性。
在一种可能的实施方式中，多个服务端中的其中一个服务端作为主节点，以第一服务端作为主节点为例，主节点用于对多个服务端进行管理，第一服务端向第二服务器发送第一触发指示，第一触发指示用于指示第二服务端向第一服务端发送第二聚合信息，以使第 一服务端接收第二服务端发送的第二聚合信息。
因此，本申请实施方式中，联邦学习系统中的联邦学习过程可以由主节点来进行管理，在主节点的触发下各个服务端之间同步数据，从而实现联邦学习系统内的数据一致性。
在一种可能的实施方式中，主节点中包括计数器，计数器用于对发送请求消息的客户端进行计数，上述的第一服务端在主节点的触发下获取第二服务端中的第二聚合信息，可以包括：当计数器的数值满足第一阈值时，则第一服务端向第二服务器发送的第二触发指示，第二触发指示用于指示第二服务器进行下一轮迭代。
本申请实施方式中，可以通过计数器来对接收到的请求消息进行计数，当计数达到一定数值时，即可触发各个服务端之间同步数据，相当于限制每轮联邦学习迭代过程中参与的客户端的数量，避免过多客户端参与而导致的长尾效应，使各个服务端可以及时地进行数据同步。
在一种可能的实施方式中，第一阈值包括预设值，或者，该第一阈值与上一轮迭代过程中接入联邦学习系统内所有服务端的客户端的数量确定的值。因此，本申请实施方式中，计数器对应的第一阈值可以是预先设定的值，也可以根据接入联邦学习系统内所有服务端的客户端的数量确定的值，通常该第一阈值不大于接入联邦学习系统的客户端的数量，从而避免等待客户端的接入而导致联邦学习的等待时间过长，提高联邦学习的整体效率。
在一种可能的实施方式中，主节点中还设置有计时器，上述方法还可以包括：当计时器超过第二阈值时，即计时器超时时，第一服务端向每个服务端发送第二触发指示，该第二触发指示用于触发各个服务端进入下一轮迭代。因此，本申请实施方式中，可以通过计时器来设置每轮迭代的时间窗口，当计时器超时时即可进入下一轮迭代学习，从而避免因等待时长过长而导致的长尾效应，即使客户端在进行联邦学习时中途退出也不影响联邦学习的整体训练效率。
在一种可能的实施方式中，第二阈值为预设值、根据接入联邦学习系统的客户端的数量确定的值或者联邦学习系统内的所有服务端和客户端之间进行通信的数据量的大小确定的值。
因此，本申请实施方式中，可以根据客户端的数量或者通信量来确定时间窗口，使计时器对应的时间窗口与实际场景匹配，更合理地设置每轮迭代的时间窗口。
在一种可能的实施方式中，上述方法还可以包括：第一服务端接收第三客户端发送的查询消息，第三客户端是多个客户端中的任意一个；第一服务端针对查询消息向第三客户端发送更新后的全局模型的信息。
因此，本申请实施方式中，客户端可以向服务端发送查询消息以查询最新的全局模型。因每轮迭代学习过程中各个服务端之间进行了数据同步，因此每个服务端上都具有比自身的聚合信息更多的数据，客户端无论接入哪个服务端都可以获取到更多的数据，使客户端获取到更准确的全局模型。
第二方面，本申请提供一种联邦学习方法，应用于联邦学习系统，联邦学习系统包括多个服务端，多个服务端中的其中一个服务端作为主节点，多个服务端用于进行迭代学习以实现联邦学习；其中迭代学习中的任意一轮迭代的过程包括：
在多个服务端中任意一个服务端接收到第一个请求消息之后，主节点启动计数器和计时器，计数器用于对多个服务端在一轮迭代中接收到的请求消息进行计数，请求消息用于请求获取多个服务端中对应的服务端中存储的全局模型；若计数器的值达到第一阈值，主节点向多个服务端中的每个服务端发送第一触发指示，第一触发指示用于指示多个服务端之间互相传输本地存储的信息；若计数器的值未达到第一阈值，计时器的值达到第二阈值，则主节点向每个服务端发送第二触发指示，第二触发指示用于指示每个服务端进行下一轮迭代。
本申请实施方式中，将联邦学习系统中的其中一个服务端作为主节点，由主节点来维护计时器和计数器，在每轮迭代过程中，通过计数器对接收到的请求消息进行计数，并同时使用计时器来进行计时，从而限制每轮迭代参与的客户端的数量以及每轮迭代的训练时长，从而避免因客户端掉队而导致的无意义的等待，提高联邦学习的整体效率，避免出现长尾效应。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。因此，本申请实施方式中，计数器对应的第一阈值可以是预先设定的值，也可以根据接入联邦学习系统内所有服务端的客户端的数量确定的值，通常该第一阈值不大于接入联邦学习系统的客户端的数量，从而避免等待客户端的接入而导致联邦学习的等待时间过长，提高联邦学习的整体效率。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关或者第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。因此，本申请实施方式中，可以通过计时器来设置每轮迭代的时间窗口，当计时器超时时即可进入下一轮迭代学习，从而避免因等待时长过长而导致的长尾效应，即使客户端在进行联邦学习时中途退出也不影响联邦学习的整体训练效率。
第三方面，本申请提供一种联邦学习系统，包括多个服务端和多个客户端，多个服务端包括第一服务端和第二服务端，第一服务端和第二服务端上均存储了全局模型的信息，多个服务端用于进行迭代学习以实现联邦学习，其中迭代学习中的任意一轮迭代过程中：
第一服务端用于接收至少一个第一客户端分别发送的请求消息；
第一服务端用于针对至少一个第一客户端分别发送的请求消息向至少一个第一客户端发送全局模型的信息和训练配置参数，全局模型的信息和训练配置参数用于指示至少一个第一客户端使用训练配置参数对全局模型进行训练；
第一服务端用于接收至少一个第一客户端分别反馈的第一模型更新参数，第一模型更新参数为至少一个第一客户端训练后得到的全局模型的参数；
第一服务端用于对至少一个第一客户端分别反馈的第一模型更新参数进行聚合，得到第一聚合信息；
第二服务端用于接收对应的至少一个第二客户端发送的第二模型更新参数，并对至少一个第二客户端发送的第二模型更新参数进行聚合，以得到第二聚合信息；
第一服务端用于接收每个第二服务端发送的第二聚合信息；
第一服务端用于基于第一聚合信息和每个第二服务端发送的第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
本申请实施方式提供的联邦学习系统中，在进行每轮迭代学习时，客户端可以主动向服务端发送请求来请求参与联邦学习，并将训练的全局模型的参数反馈给服务端，服务端对接收到的模型更新参数进行聚合，从而得到聚合信息，并且，服务端还接收其他服务端发送的聚合信息，从而使服务端具有比自身的聚合信息更多的信息。使客户端和服务端无需保持长连接，客户端在任意时刻接入服务端都可以获取到更多的数据，可以提高客户端获取到的数据的准确度，从而可以基于更多的数据来进行每轮迭代训练，提高联邦学习的整体效率。
在一种可能的实施方式中，多个服务端中包括的第二服务端的数量为多个；多个第二服务端中的每个第二服务端用于接收对应的至少一个第二客户端发送的第二模型更新参数，并对至少一个第二客户端发送的第二模型更新参数进行聚合，以得到第二聚合信息；第一服务端具体用于接收多个第二服务端分别发送的第二聚合信息；第一服务端具体用于基于第一聚合信息和多个第二服务端分别发送的第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
本申请实施方式中，联邦学习系统内的服务端之间可以互相传输数据，使每个服务端上都具有全量数据，从而使客户端在任意时刻接入服务端，都可以获取到全量数据，提高客户端获取到的数据的准确度，且在联邦学习的过程中客户端获取到全量数据，可以提高联邦学习的整体效率。
在一种可能的实施方式中，多个服务端中还包括作为主节点的第三服务端，主节点用于对多个服务端进行管理；主节点用于向多个服务端分别发送第一触发指示；第二服务器具体用于基于第一触发指示向第一服务器发送第二聚合信息。
因此，本申请实施方式中，可以由主节点来触发服务端之间互相传输聚合信息，从而使每个服务端都可以具有更全面的全量数据，从而使客户端在任意时刻接入服务端，都可以获取到全量数据，提高客户端获取到的数据的准确度，且在联邦学习的过程中客户端获取到全量数据，可以提高联邦学习的整体效率。
在一种可能的实施方式中，主节点中包括计数器，计数器用于对多个服务端接收到的请求消息的数量进行计数，请求消息用于请求获取多个服务端中对应的服务端中存储的全局模型；主节点具体用于当计数器的数值满足第一阈值时，向多个服务端中的每个服务端发送第一触发指示，第一触发指示用于触发每个第二服务器向第一服务器发送第二聚合信息。
本申请实施方式中，可以通过计数器来对接收到的请求消息进行计数，当计数达到一定数值时，即可触发各个服务端之间同步数据，相当于限制每轮联邦学习迭代过程中参与的客户端的数量，避免过多客户端参与而导致的长尾效应，使各个服务端可以及时地进行数据同步。
在一种可能的实施方式中，主节点还包括计时器，计时器从每一轮迭代过程中接收到的第一个请求消息开始计时；主节点还用于当计时器超过第二阈值时，向多个服务端中的 每个服务端分别发送第二触发指示，第二触发指示用于指示多个服务端进行下一轮迭代。
因此，本申请实施方式中，可以通过计时器来设置每轮迭代的时间窗口，当计时器超时时即可进入下一轮迭代学习，从而避免因等待时长过长而导致的长尾效应，即使客户端在进行联邦学习时中途退出也不影响联邦学习的整体训练效率。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代过程中接入联邦学习系统中每个服务端的客户端的数量相关，或者，第二阈值与上一轮迭代过程中多个服务端和多个客户端之间进行通信的数据量的大小相关。因此，本申请实施方式中，可以根据客户端的数量或者通信量来确定时间窗口，使计时器对应的时间窗口与实际场景匹配，更合理地设置每轮迭代的时间窗口。
在一种可能的实施方式中，第一服务端接收第三客户端发送的查询消息，第三客户端包括接入联邦学习系统的任意一个客户端；第一服务端针对查询消息向第三客户端发送更新后的全局模型的信息。因此，本申请实施方式中，客户端可以向服务端发送查询消息以查询最新的全局模型。因每轮迭代学习过程中各个服务端之间进行了数据同步，因此每个服务端上都具有比自身的聚合信息更多的数据，客户端无论接入哪个服务端都可以获取到更多的数据，使客户端获取到更准确的全局模型。
在一种可能的实施方式中，第一服务端还用于向第二服务端发送第一聚合信息；第二服务器端具体用于结合第一聚合信息和第二聚合信息，更新本地存储的全局模型，以得到更新后的全局模型。因此，本申请实施方式中，联邦学习系统内的服务端之间可以互相传输聚合信息，从而使每个服务端上都具有更多数据，从而保持联邦学习系统内的数据一致性。
第四方面，本申请提供一种服务端，应用于联邦学习系统，联邦学习系统中包括多个服务端和多个客户端，多个服务端用于进行迭代联邦学习，多个服务端中的任意一个服务端包括：
收发模块，用于接收至少一个第一客户端发送的请求消息，请求消息用于请求第一服务器中存储的全局模型，多个客户端包括至少一个第一客户端；
收发模块，还用于向至少一个第一客户端发送全局模型的信息和训练配置参数，全局模型的信息和训练配置参数用于指示至少一个第一客户端使用训练配置参数对全局模型进行训练；
收发模块，还用于接收至少一个第一客户端分别反馈的第一模型更新参数，第一模型更新参数是至少一个第一客户端训练后得到的全局模型的参数；
聚合模块，用于对至少一个第一客户端反馈的第一模型更新参数进行聚合，以得到在本轮迭代中的第一聚合信息；
收发模块，还用于获取第二服务端发送的第二聚合信息，第二聚合信息为第二服务端在本轮迭代中，对接收到的第二模型更新参数进行聚合得到的信息；
更新模块，用于基于第一聚合信息和第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
在一种可能的实施方式中，收发模块，还用于接收多个第二服务端分别发送的第二聚 合信息；
更新模块，具体用于基于第一聚合信息和多个第二服务端分别发送的第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
在一种可能的实施方式中，第一服务器为联邦学习系统中的主节点，主节点用于对多个服务端进行管理，
收发模块，还用于向第二服务器发送第一触发指示，第一触发指示用于指示第二服务端向第一服务端发送第二聚合信息；
收发模块，还用于接收第二服务端中的第二聚合信息。
在一种可能的实施方式中，第一服务器中设置有计数器，计数器用于对多个服务端接收到的请求消息的数量进行计数，
收发模块，具体用于当计数器的数值满足第一阈值时，向第二服务端发送第一触发指示。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。
在一种可能的实施方式中，主节点中包括计时器，计时器从每一轮迭代中接收到第一个请求消息之后开始计时，
收发模块，具体用于当计时器超过第二阈值时，接收主节点发送的第二触发指示，第二触发指示用于指示进行下一轮迭代学习。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关，或者，第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。
在一种可能的实施方式中，收发模块，还用于接收第三客户端发送的查询消息，第三客户端包括接入联邦学习系统中的客户端中的任意一个；
收发模块，还用于针对查询消息向第三客户端发送更新后的全局模型的信息。
在一种可能的实施方式中，收发模块，还用于向第二服务端发送第一聚合信息，以使第二服务端基于第一聚合信息和第二聚合信息，更新本地存储的全局模型以得到更新后的全局模型。
第五方面，本申请提供一种服务端，应用于联邦学习系统，联邦学习系统包括多个服务端，多个服务端中的其中一个服务端作为主节点，多个服务端用于进行迭代学习以实现联邦学习，主节点包括：
启动模块，用于在多个服务端中任意一个服务端接收到第一个请求消息之后，主节点启动计数器和计时器，计数器用于对多个服务端在一轮迭代中接收到的请求消息进行计数，请求消息用于请求获取多个服务端中对应的服务端中存储的全局模型；
收发模块，用于若计数器的值达到第一阈值，主节点向多个服务端中的每个服务端发送第一触发指示，第一触发指示用于指示多个服务端之间互相传输本地存储的信息；
收发模块，还用于若计数器的值未达到第一阈值，计时器的值达到第二阈值，则主节点向每个服务端发送第二触发指示，第二触发指示用于指示每个服务端进行下一轮迭代。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关或者第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。第六方面，本申请实施例提供一种联邦学习装置，包括：处理器和存储器，其中，处理器和存储器通过线路互联，处理器调用存储器中的程序代码用于执行上述第一方面或者第二方面任一项所示的用于联邦学习方法中与处理相关的功能。可选地，该联邦学习装置可以是芯片。
第七方面，本申请实施例提供了一种联邦学习装置，该联邦学习装置也可以称为数字处理芯片或者芯片，芯片包括处理单元和通信接口，处理单元通过通信接口获取程序指令，程序指令被处理单元执行，处理单元用于执行如上述第一方面或第二方面中任一可选实施方式中与处理相关的功能。
第八方面，本申请实施例提供了一种计算机可读存储介质，包括指令，当其在计算机上运行时，使得计算机执行上述第一方面或第二方面中任一可选实施方式中的方法。
第九方面，本申请实施例提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述第一方面或第二方面中任一可选实施方式中的方法。
附图说明
图1为本申请应用的一种人工智能主体框架示意图；
图2为本申请实施例提供的一种联邦学习系统的架构示意图；
图3为本申请实施例提供的一种服务器的结构示意图；
图4为本申请实施例提供的一种联邦学习方法的流程示意图；
图5为本申请实施例提供的另一种联邦学习系统的架构示意图；
图6为本申请实施例提供的另一种联邦学习方法的流程示意图；
图7为本申请实施例提供的另一种联邦学习方法的流程示意图；
图8为本申请实施例提供的一种聚合方式示意图；
图9为本申请实施例提供的一种服务器的结构示意图；
图10为本申请实施例提供的另一种服务器的结构示意图；
图11为本申请实施例提供的另一种服务器的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
首先对人工智能系统总体工作流程进行描述，请参见图1，图1示出的为人工智能主体框架的一种结构示意图，下面从“智能信息链”(水平轴)和“IT价值链”(垂直轴)两个 维度对上述人工智能主题框架进行阐述。其中，“智能信息链”反映从数据的获取到处理的一列过程。举例来说，可以是智能信息感知、智能信息表示与形成、智能推理、智能决策、智能执行与输出的一般过程。在这个过程中，数据经历了“数据—信息—知识—智慧”的凝练过程。“IT价值链”从人智能的底层基础设施、信息(提供和处理技术实现)到系统的产业生态过程，反映人工智能为信息技术产业带来的价值。
(1)基础设施
基础设施为人工智能系统提供计算能力支持，实现与外部世界的沟通，并通过基础平台实现支撑。通过传感器与外部沟通；计算能力由智能芯片，如中央处理器(central processing unit，CPU)、网络处理器(neural-network processing unit，NPU)、图形处理器(英语：graphics processing unit，GPU)、专用集成电路(application specific integrated circuit,ASIC)或现场可编程逻辑门阵列(field programmable gate array，FPGA)等硬件加速芯片)提供；基础平台包括分布式计算框架及网络等相关的平台保障和支持，可以包括云存储和计算、互联互通网络等。举例来说，传感器和外部沟通获取数据，这些数据提供给基础平台提供的分布式计算系统中的智能芯片进行计算。
(2)数据
基础设施的上一层的数据用于表示人工智能领域的数据来源。数据涉及到图形、图像、语音、文本，还涉及到传统设备的物联网数据，包括已有系统的业务数据以及力、位移、液位、温度、湿度等感知数据。
(3)数据处理
数据处理通常包括数据训练，机器学习，深度学习，搜索，推理，决策等方式。
其中，机器学习和深度学习可以对数据进行符号化和形式化的智能信息建模、抽取、预处理、训练等。
推理是指在计算机或智能系统中，模拟人类的智能推理方式，依据推理控制策略，利用形式化的信息进行机器思维和求解问题的过程，典型的功能是搜索与匹配。
决策是指智能信息经过推理后进行决策的过程，通常提供分类、排序、预测等功能。
(4)通用能力
对数据经过上面提到的数据处理后，进一步基于数据处理的结果可以形成一些通用的能力，比如可以是算法或者一个通用系统，例如，翻译，文本的分析，计算机视觉的处理，语音识别，图像的识别等等。
(5)智能产品及行业应用
智能产品及行业应用指人工智能系统在各领域的产品和应用，是对人工智能整体解决方案的封装，将智能信息决策产品化、实现落地应用，其应用领域主要包括：智能终端、智能交通、智能医疗、自动驾驶、平安城市等。
本申请实施例可以应用于联邦学习领域，可以对通过客户端和服务器协同对神经网络进行训练，因此涉及了大量神经网络的相关应用，如在进行联邦学习时客户端训练的神经网络。为了更好地理解本申请实施例的方案，下面先对本申请实施例可能涉及的神经网络的相关术语和概念进行介绍。
(1)神经网络
神经网络可以是由神经单元组成的，神经单元可以是指以xs和截距1为输入的运算单元，该运算单元的输出可以如公式(1-1)所示：
Figure PCTCN2022087647-appb-000001
其中，s＝1、2、……n，n为大于1的自然数，Ws为xs的权重，b为神经单元的偏置。f为神经单元的激活函数(activation functions)，用于将非线性特性引入神经网络中，来将神经单元中的输入信号转换为输出信号。该激活函数的输出信号可以作为下一层卷积层的输入，激活函数可以是sigmoid函数。神经网络是将多个上述单一的神经单元联结在一起形成的网络，即一个神经单元的输出可以是另一个神经单元的输入。每个神经单元的输入可以与前一层的局部接受域相连，来提取局部接受域的特征，局部接受域可以是由若干个神经单元组成的区域。
(2)深度神经网络
深度神经网络(deep neural network，DNN)，也称多层神经网络，可以理解为具有多层中间层的神经网络。按照不同层的位置对DNN进行划分，DNN内部的神经网络可以分为三类：输入层，中间层，输出层。一般来说第一层是输入层，最后一层是输出层，中间的层数都是中间层，或者称为隐层。层与层之间是全连接的，也就是说，第i层的任意一个神经元一定与第i+1层的任意一个神经元相连。
虽然DNN看起来很复杂，其每一层可以表示为线性关系表达式：
Figure PCTCN2022087647-appb-000002
其中，
Figure PCTCN2022087647-appb-000003
是输入向量，
Figure PCTCN2022087647-appb-000004
是输出向量，
Figure PCTCN2022087647-appb-000005
是偏移向量或者称为偏置参数，w是权重矩阵(也称系数)，α()是激活函数。每一层仅仅是对输入向量
Figure PCTCN2022087647-appb-000006
经过如此简单的操作得到输出向量
Figure PCTCN2022087647-appb-000007
由于DNN层数多，系数W和偏移向量
Figure PCTCN2022087647-appb-000008
的数量也比较多。这些参数在DNN中的定义如下所述：以系数w为例：假设在一个三层的DNN中，第二层的第4个神经元到第三层的第2个神经元的线性系数定义为
Figure PCTCN2022087647-appb-000009
上标3代表系数W所在的层数，而下标对应的是输出的第三层索引2和输入的第二层索引4。
综上，第L-1层的第k个神经元到第L层的第j个神经元的系数定义为
Figure PCTCN2022087647-appb-000010
需要注意的是，输入层是没有W参数的。在深度神经网络中，更多的中间层让网络更能够刻画现实世界中的复杂情形。理论上而言，参数越多的模型复杂度越高，“容量”也就越大，也就意味着它能完成更复杂的学习任务。训练深度神经网络的也就是学习权重矩阵的过程，其最终目的是得到训练好的深度神经网络的所有层的权重矩阵(由很多层的向量W形成的权重矩阵)。
(3)卷积神经网络
卷积神经网络(convolutional neuron network，CNN)是一种带有卷积结构的深度神经网络。卷积神经网络包含了一个由卷积层和子采样层构成的特征抽取器，该特征抽取器 可以看作是滤波器。卷积层是指卷积神经网络中对输入信号进行卷积处理的神经元层。在卷积神经网络的卷积层中，一个神经元可以只与部分邻层神经元连接。一个卷积层中，通常包含若干个特征平面，每个特征平面可以由一些矩形排列的神经单元组成。同一特征平面的神经单元共享权重，这里共享的权重就是卷积核。共享权重可以理解为提取图像信息的方式与位置无关。卷积核可以以随机大小的矩阵的形式初始化，在卷积神经网络的训练过程中卷积核可以通过学习得到合理的权重。另外，共享权重带来的直接好处是减少卷积神经网络各层之间的连接，同时又降低了过拟合的风险。
本申请以下提及的用于提取特征的网络，即可以包括一层或者多层卷积层，示例性地，该用于提取特征的网络即可以采用CNN来实现。
(4)损失函数
在训练深度神经网络的过程中，因为希望深度神经网络的输出尽可能的接近真正想要预测的值，所以可以通过比较当前网络的预测值和真正想要的目标值，再根据两者之间的差异情况来更新每一层神经网络的权重向量(当然，在第一次更新之前通常会有初始化的过程，即为深度神经网络中的各层预先配置参数)，比如，如果网络的预测值高了，就调整权重向量让它预测低一些，不断地调整，直到深度神经网络能够预测出真正想要的目标值或与真正想要的目标值非常接近的值。因此，就需要预先定义“如何比较预测值和目标值之间的差异”，这便是损失函数(loss function)或目标函数(objective function)，它们是用于衡量预测值和目标值的差异的重要方程。其中，以损失函数举例，损失函数的输出值(loss)越高表示差异越大，那么深度神经网络的训练就变成了尽可能缩小这个loss的过程。该损失函数通常可以包括误差平方均方、交叉熵、对数、指数等损失函数。例如，可以使用误差均方作为损失函数，定义为
Figure PCTCN2022087647-appb-000011
具体可以根据实际应用场景选择具体的损失函数。
(5)反向传播算法
神经网络可以采用误差反向传播(back propagation，BP)算法在训练过程中修正初始的神经网络模型中参数的大小，使得神经网络模型的重建误差损失越来越小。具体地，前向传递输入信号直至输出会产生误差损失，通过反向传播误差损失信息来更新初始的神经网络模型中参数，从而使误差损失收敛。反向传播算法是以误差损失为主导的反向传播运动，旨在得到最优的神经网络模型的参数，例如权重矩阵。
在本申请中，客户端在进行模型训练时，即可通过损失函数或者通过BP算法来对全局模型进行训练，得到训练后的全局模型。
(6)联邦学习(federated learning，FL)
一种分布式机器学习算法，通过多个客户端(client或者FL-Client)，如移动设备或边缘服务器，和服务器在数据不出域的前提下，协作式完成模型训练和算法更新，得到训练后的全局模型。如客户端使用本地存储的训练样本对全局模型进行训练并将训练后得到的模型参数反馈给服务端，由服务端对接收到的模型参数进行聚合，并基于聚合得到的信息更新本地存储的全局模型，得到最新的全局模型。
(7)横向联邦
在两个数据集的用户特征重叠较多而用户重叠较少的情况下，把数据集按照用户维度切分，并去除双方用户特征相同而用户不完全相同的部分数据进行训练。
通常，在进行联邦学习时，需要多个服务器和客户端参与，由客户端对模型进行训练，然后上传至服务器，由服务器聚合后得到最终的模型。
例如，在机器学习和深度学习领域，分布式的优化已经成了一种先决条件。常用的分布式训练是参数服务器(Parameter Server，PS)模式，包括了Worker(参数服务器训练模式中进行训练的计算节点)和PS(参数服务器)角色，训练数据由计算中心分配到各Worker，若干Worker与PS通过网络通信，交换模型或者梯度等信息，由PS进行聚合以及模型优化，从而高效训练得到最终模型。在PS模式中，每个服务器实际上都只负责分到的部分参数(多个服务器共同维持一个全局的共享参数)，而每个计算节点也只分到部分数据和处理任务。然而，PS模式虽然解决了大规模分布式训练的问题，但是其每个服务器(server)只能够处理部分参数的计算，并且要求参与方和中心服务器保持稳定长连接，在Cross-Device场景中，FL-Client大多是无状态，不可靠，且不能寻址的，而且中心服务器需要聚合全量的参数并完成聚合(aggregation)计算，PS模式训练模式显然也处理不了跨设备终端的挑战。
又例如，在已有的一些联邦学习框架中，如TensorFlow Federated(TFF)框架或者PaddleFL框架，TFF聚焦于提供联邦学习的仿真接口，支持在本地模拟任意数量的FL-Client和FL-Server，但是其框架本身不存在跨网络通信，也就无法在真实生产环境中使用。PaddleFL在提供了一种基于数据并行的分布式联邦学习的框架，包含有FL-Server，FL-Trainer，FL-Scheduler三种角色，但是FL-Client和FL-Server通过远程过程调用系统(remote procedure calls，RPC)进行跨进程通信，两者存在的稳定的长连接，并且FL-Scheduler可以调度Worker，在每个更新周期前，决定哪些Worker可以参与训练。这在跨设备的部署场景中是无法实现的。
因此，本申请提供一种联邦学习架构，通过更松散的接入方式，使客户端接入服务器，各个服务器之间保持数据一致性，使客户端在任何时候接入服务器都可以获取到全量数据，提高客户端的模型训练效率。
本申请提供的联邦学习系统可以如图2所述，该系统(或者也可以简称为集群)中可以包括多个服务端，该多个服务端之间互相建立连接，每个服务端可以和一个或者多个客户端通信。
其中，服务端通常可以部署于服务器中，在一些场景中，服务端也可以部署于终端设备中，具体可以根据实际应用场景调整，本申请对此不作限定。客户端可以部署于终端中，如手机、手环、智能电视等终端中，也可以部署于服务器中，具体可以根据实际应用场景调整，本申请对此也不作限定。
该多个服务端中的其中一个服务端可以作为主节点，通过主节点来使集群内保持事务一致性，从而使各个服务端之间可以协同工作。
该框架中还可以包括调度器(图中未示出)或者弹性负载均衡(elastic load balance，ELB)(图中未示出)等，此处不再一一赘述。如调度器可以部署于任意一个服务端上，可 以用于对集群内的通信资源进行调度，如发送效率所使用的时间窗口、通道等。ELB可以部署于任意一个服务端上，可以用于对服务端与客户端之间的通信进行路由，由ELB根据各个服务端的负载确定客户端接入的服务端等，如ELB可以将发送请求消息的客户端接入当前负载较少的服务端中，从而使各个服务端可以保持合适的负载，避免出现某一个或者多个服务端负载过大的情况。
具体地，在进行联邦学习时，每个服务端可以向与其建立了连接的客户端下发待训练的模型，客户端可以使用本地存储的训练样本对该模型进行训练，并将训练后的模型的数据反馈至服务端，服务端在接收到一个或者多个客户端反馈的训练后的一个或者多个模型之后，即可对接收到的一个或者多个模型的数据进行聚合，得到聚合后的数据，相当于聚合后的模型。在满足停止条件之后，即可输出最终的模型，完成联邦学习。
通常，随着客户端的增加，客户端和服务端之间的通信量也增加。例如，当千万级客户端同时进行联邦学习请求时，需要大量网络带宽的支持，并且单台服务端不能承接庞大的数据和参数，因此如何得到一种高效的联邦学习系统来应对跨设备的系统异构和远程通信挑战成为一个关键问题。
示例性地，本申请提供的联邦学习系统中的服务端的结构可以参阅图3，该服务端中可以包括训练引擎、存储模块、限时通信模块、设备管理模块、加密模块、协议解析模块或通信模块等。
通信模块，用于FL-Server和FL-Client之间的数据通信，以及FL-Server集群动态组网实现中心服务端资源的弹性伸缩。其中在FL-Server与FL-Client进行通信时，可以采用超文本传输协议(hypertext transfer protocol，HTTP)协议进行传输，实现稳定的短时连接。而FL-Server与FL-Server进行通信时，则可以通过远程过程调用系统(remote procedure calls，RPC)协议进行跨进程通信，实现稳定的长连接。
设备管理模块：负责联邦学习设备选择等设备相关业务，如在接收到客户端的请求消息之后，筛选该客户端是否参与联邦学习。
加解密模块：对上传下载的模型权重加密，降低通过模型逆向攻击获得原始样本的可能性。
协议解析模块：对发送的通信消息进行序列化或者对接收到的消息反序列化等。
训练引擎：对FL-Client上传数据进行联邦学习计算的模块，包括聚合器和优化器。聚合器可以用于对数据进行聚合，如对FL-Client上传的数据进行求和、加权求和等。优化器可以用于计算模型梯度，并基于计算出来的梯度对模型进行优化等，得到输出更准确的模型。
存储模块：存储联邦学习需要的元数据以及训练所得的全局最优模型等数据。
限时通信装置：包括计时器和计数器，在计时器超时或者计数器值达到阈值时进行后续的处理。
下面结合前述图2的联邦学习系统以及图3所示出的系统中的服务端的结构，对本申请提供的联邦学习方法方法的流程进行介绍。
参与图4，下面结合前述的架构对本申请提供的联邦学习方法的流程示意图。
401、第一服务端接收至少一个客户端发送的至少一个请求消息。
该第一服务端可以是前述图2中所示出的多个第一服务端中的任意一个，该至少一个客户端可以是与该第一服务端建立了连接的客户端，如该至少一个客户端发送的请求消息可以由ELB路由至该第一服务端中。
客户端可以通过该请求消息向第一服务端请求参与联邦学习，以请求使用本地的数据进行模型训练，通过本地数据影响全局模型，得到携带客户端的个性化特征的模型。
例如，当客户端上存储的数据超过预设数据量时，即可向联邦学习系统发送请求消息，由该系统中的ELB将该请求消息路由到适配的第一服务端中，以通过该第一服务端和客户端协同实现联邦学习。
又例如，客户端可以采集到与其他设备不同的个性化的训练数据，客户端在采集到训练数据后，可以主动向联邦学习系统发送请求消息，以请求使用本地存储的数据来进行联邦学习，从而得到与训练数据适配的模型，并将该模型同步至系统中，完成联邦学习，从而使集群内的全局模型可以适应性地学习到各个客户端采集到的各种数据，提高全局模型的输出精度。
402、第一服务端针对请求消息向至少一个客户端分别发送全局模型以及训练配置参数。
在第一服务端接收到至少一个客户端发送的请求消息之后，若确认允许该至少一个客户端参与联邦学习，则可以向该至少一个客户端分别发送本地存储的全局模型的信息以及训练配置参数，该全局模型即待训练的模型，该全局模型的信息可以包括全局模型的结构参数(如全局模型的深度或者宽度等)或者权重参数等，该训练配置参数即对该全局模型进行训练时使用的参数，如学习率、epoch数量或者安全算法中类别等信息。
在一种可能的方式中，若第一服务端接收到多个客户端发送的多个请求消息，则第一服务端可以对该多个客户端进行筛选，筛选出其中的至少一个客户端来参与联邦学习。例如，请求消息中可以携带终端的信息，如网络连接状态、电量、负载等信息，第一服务端可以根据终端的信息筛选出连接稳定的客户端来进行联邦学习，并向筛选出来的至少一个客户端下发模型和训练配置参数，从而使客户端可以基于该训练配置参数对该模型进行训练。
403、客户端使用训练配置参数对全局模型进行训练。
客户端(或者称为第一客户端)接收到第一服务端下发的全局模型和训练配置参数之后，即可使用训练集对该全局模型进行训练，得到训练后的全局模型。该训练集中可以包括多个训练样本，该训练样本可以是客户端采集到的数据，也可以是客户端接收到的数据，此处不作限定。
例如，客户端可以是部署于用户的终端上的应用程序，在用户启用客户端的过程中，客户端可以采集用户启用客户端的期间产生的数据，并将该数据作为训练样本，向服务端请求参与联邦学习，从而将采集到的数据应用到模型更新中，从而使联邦学习系统中的全局模型可以与各个客户端适配。
404、客户端向第一服务端反馈模型更新参数。
在客户端对全局模型进行训练得到训练后的模型之后，即可向第一服务端反馈训练后 的模型的信息，即模型更新参数。
具体地，该模型更新参数中可以包括模型的权重参数。例如，在第t次迭代联邦学习的过程中，服务端j接收到的模型更新参数可以表示为：
Figure PCTCN2022087647-appb-000012
n k表示第k个客户端的数据集大小，
Figure PCTCN2022087647-appb-000013
即为客户端训练后得到的模型的权重参数。
405、第一服务端对至少一个客户端反馈的模型更新参数进行聚合，得到第一聚合信息。
在第一服务端接收到至少一个客户端分别反馈的模型更新参数之后，即可对该至少一个客户端分别反馈的模型更新参数进行聚合，得到第一聚合信息。相当于对该至少一个客户端反馈的模型进行聚合，得到聚合后的模型的信息。
具体地，具体方式可以包括多种，如平均、加权融合等，本申请对于聚合方式不作限定，具体可以根据实际应用场景调整。
406、第一服务端获取第二服务端中的第二聚合信息。
其中，第二服务端包括联邦学习系统中除第一服务端之外的服务端，第二服务端的数量可以是一个或者多个。第一服务端除了可以聚合与其连接的至少一个客户端的模型更新参数并进行聚合，还可以接收到其他服务端对接收到的模型更新参数聚合后得到的第二聚合信息，从而使联邦学习系统中的各个服务端中的聚合信息可以保持一致性。并且，每个服务端都可以获取到其他服务端的聚合信息，从而使每个服务端都可以保持更多的数据，每个客户端在接入服务端时，都可以获取到更全面的数据，而无需与服务端保持长连接状态。因此，本申请提供的联邦学习系统可以适应客户端部署在移动终端的场景。
在一种可能的场景中，第一服务端可以接收联邦学习系统内除自身外的所有服务器发送的聚合信息，即该联邦学习系统内的各个服务器之间可以互相传输聚合信息，从而使每个服务端上都可以具有全量数据。客户端在任意时刻接入服务端，都可以获取到全量数据，从而提高客户端获取到的数据的准确性。
需要说明的是，本申请以下实施方式中，以联邦学习系统内的所有服务端互相传输聚合信息为例进行示例性说明，在实际应用场景中，也可以是部分服务端之间传输数据，或者部分服务端仅接收聚合信息而不发送聚合信息等，具体可以根据实际应用场景调整，本申请对此并不作限定。
还需要说明的是，本申请实施方式中，第二服务端或者第二聚合信息的方式与第一服务端获取第一聚合信息的方式类似，如第二服务端接收到第二客户端发送的请求消息之后，向第二客户端下发全局模型和训练配置参数，该全局模型和训练配置参数通常与第一服务器向第一客户端下发的全局模型和训练配置参数相同。然后第二客户端使用本地存储的训练数据基于该训练配置参数对全局模型进行训练，并将训练的全局模型的参数(如权重参数、模型的宽度或深度等)反馈至第二服务端，第二服务端对接收到的模型更新参数进行聚合，即可得到第二聚合信息。本申请对第二服务端获取第二聚合信息的具体过程不再赘述。
在一种可能的实施方式中，可以由主节点来触发集群内的各个服务之间互相传输聚合 信息，来实现集群内的数据同步，从而使集群内的服务端都可以具有更全面的数据。
该主节点可以是联邦学习系统中的其中一个服务端，如前述的第一服务端或者第二服务端等。
在一种可能的实施方式中，主节点中可以设置计数器或者计时器中的至少一种，用于对每一轮迭代中参与的客户端数量或者时间窗口进行监控，从而有效控制参与联邦学习的客户端的数量，以及每一轮迭代过程的时长，避免无效的等待，减少“掉尾”效应。
具体地，计数器用于对联邦学习系统内接收到的由客户端发送的请求消息进行计数，在每一轮迭代完成之后，主节点即可对计数器进行初始化。在下一轮迭代中，联邦学习系统内的任意一个服务端接收到第一个请求消息时，主节点即可启动计时器，每接收到一个请求消息即进行计数。当计数器的值达到第一阈值时，即可触发第一服务端向客户端发送模型和训练配置参数，如主节点向各个服务端发送第一触发指示，以指示各个服务端将自身的聚合信息传输给其他服务端，如指示第二服务端将第二聚合信息发送给第一服务端。若主节点为第一服务端，则主节点可以触发自身将第一聚合信息发送给其他服务端(如所有的第二服务端)，还向第二服务端发送第一触发指示，以指示各个第二服务端将自身的第二聚合信息发送给其他服务端。
因此，可以通过计数器来对参与联邦学习的客户端进行监测，避免因客户端太多而导致的学习效率低。例如，ELB在将每个请求消息路由至服务端之后，可以向主节点发送通知消息，以通知主节点接收到的主节点的情况，如请求消息的数量、接收时间戳等，从而使主节点对请求消息的数量进行计数。当计数器的计数超过第一阈值时，主节点向各个服务器发送第一触发指示，触发联邦学习系统内各个服务端(包括主节点本身)之间互相传输数据，从而完成服务端之间的数据同步，使每个服务端上都具有更全面的数据，客户端在接入服务端之后，即可获取到更全面的数据，从而可以基于更全面的数据进行联邦学习或者应用等，提高客户端的用户的使用体验。
可选地，第一阈值包括预设值，或者，与本轮迭代或者上一轮迭代接入多个服务端的客户端的数量相关等。因此，本申请实施方式中，计数器对应的第一阈值可以是预先设定的值，也可以根据接入服务端的客户端的数量确定的值，通常该第一阈值不大于接入服务端的客户端的数量，从而避免等待客户端的接入而导致联邦学习的等待时间过长，提高联邦学习的整体效率。
计时器可以用于在联邦学习的过程中设置时间窗口，即仅在时间窗口内进行一轮联邦学习，当计时器的计时超过第二阈值时，主节点即可触发服务端进行下一轮学习，若当前学习过程中获取到了模型更新参数，则可以丢弃该模型更新参数，当然也可以保留并具体该模型更新参数，以提高数据的利用率。因此，本申请实施方式中，通过计时器来对每一轮迭代学习的时间窗口进行监控，每一轮迭代学习可以在计时器的时间窗口内进行，从而在某些客户端掉线或者训练效率过低等时，降低这些客户端对联邦学习过程的影响，提高联邦学习的效率。
具体地，当联邦学习系统内的任意一个服务端接收到第一个请求消息时，主节点即可启动计时器，在计时器的时间窗口内完成本轮迭代。当计时器超时时，主节点即可向联邦 学习系统中的各个服务端(包括主节点本身)发送第二触发指示，触发各个服务端进入下一轮迭代。针对主节点本身，主节点在确定计时器超时时，即可直接确定进入下一轮迭代，可以无需针对自身发送第二触发指示，当然也可以针对自身生成第二触发指示，本申请对此不作限定。
可选地，该第二阈值可以是预先设定的值，也可以是与本轮或者上一轮接入联邦学习系统内的客户端的数量相关，或者，与本轮或者上一轮客户端和服务端之间的通信的数据量的大小相关。例如，该第二阈值可以和接入联邦学习系统内的客户端的数量确定的值或者根据客户端和服务端之间的通信的数据量的大小等呈正相关关系，如上一轮迭代中接入联邦学习系统的客户端的数量越多，该第二阈值也就越大，上一轮迭代中客户端和服务端之间的通信的数据量越大，该第二阈值也就越大等，具体可以根据实际应用场景进行调整。因此，本申请实施方式中，可以为计时器设置和实际应用场景匹配的时间窗口，可以避免时间窗口过长或者过短，从而提高联邦学习的整体效率。
在一种可能的场景中，当计时器和计数器协同工作时，在每轮迭代过程中，联邦学习系统内的服务端接收到第一个请求消息时，主节点启动计时器和计数器。若计数器的数值达到第一阈值，而计时器未超时，则可以关闭计时器，在对本轮迭代过程中所有的客户端反馈的模型更新参数进行聚合完成之后，即可进入下一轮迭代。若计数器的数值未达到第一阈值，而计时器超时，则可以确定结束本轮迭代，进入下一轮迭代，本轮迭代中的数据可以丢弃也可以保留。若计数器的数值未达到第一阈值，计时器也未超时，则可以继续本轮迭代。
因此，本申请实施方式中，可以通过计时器和计数器协同工作，对本轮迭代中参与联邦学习的客户端的数量进行限制，以及对本轮迭代的时长进行限制，从而避免因等待客户端接入或者客户端中途退出等情况而导致的长尾效应。
407、第一服务端对第一聚合信息和第二聚合信息进行聚合，得到第三聚合信息。
在第一服务端得到第一聚合信息和其他服务端发送的第二聚合信息之后，即可基于接收到的第一聚合信息和第二聚合信息更新本地存储的全局模型，即步骤407和步骤408。
其中，第一服务端除了可以对自身接收到的一个或者多个客户端反馈的模型更新参数进行聚合，还可以接收其他服务端得到的聚合信息，并对自身计算得到的第一聚合信息和接收到的第二聚合信息进行再次聚合，得到数据更完整的第三聚合信息。
因此，本申请实施方式中，各个服务端不仅对自身接收到的模型更新参数进行聚合，还对联邦学习系统内的其他服务端聚合得到的信息进行再次聚合，从而每个服务端上都可以具有联邦系统内的每个服务端中包括的数据，从而使各个服务端的数据保持一致，客户端在任意时刻接入服务端都可以获取到全量数据。
例如，在t+1次Iteration时，联邦学习系统内总的客户端的数量记为K，服务端j收集到部分客户端(记为集合S j)上传的权重更新信息表示为
Figure PCTCN2022087647-appb-000014
n k表示第k个客户端的数据集大小。每个服务端先计算出各自的
Figure PCTCN2022087647-appb-000015
和
Figure PCTCN2022087647-appb-000016
然后由主节点触发服务端之间的数据聚合，得到聚合计算后的
Figure PCTCN2022087647-appb-000017
和n j，各服务端都可以 得到集群维度的计算结果，即全量数据。在下一轮迭代中，客户端请求任意服务端都能得到正确的全量数据。
此外，第一服务端也可以将第一聚合信息发送给第二服务端，以使第二服务端对自身聚合得到的第二聚合信息和接收到的第一聚合信息进行再次聚合，得到第三聚合信息。通常，为保持各个服务端中的数据的异质性，第二服务端在进行聚合时可以采用与第一服务端相同的聚合方式，从而最终得到相同的第三聚合信息，使每个服务端中可以具有全量数据。
408、第一服务端更新存储的全局模型。
在得到第三聚合信息之后，该第三聚合信息中可以包括聚合后的模型的权重参数，第一服务端可以使用最终聚合后得到权重参数来更新本地存储的全局模型，从而得到更新后的全局模型。
因此，本申请实施方式中，客户端可以通过向服务端发送请求的方式来接入服务端，在服务端允许之后即可参与联邦学习，相当于客户端以松散的方式接入服务端，而无需与服务端保持长连接，在可以适应一些客户端部署于移动终端的场景。并且，各个服务端除了对与其连接的客户端反馈的模型更新参数进行聚合，还可以聚合其他服务端聚合后得到的信息，从而使各个服务端都可以具有联邦学习系统内的全量数据，客户端在任何时刻接入服务端，都可以获取到全量数据，从而使最终训练得到的模型的输出精度更高。
并且，本申请还可以通过设置计时器和/或计数器的方式，来对接入服务端的客户端进行计数，并在进行联邦学习的过程中进行计时，在计时器的时间窗口内进行联邦学习，若超出时间窗口未完成一个迭代学习过程，则可以直接进行下一轮学习，从而避免存在部分客户端因连接不稳定或者训练效率低而影响整体的训练过程，提高联邦学习的整体学习效率。并且，客户端在进行联邦训练的过程中可以随时退出，而不影响联邦学习的效率，避免了对客户端的绑定。
需要说明的是，前述的步骤401-408可以是联邦学习过程中的其中一个迭代过程，在联邦学习的过程中，可以包括一次或者多次迭代，本申请仅以一次迭代为例进行示例性说明，并不作为限定，具体可以根据实际应用场景确定迭代次数。
409、客户端向第一服务端发送查询消息。
其中，客户端(或者称为第二客户端)在需要获取最新的模型时，可以向服务端发送查询消息，以请求查询最新的全局模型。
具体地，客户端可以通过联邦学习系统的地址来访问联邦学习系统，可以由ELB来对客户端发送的查询消息进行路由，从而将请求消息路由至适配的服务端中，此处的客户端向第一服务端发送查询消息仅仅是示例性说明，此处的第一服务端可以替换为其他服务端，如第二服务端或者第三服务端等等，此处并不作限定。
410、第一服务端向客户端下发更新的全局模型。
第一服务端在接收到客户端的查询消息之后，即可将本地最新的模型下发给客户端。
在前述步骤408中，第一服务端使用最新的聚合信息更新了存储的全局模型，在接收到查询消息之后，即可将最新的全局模型下发给客户端。具体可以是向该客户端下发模型 的结构和参数，如神经网络的层数、卷积核大小、输入输出的分辨率、各个网络层内的权重参数等等，从而使客户端获取到最新的模型结构。
需要说明的是，本申请中的步骤409和步骤410为可选步骤。
本申请实施方式中，各个服务端之间可以互相传输最新的聚合信息，从而使联邦学习系统内的各个服务端都具有全量数据，客户端在任意时刻接入服务端都可以获取到全量数据，而无需与服务端保持长连接，也可以获取到全量数据，提高客户端能获取到的模型的准确度。
前述对本申请提供的联邦学习方法的流程进行了介绍，为便于理解，下面结合具体的应用场景对本申请提供的联邦学习方法进行更详细的说明。
首先，对本申请提供的联邦学习方法应用的联邦学习系统进行示例性介绍。
示例性地，本申请提供的联邦学习系统可以如图5所示，此处以三个服务端(即FL-Server0、FL-Server1和FL-Server2)进行示例性说明，也可以替换为更多的服务端。调度器(即FL-Scheduler)可以用于对服务端集群内的通信资源或者存储资源进行调度，使各个服务端之间可以进行数据交换。
服务端之间或者服务端与调度器可以建立传输控制协议(transmission control protocol，TCP)连接，并保持稳定的长期连接，通过TCP私有协议进行通信。当然，此处的TCP也可以替换为其他协议，如因特网分组交换协议(internet packet exchange，IPX)、顺序分组交换(sequenced Packet Exchange，SPX)等协议，本申请对此并不作限定。
客户端可以通过IP地址接入联邦学习系统，服务端对外部暴露超文本传输协议(hypertext transfer protocol，HTTP)端口，对外部提供分布式服务。客户端可以与服务端建立HTTP连接，即短连接。
服务端0可以作为主节点，其他服务端作为从节点，主节点负责集群范围内的数据同步，如联邦学习过程中的触发所有节点进行聚合、全局计时或者计数等，从而保持集群内的数据一致性。
通常，若客户端部署于移动终端中，则服务端可能无法随时获取客户端的状态，如移动终端可能与服务端之间的连接不稳定，检索不到移动终端或者获取不到移动终端的连接状态等，可能导致服务端和移动终端之间通信中断，可能影响联邦学习的效率。
参阅图6，本申请提供的另一种联邦学习方法的流程示意图。
首先，在每轮迭代之前，即可初始化计时器和计数器，如设置为0。在每轮迭代过程中，联邦学习系统接收到由客户端发送的第一个请求消息之后，主节点上的计数器即可进行计数。在计数器开始计数时，可以同步开启计时器，在计时器的时间窗口内进行一轮联邦学习。
若计数器超过第一阈值，则可以关闭计时器，将当前轮迭代中发送请求的客户端作为当前轮次联邦学习中进行训练的客户端。该第一阈值可以是根据上一轮参与联邦学习的客户端的数量确定的，例如，上一轮参与联邦学习的客户端相对更上一轮迭代中的客户端减少，则可以在已设置的第一阈值的基础上进行降低，若上一轮参与联邦学习的客户端相对更上一轮迭代中的客户端增多，则可以在原设置的第一阈值的基础上进行增加，从而使计 数器对应的阈值和每轮迭代中参与训练的客户端适配，最大化训练效率。
以其中一个FL-server 1和FL-client为例，FL-server 1可以向FL-client下发待训练的全局模型和训练配置参数。该全局模型可以包括各种神经网络，如DNN、CNN、残差网络(ResNet)或者构建得到的网络等，具体可以根据实际应用场景进行调整。训练配置参数可以包括学习率、epoch数量或者安全算法中类别等训练模型时涉及到的参数，用于在客户端进行模型训练时，确定训练中的学习率、迭代次数或者传输数据时的加密算法类型等。
在FL-client使用训练配置参数以及本地存储的训练样本进行模型训练之后，即可将得到的模型更新参数反馈至FL-server 1，该模型更新参数可以包括训练后的模型相关的参数，如更新的全局模型的结构(如网络层的深度或者宽度等)、权重参数、或者降低稀疏性后的权重参数等。
本实施例仅以一个FL-client进行示例性说明，通常可以具有多个FL-client，每个FL-client都可以使用本地存储的训练样本对接收到的全局模型进行训练。例如，每个训练样本中包括样本和对应的真值标签，在一次训练过程中，将样本作为模型的输入，然后通过损失函数来衡量模型的输出和真值标签之间的差异，并通过差异计算全局模型的更新梯度，使用梯度对全局模型的权重参数进行反向更新，从而得到更新后的模型。
若FL-server 1接收到多个FL-client反馈的模型更新参数，则FL-server 1可以对该多个FL-client反馈的模型更新参数进行聚合，相当于数据归约。具体的聚合方式可以包括求和、求均值、加权求和等，具体可以根据实际应用场景进行调整。例如，FL-server 1对模型跟新数据进行聚合得到的数据可以表示为：
Figure PCTCN2022087647-appb-000018
其中S为客户端集合，
Figure PCTCN2022087647-appb-000019
表示各个客户端反馈的模型更新参数，n k表示每个客户端反馈的模型更新参数的大小。
在计数器的值超过第一阈值时，由主节点，即FL-server 0触发自身所在的服务端和从节点(如FL-server 1、FL-server 2等等)对各个服务端之间互相传输的聚合信息进行再次聚合，需要说明的是，此处为便于理解，仅以FL-server 0触发从节点进行聚合为例进行示例性说明。例如，在FL-server 0的触发下，FL-server 2将对接收到的模型更新参数聚合得到的聚合信息发送至FL-server 1，FL-server 1对接收到的聚合信息和本地存储的聚合信息进行聚合，得到最终的聚合信息。
在进行联邦学习的一轮迭代过程中，若计数器不超过第一阈值，而计时器超时，则可以终止当前次迭代，进入下一轮迭代中，即初始化计数器。如FL-server 0向FL-server 1、FL-server 2发送第二触发指示，以指示进入下一轮迭代。此时，当前次迭代过程中的数据可以丢弃或者启用。例如，若数据完整也可以继续采用，仅丢弃部分未响应的客户端的数据，具体可以根据实际应用场景调整。
因此，在本申请实施方式中，可以由主节点维护计数器和计时器，对发送了请求的客户端进行计数，在计时器达到阈值时即可触发集群维度的聚合操作，即各个服务端对接收到的其他服务端的聚合信息进行再次聚合，从而使各个服务端都可以具有其他服务端的聚 合信息，每个服务端都可以具有全量数据，当客户端在任意时刻接入任意服务端都可以获取到全量数据，提高联邦学习的效率。并且，主节点还维护计时器，当计时器超时时，即可进行下一轮迭代，相当于仅在计时器的时间窗口内进行一轮迭代学习，防止因等待客户端响应的时长太长而影响联邦学习的效率，解决因客户端掉队而导致的长尾效应。
此外，本申请提供的联邦学习系统中，主节点作为对每轮迭代时长和参与迭代的客户端的数量进行监控的重要节点，主要可以用于执行以下步骤：
在多个服务端中任意一个服务端接收到第一个请求消息之后，主节点启动计数器和计时器，计数器用于对多个服务端在一轮迭代中接收到的请求消息进行计数，请求消息用于请求获取多个服务端中对应的服务端中存储的全局模型；若计数器的值达到第一阈值，主节点向多个服务端中的每个服务端发送第一触发指示，第一触发指示用于指示多个服务端之间互相传输本地存储的信息；若计数器的值未达到第一阈值，计时器的值达到第二阈值，则主节点向每个服务端发送第二触发指示，第二触发指示用于指示每个服务端进行下一轮迭代。
本申请实施方式中，将联邦学习系统中的其中一个服务端作为主节点，由主节点来维护计时器和计数器，在每轮迭代过程中，通过计数器对接收到的请求消息进行计数，并同时使用计时器来进行计时，从而限制每轮迭代参与的客户端的数量以及每轮迭代的训练时长，从而避免因客户端掉队而导致的无意义的等待，提高联邦学习的整体效率，避免出现长尾效应。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。因此，本申请实施方式中，计数器对应的第一阈值可以是预先设定的值，也可以根据接入联邦学习系统内所有服务端的客户端的数量确定的值，通常该第一阈值不大于接入联邦学习系统的客户端的数量，从而避免等待客户端的接入而导致联邦学习的等待时间过长，提高联邦学习的整体效率。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关或者第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。因此，本申请实施方式中，可以通过计时器来设置每轮迭代的时间窗口，当计时器超时时即可进入下一轮迭代学习，从而避免因等待时长过长而导致的长尾效应，即使客户端在进行联邦学习时中途退出也不影响联邦学习的整体训练效率。
为便于理解，参阅图7，下面结合联邦学习系统内由主节点和各个服务端执行的流程进行示例性介绍。
701、初始化计数器和计时器。
其中，在进行每轮迭代之前，对计数器和计时器进行初始化，如将计数器和计时器都设置为0。计数器和计时器可以由主节点维护，如前述图5中所述示出的FL-Server 0。
例如，在每轮迭代开始时，初始化计数器，设置本轮迭代参与的客户端的请求消息的数量，如设置为C，通常这个值可以设置为与服务端有通信连接的客户端的数量的子集，即通常不超过与服务端建立了连接的客户端的数量，以避免每轮迭代中对客户端的等待时 长过长。如若在联邦学习系统中，有100个客户端和服务端建立了连接，而仅有10个客户端需要参与联邦学习，则此时若C设置过大，可能导致等待客户端参与的时长过长，降低联邦学习的效率。
702、FL-Server接收FL-Client的请求消息。
FL-Client发送的请求消息可以由ELB路由至FL-Server，此处的FL-Server可以是联邦学习系统内的任意一个服务端，如前述图5中的FL-Server0、FL-Server1或者FL-Server2等服务端中的任意一个。
此外，在FL-Server接收到FL-Client请求消息之后，FL-Server可以确认是否允许该FL-Client参与联邦学习。如可以采用随机筛选、先到先得或者其他筛选方式来确定是否允许FL-Client参与联邦学习。例如，还可以通过请求消息确定FL-Client的连接状态，如时延、通信时长等，或者FL-Client的状态，如电量过低或者负载过高等，则可以不允许该FL-Client参与联邦学习，以避免FL-Client影响联邦学习的效率。
703、计数器累加。
在FL-Server接收到FL-Client请求消息时，主节点也可以对发送了请求消息的FL-Client数量进行记录，计数器进行累加。
704、判断计数器是否为1，若是，则执行步骤708，若否，则执行步骤705。
其中，在对计数器进行累加之后，可以对计数器的值进行监控，若计数器的值为1，即在本轮迭代过程中，接收到了第一个客户端发送的请求消息，此时即可开启计时器，即执行步骤709，而此时仅因仅接收到一个客户端发送的请求消息，因此无需进行数据同步，在接收到多个客户端发送的请求消息之后，即可后在集群维度进行数据同步，使集群内的服务端保持数据一致性，即执行步骤705。
705、FL-Server处理请求并在集群维度同步数据。
在接收到多个FL-Client发送的请求消息之后，即可向该多个FL-Client中的全部或者部分FL-Client下发模型以及训练配置参数。FL-Client基于训练配置参数，使用本地存储的训练样本对该模型进行训练，并将训练后的模型的相关数据反馈至FL-Server。
若多个FL-Client参与了联邦学习，则FL-Server可以接收多个FL-Client反馈的模型更新参数，FL-Client可以对接收到的多个FL-Client反馈的模型更新参数进行聚合，得到聚合信息(即第一聚合信息)。各个FL-Server之间可以同步各自计算得到的聚合信息，从而实现集群维度的数据同步。
706、FL-Server之间进行聚合。
在各个FL-Server之间同步了各自的聚合信息之后，即可对自身聚合得到的聚合信息(即第一聚合信息)其他FL-Server发送的聚合信息(即第二聚合信息)进行聚合，得到最终的全局聚合信息(即第三聚合信息)，从而使各个FL-Server都可以具有全量数据。
示例性地，FL-Server之间进行聚合的方式具体可以如图8所示，其中，联邦学习系统内包括了多个服务端，该多个服务端呈环形互相连接，即如图8中所示出的Server 0、Server 1、Server 2和Server 3等，此处仅以4个服务端作为示例性说明，实际应用场景中也可以替换为更多或者更少的服务端数量，此处不作限定。客户端包括与服务端连接的K个客 户端，如图8中所示出的client 1至client K，每个客户端接收接入的服务端发送的初始权重w t，然后使用本地存储的训练样本对初始权重对应的模型进行训练，得到训练后的模型的权重
Figure PCTCN2022087647-appb-000020
以及对应的数据量n 1-n k，n k表示第k个客户端的数据集大小，然后将训练后的模型的权重以及数据量反馈给接入的服务端，如client 1将训练后的模型权重反馈给Server 0，client K将训练后得到的模型的权重反馈给Server 2等。服务端j对各自接收到的客户端发送的权重进行聚合，得到
Figure PCTCN2022087647-appb-000021
然后主节点触发各个服务端对服务端之间传输的数据进行聚合，得到聚合信息，如每个服务端聚合后的数据表示为
Figure PCTCN2022087647-appb-000022
和n j。
707、计数是否达到第一阈值，若是，则执行步骤710，若否，则执行步骤702。
在本轮的联邦学习迭代过程层中，可以实时对计数器的数值进行监测，若计数器的计数达到第一阈值，则表示本轮联邦学习迭代过程中参与的客户端的数量已足够，无需在增加更多的客户端来参与本轮联邦学习，即停止计时，并进入下一轮联邦学习的迭代过程，即执行步骤710。
708、触发计时器计时。
在联邦学习的每轮迭代过程中，在接收到第一个由客户端发送的请求消息之后，即可触发计时器开始计时。
例如，在ELB将第一个请求消息路由至对应的服务器的同时或者之后，ELB还通知主节点已接收到第一个请求消息，主节点在接收到ELB的通知消息之后即可启动计时器和计数器，以对联邦学习系统接收到的请求消息进行计数，并对本轮迭代的过程进行计时。
709、是否超时，若是，则执行步骤710。
在每轮迭代学习的过程中，可以对计时器的状态进行实时监测，判断计时器是否超时，若超时，则停止计时，并进行下一轮迭代学习。
可以理解为，相当于通过计时器来设置时间窗口，在计时器的时间窗口内进行本轮迭代学习过程，若超出该时间窗口，则可以进入下一轮迭代中，从而避免每轮迭代消耗的时长过长。
在计时器的时间窗口T内所接收到的客户端发送的请求消息或者数据可以被服务端正常处理，如执行加解密或者对接收到的数据进行聚合等，若在时间窗口T之外若计数器还未达到C，则计时器和计数器都重置为初始状态，进入下一轮迭代；而若在时间窗口T之内计数器达到C，则停止计时器的计时，并在处理完本轮迭代中客户端发送的数据之后，进入下一轮迭代。因此，在本申请实施方式中，在计数器和计时器相互配合的基础上，每轮迭代耗时最长即为时间窗口T的时长，并且在迭代过程中，可以记录每轮迭代的通信时长，从而动态调整时间窗口的时长，从而消除客户端的长尾效应，提升训练效率。
此外，可选地，该计时器的作用也可以是触发集群内的服务端之间进行数据同步，即前述步骤705-706，在完成步骤705-706之后，即可进入下一轮迭代，具体可以根据实际应用场景进行调整。
710、停止计时。
在计时器停止计时之后，即表示结束本轮迭代学习，在初始化计时器和计数器之后，即可进入下一轮迭代，即执行步骤701。
其中，步骤702、步骤705和步骤706可以由集群内的任意一个FL-Server执行，步骤701、步骤703、步骤704、步骤707-步骤710则由主节点来执行。
因此，本申请实施方式中，客户端可以以请求和响应结合的方式接入服务端，通过设置计数器和计时器的方式来对每轮迭代参与的客户端和迭代时长进行监控，即使客户端在联邦学习的中途退出，也不影响集群的联邦学习的整体学习过程，避免因寻址不到客户端而导致的掉尾问题，提高联邦学习的整体效率。并且，各个服务端之间可以进行数据同步，即每个服务端中都具有集群内的全量数据，客户端在任意时刻接入任意服务端都可以获取到全量数据，从而提高客户端得到的模型的准确度。
示例性地，以一个训练简单的分类网络为例。联邦学习系统可以由多台服务端组成，客户端部署于用户的移动终端，如手机、相机或平板等移动终端。移动终端可以通过摄像头采集到图像，用户可以在相册手动标记出拍摄到的图像中的对象的类型，如图像中的“猫”、“狗”等，移动终端可以将用户手动标记后的图像作为训练样本，当训练样本的数量达到一定量时，移动终端可以向联邦学习系统请求参与联邦学习，从而将自身采集到的数据应用到分类网络中。
移动终端可以获取到联邦学习系统对外暴露的HTTP端口，并通过该HTTP端口向集群发送请求消息，主节点上设置的计数器对该请求消息进行计数，且主节点还维护计时器。计数器的上限值可以是预先设定的，也可以是根据上一轮迭代中参与的客户端的数量确定的，或者根据与集群建立了连接的客户端的数量确定的。计时器的上限值可以是预先设定的，也可以是根据上一轮迭代所花费的时长确定的等。
且若该移动终端发送的请求消息为本轮迭代过程中发送的第一个请求消息，则启用计时器。若移动终端发送的请求消息为本轮迭代中的最后一个请求消息，即计数器的值达到预先设定的第一阈值，则该移动终端为本轮迭代中确定的最后一个参与联邦学习的客户端，在移动终端完成模型训练之后，即可进入下一轮迭代学习。
集群内部的ELB将移动终端的请求消息路由至适配的服务端中，由该服务端确定是否允许移动终端参与联邦学习。若服务端允许该移动终端参与联邦学习，服务端将本地存储的分类网络下发给移动终端，同时还下发训练配置参数，如训练时所使用的学习率、数据的加解密方式等。移动终端使用本地采集到的训练样本，根据训练配置参数对分类网络进行训练，得到训练后的分类网络。移动终端将训练后的分类网络的权重参数反馈给服务端，若服务端接收到该权重参数的时刻在计时器的时间窗口内，则服务端对接收到的权重参数进行聚合，得到第一聚合信息。
服务端之间可以互相传输聚合得到的信息，从而使每个服务端上都具有全量数据。移动终端可以在任何时刻向服务端发送查询消息，以查询最新的分类网络，服务端可以将存储的最新的分类网络发送给移动终端，从而时移动终端在任意时刻接入任意服务端都可以获取到最新的全量数据。
因此，在本申请实施方式中，当客户端部署于移动终端时，可以通过请求和响应结合的方式，使客户端以松散的方式接入联邦学习系统并参与联邦学习，无需保持稳定的长连接，适用于客户端部署于移动终端的场景。并且，联邦学习系统内部还部署了计时器和计数器，通过计时器和计数器配合的方式，避免因等待客户端而导致的联邦学习效率低，减少长尾效应，提高了联邦学习的整体效率。此外，计时器或者计数器的上限值可以根据上一轮迭代或者前几轮迭代训练过程中的信息来调整的，可以根据实际场景动态调整计时器或者计数器的上限值，进一步提高联邦学习的效率。
前述对本申请提供的联邦学习系统和联邦学习方法的流程进行了详细介绍，下面对本申请提供的服务端的的结构进行介绍。
参阅图9，本申请提供一种服务端的结构示意图，该服务端应用于联邦学习系统，联邦学习系统中包括多个服务端和多个客户端，多个服务端用于进行迭代联邦学习，多个服务端中的任意一个服务端包括：
收发模块901，用于接收至少一个第一客户端发送的请求消息，请求消息用于请求第一服务器中存储的全局模型，多个客户端包括至少一个第一客户端；
收发模块901，还用于向至少一个第一客户端发送全局模型的信息和训练配置参数，全局模型的信息和训练配置参数用于指示至少一个第一客户端使用训练配置参数对全局模型进行训练；
收发模块901，还用于接收至少一个第一客户端分别反馈的第一模型更新参数，第一模型更新参数是至少一个第一客户端训练后得到的全局模型的参数；
聚合模块902，用于对至少一个第一客户端反馈的第一模型更新参数进行聚合，以得到在本轮迭代中的第一聚合信息；
收发模块901，还用于获取第二服务端发送的第二聚合信息，第二聚合信息为第二服务端在本轮迭代中，对接收到的第二模型更新参数进行聚合得到的信息；
更新模块903，用于基于第一聚合信息和第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
在一种可能的实施方式中，收发模块，还用于接收多个第二服务端分别发送的第二聚合信息；
更新模块903，具体用于基于第一聚合信息和多个第二服务端分别发送的第二聚合信息，更新第一服务端上存储的全局模型，以得到更新后的全局模型。
在一种可能的实施方式中，第一服务器为联邦学习系统中的主节点，主节点用于对多个服务端进行管理，
收发模块，还用于向第二服务器发送第一触发指示，第一触发指示用于指示第二服务端向第一服务端发送第二聚合信息；
收发模块901，还用于接收第二服务端中的第二聚合信息。
在一种可能的实施方式中，第一服务器中设置有计数器，计数器用于对多个服务端接收到的请求消息的数量进行计数，
收发模块901，具体用于当计数器的数值满足第一阈值时，向第二服务端发送第一触 发指示。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。
在一种可能的实施方式中，主节点中包括计时器，计时器从每一轮迭代中接收到第一个请求消息之后开始计时，
收发模块901，具体用于当计时器超过第二阈值时，接收主节点发送的第二触发指示，第二触发指示用于指示进行下一轮迭代学习。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关，或者，第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。
在一种可能的实施方式中，收发模块，还用于接收第三客户端发送的查询消息，第三客户端包括接入联邦学习系统中的客户端中的任意一个；
收发模块901，还用于针对查询消息向第三客户端发送更新后的全局模型的信息。
在一种可能的实施方式中，收发模块，还用于向第二服务端发送第一聚合信息，以使第二服务端基于第一聚合信息和第二聚合信息，更新本地存储的全局模型以得到更新后的全局模型。
参阅图10，本申请提供另一种服务端的结构示意图，即前述的主节点。该服务端应用于联邦学习系统，联邦学习系统包括多个服务端，多个服务端中的其中一个服务端作为主节点，多个服务端用于进行迭代学习以实现联邦学习，主节点包括：
启动模块1001，用于在多个服务端中任意一个服务端接收到第一个请求消息之后，主节点启动计数器和计时器，计数器用于对多个服务端在一轮迭代中接收到的请求消息进行计数，请求消息用于请求获取多个服务端中对应的服务端中存储的全局模型；
收发模块1002，用于若计数器的值达到第一阈值，主节点向多个服务端中的每个服务端发送第一触发指示，第一触发指示用于指示多个服务端之间互相传输本地存储的信息；
收发模块1002，还用于若计数器的值未达到第一阈值，计时器的值达到第二阈值，则主节点向每个服务端发送第二触发指示，第二触发指示用于指示每个服务端进行下一轮迭代。
在一种可能的实施方式中，第一阈值为预设值，或者，第一阈值与上一轮迭代过程中接入联邦学习系统的客户端的数量相关。
在一种可能的实施方式中，第二阈值为预设值，或者，第二阈值与上一轮迭代中接入联邦学习系统的客户端的数量相关或者第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。
本申请实施例还提供一种服务端，请参阅图11，图11为本申请实施例提供的服务端的一种结构示意图。服务端1100上可以部署有图9对应实施例中所描述的服务端，用于实现图4至图8对应实施例中服务端的功能。具体的，服务端1100可因配置或性能不同而产生比较大的差异，可以包括一个或一个以上中央处理器CPU1122(例如，一个或一个以上处理器)和存储器1132，一个或一个以上存储应用程序1142或数据1144的存储介质1130 (例如一个或一个以上海量存储设备)。
其中，存储器1132和存储介质1130可以是短暂存储或持久存储。在一个实施例中，存储器1132为随机存储存储器RAM，可以与中央处理器1122直接交换数据，用于加载数据1144和应用程序1142和/或操作系统1141以供中央处理器1122直接运行与运用，通常作为操作系统或其他正在运行中的程序的临时数据存储媒介。存储在存储介质1130的程序可以包括一个或一个以上模块(图11中未示出)，每个模块可以包括对服务端中的一系列指令操作。
更进一步地，中央处理器1122可以设置为与存储介质1130通信，在服务端1100上执行存储介质1130中的一系列指令操作。在一种可能的实施方式中，存储介质1130中存储了前述图4-图8中任一实施例所示的方法步骤对应的程序指令以及数据。
服务端1100还可以包括一个或一个以上电源1126，一个或一个以上有线或无线网络接口1150，一个或一个以上输入输出接口1158，和/或，一个或一个以上操作系统1141，例如Windows Server TM，Mac OS X TM，Unix TM，Linux TM，FreeBSD TM等等。
在一种可能的实施方式中，中央处理器1122用于执行前述图4-图8中任一实施例所示的服务端执行的方法步骤。
本申请实施例还提供了一种联邦学习装置，该联邦学习装置也可以称为数字处理芯片或者芯片，芯片包括处理单元和通信接口，处理单元可以通过通信接口获取程序指令，程序指令被处理单元执行，处理单元用于执行前述图4-图8中任一实施例所示的服务端执行的方法步骤。
本申请实施例还提供一种数字处理芯片。该数字处理芯片中集成了用于实现上述中央处理器1122，或者中央处理器1122的功能的电路和一个或者多个接口。当该数字处理芯片中集成了存储器时，该数字处理芯片可以完成前述实施例中的任一个或多个实施例的方法步骤。当该数字处理芯片中未集成存储器时，可以通过通信接口与外置的存储器连接。该数字处理芯片根据外置的存储器中存储的程序代码来实现上述实施例中服务端执行的动作。
更详细地，当本申请实施例提供的联邦学习装置为芯片时，该芯片具体包括：处理单元和通信单元，所述处理单元例如可以是处理器，所述通信单元例如可以是输入/输出接口、管脚或电路等。该处理单元可执行存储单元存储的计算机执行指令，以使服务端内的芯片执行上述图4-图8所示实施例描述的联邦学习方法。
可选地，前述的存储单元可以为芯片内的存储单元，如寄存器、缓存等，所述存储单元还可以是所述无线接入设备端内的位于所述芯片外部的存储单元，如只读存储器(read-only memory，ROM)或可存储静态信息和指令的其他类型的静态存储设备，随机存取存储器RAM等。
具体地，前述的处理单元或者处理器可以是中央处理器、网络处理器(neural-network processing unit，NPU)、图形处理器(graphics processing unit，GPU)、数字信号处理器(digital signal processor，DSP)、专用集成电路(application specific integrated circuit,ASIC)或现场可编程逻辑门阵列(field programmable gate array，FPGA)或 者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者也可以是任何常规的处理器等。
其中，上述任一处提到的处理器，可以是一个通用中央处理器，微处理器，ASIC，或一个或多个用于控制上述图4-图8的方法的程序执行的集成电路。
本申请实施例中还提供一种计算机可读存储介质，该计算机可读存储介质中存储有程序，当其在计算机上运行时，使得计算机执行如前述图4-图8所示实施例描述的方法中的步骤。
本申请实施例中还提供一种计算机程序产品，当其在计算机上运行时，使得计算机执行如前述图4-图8所示实施例描述的方法中服务端所执行的步骤。
另外需说明的是，以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。另外，本申请提供的装置实施例附图中，模块之间的连接关系表示它们之间具有通信连接，具体可以实现为一条或多条通信总线或信号线。
通过以上的实施方式的描述，所属领域的技术人员可以清楚地了解到本申请可借助软件加必需的通用硬件的方式来实现，当然也可以通过专用硬件包括专用集成电路、专用CPU、专用存储器、专用元器件等来实现。一般情况下，凡由计算机程序完成的功能都可以很容易地用相应的硬件来实现，而且，用来实现同一功能的具体硬件结构也可以是多种多样的，例如模拟电路、数字电路或专用电路等。但是，对本申请而言更多情况下软件程序实现是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在可读取的存储介质中，如计算机的软盘、U盘、移动硬盘、只读存储器ROM、随机存取存储器RAM、磁碟或者光盘等，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务端，或者网络设备等)执行本申请各个实施例所述的方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。
所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务端或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务端或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存储的任何可用介质或者是包含一个或多个可用介质集成的服务端、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘(solid state disk，SSD))等。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”、“第四”等(如果存在)是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的实施例能够以除了在这里图示或描述的内容以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
最后应说明的是：以上，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。

Claims (23)

  1. 一种联邦学习方法，其特征在于，
    应用于联邦学习系统，所述联邦学习系统中包括多个服务端和多个客户端，所述多个服务端包括第一服务端和第二服务端，所述多个服务端用于进行迭代学习以实现联邦学习，其中所述迭代学习中的任意一轮迭代的过程包括：
    所述第一服务端接收至少一个第一客户端发送的请求消息，所述请求消息用于请求所述第一服务器中存储的全局模型；
    所述第一服务端向所述至少一个第一客户端发送所述全局模型的信息和训练配置参数；
    所述第一服务端接收至少一个第一客户端分别反馈的第一模型更新参数，所述第一模型更新参数是所述至少一个第一客户端训练的全局模型的参数；
    所述第一服务端对所述至少一个第一客户端反馈的第一模型更新参数进行聚合，以得到在本轮迭代中的第一聚合信息；
    所述第一服务端获取所述第二服务端发送的第二聚合信息，所述第二聚合信息为所述第二服务端在本轮迭代中，对接收到的第二模型更新参数进行聚合得到的信息；
    所述第一服务端基于所述第一聚合信息和所述第二聚合信息，更新所述第一服务端上存储的所述全局模型，以得到更新后的全局模型。
  2. 根据权利要求1所述的方法，其特征在于，所述多个服务端中包括的所述第二服务端的数量为多个，所述第一服务端获取所述第二服务端发送的第二聚合信息，包括：
    所述第一服务端接收所述多个第二服务端分别发送的所述第二聚合信息；
    所述第一服务端基于所述第一聚合信息和所述第二聚合信息，更新所述第一服务端上存储的所述全局模型，包括：
    所述第一服务端基于所述第一聚合信息和所述多个第二服务端分别发送的所述第二聚合信息，更新所述第一服务端上存储的所述全局模型，以得到更新后的全局模型。
  3. 根据权利要求1或2所述的方法，其特征在于，所述第一服务端为所述联邦学习系统中的主节点，所述主节点用于对所述多个服务端进行管理，所述第一服务端获取第二服务端发送的第二聚合信息，还包括：
    所述第一服务端向所述第二服务器发送第一触发指示，所述第一触发指示用于指示所述第二服务端向所述第一服务端发送所述第二聚合信息；
    所述第一服务端接收所述第二服务端中的第二聚合信息。
  4. 根据权利要求2所述的方法，其特征在于，所述第一服务器中设置有计数器，所述计数器用于对所述多个服务端接收到的所述请求消息的数量进行计数，所述第一服务端向所述第二服务器发送第一触发指示，包括：
    当所述计数器的数值满足第一阈值时，所述第一服务端向所述第二服务端发送所述第一触发指示。
  5. 根据权利要求4所述的方法，其特征在于，所述第一阈值为预设值，或者，所述第一阈值与上一轮迭代过程中接入所述联邦学习系统的客户端的数量相关。
  6. 根据权利要求3-5中任一项所述的方法，其特征在于，所述主节点中包括计时器， 所述计时器从每一轮迭代中接收到第一个请求消息之后开始计时，所述方法还包括：
    当所述计时器超过第二阈值时，所述第一服务端向所述第二服务器发送的第二触发指示，所述第二触发指示用于指示所述第二服务器进行下一轮迭代。
  7. 根据权利要求6所述的方法，其特征在于，所述第二阈值为预设值，或者，所述第二阈值与上一轮迭代中接入所述联邦学习系统的客户端的数量相关，或者，所述第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。
  8. 根据权利要求1-7中任一项所述的方法，其特征在于，所述方法还包括：
    所述第一服务端接收第三客户端发送的查询消息，所述第三客户端包括接入所述联邦学习系统中的客户端中的任意一个；
    所述第一服务端针对所述查询消息向所述第三客户端发送更新后的全局模型的信息。
  9. 根据权利要求1-8中任一项所述的方法，其特征在于，所述方法还包括：
    所述第一服务端向所述第二服务端发送所述第一聚合信息，以使所述第二服务端基于所述第一聚合信息和所述第二聚合信息，更新本地存储的所述全局模型以得到所述更新后的全局模型。
  10. 一种联邦学习方法，其特征在于，应用于联邦学习系统和多个客户端，所述联邦学习系统包括多个服务端，所述多个服务端中的其中一个服务端作为主节点，所述多个服务端用于进行迭代学习以实现联邦学习；
    其中所述迭代学习中的任意一轮迭代的过程包括：
    在所述多个服务端中任意一个服务端接收到第一个请求消息之后，所述主节点启动计数器和计时器，所述计数器用于对所述多个服务端在一轮迭代中接收到的请求消息进行计数，所述请求消息用于客户端请求获取所述多个服务端中对应的服务端中存储的全局模型；
    若所述计数器的值达到第一阈值，所述主节点向所述多个服务端中的每个服务端发送第一触发指示，所述第一触发指示用于指示所述多个服务端之间互相传输本地存储的信息；
    若所述计数器的值未达到所述第一阈值，所述计时器的值达到第二阈值，则所述主节点向所述每个服务端发送第二触发指示，所述第二触发指示用于指示所述每个服务端进行下一轮迭代。
  11. 根据权利要求10所述的方法，其特征在于，所述第一阈值为预设值，或者，所述第一阈值与上一轮迭代过程中接入所述联邦学习系统的客户端的数量相关。
  12. 根据权利要求10或11所述的方法，其特征在于，所述第二阈值为预设值，或者，所述第二阈值与上一轮迭代中接入所述联邦学习系统的客户端的数量相关或者所述第二阈值与上一轮迭代中联邦学习系统内的每个服务端和对应的客户端之间进行通信的数据量的大小相关。
  13. 一种联邦学习系统，其特征在于，包括多个服务端和多个客户端，所述多个服务端包括第一服务端和第二服务端，所述第一服务端和所述第二服务端上均存储有全局模型的信息，所述多个服务端用于进行迭代学习以实现联邦学习，其中所述迭代学习中的任意一轮迭代过程中：
    所述第一服务端用于接收至少一个第一客户端分别发送的请求消息；
    所述第一服务端用于针对所述至少一个第一客户端分别发送的请求消息向所述至少一个第一客户端发送所述全局模型的信息和训练配置参数；
    所述第一服务端用于接收所述至少一个第一客户端分别反馈的第一模型更新参数，所述第一模型更新参数为所述至少一个第一客户端训练后得到的全局模型的参数；
    所述第一服务端用于对所述至少一个第一客户端分别反馈的第一模型更新参数进行聚合，得到第一聚合信息；
    所述第二服务端用于接收对应的至少一个第二客户端发送的第二模型更新参数，并对所述至少一个第二客户端发送的第二模型更新参数进行聚合，以得到第二聚合信息；
    所述第一服务端用于接收所述每个第二服务端发送的第二聚合信息；
    所述第一服务端用于基于所述第一聚合信息和所述每个第二服务端发送的第二聚合信息，更新所述第一服务端上存储的所述全局模型，以得到更新后的全局模型。
  14. 根据权利要求13所述的系统，其特征在于，所述多个服务端中包括的所述第二服务端的数量为多个；
    所述多个第二服务端中的每个第二服务端用于接收对应的至少一个第二客户端发送的第二模型更新参数，并对所述至少一个第二客户端发送的第二模型更新参数进行聚合，以得到所述第二聚合信息；
    所述第一服务端具体用于接收所述多个第二服务端分别发送的所述第二聚合信息；
    所述第一服务端具体用于基于所述第一聚合信息和所述多个第二服务端分别发送的所述第二聚合信息，更新所述第一服务端上存储的所述全局模型，以得到更新后的全局模型。
  15. 根据权利要求13或14所述的系统，其特征在于，所述多个服务端中还包括作为主节点的第三服务端，所述主节点用于对所述多个服务端进行管理；
    所述主节点用于向所述多个服务端分别发送第一触发指示；
    所述第二服务器具体用于基于所述第一触发指示向所述第一服务器发送所述第二聚合信息。
  16. 根据权利要求15所述的系统，其特征在于，所述主节点中包括计数器，所述计数器用于对所述多个服务端接收到的请求消息的数量进行计数，所述请求消息用于请求获取所述多个服务端中对应的服务端中存储的全局模型；
    所述主节点具体用于当所述计数器的数值满足第一阈值时，向所述多个服务端中的每个服务端发送所述第一触发指示，所述第一触发指示用于触发所述每个第二服务器向所述第一服务器发送所述第二聚合信息。
  17. 根据权利要求15或16所述的系统，其特征在于，所述主节点还包括计时器，所述计时器从每一轮迭代过程中接收到的第一个所述请求消息开始计时；
    所述主节点还用于当所述计时器超过第二阈值时，向所述多个服务端中的每个服务端分别发送第二触发指示，所述第二触发指示用于指示所述多个服务端进行下一轮迭代。
  18. 根据权利要求17所述的系统，其特征在于，所述第二阈值为预设值，或者，所述第二阈值与上一轮迭代过程中接入所述联邦学习系统中每个服务端的客户端的数量相关， 或者，所述第二阈值与上一轮迭代过程中所述多个服务端和所述多个客户端之间进行通信的数据量的大小相关。
  19. 根据权利要求13-18中任一项所述的系统，其特征在于，
    所述第一服务端接收第三客户端发送的查询消息，所述第三客户端包括接入所述联邦学习系统的任意一个客户端；
    所述第一服务端针对所述查询消息向所述第三客户端发送更新后的全局模型的信息。
  20. 根据权利要求13-19中任一项所述的系统，其特征在于，
    所述第一服务端还用于向所述第二服务端发送所述第一聚合信息；
    所述第二服务器端具体用于结合所述第一聚合信息和所述第二聚合信息，更新本地存储的所述全局模型，以得到所述更新后的全局模型。
  21. 一种联邦学习装置，其特征在于，包括处理器，所述处理器和存储器耦合，所述存储器存储有程序，当所述存储器存储的程序指令被所述处理器执行时实现权利要求1-9或者10-12中任一项所述的方法的步骤。
  22. 一种计算机可读存储介质，包括程序，当其被处理单元所执行时，执行如权利要求1至9或者10-12中任一项所述的方法的步骤。
  23. 一种计算机程序产品，其特征在于，所述计算机程序产品包括软件代码，所述软件代码用于执行如权利要求1至9或者10-12中任一项所述的方法的步骤。
PCT/CN2022/087647 2021-04-25 2022-04-19 一种联邦学习方法以及装置 Ceased WO2022228204A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP22794675.3A EP4318338A4 (en) 2021-04-25 2022-04-19 FEDERATE LEARNING METHOD AND APPARATUS
US18/493,136 US20240054354A1 (en) 2021-04-25 2023-10-24 Federated learning method and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110450585.7A CN115249073B (zh) 2021-04-25 2021-04-25 一种联邦学习方法以及装置
CN202110450585.7 2021-04-25

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/493,136 Continuation US20240054354A1 (en) 2021-04-25 2023-10-24 Federated learning method and apparatus

Publications (1)

Publication Number Publication Date
WO2022228204A1 true WO2022228204A1 (zh) 2022-11-03

Family

ID=83696603

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/087647 Ceased WO2022228204A1 (zh) 2021-04-25 2022-04-19 一种联邦学习方法以及装置

Country Status (4)

Country Link
US (1) US20240054354A1 (zh)
EP (1) EP4318338A4 (zh)
CN (1) CN115249073B (zh)
WO (1) WO2022228204A1 (zh)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115775010A (zh) * 2022-11-23 2023-03-10 国网江苏省电力有限公司信息通信分公司 基于横向联邦学习的电力数据共享方法
CN115905978A (zh) * 2022-11-18 2023-04-04 安徽工业大学 基于分层联邦学习的故障诊断方法及系统
CN115983409A (zh) * 2022-11-11 2023-04-18 北京大学 基于差分隐私的联邦学习训练方法、装置、系统及设备
CN116384506A (zh) * 2023-03-28 2023-07-04 支付宝(杭州)信息技术有限公司 一种模型训练的方法、装置、存储介质及电子设备
CN117035009A (zh) * 2023-08-24 2023-11-10 重庆大学 一种基于数据集压缩的个性化通信方法
CN117196010A (zh) * 2023-08-03 2023-12-08 西安电子科技大学杭州研究院 一种面向网联汽车的区块链分布式联邦学习方法、系统及终端
CN117390448A (zh) * 2023-10-25 2024-01-12 西安交通大学 一种用于云际联邦学习的客户端模型聚合方法及相关系统
CN117436515A (zh) * 2023-12-07 2024-01-23 四川警察学院 联邦学习方法、系统、装置以及存储介质
CN117472866A (zh) * 2023-12-27 2024-01-30 齐鲁工业大学(山东省科学院) 一种区块链监管与激励下的联邦学习数据共享方法
WO2024263820A3 (en) * 2023-06-20 2025-02-06 Xeba Technologies, LLC Object messaging and intelligent objects (omio)
CN119558423A (zh) * 2023-09-01 2025-03-04 安徽工业大学 一种基于客户端目标增强的个性化联邦学习方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20240028870A1 (en) * 2022-07-22 2024-01-25 Capital One Services, Llc Selective reporting of machine learning parameters for federated learning
US20240104426A1 (en) * 2022-09-26 2024-03-28 Qualcomm Incorporated MACHINE LEARNING SIGNALING AND OPERATIONS FOR WIRELESS LOCAL AREA NETWORKS (WLANs)
CN116383755B (zh) * 2022-12-30 2026-04-17 科大讯飞股份有限公司 节点协同方法、装置、网络、节点设备和存储介质
CN116011991B (zh) * 2022-12-30 2023-12-19 中国电子科技集团公司第三十八研究所 基于代理和备份技术的多人协同任务保障方法
WO2025054971A1 (en) * 2023-09-15 2025-03-20 Huawei Technologies Co., Ltd. Apparatus, method and readable storage medium for model training
CN118446335A (zh) * 2024-05-07 2024-08-06 中国电信股份有限公司技术创新中心 联邦学习方法、装置、系统、客户端及服务器
CN121262286B (zh) * 2025-12-04 2026-04-24 厦门理工学院 一种冲突感知的多服务器联邦学习客户端调度方法、装置

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180331897A1 (en) * 2016-01-30 2018-11-15 Huawei Technologies Co.,Ltd. Method and device for training model in distributed system
US10372504B2 (en) * 2017-08-03 2019-08-06 Akamai Technologies, Inc. Global usage tracking and quota enforcement in a distributed computing system
CN110175283A (zh) * 2019-05-10 2019-08-27 深圳前海微众银行股份有限公司 一种推荐模型的生成方法及装置
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
US20200027033A1 (en) * 2018-07-19 2020-01-23 Adobe Inc. Updating Machine Learning Models On Edge Servers
CN112257874A (zh) * 2020-11-13 2021-01-22 腾讯科技（深圳）有限公司 分布式机器学习系统的机器学习方法、装置、系统
CN112686393A (zh) * 2020-12-31 2021-04-20 华南理工大学 一种联邦学习系统

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111444021B (zh) * 2020-04-02 2023-03-24 电子科技大学 基于分布式机器学习的同步训练方法、服务器及系统
CN112181666B (zh) * 2020-10-26 2023-09-01 华侨大学 一种基于边缘智能的设备评估和联邦学习重要性聚合方法
CN112532451B (zh) * 2020-11-30 2022-04-26 安徽工业大学 基于异步通信的分层联邦学习方法、装置、终端设备及存储介质
CN112508205B (zh) * 2020-12-04 2024-07-16 中国科学院深圳先进技术研究院 一种联邦学习调度方法、装置及系统
CN112668726B (zh) * 2020-12-25 2023-07-11 中山大学 一种高效通信且保护隐私的个性化联邦学习方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180331897A1 (en) * 2016-01-30 2018-11-15 Huawei Technologies Co.,Ltd. Method and device for training model in distributed system
US10372504B2 (en) * 2017-08-03 2019-08-06 Akamai Technologies, Inc. Global usage tracking and quota enforcement in a distributed computing system
US20200027033A1 (en) * 2018-07-19 2020-01-23 Adobe Inc. Updating Machine Learning Models On Edge Servers
CN110175283A (zh) * 2019-05-10 2019-08-27 深圳前海微众银行股份有限公司 一种推荐模型的生成方法及装置
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
CN112257874A (zh) * 2020-11-13 2021-01-22 腾讯科技（深圳）有限公司 分布式机器学习系统的机器学习方法、装置、系统
CN112686393A (zh) * 2020-12-31 2021-04-20 华南理工大学 一种联邦学习系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4318338A4

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115983409A (zh) * 2022-11-11 2023-04-18 北京大学 基于差分隐私的联邦学习训练方法、装置、系统及设备
CN115905978A (zh) * 2022-11-18 2023-04-04 安徽工业大学 基于分层联邦学习的故障诊断方法及系统
CN115775010B (zh) * 2022-11-23 2024-03-19 国网江苏省电力有限公司信息通信分公司 基于横向联邦学习的电力数据共享方法
CN115775010A (zh) * 2022-11-23 2023-03-10 国网江苏省电力有限公司信息通信分公司 基于横向联邦学习的电力数据共享方法
CN116384506A (zh) * 2023-03-28 2023-07-04 支付宝(杭州)信息技术有限公司 一种模型训练的方法、装置、存储介质及电子设备
WO2024263820A3 (en) * 2023-06-20 2025-02-06 Xeba Technologies, LLC Object messaging and intelligent objects (omio)
CN117196010A (zh) * 2023-08-03 2023-12-08 西安电子科技大学杭州研究院 一种面向网联汽车的区块链分布式联邦学习方法、系统及终端
CN117035009A (zh) * 2023-08-24 2023-11-10 重庆大学 一种基于数据集压缩的个性化通信方法
CN119558423A (zh) * 2023-09-01 2025-03-04 安徽工业大学 一种基于客户端目标增强的个性化联邦学习方法
CN117390448A (zh) * 2023-10-25 2024-01-12 西安交通大学 一种用于云际联邦学习的客户端模型聚合方法及相关系统
CN117390448B (zh) * 2023-10-25 2024-04-26 西安交通大学 一种用于云际联邦学习的客户端模型聚合方法及相关系统
CN117436515B (zh) * 2023-12-07 2024-03-12 四川警察学院 联邦学习方法、系统、装置以及存储介质
CN117436515A (zh) * 2023-12-07 2024-01-23 四川警察学院 联邦学习方法、系统、装置以及存储介质
CN117472866B (zh) * 2023-12-27 2024-03-19 齐鲁工业大学(山东省科学院) 一种区块链监管与激励下的联邦学习数据共享方法
CN117472866A (zh) * 2023-12-27 2024-01-30 齐鲁工业大学(山东省科学院) 一种区块链监管与激励下的联邦学习数据共享方法

Also Published As

Publication number Publication date
EP4318338A1 (en) 2024-02-07
US20240054354A1 (en) 2024-02-15
CN115249073B (zh) 2026-03-24
CN115249073A (zh) 2022-10-28
EP4318338A4 (en) 2024-09-25

Similar Documents

Publication Publication Date Title
WO2022228204A1 (zh) 一种联邦学习方法以及装置
JP7383803B2 (ja) 不均一モデルタイプおよびアーキテクチャを使用した連合学習
US10163420B2 (en) System, apparatus and methods for adaptive data transport and optimization of application execution
US11716271B2 (en) Automated data flows using flow-based data processor blocks
CN110390246A (zh) 一种边云环境中的视频分析方法
CN116685985A (zh) 具有多样化反馈的联合学习系统与方法
Rani et al. Blockchain-based IoT enabled health monitoring system: P. Rani et al.
CN112286666B (zh) 基于回调机制的细粒度数据流可靠卸载方法
WO2020053887A1 (en) Distributed training of systems for medical image analysis
US20220156235A1 (en) Automatic generation of labeled data in iot systems
CN113537495A (zh) 基于联邦学习的模型训练系统、方法、装置和计算机设备
US20220172054A1 (en) Intermediate network node and method performed therein for handling data of communication networks
US12045695B2 (en) System and methods for privacy preserving cross-site federated learning
Zhang et al. Toy-IoT-oriented data-driven CDN performance evaluation model with deep learning
Dunne et al. A comparison of data streaming frameworks for anomaly detection in embedded systems
Kaur et al. Federated learning in IoT: A survey from a resource-constrained perspective
WO2024041119A1 (zh) 数据备份方法和装置
Rashid et al. Edgestore: Towards an edge-based distributed storage system for emergency response
KR20240114208A (ko) 다층 건물에서 상황 인식 학습 모델의 성능 향상을 위한 군집과 연합 결합 학습 방법 및 시스템
Agarwal et al. ANN-based scalable video encoding method for crime surveillance-intelligence of things applications
CN116389175B (zh) 流量数据检测方法、训练方法、系统、设备及介质
Latif et al. Cloudlet federation based context-aware federated learning approach
Challoob et al. Enhancing the performance assessment of network-based and machine learning for module availability estimation
CN116405516A (zh) 一种基于云边协同的工业数据库管控系统及方法
Miao et al. Data Exchange Mechanism for Real‐Time Object Detection in Cloud‐Edge IoT System

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22794675

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2022794675

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2022794675

Country of ref document: EP

Effective date: 20231026

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 11202307990V

Country of ref document: SG