Communication cost control method and system for homomorphic encryption federated learning based on critical learning period identification

Abstract

The present invention provides a homomorphic encryption federated learning communication cost control method and system based on critical learning period identification. The method identifies the client clusters allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, monitors the client clusters, and obtains the working status characteristic information of all clients. The method also monitors the loss value change data of the model during the training process to distinguish and determine the critical learning period and non-critical learning period of the model during the training process. The method determines the client connection change information of the model during the training process based on the time attribute information of each of the critical learning period and the non-critical learning period, and quantitatively calibrates the client connection change. The method also adjusts the communication connection status between the server and the client based on the client connection change information and the working status characteristic information, effectively reducing the communication overhead of the model training without affecting the accuracy of the model, thereby ensuring the working reliability and security of the homomorphic encryption federated learning model.

Description

Technical Field

[0001] The present invention relates to the field of federated learning, and in particular to a method and system for controlling communication costs of homomorphic encrypted federated learning based on identification of key learning periods. Background Art

[0002] Homomorphically encrypted federated learning models can effectively protect the data model security during the federated learning process. However, to ensure the normal operation of homomorphically encrypted federated learning models, they require a corresponding amount of communication overhead. Currently, compression of homomorphically encrypted federated learning models is used to improve their training efficiency. However, this also causes information loss, increases the difficulty of model convergence, and affects model accuracy. Therefore, existing traditional compression methods for homomorphically encrypted federated learning models cannot simultaneously maintain model accuracy and reduce communication overhead, and cannot improve the reliability and security of homomorphically encrypted federated learning models. Summary of the Invention

[0003] The purpose of the present invention is to provide a homomorphic encryption federated learning communication cost control method and system based on critical learning period identification, which identifies the client cluster allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, and monitors the client cluster to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process, thereby distinguishing and determining the critical learning period and non-critical learning period of the model in the training process, and quantifying and distinguishing the importance of the model at different stages in the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection change is quantitatively calibrated; it also adjusts the communication connection status between the server and the client based on the client connection change information and the working status characteristic information, effectively reducing the communication overhead of the model training without affecting the accuracy of the model, and ensuring the working reliability and security of the homomorphic encryption federated learning model.

[0004] The present invention is achieved through the following technical solutions:

[0005] A communication cost control method for homomorphic encryption federated learning based on critical learning period identification includes:

[0006] Identify the server where the homomorphically encrypted federated learning model is trained to obtain communication attribute information of the server within the network; based on the communication attribute information, determine the client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client;

[0007] Monitoring the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; determining all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; and determining client connection change information of the homomorphically encrypted federated learning model during the training process based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model;

[0008] Based on the client connection change information and the working status characteristic information, the communication connection status between the server and the client is adjusted; and based on the real-time training status of the homomorphic encryption federated learning model, the communication authority of the server within the network is adjusted.

[0009] Optionally, a server where the homomorphically encrypted federated learning model is trained is identified to obtain communication attribute information of the server within the network; based on the communication attribute information, a client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model is determined, and all clients under the client cluster are monitored to obtain working status characteristic information of each of the clients, including:

[0010] Comparing the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection lists of all gateways within the network to determine the location information and gateway bandwidth information of the gateway accessed by the server within the network; based on the gateway location information, determining all client clusters within a preset distance range from the server within the network; and then, based on the gateway bandwidth information, selecting a client cluster with a matching communication bandwidth value from all client clusters within the preset distance range as the client cluster allowed to connect to the server;

[0011] Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time to-be-processed task volume of all the clients, which is used as the working status feature information.

[0012] Optionally, monitoring all clients under the client cluster to obtain working status characteristic information of all clients includes:

[0013] In step S1, monitoring all clients under the client cluster is performed by allocating a monitoring time within a corresponding cycle to each client under the client cluster by the server, thereby achieving the effect of time-division multiplexing monitoring. First, the following formula (1) is used to control the duration of a cycle of time-division multiplexing monitoring by the server according to the number of all clients under the client cluster and the maximum amount of data that each client can generate.

[0014]

[0015] In the above formula (1), T represents the duration of a period of time-division multiplexing monitoring by the server; n represents the number of all clients under the client cluster; T0 represents the duration of processing unit data volume by the server; S0 represents unit data volume; e represents a natural constant; S max (r) represents the maximum amount of data that can be generated by the rth client;

[0016] Step S2, using the following formula (2), according to the usage level of each client under the client cluster and the maximum amount of data that each client can generate, controls the listening time within the cycle allocated to each client during the time-division multiplexing monitoring of the server,

[0017]

[0018] In the above formula (2), T # (r) represents the listening time in the period allocated to the r-th client; D(r) represents the usage degree of the r-th client;

[0019] Step S3, using the following formula (3), based on the amount of characteristic information data of the working status of each client monitored in each monitoring cycle, the usage level of each client under the client cluster is updated.

[0020]

[0021] In the above formula (3), D′(r) represents the updated usage level of the r-th client; J(r) represents the amount of characteristic information data of the working status of the r-th client monitored during the monitoring period.

[0022] Optionally, monitoring the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; determining all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; and determining client connection change information of the homomorphically encrypted federated learning model during the training process based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, including:

[0023] Determine a monitoring frequency of the training process of the homomorphically encrypted federated learning model based on a data training speed during the training process, thereby obtaining loss value change data of the homomorphically encrypted federated learning model during the training process; determine an average loss rate of the homomorphically encrypted federated learning model in all training stages based on the loss value change data; if the average loss rate is greater than a preset loss rate threshold, determine the corresponding training stage as a critical learning period of the homomorphically encrypted federated learning model during the training process; otherwise, determine the corresponding training stage as a non-critical learning period of the homomorphically encrypted federated learning model during the training process;

[0024] Obtain the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model, and based on the time length information, determine the client connection increase information in each critical learning period and the client connection decrease information in each non-critical learning period of the homomorphic encryption federated learning model, and use this as the client connection change information.

[0025] Optionally, adjusting the communication connection state between the server and the client based on the client connection change information and the working state characteristic information; and adjusting the communication authority of the server within the network based on the real-time training state of the homomorphic encryption federated learning model, including:

[0026] When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet a preset remaining available computing power condition are selected from all clients currently connected to the server, thereby interrupting the connections between all selected clients and the server;

[0027] When the homomorphic encryption federated learning model is in a critical learning period, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect based on the client connection increase information and the work status characteristic information, so as to connect all the selected clients to the server;

[0028] Based on the real-time training convergence change state of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the communication authority of the server within the network; if not, keep the current communication authority of the server within the network unchanged.

[0029] A communication cost control system for homomorphic encrypted federated learning based on critical learning period identification, including:

[0030] A communication status identification module is used to identify the server where the homomorphic encryption federated learning model is trained and obtain the communication attribute information of the server within the network;

[0031] A client selection and monitoring module is configured to determine, based on the communication attribute information, a client cluster that the server is allowed to connect to during training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client;

[0032] a critical learning period identification module, configured to monitor the training process of the homomorphically encrypted federated learning model and obtain loss value change data of the homomorphically encrypted federated learning model during the training process; and determine all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data;

[0033] a client connection change determination module, configured to determine client connection change information of the homomorphically encrypted federated learning model during training based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model;

[0034] a communication connection state adjustment module, configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information;

[0035] A communication authority adjustment module is used to adjust the communication authority of the server within the network based on the real-time training status of the homomorphic encryption federated learning model.

[0036] Optionally, the communication state identification module is used to identify the server where the homomorphic encryption federated learning model is trained, and obtain communication attribute information of the server within the network, including:

[0037] Compare the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection list of all gateways within the network to determine the location information and gateway bandwidth information of the gateway connected to the server within the network;

[0038] The client selection and monitoring module is used to determine, based on the communication attribute information, a client cluster that the server is allowed to connect to during the training of the homomorphic encryption federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client, including:

[0039] Based on the gateway location information, determining all client clusters whose connection paths to the server within the network are within a preset distance range; and based on the gateway bandwidth information, selecting a client cluster with a matching communication bandwidth value from all client clusters within the preset distance range as the client cluster allowed to connect to the server;

[0040] Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time to-be-processed task volume of all the clients, which is used as the working status feature information.

[0041] Optionally, the critical learning period identification module is configured to monitor the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; and based on the loss value change data, determine all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process, including:

[0042] Determine a monitoring frequency of the training process of the homomorphically encrypted federated learning model based on a data training speed during the training process, thereby obtaining loss value change data of the homomorphically encrypted federated learning model during the training process; determine an average loss rate of the homomorphically encrypted federated learning model in all training stages based on the loss value change data; if the average loss rate is greater than a preset loss rate threshold, determine the corresponding training stage as a critical learning period of the homomorphically encrypted federated learning model during the training process; otherwise, determine the corresponding training stage as a non-critical learning period of the homomorphically encrypted federated learning model during the training process;

[0043] The client connection change determination module is used to determine client connection change information of the homomorphically encrypted federated learning model during training based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, including:

[0044] Obtain the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model, and based on the time length information, determine the client connection increase information in each critical learning period and the client connection decrease information in each non-critical learning period of the homomorphic encryption federated learning model, and use this as the client connection change information.

[0045] Optionally, the connection state adjustment module is configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information, including:

[0046] When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet a preset remaining available computing power condition are selected from all clients currently connected to the server, thereby interrupting the connections between all selected clients and the server;

[0047] When the homomorphic encryption federated learning model is in a critical learning period, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect based on the client connection increase information and the work status characteristic information, so as to connect all the selected clients to the server;

[0048] The communication authority adjustment module is used to adjust the communication authority of the server within the network based on the real-time training status of the homomorphic encryption federated learning model, including:

[0049] Based on the real-time training convergence change state of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the communication authority of the server within the network; if not, keep the current communication authority of the server within the network unchanged.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The homomorphic encryption federated learning communication cost control method and system based on critical learning period identification provided in the present application identifies the client cluster allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, and monitors the client cluster to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process, thereby distinguishing and determining the critical learning period and non-critical learning period of the model in the training process, and quantifying and distinguishing the importance of the model at different stages in the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection change is quantitatively calibrated; it also adjusts the communication connection status between the server and the client based on the client connection change information and the working status characteristic information, effectively reducing the communication overhead of the model training without affecting the accuracy of the model, and ensuring the working reliability and security of the homomorphic encryption federated learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. Among them:

[0053] Figure 1 A flowchart of the communication cost control method for homomorphic encryption federated learning based on key learning period identification provided by the present invention.

[0054] Figure 2 This is a structural diagram of the homomorphic encryption federated learning communication cost control system based on key learning period identification provided by the present invention. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. It will be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some, rather than all, structures related to the present application are shown in the accompanying drawings. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0056] As used herein, the terms "comprise," "comprising," and "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0057] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0058] See also Figure 1 As shown, an embodiment of the present application provides a method for controlling communication costs of homomorphic encrypted federated learning based on key learning period identification. The method for controlling communication costs of homomorphic encrypted federated learning based on key learning period identification includes:

[0059] Identify the server where the homomorphically encrypted federated learning model is trained and obtain the communication attribute information of the server within the network. Based on this communication attribute information, determine the client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model. Then monitor all clients under this client cluster to obtain the working status characteristic information of each client.

[0060] Monitoring the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; determining all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; and determining client connection change information of the homomorphically encrypted federated learning model during the training process based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model.

[0061] Based on the client connection change information and the working status characteristic information, the communication connection status between the server and the client is adjusted; and based on the real-time training status of the homomorphic encryption federated learning model, the communication authority of the server within the network is adjusted.

[0062] The beneficial effects of the above embodiments are as follows: the method for controlling communication cost of homomorphic encryption federated learning based on identification of critical learning periods identifies the client clusters allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, and monitors the client clusters to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process to distinguish and determine the critical learning period and non-critical learning period of the model in the training process, and quantifies and identifies the importance of the model at different stages in the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection changes are quantitatively calibrated; it also adjusts the communication connection status between the server and the client based on the client connection change information and the working status characteristic information, effectively reducing the communication overhead of the model training without affecting the accuracy of the model, thereby ensuring the working reliability and security of the homomorphic encryption federated learning model.

[0063] In another embodiment, a server where a homomorphically encrypted federated learning model is trained is identified to obtain communication attribute information of the server within the network. Based on the communication attribute information, a client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model is determined, and all clients under the client cluster are monitored to obtain working status characteristic information of each client, including:

[0064] Compare the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection lists of all gateways within the network to determine the location information and gateway bandwidth information of the gateway connected to the server within the network; based on the gateway location information, determine all client clusters within a preset distance range from the server's connection path within the network; then, based on the gateway bandwidth information, select a client cluster with a matching communication bandwidth value from all client clusters within the preset distance range as the client cluster allowed to connect to the server;

[0065] Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time pending task volume of all the clients, which is used as the working status feature information.

[0066] The beneficial effects of the above embodiments are that the homomorphic encryption federated learning model needs to be trained with the help of the corresponding server before being applied to different occasions. It can only be used in the corresponding task calculation after the convergence of the trained model meets the corresponding numerical range conditions. The homomorphic encryption federated learning model needs the assistance of different external clients during the training process, that is, the model needs to communicate with different clients, which will put forward corresponding communication overhead requirements for the model. In order to ensure that the model can establish a reliable and stable communication connection with external clients during the training process, the device identity information of the server where the model is trained is compared with the device connection list of all gateways in the network where the server is located, so as to determine the gateway that the server is currently connected to in the network, and then further determine the location information and gateway bandwidth information of the accessed gateway, so as to perform a targeted search for the clients that the server can connect to. Based on the gateway location information, all client clusters whose connection paths to the server within the network are within a preset distance range are determined, thereby ensuring communication connections between the server and clients that are relatively close to the network. Furthermore, based on the gateway bandwidth information, a client cluster with a matching communication bandwidth value is selected from all client clusters within the preset distance range (i.e., the communication bandwidth value required for all clients under the selected client cluster to communicate with the server is less than or equal to the gateway bandwidth), and this is used as the client cluster allowed to connect to the server, thereby ensuring that the server obtains a good and stable client communication connection. Furthermore, based on the device identity information of all clients under the client cluster allowed to connect, all subordinate clients are monitored to obtain the real-time pending task volume of each client. This allows for quantitative identification of the task processing load of all clients, providing a reliable basis for subsequent adjustments to the connection status between the server and the clients.

[0067] In another embodiment, all clients under the client cluster are monitored to obtain the working status characteristic information of all clients, including:

[0068] Step S1, monitoring all clients under the client cluster is achieved by allocating a monitoring time within a corresponding cycle to each client under the client cluster by the server, thereby achieving the effect of time-division multiplexing monitoring. First, the following formula (1) is used to control the length of a cycle of time-division multiplexing monitoring of the server according to the number of all clients under the client cluster and the maximum amount of data that each client can generate.

[0069]

[0070] In the above formula (1), T represents the duration of a period of time-division multiplexing monitoring of the server; n represents the number of all clients under the client cluster; T0 represents the time it takes for the server to process a unit of data; S0 represents the unit of data; e represents a natural constant; S max (r) represents the maximum amount of data that can be generated by the rth client;

[0071] Step S2, using the following formula (2), according to the usage level of each client under the client cluster and the maximum amount of data that each client can generate, controls the listening time within the cycle allocated to each client during the time-division multiplexing monitoring of the server.

[0072]

[0073] In the above formula (2), T # (r) represents the listening time in the period allocated to the r-th client; D(r) represents the usage degree of the r-th client;

[0074] Step S3, using the following formula (3), based on the amount of characteristic information data of each client's working status monitored in each monitoring cycle, the usage level of each client under the client cluster is updated.

[0075]

[0076] In the above formula (3), D′(r) represents the updated usage level of the r-th client; J(r) represents the amount of characteristic information data of the working status of the r-th client monitored during the monitoring period.

[0077] The beneficial effect of the above embodiment is that, using the above formula (1), according to the number of all clients under the client cluster and the maximum amount of data that each client can generate, the length of a period of time-sharing multiplexing monitoring of the server is controlled, thereby ensuring that the length of the monitoring period can meet the monitoring of each client; then using the above formula (2), according to the usage level of each client under the client cluster and the maximum amount of data that each client can generate, the listening time within the period allocated to each client during the time-sharing multiplexing monitoring of the server is controlled, thereby reasonably allocating the listening time according to the usage level, thereby ensuring the stability of the system; then using the above formula (3), according to the amount of working status characteristic information data of each client monitored in each monitoring period, the usage level of each client under the client cluster is updated, thereby ensuring the reliability of the system without iterative updates of the listening.

[0078] In another embodiment, the training process of the homomorphically encrypted federated learning model is monitored to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process are determined; based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, client connection change information of the homomorphically encrypted federated learning model during the training process is determined, including:

[0079] Based on the data training speed of the homomorphically encrypted federated learning model during the training process, a monitoring frequency of the training process of the homomorphically encrypted federated learning model is determined, thereby obtaining loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, an average loss rate of the homomorphically encrypted federated learning model in all training stages is determined; if the average loss rate is greater than a preset loss rate threshold, the corresponding training stage is determined as a critical learning period of the homomorphically encrypted federated learning model during the training process; otherwise, the corresponding training stage is determined as a non-critical learning period of the homomorphically encrypted federated learning model during the training process;

[0080] Obtain the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model, and based on the time length information, determine the increase in client connections in each critical learning period and the decrease in client connections in each non-critical learning period of the homomorphic encryption federated learning model, as the client connection change information.

[0081] The beneficial effect of the above embodiment is that the speed of change of the loss value of the homomorphic encryption federated learning model during the training process will change with the data training speed of the model. In order to accurately obtain the change of the loss value of the model during the training process, the monitoring frequency of the training process of the homomorphic encryption federated learning model is determined based on the data training speed of the homomorphic encryption federated learning model during the training process. Generally speaking, the greater the data training speed, the greater the corresponding monitoring frequency. Based on the loss value change data, the average loss rate of the homomorphic encryption federated learning model in all training stages is determined, and the average loss rate is compared with the threshold value to distinguish whether the training stage of the model belongs to the critical learning period or the non-critical learning period, so as to facilitate the subsequent differentiated client connection changes when the model is in the critical learning period or the non-critical learning period. The time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model is also obtained to determine the increase in client connections in each critical learning period and the decrease in client connections in each non-critical learning period of the homomorphic encryption federated learning model. This allows the number of client connections to be appropriately increased when the model training is in the critical learning period, providing sufficient client support for the model training, and the number of client connections to be appropriately reduced when the model training is in the non-critical learning period, providing a reliable basis for reducing communication overhead during the model training process.

[0082] In another embodiment, adjusting the communication connection state between the server and the client based on the client connection change information and the working state characteristic information; and adjusting the communication authority of the server within the network based on the real-time training state of the homomorphically encrypted federated learning model, including:

[0083] When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet the preset remaining available computing power conditions are selected from all clients currently connected to the server, thereby disconnecting the connections between all selected clients and the server;

[0084] When the homomorphic encryption federated learning model is in a critical learning period, based on the client connection increase information and the work status characteristic information, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect, thereby connecting all the selected clients to the server;

[0085] Based on the real-time training convergence change status of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the server's communication permissions within the network; if not, keep the server's current communication permissions within the network unchanged.

[0086] The beneficial effect of the above embodiment is that when the homomorphic encryption federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet the preset remaining available computing power condition are selected from all the clients currently connected to the server, thereby interrupting the connection between all the selected clients and the server, such as interrupting the connection between the clients whose remaining available computing power value is less than the preset computing power threshold among all the clients currently connected to the server, and specifically, the number of clients connected to the server can be reduced exponentially. When the homomorphic encryption federated learning model is in a critical learning period, based on the client connection increase information and the working status characteristic information, a corresponding number of clients that meet the preset pending task amount condition are selected from the client cluster that is allowed to connect, thereby connecting all the selected clients to the server, such as connecting the clients whose pending task amount is less than the preset task amount threshold in the client cluster that is allowed to connect to the server, and specifically, the number of clients connected to the server can be increased exponentially. There is also a real-time training convergence change state based on the homomorphic encryption federated learning mode to determine whether the homomorphic encryption federated learning model has completed training. Specifically, when the real-time training convergence change state indicates that the difference between the convergence of the model and the preset target convergence is less than or equal to the preset threshold, it is determined that the homomorphic encryption federated learning model has completed training. At this time, the communication authority of the server within the network is terminated, which facilitates the calculation and processing of different tasks by the homomorphic encryption federated learning model; when the real-time training convergence change state indicates that the difference between the convergence of the model and the preset target convergence is greater than the preset threshold, it is determined that the homomorphic encryption federated learning model has not completed training. At this time, the current communication authority of the server within the network is kept unchanged to ensure that the homomorphic encryption federated learning model can be fully trained.

[0087] See also Figure 2 As shown, an embodiment of the present application provides a homomorphic encryption federated learning communication cost control system based on key learning period identification. The homomorphic encryption federated learning communication cost control system based on key learning period identification includes:

[0088] The communication state identification module is used to identify the server where the homomorphic encryption federated learning model is trained and obtain the communication attribute information of the server within the network;

[0089] A client selection and monitoring module is used to determine, based on the communication attribute information, the client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain the working status characteristic information of each client;

[0090] A critical learning period identification module is configured to monitor the training process of the homomorphically encrypted federated learning model and obtain loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, determine all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process;

[0091] A client connection change determination module is used to determine client connection change information of the homomorphically encrypted federated learning model during training based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model;

[0092] A communication connection state adjustment module, configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information;

[0093] The communication authority adjustment module is used to adjust the communication authority of the server within the network based on the real-time training status of the homomorphic encryption federated learning model.

[0094] The beneficial effects of the above embodiments are as follows: the homomorphic encryption federated learning communication cost control system based on critical learning period identification identifies the client cluster allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, and monitors the client cluster to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process to distinguish and determine the critical learning period and non-critical learning period of the model in the training process, and quantifies and identifies the importance of the model at different stages of the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection changes are quantitatively calibrated; based on the client connection change information and working status characteristic information, the communication connection status between the server and the client is adjusted, effectively reducing the communication overhead of model training without affecting the accuracy of the model, thereby ensuring the working reliability and security of the homomorphic encryption federated learning model.

[0095] In another embodiment, the communication state identification module is used to identify the server where the homomorphic encryption federated learning model is trained, and obtain the communication attribute information of the server within the network, including:

[0096] Compare the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection list of all gateways within the network to determine the location and bandwidth of the gateway connected to the server within the network.

[0097] The client selection and monitoring module is used to determine, based on the communication attribute information, a client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client, including:

[0098] Based on the gateway location information, all client clusters whose connection paths to the server within the network are within a preset distance range are determined; and based on the gateway bandwidth information, a client cluster having a matching communication bandwidth value is selected from all client clusters within the preset distance range as the client cluster allowed to connect to the server;

[0099] Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time pending task volume of all the clients, which is used as the working status feature information.

[0100] The beneficial effects of the above embodiments are as follows: the method for controlling communication cost of homomorphic encryption federated learning based on identification of critical learning periods identifies the client clusters allowed to connect to the server for training the homomorphic encryption federated learning model from within the network, and monitors the client clusters to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process to distinguish and determine the critical learning period and non-critical learning period of the model in the training process, and quantifies and identifies the importance of the model at different stages in the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection changes are quantitatively calibrated; it also adjusts the communication connection status between the server and the client based on the client connection change information and the working status characteristic information, effectively reducing the communication overhead of the model training without affecting the accuracy of the model, thereby ensuring the working reliability and security of the homomorphic encryption federated learning model.

[0101] In another embodiment, the critical learning period identification module is used to monitor the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, determine all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process, including:

[0102] Based on the data training speed of the homomorphically encrypted federated learning model during the training process, a monitoring frequency of the training process of the homomorphically encrypted federated learning model is determined, thereby obtaining loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, an average loss rate of the homomorphically encrypted federated learning model in all training stages is determined; if the average loss rate is greater than a preset loss rate threshold, the corresponding training stage is determined as a critical learning period of the homomorphically encrypted federated learning model during the training process; otherwise, the corresponding training stage is determined as a non-critical learning period of the homomorphically encrypted federated learning model during the training process;

[0103] The client connection change determination module is used to determine client connection change information of the homomorphically encrypted federated learning model during training based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, including:

[0104] Obtain the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model, and based on the time length information, determine the increase in client connections in each critical learning period and the decrease in client connections in each non-critical learning period of the homomorphic encryption federated learning model, as the client connection change information.

[0105] The beneficial effects of the above embodiments are that the homomorphic encryption federated learning model needs to be trained with the help of the corresponding server before being applied to different occasions. It can only be used in the corresponding task calculation after the convergence of the trained model meets the corresponding numerical range conditions. The homomorphic encryption federated learning model needs the assistance of different external clients during the training process, that is, the model needs to communicate with different clients, which will put forward corresponding communication overhead requirements for the model. In order to ensure that the model can establish a reliable and stable communication connection with external clients during the training process, the device identity information of the server where the model is trained is compared with the device connection list of all gateways in the network where the server is located, so as to determine the gateway that the server is currently connected to in the network, and then further determine the location information and gateway bandwidth information of the accessed gateway, so as to perform a targeted search for the clients that the server can connect to. Based on the gateway location information, all client clusters whose connection paths to the server within the network are within a preset distance range are determined, thereby ensuring communication connections between the server and clients that are relatively close to the network. Furthermore, based on the gateway bandwidth information, a client cluster with a matching communication bandwidth value is selected from all client clusters within the preset distance range (i.e., the communication bandwidth value required for all clients under the selected client cluster to communicate with the server is less than or equal to the gateway bandwidth), and this is used as the client cluster allowed to connect to the server, thereby ensuring that the server obtains a good and stable client communication connection. Furthermore, based on the device identity information of all clients under the client cluster allowed to connect, all subordinate clients are monitored to obtain the real-time pending task volume of each client. This allows for quantitative identification of the task processing load of all clients, providing a reliable basis for subsequent adjustments to the connection status between the server and the clients.

[0106] In another embodiment, the connection state adjustment module is configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information, including:

[0107] When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet the preset remaining available computing power conditions are selected from all clients currently connected to the server, thereby disconnecting the connections between all selected clients and the server;

[0108] When the homomorphic encryption federated learning model is in a critical learning period, based on the client connection increase information and the work status characteristic information, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect, thereby connecting all the selected clients to the server;

[0109] The communication permission adjustment module is used to adjust the communication permission of the server within the network based on the real-time training status of the homomorphically encrypted federated learning model, including:

[0110] Based on the real-time training convergence change status of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the server's communication permissions within the network; if not, keep the server's current communication permissions within the network unchanged.

[0111] The beneficial effect of the above embodiment is that when the homomorphic encryption federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet the preset remaining available computing power condition are selected from all the clients currently connected to the server, thereby interrupting the connection between all the selected clients and the server, such as interrupting the connection between the clients whose remaining available computing power value is less than the preset computing power threshold among all the clients currently connected to the server, and specifically, the number of clients connected to the server can be reduced exponentially. When the homomorphic encryption federated learning model is in a critical learning period, based on the client connection increase information and the working status characteristic information, a corresponding number of clients that meet the preset pending task amount condition are selected from the client cluster that is allowed to connect, thereby connecting all the selected clients to the server, such as connecting the clients whose pending task amount is less than the preset task amount threshold in the client cluster that is allowed to connect to the server, and specifically, the number of clients connected to the server can be increased exponentially. There is also a real-time training convergence change state based on the homomorphic encryption federated learning mode to determine whether the homomorphic encryption federated learning model has completed training. Specifically, when the real-time training convergence change state indicates that the difference between the convergence of the model and the preset target convergence is less than or equal to the preset threshold, it is determined that the homomorphic encryption federated learning model has completed training. At this time, the communication authority of the server within the network is terminated, which facilitates the calculation and processing of different tasks by the homomorphic encryption federated learning model; when the real-time training convergence change state indicates that the difference between the convergence of the model and the preset target convergence is greater than the preset threshold, it is determined that the homomorphic encryption federated learning model has not completed training. At this time, the current communication authority of the server within the network is kept unchanged to ensure that the homomorphic encryption federated learning model can be fully trained.

[0112] In general, the communication cost control method and system for homomorphic encryption federated learning based on critical learning period identification identifies the client cluster allowed to connect to the server training the homomorphic encryption federated learning model from within the network, and monitors the client cluster to obtain the working status characteristic information of all clients; it also monitors the loss value change data of the model during the training process to distinguish and determine the critical learning period and non-critical learning period of the model in the training process, and quantifies and identifies the importance of the model at different stages of the entire training process, so as to facilitate the subsequent adjustment of the number of connected clients for different learning periods and reduce unnecessary client connections; based on the time attribute information of the critical learning period and the non-critical learning period, the client connection change information of the model during the training process is determined, and the client connection changes are quantitatively calibrated; based on the client connection change information and working status characteristic information, the communication connection status between the server and the client is adjusted, effectively reducing the communication overhead of model training without affecting the accuracy of the model, ensuring the working reliability and security of the homomorphic encryption federated learning model.

[0113] The above is only a specific embodiment of the present invention, and any other improvements made based on the concept of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A homomorphic encryption federated learning communication cost control method based on key learning period identification, characterized by: include: Identify the server where the homomorphically encrypted federated learning model is trained and obtain the communication attribute information of the server within the network; Based on the communication attribute information, determining a client cluster that the server is allowed to connect to during training of the homomorphically encrypted federated learning model, and monitoring all clients under the client cluster to obtain working status characteristic information of each of the clients; Monitoring the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; determining all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; and determining client connection change information of the homomorphically encrypted federated learning model during the training process based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model; Adjusting the communication connection state between the server and the client based on the client connection change information and the working state characteristic information; And based on the real-time training status of the homomorphic encryption federated learning model, the communication authority of the server within the network is adjusted.

2. The method for controlling communication costs of homomorphic encrypted federated learning based on critical learning period identification according to claim 1, characterized in that: Identify the server where the homomorphically encrypted federated learning model is trained to obtain communication attribute information of the server within the network; based on the communication attribute information, determine the client cluster that the server is allowed to connect to during the training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain the working status characteristic information of each client, including: Comparing the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection lists of all gateways within the network to determine the location information and gateway bandwidth information of the gateway accessed by the server within the network; based on the gateway location information, determining all client clusters within a preset distance range from the server within the network; and then, based on the gateway bandwidth information, selecting a client cluster with a matching communication bandwidth value from all client clusters within the preset distance range as the client cluster allowed to connect to the server; Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time to-be-processed task volume of all the clients, which is used as the working status feature information.

3. The method for controlling communication costs of homomorphic encrypted federated learning based on critical learning period identification according to claim 1, characterized in that: Monitor all clients under the client cluster to obtain the working status characteristic information of all clients, including: In step S1, monitoring all clients under the client cluster is performed by allocating a monitoring time within a corresponding cycle to each client under the client cluster by the server, thereby achieving the effect of time-division multiplexing monitoring. First, the following formula (1) is used to control the duration of a cycle of time-division multiplexing monitoring by the server according to the number of all clients under the client cluster and the maximum amount of data that each client can generate. In the above formula (1), T represents the duration of a period of time-division multiplexing monitoring by the server; n represents the number of all clients under the client cluster; T0 represents the duration of processing unit data volume by the server; S0 represents unit data volume; e represents a natural constant; S max (r) represents the maximum amount of data that can be generated by the rth client; Step S2, using the following formula (2), according to the usage level of each client under the client cluster and the maximum amount of data that each client can generate, controls the listening time within the cycle allocated to each client during the time-division multiplexing monitoring of the server, In the above formula (2), T # (r) represents the listening time in the cycle allocated to the rth client; D(r) represents the usage degree of the rth client; Step S3, using the following formula (3), based on the amount of characteristic information data of the working status of each client monitored in each monitoring cycle, the usage level of each client under the client cluster is updated. In the above formula (3), D′(r) represents the updated usage level of the r-th client; J(r) represents the amount of characteristic information data of the working status of the r-th client monitored during the monitoring period.

4. The method for controlling communication costs of homomorphic encrypted federated learning based on critical learning period identification according to claim 2, characterized in that: Monitoring the training process of the homomorphically encrypted federated learning model to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; determining all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; and determining client connection change information of the homomorphically encrypted federated learning model during the training process based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, including: Based on the data training speed of the homomorphic encryption federated learning model during the training process, the monitoring frequency of the training process of the homomorphic encryption federated learning is determined to obtain the loss value change data of the homomorphic encryption federated learning model during the training process; based on the loss value change data, the average loss rate of the homomorphic encryption federated learning model in all training stages is determined; if the average loss rate is greater than the preset loss rate threshold, the corresponding training stage is determined as the critical learning period of the homomorphic encryption federated learning model during the training process; otherwise, the corresponding training stage is determined as the non-critical learning period of the homomorphic encryption federated learning model during the training process; the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model is obtained, and based on the time length information, the client connection increase information of the homomorphic encryption federated learning model in each critical learning period and the client connection decrease information in each non-critical learning period are determined, which are used as the client connection change information.

5. The method for controlling communication costs of homomorphic encrypted federated learning based on critical learning period identification according to claim 3 is characterized in that: Adjusting the communication connection state between the server and the client based on the client connection change information and the working state characteristic information; and adjusting the communication authority of the server within the network based on the real-time training status of the homomorphically encrypted federated learning model, including: When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet a preset remaining available computing power condition are selected from all clients currently connected to the server, thereby interrupting the connections between all selected clients and the server; When the homomorphic encryption federated learning model is in a critical learning period, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect based on the client connection increase information and the work status characteristic information, so as to connect all the selected clients to the server; Based on the real-time training convergence change state of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the communication authority of the server within the network; if not, keep the current communication authority of the server within the network unchanged.

6. A homomorphic encryption federated learning communication cost control system based on key learning period identification, characterized by: include: A communication status identification module is used to identify the server where the homomorphic encryption federated learning model is trained and obtain the communication attribute information of the server within the network; A client selection and monitoring module is configured to determine, based on the communication attribute information, a client cluster that the server is allowed to connect to during training of the homomorphically encrypted federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client; a critical learning period identification module, configured to monitor the training process of the homomorphically encrypted federated learning model and obtain loss value change data of the homomorphically encrypted federated learning model during the training process; and determine all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process based on the loss value change data; a client connection change determination module, configured to determine client connection change information of the homomorphically encrypted federated learning model during training based on time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model; a communication connection state adjustment module, configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information; A communication authority adjustment module is used to adjust the communication authority of the server within the network based on the real-time training status of the homomorphic encryption federated learning model.

7. The homomorphic encryption federated learning communication cost control system based on key learning period identification according to claim 5, characterized in that: The communication state identification module is used to identify the server where the homomorphic encryption federated learning model is trained and obtain the communication attribute information of the server within the network, including: Compare the device identity information of the server where the homomorphically encrypted federated learning model is trained with the device connection list of all gateways within the network to determine the location information and gateway bandwidth information of the gateway connected to the server within the network; The client selection and monitoring module is used to determine, based on the communication attribute information, a client cluster that the server is allowed to connect to during the training of the homomorphic encryption federated learning model, and monitor all clients under the client cluster to obtain working status characteristic information of each client, including: Based on the gateway location information, determining all client clusters whose connection paths to the server within the network are within a preset distance range; and based on the gateway bandwidth information, selecting a client cluster with a matching communication bandwidth value from all client clusters within the preset distance range as the client cluster allowed to connect to the server; Based on the device identity information of all clients under the client cluster that is allowed to connect, all the clients are monitored to obtain the real-time to-be-processed task volume of all the clients, which is used as the working status feature information.

8. The homomorphic encryption federated learning communication cost control system based on key learning period identification according to claim 6, characterized in that: The critical learning period identification module is used to monitor the training process of the homomorphic encryption federated learning model and obtain loss value change data of the homomorphic encryption federated learning model during the training process; Based on the loss value change data, all critical learning periods and all non-critical learning periods of the homomorphically encrypted federated learning model during the training process are determined, including: Based on the data training speed of the homomorphically encrypted federated learning model during the training process, a monitoring frequency of the training process of the homomorphically encrypted federated learning model is determined to obtain loss value change data of the homomorphically encrypted federated learning model during the training process; based on the loss value change data, an average loss rate of the homomorphically encrypted federated learning model in all training stages is determined; if the average loss rate is greater than a preset loss rate threshold, the corresponding training stage is determined as a critical learning period of the homomorphically encrypted federated learning model during the training process; otherwise, the corresponding training stage is determined as a non-critical learning period of the homomorphically encrypted federated learning model during the training process; the client connection change determination module is used to determine client connection change information of the homomorphically encrypted federated learning model during the training process based on the time attribute information of all critical learning periods and all non-critical learning periods under the homomorphically encrypted federated learning model, including: Obtain the time length information of all critical learning periods and all non-critical learning periods under the homomorphic encryption federated learning model, and based on the time length information, determine the client connection increase information in each critical learning period and the client connection decrease information in each non-critical learning period of the homomorphic encryption federated learning model, and use this as the client connection change information.

9. The homomorphic encryption federated learning communication cost control system based on key learning period identification according to claim 7, characterized in that: The connection state adjustment module is configured to adjust the communication connection state between the server and the client based on the client connection change information and the working state characteristic information, including: When the homomorphically encrypted federated learning model is in a non-critical learning period, based on the client connection reduction information, a corresponding number of clients that meet a preset remaining available computing power condition are selected from all clients currently connected to the server, thereby interrupting the connections between all selected clients and the server; When the homomorphic encryption federated learning model is in a critical learning period, a corresponding number of clients that meet a preset pending task amount condition are selected from the client cluster allowed to connect based on the client connection increase information and the work status characteristic information, so as to connect all the selected clients to the server; The communication authority adjustment module is used to adjust the communication authority of the server within the network based on the real-time training status of the homomorphic encryption federated learning model, including: Based on the real-time training convergence change state of the homomorphic encryption federated learning model, determine whether the homomorphic encryption federated learning model has completed training; if so, terminate the communication authority of the server within the network; if not, keep the current communication authority of the server within the network unchanged.

Images