Trusted federal learning method and device, terminal, server and storage medium

Abstract

The invention relates to a trusted federal learning method and device, a terminal, a server and a storage medium. The method comprises the following steps: receiving an initial global model sent by a server; executing model processing operation; wherein the model processing operation comprises the steps of training the initial global model, and sending related information of the global model obtained by training to the server under the condition that parameters of the global model obtained by training accord with first multi-objective optimization; the first multi-objective optimization comprises the precision of the global model, the privacy leakage rate of the terminal and the storage resource consumption of the terminal. By adopting the method, the performance, the privacy protection effect and the training efficiency of the credible federal learning model can be balanced.

Description

Technical Field

[0001] The present application relates to the technical field of trusted federated learning, and in particular to a trusted federated learning method, device, terminal, server and storage medium. Background Art

[0002] Trusted federated learning is a federated learning method that balances the privacy protection effect and learning efficiency of terminal participants under the premise of ensuring model availability. Trusted federated learning must have certain privacy protection and security anti-attack capabilities under the premise of meeting the agreed model availability. At the same time, the complexity of the entire machine learning is controllable and the learning efficiency is guaranteed. However, for heterogeneous learning networks, since the data on the terminal is not independent and identically distributed, the local storage and computing resources are limited and differentiated, and the terminals participating in the training will not be able to achieve the common training goals synchronously.

[0003] In traditional technology, asynchronous training is mainly achieved through personalized federated learning. At present, how to balance model performance, privacy protection effect and training efficiency in heterogeneous federated learning systems is an urgent problem that needs to be solved in trusted federated learning. Summary of the invention

[0004] The embodiments of the present application provide a trusted federated learning method, device, terminal, server and storage medium, which can balance the performance, privacy protection effect and training efficiency of a trusted federated learning model.

[0005] In a first aspect, the present application provides a trusted federated learning method, applied to a terminal, the method comprising:

[0006] receiving the initial global model sent by the server;

[0007] Execute model processing operations; wherein, the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0008] In one embodiment, the parameters of the trained global model include a local benefit value of the global model; and sending the relevant information of the trained global model to the server includes:

[0009] According to the local profit value and a first preset profit threshold, relevant information of the trained global model is sent to the server.

[0010] In one embodiment, the initial global model includes a global layer and a personality layer.

[0011] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0012] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

[0013] In one embodiment, the parameters of the global model obtained through training conform to the first multi-objective optimization, including:

[0014] The accuracy of the global model obtained by training is greater than a preset first accuracy value;

[0015] The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold;

[0016] The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0017] In one embodiment, the method further comprises:

[0018] Receiving a converged model sent by the server; the converged model is obtained by the server according to the global model obtained by the training;

[0019] The converged model is used as a new initial global model, and the model processing operation is repeatedly performed until a preset convergence condition is reached.

[0020] In one of the embodiments, the convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the profit value of the federated learning system is greater than a second preset profit threshold.

[0021] In one embodiment, the method further comprises:

[0022] Receive the first preset revenue threshold sent by the server.

[0023] In a second aspect, the present application also provides a trusted federated learning method, which is applied to a server, and the method includes:

[0024] Sending an initial global model to the terminal;

[0025] Receive relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is sent when the terminal trains the initial global model and the parameters of the trained global model meet the first multi-objective optimization; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

[0026] In one of the embodiments, the parameters of the trained global model include a local benefit value of the global model; the relevant information of the trained global model is sent by the terminal to the server based on the local benefit value and a first preset benefit threshold.

[0027] In one embodiment, the initial global model includes a global layer and a personality layer.

[0028] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0029] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

[0030] In one embodiment, the method further comprises:

[0031] Obtaining a converged model according to the global model obtained through the training;

[0032] The converged model is sent to the terminal, so that the terminal uses the converged model as a new initial global model and repeatedly performs the model processing operation until a preset convergence condition is reached.

[0033] In one embodiment, obtaining a converged model according to the global model obtained through training includes:

[0034] According to the second multi-objective optimization, the global model obtained by the training is optimized to obtain the converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0035] In one embodiment, sending the converged model to the terminal includes:

[0036] The global layer of the converged model is sent to the terminal.

[0037] In one embodiment, the method further comprises:

[0038] If the relevant information includes the individual layer of the trained global model, the individual layer of the converged model is sent to the terminal.

[0039] In one embodiment, the method further comprises:

[0040] The first preset profit threshold is sent to the terminal.

[0041] In a third aspect, the present application further provides a trusted federated learning device, applied to a terminal, the device comprising:

[0042] A first receiving module, used for receiving an initial global model sent by a server;

[0043] An execution module is used to perform model processing operations; wherein the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0044] In a fourth aspect, the present application further provides a trusted federated learning device, which is applied to a server, and the device includes:

[0045] A first sending module, used for sending an initial global model to a terminal;

[0046] A receiving module is used to receive relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is sent when the terminal trains the initial global model and the parameters of the trained global model meet the first multi-objective optimization; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

[0047] In a fifth aspect, the present application further provides a terminal, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in the first aspect when executing the computer program.

[0048] In a sixth aspect, the present application further provides a server, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in the second aspect when executing the computer program.

[0049] In a seventh aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect and the second aspect are implemented.

[0050] In an eighth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect and the second aspect.

[0051] The above-mentioned trusted federated learning method, device, terminal, server and storage medium, the terminal receives the initial global model sent by the server, and trains the received initial global model, and sends relevant information of the trained global model to the server when the parameter symbols of the trained global model are optimized for the first multi-objective optimization, wherein the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal. Since the terminal sends relevant information of the trained global model to the server when the parameters of the trained global model meet the first multi-objective optimization, for heterogeneous federated learning systems with differences in data quality and local storage and computing resources, the first multi-objective optimization can balance different objective constraints, and can protect the privacy of the terminal and reduce training overhead while ensuring the accuracy of the trained global model, thereby improving the efficiency of trusted federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is an application environment diagram of a trusted federated learning method in one embodiment;

[0053] Figure 2 is a flowchart of a trusted federated learning method in one embodiment;

[0054] Figure 3 is a flowchart of a trusted federated learning method in another embodiment;

[0055] Figure 4 is a flowchart of a trusted federated learning method in one embodiment;

[0056] Figure 5 is a flowchart of a trusted federated learning method in another embodiment;

[0057] Figure 6 is a structural block diagram of a trusted federated learning device in one embodiment;

[0058] Figure 7 is a structural block diagram of a trusted federated learning device in another embodiment;

[0059] Figure 8 is a structural block diagram of a trusted federated learning device in another embodiment;

[0060] Fig. 9 is a structural block diagram of a trusted federated learning device in another embodiment;

[0061] Fig.10 is a structural block diagram of a trusted federated learning device in one embodiment;

[0062] Fig.11 is a structural block diagram of a trusted federated learning device in another embodiment;

[0063] Fig.12 is a structural block diagram of a trusted federated learning device in another embodiment;

[0064] Fig.13 is a structural block diagram of a trusted federated learning device in another embodiment;

[0065] Fig.14 is a structural block diagram of a trusted federated learning device in another embodiment;

[0066] Fig.15 is a structural block diagram of a trusted federated learning device in another embodiment;

[0067] Fig.16 is an internal structure diagram of a terminal in one embodiment;

[0068] Fig.17 FIG. 4 is a diagram showing the internal structure of a server in one embodiment. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0070] The trusted federated learning method provided by the embodiments of the present disclosure can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 through a network. The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices, and the server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0071] In traditional technologies, for heterogeneous learning networks, since the data on the terminals are not independent and identically distributed, and the local storage and computing resources are limited and differentiated, the terminals participating in the training will not be able to achieve the common training goals synchronously. Asynchronous training is mainly achieved through personalized federated learning, but asynchronous personalized federated learning requires increasing the local computing complexity of the terminal, and the increase in terminal computing complexity is limited by the bottleneck of local resources of the terminal. The local resource bottleneck of the terminal will further lead to a decrease in the efficiency of federated learning. Based on this, how to balance model performance, privacy protection effect and training efficiency in a heterogeneous federated learning system is an urgent problem to be solved in trusted federated learning.

[0072] Based on the above-mentioned traditional technology, the trusted federated learning method provided by the embodiment of the present application, in each round of model training, the terminal participant first builds a local multi-objective optimization training model, conducts local training under the constraints of local storage and computing resources and privacy protection goals, and uploads different parts of the local model according to different constraint goals; the server builds a federal multi-objective optimization model, and generates different global models under different constraint goals. The entire system balances the constraints of different goals, protects the privacy of the client and reduces training costs while ensuring the accuracy of the model, thereby improving the efficiency of trusted federated learning.

[0073] It should be noted that the beneficial effects brought about by the embodiments of the present application or the technical problems solved are not limited to this one, but may also include other implicit or related problems. For details, please refer to the description of the following embodiments.

[0074] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0075] In one embodiment, Figure 2 As shown in the figure, a trusted federated learning method is provided, which can be applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0076] S201, receiving an initial global model sent by a server.

[0077] Among them, the terminal can be connected to the server through a communication network, and the terminal can receive the initial global model sent by the server through the communication network. Optionally, in this embodiment, the initial global model can be sent to the terminal when the server initiates a federated learning task, that is, in this embodiment, the initial global model is actively sent to the terminal by the server. Optionally, the initial global model in this embodiment can be any one of the neural network models of a convolutional neural network model, a generative adversarial network model, a recurrent neural network model, and a long short-term memory network model.

[0078] S202, execute model processing operations; wherein, the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0079] In this embodiment, after receiving the initial global model sent by the server, the terminal can perform a model processing operation on the received initial global model, wherein the model processing operation performed by the terminal on the initial global model may include training the initial global model. It can be understood that in the process of training the initial global model, after each round of training, the accuracy of the global model obtained in this round of training, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal can be obtained. Information such as the information further includes sending the relevant information of the trained global model to the server when the parameters of the trained global model meet the first multi-objective optimization of the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. As an optional implementation, in this process, the privacy leakage rate of the terminal can be estimated by each terminal based on the potential attack model in this round of training.

[0080] Optionally, in this embodiment, for each terminal in the federated learning system, each terminal can train the received initial global model in combination with its own training data. Optionally, in this embodiment, the information related to the trained global model sent by the terminal to the server may include information such as the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0081] In the above-mentioned trusted federated learning method, the terminal receives the initial global model sent by the server, and trains the received initial global model. When the parameter symbols of the trained global model are optimized for the first multi-objective, the terminal sends relevant information of the trained global model to the server, wherein the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. Since the terminal sends relevant information of the trained global model to the server when the parameters of the trained global model meet the first multi-objective optimization, for heterogeneous federated learning systems with differences in data quality and local storage and computing resources, the first multi-objective optimization can balance different objective constraints, and can protect the privacy of the terminal and reduce training overhead while ensuring the accuracy of the trained global model, thereby improving the efficiency of trusted federated learning.

[0082] In this embodiment, the detailed process of the terminal sending the information related to the trained global model to the server is described. In one embodiment, the method may further include: sending the information related to the trained global model to the server according to the local profit value and the first preset profit threshold.

[0083] Among them, the first preset profit threshold can be set by the terminal itself or sent to the terminal by the server. For example, the first preset profit threshold can be the system profit value of the federated learning system obtained in the previous round of training sent by the server to the terminal. However, it should be noted that in the initial first round of training, if the first preset profit threshold is sent by the server to the terminal, the server can set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload relevant information of the global model obtained through local training as much as possible after the first local training.

[0084] As an optional implementation, the local benefit value of the global model trained in this embodiment can be determined based on the difference between the accuracy of the global model trained and the privacy leakage of the terminal and the storage and computing resource consumption of the terminal.

[0085] In addition, it should be noted that the initial global model sent by the server to the terminal can be a hierarchical model. The initial global model can include a global layer and a personalized layer, wherein the global layer can have common features of data of most terminals, and the personalized layer has only features unique to data of different terminals. For example, the initial global model sent by the server to the terminal can be expressed as In the formula, represents the global layer of the initial global model sent by the server to the terminal, Represents the personality layer of the initial global model sent by the server to the terminal.

[0086] Optionally, in this embodiment, the local benefit value of the trained global model can be compared with a first preset benefit threshold, and the relevant information of the trained global model can be sent to the server based on the comparison result. Optionally, if the comparison result of the local benefit value of the trained global model with the first preset benefit threshold is that the local benefit value of the trained global model is greater than the first preset benefit threshold, then the relevant information of the trained global model sent by the terminal to the server may include the global layer and individual layer of the trained global model, the local benefit value of the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal; if the comparison result of the local benefit value of the trained global model with the first preset benefit threshold is that the local benefit value of the trained global model is less than or equal to the first preset benefit threshold, then the relevant information of the trained global model sent by the terminal to the server may include the global layer of the trained global model, the local benefit value of the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. It should be noted here that in the initial first round of training, if the first preset profit threshold is sent by the server to the terminal, the server can set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload the personality layer of the global model obtained through local training as much as possible after the first local training.

[0087] In the present embodiment, since the terminal sends relevant information of the trained global model to the server based on the local profit value of the trained global model and the first preset profit threshold, the relevant information of the trained global model can be selectively sent to the server through the first preset profit threshold and the local profit value of the trained global model, and the trained global model can be sent to the server after being screened, thereby reducing the sending of unnecessary information and improving the efficiency of the relevant information of the global model sent to the server.

[0088] In the scenario where the parameters of the global model obtained through training above meet the first multi-objective optimization, the first multi-objective optimization may include optimizing the accuracy of the global model obtained through training, optimizing the privacy leakage rate of the terminal, optimizing the storage and computing resource consumption of the terminal, etc. In one embodiment, the "parameters of the global model obtained through training meet the first multi-objective optimization" in the above S202 may include: the accuracy of the global model obtained through training is greater than a preset first accuracy value; the privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold; the storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0089] In an embodiment of the present application, for the i-th terminal in the federated learning system in the t-th round of training, the terminal i can train the received global model until the parameters of the trained global model meet the first multi-objective optimization including the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. The trained global model is obtained. As an optional implementation, in this embodiment, the parameters of the trained global model meet the first multi-objective optimization, which may include that the accuracy of the trained global model is greater than a preset first accuracy value, the privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold, and the storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption. Optionally, in the embodiment, as an example expression, the above-mentioned first multi-objective optimization can be expressed as: In the formula, Indicates the accuracy of the trained global model, which can be calculated based on the test data set. represents the global model obtained through training, which can be expressed as η is the learning rate in the stochastic gradient descent algorithm, which can be dynamically modified according to the constraint target. For the terminal, the tth round of training is based on The random gradient value is determined by They represent the privacy leakage rate and storage and computing resource consumption on terminal i respectively. represents the preset first privacy leakage rate threshold, Indicates the preset first storage and computing resource consumption.

[0090] Optionally, the preset first precision value in this embodiment can be the maximum value determined by the terminal from the historical precision of the global model obtained from historical training, the first privacy leakage rate threshold can be a value determined based on the terminal's own privacy protection requirements, and the first storage and computing resource consumption can be a value determined based on the terminal's own storage and computing resource limitations.

[0091] In this embodiment, since the parameters of the trained global model meet the first multi-objective optimization, including the accuracy of the trained global model being greater than the preset first accuracy value, the privacy leakage rate of the terminal being less than the preset first privacy leakage rate threshold, and the storage and computing resource consumption of the terminal being less than the preset first storage and computing resource consumption, in this way, in the process of training the initial global model, the accuracy of the trained global model can be guaranteed while protecting the privacy of the terminal and reducing the training overhead, thereby improving the efficiency of trusted federated learning.

[0092] In the scenario where the terminal sends the relevant information of the trained global model to the server, the server can obtain the converged model based on the received relevant information and send it to the terminal to continue model training. Figure 3 As shown, the above method also includes:

[0093] S301, receiving a converged model sent by a server; the converged model is obtained by the server based on a global model obtained through training.

[0094] In this embodiment, after the server receives the relevant information of the trained global model sent by the terminal, the server can obtain the converged model based on the trained global model. For example, the server can obtain the parameters of the converged model according to the average value of the parameters of the trained global model sent by each terminal in the federated learning system, thereby obtaining the converged model. Alternatively, the server can obtain the parameters of the converged model according to the maximum value of the parameters of the trained global model sent by each terminal in the federated learning system, thereby obtaining the converged model, and send the converged model to the terminal in the federated learning system, and the terminal receives the converged model sent by the server.

[0095] S302, taking the converged model as a new initial global model, and repeatedly performing model processing operations until a preset convergence condition is reached.

[0096] In this embodiment, the terminal uses the received convergence model sent by the server as a new initial global model, and repeatedly performs the above-mentioned model processing operation, that is, the terminal repeatedly performs training on the new initial global model, and when the parameters of the new global model obtained by training meet the above-mentioned first multi-objective optimization, the terminal sends the relevant information of the trained global model to the server until the preset convergence condition is reached. Optionally, the preset convergence condition in this embodiment may be that the number of times the model processing operation is performed on the new initial global model reaches a certain number threshold, or, as another implementation method, the preset convergence condition in this embodiment may also include that the accuracy of the obtained convergence model reaches a preset accuracy, and the benefit value of the federated learning system is greater than a second preset benefit threshold. Optionally, the accuracy of the convergence model in this embodiment may be obtained by testing the convergence model based on a test data set. Optionally, the second preset benefit threshold in this embodiment may be the maximum benefit value of the federated learning system determined by the terminal based on historical federated learning data.

[0097] In this embodiment, the server sends the converged model obtained based on the trained global model to the terminal, so that the terminal can use the converged model as a new initial global model and repeatedly perform model processing operations until the preset convergence conditions are reached. Through this iterative training process, the terminal can continuously optimize the initial global model to ensure the accuracy of the global model finally obtained through training.

[0098] In the scenario where the terminal sends the information about the trained global model to the server based on the local profit value and the first preset profit threshold, the terminal needs to first obtain the first preset profit threshold. In one embodiment, the method further includes: receiving the first preset profit threshold sent by the server.

[0099] Optionally, in this embodiment, while sending the initial global model to the terminal, the server may also send training-related parameters to the terminal. The training-related parameters sent by the server may include the above-mentioned first preset profit threshold. Optionally, in this embodiment, the first preset profit threshold sent by the server to the terminal may be the system profit value of the federated learning system obtained in the previous round of training. It can be understood that in the initial first round of training, the server may set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload as much relevant information of the global model obtained through local training as possible after the first local training.

[0100] In this embodiment, since the first preset profit threshold is sent to the terminal in advance by the server, the efficiency of the terminal in obtaining the first preset profit threshold is saved. In this way, the terminal can promptly send the relevant information of the trained global model to the server based on the pre-acquired first preset profit threshold and local profit value, ensuring the timeliness of the terminal sending the relevant information of the trained global model to the server, so that the server can obtain the relevant information of the trained global model in a timely manner.

[0101] In one embodiment, Figure 4 As shown in the figure, a trusted federated learning method is provided, which can be applied to Figure 1 The server in the example is used to illustrate the following steps:

[0102] S401: Send an initial global model to a terminal.

[0103] The server may be connected to the terminal via a communication network, and the server may send the initial global model to the terminal via the communication network. Optionally, in this embodiment, the initial global model may be sent to the terminal when the server initiates a federated learning task. Optionally, the initial global model in this embodiment may be any one of a convolutional neural network model, a generative adversarial network model, a recurrent neural network model, and a long short-term memory network model.

[0104] S402, receiving relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is sent when the terminal trains the initial global model and the parameters of the trained global model meet the first multi-objective optimization; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

[0105] In this embodiment, after receiving the initial global model sent by the server, the terminal can perform a model processing operation on the received initial global model, wherein the model processing operation performed by the terminal on the initial global model may include training the initial global model. It can be understood that in the process of training the initial global model, after each round of training, the accuracy of the global model obtained in this round of training, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal can be obtained. Information such as the information further includes sending the relevant information of the trained global model to the server based on the information obtained in each round of training when the parameters of the trained global model meet the first multi-objective optimization of the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal, so that the server receives the relevant information of the trained global model sent by the terminal. As an optional implementation, in this process, the privacy leakage rate of the terminal can be estimated by each terminal based on the potential attack model in this round of training.

[0106] Optionally, in this embodiment, for each terminal in the federated learning system, each terminal can train the received initial global model in combination with its own training data. Optionally, in this embodiment, the information related to the trained global model sent by the terminal received by the server may include information such as the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0107] In the above-mentioned trusted federated learning method, the server sends an initial global model to the terminal, so that the terminal can train the received initial global model. When the parameter symbols of the trained global model are optimized for the first multi-objective, the terminal sends relevant information of the trained global model to the server, wherein the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. Since the terminal sends relevant information of the trained global model to the server when the parameters of the trained global model meet the first multi-objective optimization, for heterogeneous federated learning systems with differences in data quality and local storage and computing resources, the first multi-objective optimization can balance different objective constraints, protect the privacy of the terminal and reduce training overhead while ensuring the accuracy of the trained global model, thereby improving the efficiency of trusted federated learning.

[0108] In the scenario where the server receives relevant information of the trained global model sent by the terminal, in one embodiment, the parameters of the trained global model include the local profit value of the global model; the relevant information of the trained global model is sent to the server by the terminal based on the local profit value and the first preset profit threshold.

[0109] Among them, the first preset profit threshold can be set by the terminal itself or sent to the terminal by the server. For example, the first preset profit threshold can be the system profit value of the federated learning system obtained in the previous round of training sent by the server to the terminal. However, it should be noted that in the initial first round of training, if the first preset profit threshold is sent by the server to the terminal, the server can set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload relevant information of the global model obtained through local training as much as possible after the first local training.

[0110] As an optional implementation, the local benefit value of the global model trained in this embodiment can be determined based on the difference between the accuracy of the global model trained and the privacy leakage of the terminal and the storage and computing resource consumption of the terminal.

[0111] In addition, it should be noted that the initial global model sent by the server to the terminal can be a hierarchical model. The initial global model can include a global layer and a personalized layer, wherein the global layer can have common features of data of most terminals, and the personalized layer has only features unique to data of different terminals. For example, the initial global model sent by the server to the terminal can be expressed as In the formula, represents the global layer of the initial global model sent by the server to the terminal, Represents the personality layer of the initial global model sent by the server to the terminal.

[0112] Optionally, in this embodiment, the terminal may compare the local benefit value of the trained global model with the first preset benefit threshold, and send relevant information of the trained global model to the server according to the comparison result. Optionally, in this embodiment, if the comparison result of the local benefit value of the trained global model with the first preset benefit threshold is that the local benefit value of the trained global model is greater than the first preset benefit threshold, then the relevant information of the trained global model sent by the terminal received by the server may include the global layer and individual layer of the trained global model, the local benefit value of the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal; if the comparison result of the local benefit value of the trained global model with the first preset benefit threshold is that the local benefit value of the trained global model is less than or equal to the first preset benefit threshold, then the relevant information of the trained global model sent by the terminal received by the server may include the global layer of the trained global model, the local benefit value of the trained global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal. It should be noted here that in the initial first round of training, if the first preset profit threshold is sent by the server to the terminal, the server can set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload the personality layer of the global model obtained through local training as much as possible after the first local training.

[0113] In the present embodiment, since the relevant information of the trained global model received by the server is sent by the terminal according to the local benefit value of the trained global model and the first preset benefit threshold, the first preset benefit threshold and the local benefit value of the trained global model enable the terminal to selectively send the relevant information of the trained global model to the server, and the trained global model can be screened and sent to the server, thereby reducing the sending of unnecessary information and improving the efficiency of sending the relevant information of the global model to the server, thereby also improving the efficiency of the server in receiving the relevant information of the trained global model.

[0114] After the server receives the information about the trained global model sent by the terminal, the server can further obtain the converged model corresponding to the global model according to the global model. Figure 5 As shown, the above method also includes:

[0115] S501, obtaining a converged model according to the global model obtained through training.

[0116] In this embodiment, after the server receives relevant information of the trained global model sent by the terminal, the server can obtain the converged model based on the trained global model. For example, the server can obtain the parameters of the converged model according to the average value of the parameters of the trained global model sent by each terminal in the federated learning system, thereby obtaining the converged model. Alternatively, the server can obtain the parameters of the converged model according to the maximum value of the parameters of the trained global model sent by each terminal in the federated learning system, thereby obtaining the converged model.

[0117] S502, sending the converged model to the terminal, so that the terminal uses the converged model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached.

[0118] In this embodiment, after the server obtains the convergence model, it sends the convergence model to the terminal in the federated learning system, so that the terminal can use the received convergence model sent by the server as a new initial global model and repeatedly perform the above model processing operation, that is, the terminal repeatedly performs training on the new initial global model, and when the parameters of the new global model obtained by training meet the above first multi-objective optimization, the relevant information of the trained global model is sent to the server until the preset convergence condition is reached. Optionally, the preset convergence condition in this embodiment can be that the number of times the model processing operation is performed on the new initial global model reaches a certain number threshold, or, as another implementation method, the preset convergence condition in this embodiment can also include that the accuracy of the obtained convergence model reaches a preset accuracy, and the benefit value of the federated learning system is greater than a second preset benefit threshold. Optionally, the accuracy of the convergence model in this embodiment can be obtained by testing the convergence model based on a test data set. Optionally, the second preset benefit threshold in this embodiment can be the maximum benefit value of the federated learning system determined by the terminal based on historical federated learning data.

[0119] In this embodiment, the server can obtain a converged model based on the global model obtained through training, and send the converged model to the terminal, so that the terminal can use the converged model as a new initial global model and repeatedly perform model processing operations until the preset convergence conditions are reached. Through this iterative training process, the terminal can continuously optimize the initial global model to ensure the accuracy of the global model finally obtained through training.

[0120] In the scenario where the server obtains the converged model based on the trained global model, the server can obtain the converged model by optimizing the trained global model. In one embodiment, S501 includes: optimizing the trained global model according to the second multi-objective optimization to obtain the converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0121] Among them, the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system. Optionally, the server can optimize the trained global model. When the accuracy of the optimized global model reaches the maximum value, the privacy leakage rate of the federated learning system is less than the preset privacy leakage rate, and the storage and computing resource consumption of the federated learning system is less than the preset storage and computing resource consumption, the above-mentioned convergence model is obtained.

[0122] Optionally, in an embodiment, as an example expression, the second multi-objective optimization may be expressed as: ε p ≤Ω p ε c ≤Ω c , where Represents the accuracy of the global model, which can be calculated based on the test data set. represents the global model obtained through training, Where M is the number of all terminals participating in this round of training in the federated learning system. is the convergence weight, which can be dynamically modified according to the constraint target. Where N is the number of all terminals in the federated learning system that have sent relevant information to the server. represents the local benefit value of the global model in the relevant information sent by terminal i in the federated learning system, ε p represents the privacy leakage rate of the federated learning system, 0≤α≤1, represents the privacy leakage rate on terminal i, ε c Indicates the storage and computing resource consumption of the federated learning system, 0≤β≤1, represents the storage and computing resource consumption on terminal i, Ω p represents the preset privacy leakage rate threshold, Ω c Indicates the preset storage and computing resource consumption.

[0123] In this embodiment, since the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system, when the trained global model is optimized according to the second multi-objective optimization, the negative impact of the heterogeneity of terminal data and storage and computing resources on the training efficiency can be reduced. By balancing different objective constraints, while ensuring the accuracy of the global model, the privacy of the federated learning system can be protected and the training overhead can be reduced, thereby improving the efficiency of trusted federated learning.

[0124] In this embodiment, the detailed process of the server sending the converged model to the terminal is described. In one embodiment, the above S502 includes: sending the global layer of the converged model to the terminal.

[0125] In this embodiment, after the server obtains the convergence model, it can first determine whether the accuracy of the obtained convergence model reaches the expected value. If the accuracy of the convergence model reaches the expected value, it can determine whether the benefit of the federated learning system reaches the expected value, wherein the benefit of the federated learning system can be the weighted sum of the local benefits of each terminal. If the benefit of the federated learning system reaches the expected value, the federated learning ends. If the benefit of the federated learning system does not reach the expected value, the server can send the global layer of the convergence model to the terminal. If the accuracy of the convergence model does not reach the expected value, the server can also send the global layer of the convergence model to the terminal. That is to say, after obtaining the convergence model, the server can send the global layer of the convergence model to the terminal. Further, as an optional implementation, if the relevant information sent by the terminal to the server includes the individual layer of the global model obtained by training, the server can also send the individual layer of the convergence model to the terminal, so that the server can flexibly send different information to the terminal in combination with the relevant information sent by the terminal, thereby improving the flexibility of the server in sending information to the terminal.

[0126] In this embodiment, since the global layer of the convergence model can characterize the common characteristics of the data of most terminals, the server can reduce the resource consumption of sending the convergence model to the terminal by sending the global layer of the convergence model to the terminal, thereby ensuring the efficiency of the server in sending the convergence model to the terminal.

[0127] In some scenarios, the server may also send the first preset revenue threshold to the terminal in advance. In one embodiment, the method further includes: sending the first preset revenue threshold to the terminal.

[0128] Optionally, in this embodiment, while sending the initial global model to the terminal, the server may also send parameters related to training to the terminal, wherein the parameters related to training sent by the server may include the above-mentioned first preset profit threshold. Optionally, in this embodiment, the first preset profit threshold sent by the server to the terminal may be the system profit value of the federated learning system obtained in the previous round of training. It is understandable that in the initial first round of training, the server may set the first preset profit threshold to a lower threshold, such as 0.1, to encourage the terminal to upload as much relevant information of the global model obtained through local training as possible after the first local training.

[0129] In this embodiment, since the server sends the first preset profit threshold to the terminal in advance, the efficiency of the terminal in obtaining the first preset profit threshold is saved. In this way, the terminal can promptly send the relevant information of the trained global model to the server based on the pre-acquired first preset profit threshold and local profit value, ensuring the timeliness of the terminal sending the relevant information of the trained global model to the server, so that the server can obtain the relevant information of the trained global model in a timely manner.

[0130] To facilitate understanding by those skilled in the art, the trusted federated learning method provided by this application is described in detail below in conjunction with a complete embodiment:

[0131] S1, the server sends an initial global model and a first preset profit threshold to the terminal.

[0132] S2, the terminal receives the initial global model sent by the server.

[0133] S3, the terminal trains the initial global model, and when the trained global model meets the first multi-objective optimization, if the local benefit value of the global model is greater than the first preset benefit threshold, the relevant information including the global layer and individual layer of the trained global model, the local benefit value of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal will be sent to the server; if the local benefit value of the global model is less than or equal to the first preset benefit threshold, the relevant information including the global layer of the trained global model, the local benefit value of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal will be sent to the server; wherein the parameters of the trained global model meet the first multi-objective optimization, including: the accuracy of the trained global model is greater than the preset first accuracy value; the privacy leakage rate of the terminal is less than the preset first privacy leakage rate threshold; the storage and computing resource consumption of the terminal is less than the preset first storage and computing resource consumption.

[0134] S4, the server optimizes the trained global model according to the second multi-objective optimization to obtain a converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0135] S5, the server sends the global layer of the convergence model to the terminal. If the relevant information received by the server includes the individual layer of the trained global model, the server sends the individual layer of the convergence model to the terminal so that the terminal uses the convergence model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached, wherein the convergence condition includes that the accuracy of the convergence model reaches a preset accuracy and the benefit value of the federated learning system is greater than a second preset benefit threshold.

[0136] It should be noted that for the descriptions in the above S1-S5, reference may be made to the relevant descriptions in the above embodiments, and the effects are similar, so this embodiment will not be repeated here.

[0137] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0138] Based on the same inventive concept, the embodiment of the present application also provides a trusted federated learning device for implementing the trusted federated learning method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more trusted federated learning device embodiments provided below can refer to the limitations of the trusted federated learning method above, and will not be repeated here.

[0139] In one embodiment, Figure 6 As shown, a trusted federated learning device is provided, which is applied to a terminal and includes: a first receiving module 10 and an execution module 11, wherein:

[0140] The first receiving module 10 is used to receive the initial global model sent by the server.

[0141] The execution module 11 is used to execute the model processing operation; wherein, the model processing operation includes training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending the relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0142] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0143] In one embodiment, optionally, the parameters of the global model obtained through the training include the local benefit value of the global model; Figure 7 As shown, optionally, the execution module 11 includes: a sending unit 111, wherein:

[0144] The sending unit 111 is used to send relevant information of the trained global model to the server according to the local profit value and the first preset profit threshold.

[0145] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0146] In one embodiment, optionally, the initial global model includes a global layer and a personality layer.

[0147] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0148] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0149] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0150] In one embodiment, optionally, the parameters of the global model obtained through the above training conform to the first multi-objective optimization, including:

[0151] The accuracy of the global model obtained through training is greater than the preset first accuracy value;

[0152] The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold;

[0153] The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0154] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0155] In another embodiment, optionally, Figure 8 As shown, the above device further includes: a second receiving module 12 and an operating module 13, wherein:

[0156] The second receiving module 12 is used to receive the converged model sent by the server; the converged model is obtained by the server based on the global model obtained through training.

[0157] The operation module 13 is used to use the converged model as a new initial global model and repeatedly perform model processing operations until a preset convergence condition is reached.

[0158] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0159] In one embodiment, optionally, the above-mentioned convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the profit value of the federated learning system is greater than a second preset profit threshold.

[0160] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0161] In another embodiment, optionally, Fig. 9 As shown, the above device further includes: a third receiving module 14, wherein:

[0162] The third receiving module 14 is configured to receive a first preset revenue threshold sent by the server.

[0163] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0164] Each module in the above-mentioned trusted federated learning device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0165] In one embodiment, Fig.10 As shown, a trusted federated learning device is provided, which is applied to a server and includes: a first sending module 20 and a receiving module 21, wherein:

[0166] The first sending module 20 is used to send the initial global model to the terminal.

[0167] The receiving module 21 is used to receive relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is that the terminal trains the initial global model, and the parameters of the trained global model meet the first multi-objective optimization. The first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0168] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0169] In one embodiment, optionally, the parameters of the global model obtained through the training include a local profit value of the global model; the relevant information of the global model obtained through the training is sent by the terminal to the server based on the local profit value and a first preset profit threshold.

[0170] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0171] In one embodiment, optionally, the initial global model includes a global layer and a personality layer.

[0172] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0173] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0174] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0175] In another embodiment, optionally, Fig.11 As shown, the above device also includes: an acquisition module 22 and a second sending module 23, wherein:

[0176] The acquisition module 22 is used to acquire the converged model according to the global model obtained through training.

[0177] The second sending module 23 is used to send the converged model to the terminal, so that the terminal uses the converged model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached.

[0178] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0179] In another embodiment, optionally, Fig.12 As shown, the acquisition module 22 includes: an acquisition unit 221, wherein:

[0180] The acquisition unit 221 is used to optimize the trained global model according to the second multi-objective optimization to obtain a converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0181] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0182] In another embodiment, optionally, Fig.13 As shown, the second sending module 23 includes: a sending unit 231, wherein:

[0183] The sending unit 231 is used to send the global layer of the convergence model to the terminal.

[0184] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0185] In another embodiment, optionally, Fig.14 As shown, the above device also includes: a third sending module 24, wherein:

[0186] The third sending module 24 is used to send the individual layers of the converged model to the terminal if the relevant information includes the individual layers of the trained global model.

[0187] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0188] In another embodiment, optionally, Fig.15 As shown, the above device also includes: a fourth sending module 25, wherein:

[0189] The fourth sending module 25 is configured to send the first preset revenue threshold to the terminal.

[0190] The trusted federated learning device provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated here.

[0191] Each module in the above-mentioned trusted federated learning device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0192] In one embodiment, a terminal is provided, whose internal structure diagram can be as follows: Fig.16 As shown. The terminal includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the terminal is used to provide computing and control capabilities. The memory of the terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the terminal is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a trusted federated learning method is implemented. The display screen of the terminal can be a liquid crystal display screen or an electronic ink display screen, and the input device of the terminal can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the terminal shell, or an external keyboard, touchpad or mouse, etc.

[0193] Those skilled in the art will understand that Fig.16 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0194] In one embodiment, a server is provided, whose internal structure diagram can be as follows: Fig.17 As shown. The server includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the server is used to provide computing and control capabilities. The memory of the server includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the server is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a trusted federated learning method is implemented.

[0195] Those skilled in the art will understand that Fig.17The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0196] In one embodiment, a terminal is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0197] receiving the initial global model sent by the server;

[0198] Execute model processing operations; wherein the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0199] In one embodiment, when the processor executes the computer program, the following steps are also implemented: according to the local profit value and the first preset profit threshold, the relevant information of the trained global model is sent to the server.

[0200] In one embodiment, the initial global model includes a global layer and a personalized layer, and the processor further implements the following steps when executing the computer program: if the local benefit value is greater than a first preset benefit threshold, the relevant information includes the global layer and the personalized layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0201] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0202] In one embodiment, the parameters of the global model obtained through training conform to the first multi-objective optimization, including:

[0203] The accuracy of the global model obtained through training is greater than the preset first accuracy value;

[0204] The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold;

[0205] The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0206] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0207] Receive the converged model sent by the server; the converged model is obtained by the server based on the global model obtained through training;

[0208] The converged model is used as the new initial global model and the model processing operation is repeated until the preset convergence condition is reached.

[0209] In one embodiment, the convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the profit value of the federated learning system is greater than a second preset profit threshold.

[0210] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0211] A first preset revenue threshold value sent by a receiving server is received.

[0212] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0213] Receive the initial global model sent by the server;

[0214] Execute model processing operations; wherein the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0215] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: according to the local profit value and the first preset profit threshold, relevant information of the trained global model is sent to the server.

[0216] In one embodiment, the initial global model includes a global layer and a personal layer, and when the computer program is executed by the processor, the following steps are further implemented: if the local benefit value is greater than a first preset benefit threshold, the relevant information includes the global layer and the personal layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0217] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0218] In one embodiment, the parameters of the global model obtained through training conform to the first multi-objective optimization, including:

[0219] The accuracy of the global model obtained through training is greater than the preset first accuracy value;

[0220] The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold;

[0221] The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0222] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0223] Receive the converged model sent by the server; the converged model is obtained by the server based on the global model obtained through training;

[0224] The converged model is used as the new initial global model and the model processing operation is repeated until the preset convergence condition is reached.

[0225] In one embodiment, the convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the profit value of the federated learning system is greater than a second preset profit threshold.

[0226] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0227] A first preset revenue threshold value sent by a receiving server is received.

[0228] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0229] receiving the initial global model sent by the server;

[0230] Execute model processing operations; wherein the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0231] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: according to the local profit value and the first preset profit threshold, relevant information of the trained global model is sent to the server.

[0232] In one embodiment, the initial global model includes a global layer and a personal layer, and when the computer program is executed by the processor, the following steps are further implemented: if the local benefit value is greater than a first preset benefit threshold, the relevant information includes the global layer and the personal layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0233] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0234] In one embodiment, the parameters of the global model obtained through training conform to the first multi-objective optimization, including:

[0235] The accuracy of the global model obtained through training is greater than the preset first accuracy value;

[0236] The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold;

[0237] The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

[0238] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0239] Receive the converged model sent by the server; the converged model is obtained by the server based on the global model obtained through training;

[0240] The converged model is used as the new initial global model and the model processing operation is repeated until the preset convergence condition is reached.

[0241] In one embodiment, the convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the profit value of the federated learning system is greater than a second preset profit threshold.

[0242] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0243] A first preset revenue threshold value sent by a receiving server is received.

[0244] In one embodiment, a server is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0245] Sending an initial global model to the terminal;

[0246] Relevant information of the trained global model sent by the receiving terminal; the relevant information of the trained global model is that the terminal trains the initial global model, and the parameters of the trained global model meet the first multi-objective optimization and are sent; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0247] In one embodiment, the parameters of the trained global model include the local benefit value of the global model; the relevant information of the trained global model is sent to the server by the terminal according to the local benefit value and the first preset benefit threshold.

[0248] In one embodiment, the initial global model includes a global layer and a personality layer.

[0249] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0250] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0251] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0252] According to the global model obtained through training, a converged model is obtained;

[0253] The converged model is sent to the terminal so that the terminal uses the converged model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached.

[0254] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0255] According to the second multi-objective optimization, the trained global model is optimized to obtain a converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0256] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0257] Send the global layer of the converged model to the terminal.

[0258] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0259] If the relevant information includes the individual layers of the trained global model, the individual layers of the converged model are sent to the terminal.

[0260] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0261] A first preset revenue threshold is sent to the terminal.

[0262] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0263] Sending an initial global model to the terminal;

[0264] Relevant information of the trained global model sent by the receiving terminal; the relevant information of the trained global model is that the terminal trains the initial global model, and the parameters of the trained global model meet the first multi-objective optimization and are sent; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0265] In one embodiment, the parameters of the trained global model include the local benefit value of the global model; the relevant information of the trained global model is sent to the server by the terminal according to the local benefit value and the first preset benefit threshold.

[0266] In one embodiment, the initial global model includes a global layer and a personality layer.

[0267] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0268] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0269] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0270] According to the global model obtained through training, a converged model is obtained;

[0271] The converged model is sent to the terminal so that the terminal uses the converged model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached.

[0272] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0273] According to the second multi-objective optimization, the trained global model is optimized to obtain a converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0274] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0275] Send the global layer of the converged model to the terminal.

[0276] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0277] If the relevant information includes the individual layers of the trained global model, the individual layers of the converged model are sent to the terminal.

[0278] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0279] A first preset revenue threshold is sent to the terminal.

[0280] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0281] Sending an initial global model to the terminal;

[0282] Relevant information of the trained global model sent by the receiving terminal; the relevant information of the trained global model is that the terminal trains the initial global model, and the parameters of the trained global model meet the first multi-objective optimization and are sent; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0283] In one embodiment, the parameters of the trained global model include the local benefit value of the global model; the relevant information of the trained global model is sent to the server by the terminal according to the local benefit value and the first preset benefit threshold.

[0284] In one embodiment, the initial global model includes a global layer and a personality layer.

[0285] If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal;

[0286] If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

[0287] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0288] According to the global model obtained through training, a converged model is obtained;

[0289] The converged model is sent to the terminal so that the terminal uses the converged model as a new initial global model and repeatedly performs model processing operations until a preset convergence condition is reached.

[0290] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0291] According to the second multi-objective optimization, the trained global model is optimized to obtain a converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

[0292] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0293] Send the global layer of the converged model to the terminal.

[0294] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0295] If the relevant information includes the individual layers of the trained global model, the individual layers of the converged model are sent to the terminal.

[0296] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0297] A first preset revenue threshold is sent to the terminal.

[0298] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0299] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0300] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A trusted federated learning method, characterized in that: Applied to a terminal, the method comprises: Receive the initial global model sent by the server; Execute model processing operations; wherein, the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

2. The method according to claim 1, characterized in that The parameters of the trained global model include the local benefit value of the global model; and sending the relevant information of the trained global model to the server includes: According to the local profit value and a first preset profit threshold, relevant information of the trained global model is sent to the server.

3. The method according to claim 2, characterized in that The initial global model includes a global layer and a personal layer. If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal; If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

4. The method according to any one of claims 1 to 3, characterized in that: The parameters of the global model obtained through training meet the first multi-objective optimization, including: The accuracy of the global model obtained by training is greater than a preset first accuracy value; The privacy leakage rate of the terminal is less than a preset first privacy leakage rate threshold; The storage and computing resource consumption of the terminal is less than a preset first storage and computing resource consumption.

5. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Receiving a converged model sent by the server; the converged model is obtained by the server according to the global model obtained by the training; The converged model is used as a new initial global model, and the model processing operation is repeatedly performed until a preset convergence condition is reached.

6. The method according to claim 5, characterized in that The convergence condition includes that the accuracy of the convergence model reaches a preset accuracy, and the benefit value of the federated learning system is greater than a second preset benefit threshold.

7. The method according to claim 2, characterized in that The method further comprises: Receive the first preset revenue threshold sent by the server.

8. A trusted federated learning method, characterized in that: Applied to a server, the method comprises: Sending an initial global model to the terminal; Receive relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is sent when the terminal trains the initial global model and the parameters of the trained global model meet the first multi-objective optimization; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

9. The method according to claim 8, characterized in that The parameters of the trained global model include a local benefit value of the global model; the relevant information of the trained global model is sent by the terminal to the server according to the local benefit value and a first preset benefit threshold.

10. The method according to claim 9, characterized in that The initial global model includes a global layer and a personal layer. If the local benefit value is greater than the first preset benefit threshold, the relevant information includes the global layer and the individual layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal; If the local benefit value is less than or equal to the first preset benefit threshold, the relevant information includes the global layer of the trained global model, the local benefit value, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

11. The method according to claim 10, characterized in that The method further comprises: Obtaining a converged model according to the global model obtained through the training; The converged model is sent to the terminal, so that the terminal uses the converged model as a new initial global model and repeatedly performs the model processing operation until a preset convergence condition is reached.

12. The method according to claim 11, characterized in that The step of obtaining a converged model according to the global model obtained through the training comprises: According to the second multi-objective optimization, the global model obtained by the training is optimized to obtain the converged model; the second multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the federated learning system, and the storage and computing resource consumption of the federated learning system.

13. The method according to claim 11 or 12, characterized in that: The sending the converged model to the terminal includes: The global layer of the converged model is sent to the terminal.

14. The method according to claim 13, characterized in that The method further comprises: If the relevant information includes the individual layer of the trained global model, the individual layer of the converged model is sent to the terminal.

15. The method according to any one of claims 9 to 12, characterized in that: The method further comprises: The first preset profit threshold is sent to the terminal.

16. A trusted federated learning device, characterized in that: Applied to a terminal, the device comprises: A first receiving module, used for receiving an initial global model sent by a server; An execution module is used to perform model processing operations; wherein the model processing operations include training the initial global model, and when the parameters of the trained global model meet the first multi-objective optimization, sending relevant information of the trained global model to the server; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal, and the storage and computing resource consumption of the terminal.

17. A trusted federated learning device, characterized in that: Applied to a server, the device comprises: A first sending module, used for sending an initial global model to a terminal; A receiving module is used to receive relevant information of the trained global model sent by the terminal; the relevant information of the trained global model is sent when the terminal trains the initial global model and the parameters of the trained global model meet the first multi-objective optimization; the first multi-objective optimization includes the accuracy of the global model, the privacy leakage rate of the terminal and the storage and computing resource consumption of the terminal.

18. A terminal comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

19. A server comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 8 to 15 are implemented.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 15 are implemented.

21. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 15 are implemented.

Images