Federal learning method and device, network equipment, medium and program product

Abstract

The invention provides a federated learning method and device, network equipment, a medium and a program product, and relates to the technical field of federated learning. The method comprises the following steps: for each second network element, segmenting a model according to a model segmentation point, and obtaining a first sub-model and a second sub-model corresponding to the second network element; sending corresponding model segmentation information to a second network element; obtaining a trained second sub-model sent by a second network element, the trained second sub-model being obtained by training the second sub-model by the second network element according to the first model parameter, and the first model parameter being obtained by training the first sub-model by the first network element; and obtaining a trained model of the first network element according to the trained second sub-model, and sending the trained model to the third network element. According to the scheme of the invention, federated learning is carried out through the three-layer federated learning architecture and the model segmentation points, wireless resources are fully utilized, and the calculation efficiency and the training efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a federated learning method, apparatus, network equipment, media, and program product. Background Technology

[0002] In artificial intelligence applications, machine learning models, including convolutional neural networks or deep neural networks, are trained using rich data generated at the network edge (such as sensors and IoT devices). The principle is to update the model's parameters to minimize the error in the output, ultimately establishing a mapping function to predict unknown data. However, this data is often privacy-sensitive (such as medical and financial data), so businesses or individuals may refuse to share their data with service providers for centralized training.

[0003] To ensure data privacy and security, a distributed machine learning technique called federated learning is currently being offered. Federated learning enables model training using multiple participants and their individual local data without disclosing each participant's local data. Each participant trains its own model locally and contributes it to a consortium center for joint model training. The center then makes this model available to all participants.

[0004] However, due to the limited computing and communication capabilities of devices, federated learning technology suffers from low computational efficiency and low system efficiency in wireless / edge networks. Summary of the Invention

[0005] The purpose of this invention is to provide a federated learning method, apparatus, network device, medium, and program product to solve the problems of low computational efficiency and low efficiency of federated learning systems in wireless / edge networks in the prior art.

[0006] To achieve the above objectives, the present invention is implemented as follows:

[0007] In a first aspect, embodiments of the present invention provide a federated learning method applied to a first network element, the method comprising:

[0008] For each second network element, the model is segmented according to the model segmentation point to obtain the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined based on the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element.

[0009] Send the corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model;

[0010] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model.

[0011] The trained model of the first network element is obtained based on the trained second sub-model, and the trained model of the first network element is sent to the third network element.

[0012] Optionally, the federated learning method, wherein obtaining the model trained on the first network element based on the trained second sub-model, includes:

[0013] The first sub-model and the second sub-model after training are merged to obtain the model after training the second network element, wherein the first sub-model after training is obtained based on the parameters of the first model.

[0014] The models trained by at least one of the second network elements are aggregated to obtain the model trained by the first network element.

[0015] Optionally, in the federated learning method, before segmenting the model according to model splitting points for each second network element to obtain the first sub-model and the second sub-model of the second network element, the method further includes:

[0016] Obtain the first request, which is used to request federated learning;

[0017] Based on the first request, a second request is sent to at least one second network element, the second request being used to request the second network element to participate in federated learning;

[0018] Obtain at least one first feedback message sent by the second network element, wherein the first feedback message is used to instruct the second network element to agree to participate in federated learning.

[0019] Optionally, the federated learning method, wherein sending a second request to at least one second network element according to the first request includes:

[0020] Based on the network element restriction information in the first request, at least one second network element is determined to participate in federated learning, wherein the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning.

[0021] Send a second request to at least one of the second network elements.

[0022] Optionally, in the federated learning method, obtaining the trained second sub-model sent by the second network element includes:

[0023] Obtain the second model parameters sent by the second network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0024] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0025] Send the first model parameters to the second network element;

[0026] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element through backpropagation training of the second sub-model based on the parameters of the first model.

[0027] Optionally, in the federated learning method, obtaining the trained second sub-model sent by the second network element includes:

[0028] Obtain a third request sent by the first target network element, wherein the third request is sent based on a fourth request sent by the second network element, and both the third request and the fourth request are used to request the second network element to continue federated learning;

[0029] Obtain the second model parameters of the second network element forwarded by the first target network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0030] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0031] Send the first model parameters to the first target network element;

[0032] Obtain the second sub-model trained by the second network element forwarded by the first target network element.

[0033] Optionally, in the federated learning method, the third request is sent when the first network element corresponding to the second network element is switched.

[0034] Before the handover of the first network element corresponding to the second network element occurs, the method further includes one of the following:

[0035] Send a first notification to the second network element;

[0036] Obtain the first notification sent by the second network element;

[0037] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0038] Optionally, the federated learning method further includes:

[0039] If the first network element has temporary resource occupation, a fifth request is sent to the third network element, the fifth request being used to request the transfer of federated learning;

[0040] A sixth request is sent to the second target network element, which is determined by the third network element based on the resource information of at least one first network element other than the first network element. The sixth request is used to request federated learning, including the second model parameters of the second network element.

[0041] Obtain the first model parameters sent by the second target network element, wherein the first model parameters are obtained by the second target network element training the first sub-model based on the second model parameters.

[0042] Optionally, the federated learning method further includes:

[0043] Obtain the seventh request sent by the second network element, the seventh request being used to request to exit federated learning;

[0044] The third network element sends an eighth request to the second network element, which is used to request the second network element to withdraw from federated learning. The third network element updates the credit score and network element restriction information of the second network element according to the eighth request.

[0045] Secondly, embodiments of the present invention also provide a federated learning method applied to a third network element, the method comprising:

[0046] Obtain the trained model sent by at least one first network element;

[0047] The trained models sent by at least one of the first network elements are aggregated to obtain an aggregated model.

[0048] Optionally, the federated learning method further includes, before acquiring the trained model sent by at least one first network element:

[0049] Obtain resource information sent by multiple first network elements and link information sent by multiple second network elements;

[0050] Based on the resource information of multiple first network elements and the link information of multiple second network elements, at least one first network element and at least one second network element participating in federated learning are determined.

[0051] Send a first request to at least one of the first network elements.

[0052] Optionally, the federated learning method further includes, before sending the first request to at least one of the first network elements:

[0053] Obtain the target request, which includes one of the following:

[0054] The federated learning request sent by the second network element;

[0055] The model request sent by the second network element;

[0056] The model request sent by the first network element;

[0057] Based on the target request, determine whether the model indicated by the target request needs to undergo federated learning.

[0058] Optionally, the federated learning method further includes:

[0059] Obtain the eighth request sent by the first network element, the eighth request being used to request the second network element to exit federated learning;

[0060] Update the credit score of the second network element according to the eighth request;

[0061] If the credit score of the second network element is less than the credit threshold, update the network element restriction information.

[0062] Thirdly, embodiments of the present invention also provide a federated learning method applied to a second network element, the method comprising:

[0063] Obtain model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element;

[0064] The second sub-model is trained based on the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model.

[0065] Send the trained second sub-model to the first network element.

[0066] Optionally, in the federated learning method, before obtaining the model segmentation information sent by the first network element, the method further includes:

[0067] Send the target request;

[0068] Obtain the second request sent by the first network element;

[0069] Send a first feedback message to the first network element, the first feedback message being used to instruct the second network element to agree to participate in federated learning;

[0070] The target request to be sent includes one of the following:

[0071] Send a first request to the first network element, the first request being used to request federated learning;

[0072] Send a federated learning request to the third network element;

[0073] Send a model request to the first network element;

[0074] Send a model request to the third network element.

[0075] Optionally, the federated learning method, wherein training the second sub-model based on the first model parameters to obtain the trained second sub-model includes:

[0076] The second sub-model is trained by forward propagation based on network data to obtain the parameters of the second model.

[0077] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0078] Optionally, the federated learning method, wherein training the second sub-model based on the first model parameters to obtain the trained second sub-model includes:

[0079] In the event of a switchover of the first network element corresponding to the second network element, a fourth request is sent to the first target network element, the fourth request being used to request the second network element to continue federated learning;

[0080] Send the second model parameters to the first network element after the switch. The second model parameters are obtained by forward propagation training of the second sub-model based on the network data.

[0081] Obtain the first model parameters forwarded by the first target network element;

[0082] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0083] Optionally, in the federated learning method, before the first network element corresponding to the second network element is switched over, the method further includes one of the following:

[0084] Obtain the first notification sent by the first network element;

[0085] Send a first notification to the first network element;

[0086] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0087] Fourthly, embodiments of the present invention also provide a federated learning device applied to a first network element, the device comprising:

[0088] The first acquisition module is used to segment the model according to the model segmentation point for each second network element, and acquire the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined according to the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element.

[0089] The first sending module is used to send corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model;

[0090] The second acquisition module is used to acquire the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model.

[0091] The second sending module is used to obtain the model trained by the first network element based on the trained second sub-model, and send the model trained by the first network element to the third network element.

[0092] Fifthly, embodiments of the present invention also provide a federated learning device applied to a third network element, the device comprising:

[0093] The third acquisition module is used to acquire at least one trained model sent by the first network element;

[0094] The first obtaining module is used to aggregate the trained models sent by at least one of the first network elements to obtain an aggregated model.

[0095] Sixthly, embodiments of the present invention also provide a federated learning device applied to a second network element, the device comprising:

[0096] The fourth acquisition module is used to acquire model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element;

[0097] The second acquisition module is used to train the second sub-model according to the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model.

[0098] The third sending module is used to send the trained second sub-model to the first network element.

[0099] In a seventh aspect, embodiments of the present invention also provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the federated learning method as described in the first aspect, or the federated learning method as described in the second aspect, or the federated learning method as described in the third aspect.

[0100] Eighthly, embodiments of the present invention also provide a readable storage medium storing a program that, when executed by a processor, implements the federated learning method as described in the first aspect, or the federated learning method as described in the second aspect, or the federated learning method as described in the third aspect.

[0101] In a ninth aspect, embodiments of the present invention also provide a computer program product, including computer instructions that, when executed by a processor, implement the federated learning method as described in the first aspect, or the federated learning method as described in the second aspect, or the federated learning method as described in the third aspect.

[0102] The beneficial effects of the above-described technical solution of the present invention are as follows:

[0103] The federated learning method described in this embodiment of the invention involves a first network element segmenting the model for each second network element according to model segmentation points. This segmentation points are determined based on the resource information of the first network element and the link information of the second network element. The first network element then sends corresponding model segmentation information to the second network element, indicating the first and second sub-models. The trained second sub-model, obtained by the second network element through training with first model parameters, is then acquired. Finally, the trained model from the first network element is obtained based on the trained second sub-model and sent to a third network element. This three-layer federated learning architecture fully utilizes wireless resources. Furthermore, the model segmentation points can be dynamically adjusted based on the resource information of the first network element and the link information of the second network element, improving the training efficiency of the federated learning system. The computationally intensive tasks are executed on the first network element, ensuring the privacy of the second network element is not exposed while improving computational efficiency. Attached Figure Description

[0104] Figure 1 This is a schematic diagram of the application system architecture of the federated learning method described in the embodiments of the present invention;

[0105] Figure 2This is a flowchart illustrating the federated learning method according to one embodiment of the present invention;

[0106] Figure 3 A schematic flowchart illustrating one implementation of the method described in this invention;

[0107] Figure 4 Another flowchart illustrating one of the implementation methods described in the embodiments of the present invention;

[0108] Figure 5 A schematic diagram of one embodiment of the method described in this invention;

[0109] Figure 6 Another flowchart illustrating the second implementation of the method described in this embodiment of the invention;

[0110] Figure 7 A flowchart illustrating a third embodiment of the method described in this invention;

[0111] Figure 8 A flowchart illustrating the fourth implementation method of the present invention;

[0112] Figure 9 This is a flowchart illustrating the federated learning method according to another embodiment of the present invention;

[0113] Figure 10 This is a flowchart illustrating the federated learning method according to another embodiment of the present invention;

[0114] Figure 11 This is a schematic diagram of the structure of the federated learning device according to one embodiment of the present invention;

[0115] Figure 12 This is a schematic diagram of the structure of the federated learning device according to another embodiment of the present invention;

[0116] Figure 13 This is a schematic diagram of the structure of the federated learning device according to another embodiment of the present invention;

[0117] Figure 14 This is a schematic diagram of the hardware structure of the network device described in an embodiment of the present invention. Detailed Implementation

[0118] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0119] The application system architecture of the federated learning method described in this embodiment of the invention is as follows: Figure 1As shown, it includes at least one third network element (the third network element in this embodiment can also be called a third device) 10, such as a core network, O-Cloud (cloud infrastructure platform), RIC (RAN Intelligent Controller), SMO (Service Management and Orchestration), and other high-performance computing resources. Each third network element 10 can manage at least one first network element (the first network element in this embodiment can also be called a first device) 20, such as a RIC, O-CU (centralized control unit) / O-DU (distributed unit), edge computing nodes, and other high-performance computing resources. Within the coverage area of ​​each first network element 20, there is at least one second network element (the second network element in this embodiment can also be called a second device) 30, such as a terminal, passive terminal, IoT device, and other low-performance computing resources, and it possesses private data. Therefore, the first network element, the second network element, and the third network element belong to computing network elements with different computing types. It should be noted that this application system can be an application system under the O-RAN framework.

[0120] To address the issues of low computational efficiency and low training efficiency of existing federated learning techniques in wireless / edge networks, this invention provides a federated learning method. For each second network element, the model is segmented according to a model segmentation point to obtain a first sub-model and a second sub-model corresponding to the second network element. The model segmentation point is determined based on the resource information of the first network element and the link information of the second network element. Corresponding model segmentation information is sent to the second network element, indicating the first and second sub-models. The trained second sub-model sent by the second network element is obtained, whereby the second network element trains the second sub-model using first model parameters, which are obtained by the first network element training the first sub-model. Based on the trained second sub-model, the model trained by the first network element is obtained and sent to a third network element. In this architecture, the first network element can be an edge network element, the second network element can be an end network element, and the third network element can be a cloud network element. Therefore, a three-layer federated learning architecture of cloud, edge, and end can be constructed. By using the model splitting point of each second network element, the model is divided into two segments (i.e., the first sub-model and the second sub-model). Federated learning is then performed on the second network element and the first network element, respectively. This fully utilizes wireless / edge network resources, improves computational efficiency, and enhances training efficiency while keeping the privacy data of the second network element confidential.

[0121] One embodiment of the present invention, such as Figure 2As shown, a federated learning method is provided and applied to a first network element. The method includes:

[0122] S201, for each second network element, the model is segmented according to the model segmentation point to obtain the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined according to the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element.

[0123] In this embodiment of the invention, the third network element manages at least one first network element. The first network element can segment the model according to the model segmentation point for each second network element within its coverage area to obtain the first sub-model and the second sub-model corresponding to the second network element.

[0124] Since the model segmentation point is determined based on the resource information of the first network element and the link information of the second network element, the link information of the second network element includes at least one of the following: the duration of the second network element receiving channel test information; the duration of the second network element sending channel test information; and the link status. The resource information of the first network element includes at least one of the following: processing unit (e.g., CPU, GPU) parameters; memory size; storage space; and operating system. Therefore, this embodiment of the invention achieves dynamic adjustment of the model segmentation point based on the link information of the second network element and the resource information of the first network element. Moreover, this embodiment of the invention adopts a model segmentation training method, which fully utilizes the advantage of the strong computing resources of the first network element.

[0125] Optionally, determining the model split point also requires at least one of the following: resource information of the second network element; model backhaul conditions; model size; and expected return time of the third network element. The model backhaul conditions include at least one of the following: model accuracy; loss function; time; and iteration rounds.

[0126] It should be noted that the model split point is determined based on the link information of the second network element. The link information of the second network element includes the duration of the second network element receiving and sending channel test information. The reason for this determination is to ensure that at least one second network element within the coverage area of ​​the first network element returns to the trained second sub-model within a similar time. This avoids the problem that some second network elements have completed the task in this round while others are still processing the task, thus improving the efficiency of the entire federated learning system.

[0127] It should also be noted that the model segmentation points need to be determined based on the resource information of the first network element to achieve dynamic model segmentation and further improve the efficiency of the entire federated learning system.

[0128] S202, send the corresponding model segmentation information to the second network element, the model segmentation information being used to indicate the first sub-model and the second sub-model corresponding to the second network element.

[0129] In this embodiment of the invention, the model segmentation information includes at least one of the following: a second sub-model; the return conditions of the second sub-model; and segmentation point information.

[0130] S203, obtain the trained second sub-model sent by the second network element, wherein the trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters; the first model parameters are obtained by the first network element training the first sub-model.

[0131] In this embodiment of the invention, the first network element undergoes forward and backward propagation training, which means that intensive computation is deployed on the first network element, thereby reducing the computational burden on the second network element and improving computational speed and efficiency.

[0132] S204, obtain the model trained by the first network element based on the trained second sub-model, and send the model trained by the first network element to the third network element.

[0133] In this embodiment of the invention, since both the third network element and the first network element know the complete model information, only a small amount of intermediate parameters or model segmentation point information needs to be transmitted during transmission or task transfer, thus reducing latency and energy consumption.

[0134] In one implementation, optionally, S204, obtaining the model trained on the first network element based on the trained second sub-model includes:

[0135] The trained second sub-model and the trained first sub-model are merged to obtain the model after training the second network element, wherein the trained first sub-model is obtained based on the parameters of the first model.

[0136] The models trained by at least one of the second network elements are aggregated to obtain the model trained by the first network element.

[0137] It is understandable that, for each second network element, the first network element merges the trained second sub-model and the corresponding trained first sub-model, thereby obtaining the model trained by all second network elements within the coverage area, and realizing multi-terminal model merging.

[0138] In this embodiment of the invention, the first network element sends its trained model to the third network element by sending model feedback information; wherein the model feedback information includes: a model identifier; the model trained by the first network element; and network element completion information. The network element completion information includes a second network element identifier and the corresponding number of completion rounds.

[0139] Optionally, when the model trained by the first network element meets the model backhaul conditions, the first network element sends the model trained by the first network element to the third network element.

[0140] In one implementation, optionally, before S201, for each second network element, segmenting the model according to the model segmentation point and obtaining the first sub-model and the second sub-model corresponding to the second network element, the above method further includes:

[0141] Obtain the first request, which is used to request federated learning and train the model using federated learning.

[0142] According to the first request, a second request is sent to at least one second network element within the coverage area. The second request is used to request the second network element to participate in federated learning and to train the model through federated learning.

[0143] Obtain at least one first feedback message sent by the second network element, wherein the first feedback message is used to instruct the second network element to agree to participate in federated learning.

[0144] In this embodiment of the invention, the first request is used to request federated learning, and to trigger a federated learning task by performing federated learning training on the model, including one of the following:

[0145] The federated learning request sent by the second network element includes at least one of the following: model identifier; second network element identifier; request reason; training information; wherein the training information includes at least one of the following: model training service identifier; model training service registration identifier; model identifier; model information (such as application scenario, data information); the data information is the model's existing training data information, which is different from the response to input requirements, and includes at least one of the following: data collection information; data duration; data source; model accuracy; data scale.

[0146] The model request sent by the second network element includes at least one of the following: model identifier; second network element identifier; request reason.

[0147] The federated learning request sent by the third network element includes at least one of the following: model identifier; model Uniform Resource Locator (URL); model storage identifier; model return conditions; network element restriction information; wherein the network element restriction information includes the second network element identifier and the corresponding credit score.

[0148] The reasons for the above requests include at least one of the following: network degradation; decreased accuracy; outdated version; model change; data change.

[0149] Optionally, after obtaining the first request, the above method further includes:

[0150] Based on the reason for the request in the first request, determine whether the model indicated by the first request needs to undergo federated learning;

[0151] If it is determined that federated learning is necessary, then it is decided whether to participate in federated learning.

[0152] If participation in federated learning is determined, the model is requested and obtained from the model storage unit or the location of the model based on at least one of the model identifier, model URL, and model storage identifier in the first request.

[0153] It should be noted that the need for federated learning is determined based on the reason for the request. For example, if the network deteriorates to a certain level, federated learning is required; if the accuracy drops to a certain level, federated learning is required.

[0154] Optionally, the second request includes at least one of the following: model identifier; channel test information; participation round.

[0155] Optionally, the first feedback information includes at least one of the following: the link information of the second network element (such as channel test information); whether a trained model is needed; and label data.

[0156] Therefore, the first network element can obtain the link information of the second network element through the first feedback information. Furthermore, the first network element can also request link information before obtaining the first request, thereby obtaining the link information sent by the second network element.

[0157] In one implementation, optionally, according to the first request, sending a second request to at least one second network element includes:

[0158] Based on the network element restriction information in the first request, at least one second network element is determined to participate in federated learning, wherein the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning.

[0159] Send a second request to at least one of the second network elements.

[0160] In this embodiment of the invention, since the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning, including the second network element identifier and credit score, the first network element can exclude the second network elements within its coverage area according to the network element restriction information, thereby determining at least one second network element participating in federated learning, and sending a second request to at least one second network element.

[0161] It should be noted that after obtaining the network element restriction information, the first network element can determine whether to restrict the second network element indicated by the network element restriction information. If it is determined to restrict, at least one second network element participating in federated learning is determined based on the network element restriction information.

[0162] Optionally, the network element restriction information is determined by the credit management function of the third network element. This function manages the credit score of the second network element, lowering its credit score and restricting its participation in and / or requests for federated learning when the second network element defaults. A default by the second network element refers to its withdrawal from the federated learning process midway.

[0163] Optionally, when the credit score of the second network element is less than the credit threshold, participation in federated learning and / or requests for federated learning are restricted.

[0164] Optionally, the calculation process for the credit score of the second network element is as follows:

[0165] The initial value for the credit score is:

[0166] C = A (where A is any integer).

[0167] When the second network element commits a default, its credit score is:

[0168] C new =C-1

[0169] When the second network element successfully completes round t of federated learning, its credit score is:

[0170] C new =C + t*b (b>0)

[0171] It should also be noted that if the first request does not include the network element restriction information, the first network element may request the network element restriction information from the third network element.

[0172] Of course, the second network element can also request the network element restriction information from the third network element.

[0173] In one implementation, optionally, S203, obtaining the trained second sub-model sent by the second network element includes:

[0174] Obtain the second model parameters sent by the second network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0175] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0176] Send the first model parameters to the second network element;

[0177] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element through backpropagation training of the second sub-model based on the parameters of the first model.

[0178] In this embodiment of the invention, after the second network element obtains the second request and the model segmentation information, it performs forward propagation training based on the second sub-model and network data (network data is private data owned by the second network element) corresponding to the second network element indicated by the model segmentation information to obtain the second model parameters (i.e., intermediate model parameters), and sends the second model parameters to the first network element.

[0179] The first network element obtains the second model parameters, and performs forward and backward propagation training on the first sub-model based on the second model parameters to obtain the first model parameters (i.e., the gradient of the intermediate parameters of the model), and sends the first model parameters to the corresponding second network element.

[0180] The second network element obtains the first model parameters, performs backpropagation training on the second sub-model based on the first model parameters to obtain the trained second sub-model, and sends the trained second sub-model to the first network element.

[0181] like Figure 3 This is a schematic diagram of one implementation of the federated learning method described in the embodiments of the present invention. Figure 4 This is another flowchart illustrating one implementation of the federated learning method described in this invention. In the figure, "cloud" represents the third network element, "edge 1" represents the first network element 1, "edge 2" represents the first network element 2, "user 1" represents the second network element 1, "user 2" represents the second network element 2, and "user 3" represents the second network element 3; the second network element 1 and the second network element 2 correspond to the first network element 1, and the second network element 1, the second network element 2, and the first network element 1 constitute federated group 1; the second network element 3 corresponds to the first network element 2, and the second network element 3 and the first network element 2 constitute federated group 2.

[0182] In one of these implementation methods, the federated learning method includes the following steps:

[0183] Phase 1, Federated Learning Triggered:

[0184] S301, the second network element 3 sends a federated learning request to the third network element, including: model identifier; second network element identifier; and reason for request.

[0185] S302, the third network element sends feedback information about the federated learning request to the second network element 3.

[0186] Phase Two, Federated Learning Begins:

[0187] S303, the third network element sends a first request to the first network element 1 and the first network element 2, including at least one of the following: model identifier; model URL; expected return time; network element restriction information.

[0188] S304, the first network element 1 sends feedback information for the first request to the third network element.

[0189] S305, the first network element 1 sends a model request to the model storage unit to request a model, including: model identifier.

[0190] S306, the model storage unit sends the model indicated by the model identifier to the first network element 1.

[0191] Phase 3, Federated Learning Self-Organization (Can Iterate Multiple Rounds):

[0192] S307, the first network element 1 sends a second request to the second network element 1 and the second network element 2, including: link information and model identifier.

[0193] S308, the second network element 1 sends a first feedback message to the first network element 1 agreeing to participate in federated learning, including at least one of the following: link information; resource information; label data; whether the trained model is needed (e.g., no).

[0194] S308, the second network element 2 sends the first feedback message to the first network element 1, refusing to participate in federated learning.

[0195] S309, the first network element 1 calculation model segmentation point, is calculated based on the link information and resource information of the second network element 1, as well as the resource information of the first network element 1.

[0196] Phase Four: Federated Learning Model Training (requires multiple iterations):

[0197] S310, the first network element 1 sends model segmentation information to the second network element 1, including: the second sub-model; the return conditions of the second sub-model; and the model segmentation point information.

[0198] S311, the second network element 1 is trained by forward propagation based on the network data and the second sub-model to obtain the parameters of the second model.

[0199] S312, the second network element 1 sends the second model parameters (i.e., intermediate model parameters) to the first network element 1.

[0200] S313, the first network element 1 is trained by forward propagation and backward propagation based on the second model parameters and the first sub-model to obtain the first model parameters (i.e., the gradient of the intermediate parameters of the model).

[0201] S314, the first network element 1 sends the first model parameters to the second network element 1.

[0202] S315, the second network element 1 performs backpropagation training on the second sub-model based on the parameters of the first model to obtain the trained second sub-model.

[0203] S316, the second network element 1 sends the trained second sub-model to the first network element.

[0204] S317, Multi-terminal model merging: The first network element 1 merges the trained first sub-model and the trained second sub-model to obtain the model trained by the second network element; and, Multi-terminal model aggregation: Aggregate the model trained by at least one second network element to obtain the model trained by the first network element 1.

[0205] Phase 5: Federated Learning Model Aggregation and Distribution

[0206] S318, the first network element 1 sends model feedback information to the third network element, including at least one of the following: model identifier; the model trained by the first network element 1; network element completion information.

[0207] S319, Multilateral Model Merging: The third network element merges the models trained by at least one first network element to obtain a new model.

[0208] S320, the third network element sends a new model storage request to the model storage unit, including: model identifier; new model.

[0209] S321, the model storage unit sends feedback information to the third network element regarding the new model storage request, including: model identifier; model URL.

[0210] S322, the third network element sends new model distribution information to the second network element 3, including: model identifier; model URL. It should be noted that the second network element 3 previously sent the first feedback information to the first network element 2, agreeing to participate in federated learning, including: the model that needs to be trained.

[0211] S323, the second network element 3 sends a model request to the model storage unit, including: model identifier.

[0212] S324, the model storage unit sends a new model to the second network element 3.

[0213] In one implementation, optionally, S203, obtaining the trained second sub-model sent by the second network element includes:

[0214] A third request sent by the first target network element is obtained. The third request is sent based on a fourth request sent by the second network element. Both the third request and the fourth request are used to request the second network element to continue federated learning. The first target network element and the first network element are computing network elements of the same computing type.

[0215] Obtain the second model parameters of the second network element forwarded by the first target network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0216] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0217] Send the first model parameters to the first target network element;

[0218] Obtain the second sub-model trained by the second network element forwarded by the first target network element.

[0219] In this embodiment of the invention, the first target network element is the first network element corresponding to the second network element. The first target network element receives a fourth request sent by the second network element, the fourth request including at least one of the following: model identifier; model segmentation information; first network element identifier.

[0220] Next, there are three possible scenarios:

[0221] Scenario 1: The first target network element is performing federated learning of the same model. The first target network element includes the second network element in the corresponding federated group, continues federated learning, and sends a first notification to the third network element. The first notification is used to notify the second network element to continue federated learning and includes: model identifier and second network element identifier.

[0222] Scenario 2: The first target network element does not perform federated learning with the same model and exits federated learning. The first target network element sends a ninth request to the third network element. The ninth request is used to request the second network element to exit federated learning and includes: model identifier and second network element identifier.

[0223] Scenario 3: The first target network element does not perform federated learning with the same model, but continues to perform federated learning. The first target network element sends a third request to the first network element before the handover. The third request includes: model identifier and second network element identifier. This embodiment of the invention is precisely Scenario 3. Further, the first target network element sequentially forwards the second model parameters, the first model parameters, and the trained second sub-model between the first network element before the handover and the second network element.

[0224] Optionally, the third request is sent when the first network element corresponding to the second network element is switched;

[0225] The first network element corresponding to the second network element is switched to the first target network element.

[0226] In one embodiment, optionally, before the handover of the first network element corresponding to the second network element occurs, the method further includes one of the following:

[0227] Send a first notification to the second network element;

[0228] Obtain the first notification sent by the second network element;

[0229] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0230] It should be noted that when the second network element moves, the first network element corresponding to the second network element will be switched.

[0231] Optionally, before the handover of the first network element corresponding to the second network element occurs, the following two situations exist:

[0232] Scenario 1: The second network element sends a first notification to the first network element, the first notification including at least one of the following: movement direction; first target network element information. Next, the first network element may send a request to the first target network element, the request being used to forward federated learning, including: first network element identifier; model identifier; second network element identifier.

[0233] Scenario 2: The first network element sends a first notification to the second network element.

[0234] like Figure 5 This is a schematic diagram of one implementation method of the federated learning method described in the embodiments of the present invention. Figure 6 This is another flowchart illustrating a second implementation of the federated learning method described in this invention. In the diagram, "cloud" represents the third network element, "edge 1" represents the first network element 1, "edge 2" represents the first network element 2 (i.e., the first target network element), "user 1" represents the second network element 1, "user 2" represents the second network element 2, and "user 2new" represents the second network element 2 after switching the first network element. The second network element 1 and the second network element 2 correspond to the first network element 1 before the switch. The second network element 1, the second network element 2, and the first network element 1 before the switch constitute federated group 1. The second network element 2 after the switch and the first network element 2 constitute federated group 2. Before S501, the second network element 2 is performing self-organizing federated learning of the first network element and is about to switch due to the movement of the first network element.

[0235] In this second implementation method, the federated learning method includes the following steps:

[0236] S501, the first network element 1 sends a first notification to the second network element 2 to indicate that the first network element corresponding to the second network element 2 is about to switch.

[0237] S502, the second network element 2 sends the feedback information of the first notification to the first network element 1.

[0238] Next, the second network element 2 continues to move and switches to the first network element 2.

[0239] S503, the first network element 1 removes the second network element 2 from the federation group 1.

[0240] Then, the first network element 1 reports to the third network element that the second network element 2 has withdrawn from federated learning; the third network element finds that the second network element 2 has switched to the first network element 2, and sends an inquiry message to the second network element 2 asking whether to withdraw from federated learning. If the second network element 2 withdraws directly, the second network element 2 sends a feedback message to the third network element asking whether to withdraw from federated learning; if the second network element 2 does not withdraw, the following steps are executed.

[0241] S504, the second network element 2 sends a fourth request to the first network element 2, including at least one of the following: model identifier; model segmentation information; first network element identifier.

[0242] The switching phase includes the following three scenarios.

[0243] Scenario 1: The first network element 2 is performing federated learning on the same model:

[0244] S505, the first network element 2 incorporates the second network element 2 into the federation group 2.

[0245] S506, the first network element 2 sends feedback information for the fourth request to the second network element 2, which is used to instruct to continue federated learning.

[0246] S507, the first network element 2 sends a first notification to the third network element to instruct it to continue federated learning, including: model identifier and second network element identifier.

[0247] Scenario 2: The first network element 2 does not undergo federated learning with the same model and exits the federated learning process.

[0248] S508, the first network element 2 sends feedback information for the fourth request to the second network element 2, which is used to indicate the withdrawal from federated learning.

[0249] S509, the first network element 2 sends a ninth request to the third network element, including: model identifier and second network element identifier.

[0250] S510, the third network element sends feedback information for the ninth request to the first network element 2.

[0251] Scenario 3: If the first network element 2 does not undergo federated learning with the same model, continue federated learning:

[0252] S511, the first network element 2 sends a third request to the first network element, including: model identifier and second network element identifier.

[0253] S512, the first network element sends feedback information for the third request to the first network element 2.

[0254] S513, the first network element 2 sends feedback information for the fourth request to the second network element, which is used to instruct to continue federated learning.

[0255] S514, the first network element 2 sends a first notification to the third network element to instruct it to continue federated learning, including: model identifier and second network element identifier.

[0256] Following scenario three above, the steps for continuing federated learning are as follows:

[0257] It should be noted that the second network element 2 can only participate in this round of federated learning, and will not be able to participate in subsequent rounds.

[0258] S515, the second network element 2 is trained by forward propagation based on the network data and the second sub-model to obtain the second model parameters (i.e., the intermediate parameters of the model).

[0259] S516, the second network element 2 sends the second model parameters to the first network element 2.

[0260] S517, the first network element 2 forwards the second model parameters to the first network element.

[0261] S518, the first network element is trained by forward and backward propagation based on the second model parameters and the first sub-model to obtain the first model parameters (i.e., the gradient of the intermediate parameters of the model).

[0262] S519, the first network element sends the first model parameters to the first network element 2.

[0263] S520, the first network element 2 forwards the first model parameters to the second network element 2.

[0264] S521, the second network element 2 performs backpropagation training on the second sub-model based on the parameters of the first model to obtain the trained second sub-model.

[0265] S522, the second network element 2 sends the trained second sub-model to the first network element 2.

[0266] S523 is the second sub-model trained by the first network element 2 forwarding data to the first network element.

[0267] S524, the first network element performs multi-terminal model merging and aggregation.

[0268] In this embodiment of the invention, when the second network element is switched due to the movement of the first network element, the federated learning continues by forwarding parameters between the first network element and the second network element before the switch by the switched first network element, thereby realizing a fast connection mechanism, ensuring federated learning at the lowest cost, and ensuring the stability of federated learning.

[0269] In one embodiment, optionally, the above method further includes:

[0270] If the first network element has temporary resource occupation, a fifth request is sent to the third network element, the fifth request being used to request the transfer of federated learning;

[0271] A sixth request is sent to the second target network element, which is determined by the third network element based on the resource information of at least one first network element other than the first network element. The second target network element and the first network element belong to the same computing network element. The sixth request is a request for federated learning, including the second model parameters of the second network element.

[0272] Obtain the first model parameters sent by the second target network element, wherein the first model parameters are obtained by the second target network element training the first sub-model based on the second model parameters.

[0273] In this embodiment of the invention, when the first network element has temporary resource occupation, the federated learning task of the first network element can be transferred to the second target network element to realize a fast connection mechanism, so as to ensure the federated learning process at the lowest cost and ensure the stability of federated learning.

[0274] like Figure 7 This is a flowchart illustrating a third implementation of the federated learning method described in this invention. In the diagram, "cloud" represents the third network element, "edge 1" represents the first network element 1, and "edge 2" represents the first network element 2 (i.e., the second target network element). The first network element 1 is undergoing federated learning before S701, but due to temporary resource occupation, it cannot perform subsequent first sub-model training.

[0275] In this third implementation method, the federated learning method includes the following steps:

[0276] S701, the first network element 1 sends a fifth request to the third network element to request the transfer of federated learning, including: model identifier and model segmentation information.

[0277] S702, the third network element calculates the required resources and queries the appropriate first network element.

[0278] S703, the third network element sends the tenth request to the first network element 2 to request the transfer of federated learning, including: model identifier; model segmentation information; model URL.

[0279] S704, the first network element 2 sends feedback information for the sixth request to the third network element.

[0280] S705, the third network element sends feedback information of the fifth request to the first network element 1 to indicate that federated learning has been transferred, including: model identifier; segmentation point information; and second target network element identifier.

[0281] S706, the first network element 2 sends a model request to the model storage unit to request a model, including: model identifier.

[0282] S707, the model storage unit sends the model indicated by the model identifier to the first network element 2.

[0283] S708, the first network element 1 sends a sixth request to the first network element 2 to request federated learning, including: model identifier; second network element identifier; second model parameters (i.e., intermediate model parameters).

[0284] S709, the first network element 2 performs forward and backward propagation training on the first sub-model based on the second model parameters to obtain the first model parameters.

[0285] S710, the first network element 2 sends feedback information of the sixth request to the first network element 1, including: model identifier; second network element identifier; first model parameters.

[0286] It should be noted that S706 and S707 mentioned above can also be placed after S708.

[0287] In one embodiment, optionally, the above method further includes:

[0288] Obtain the seventh request sent by the second network element, the seventh request being used to request to exit federated learning, including: the second network element identifier and the model identifier;

[0289] The third network element sends an eighth request to the second network element, which is used to request the second network element to exit federated learning. The eighth request includes the second network element identifier and the model identifier. The third network element updates the credit score and network element restriction information of the second network element according to the eighth request.

[0290] In this embodiment of the invention, after the first network element obtains the seventh request from the second network element, it deletes the second network element from the corresponding federation group and reports it to the third network element. Thus, the third network element can record the second network element's default behavior for credit management.

[0291] like Figure 8This is a flowchart illustrating a fourth implementation of the federated learning method described in this invention. In the diagram, "cloud" represents the third network element, "edge 1" represents the first network element, "user 1" represents the second network element 1, and "user 2" represents the second network element 2. The second network element 1, the second network element 2, and the first network element constitute federated group 1. The second network element 2 is performing self-organizing federated learning of the first network element before step S801.

[0292] In this fourth implementation method, the federated learning method includes the following steps:

[0293] S801, the second network element 2 sends a seventh request to the first network element to request to exit federated learning, including: the second network element identifier and the model identifier.

[0294] S802, the first network element removes the second network element 2 from the federation group 1.

[0295] S803, the first network element sends the feedback information of the seventh request to the second network element 2.

[0296] S804, the first network element sends the eighth request to the third network element to request the second network element 2 to exit federated learning, including: the second network element identifier and the model identifier.

[0297] S805, the third network element sends feedback information for the eighth request to the first network element.

[0298] S806, the third network element sends a credit score update request to the credit management function, including: the second network element identifier and the update direction.

[0299] S807, the credit management function updates the credit score of the second network element 2.

[0300] S808, the credit management function sends credit score feedback information to the third network element, including: the identifier of the second network element 2 and the credit score.

[0301] S809, the third network element updates network element restriction information based on credit score.

[0302] In this embodiment of the invention, the above-mentioned federated learning method can be executed by the first network function or the first network virtual function of the first network element. This function can be called an Artificial Intelligence or Machine Learning Model Training Management Function (AI / ML ModelTraining Management Function). This function provides management services for AI / ML model training, especially management services for federated learning training. This service enables consumers to request federated learning training or request models for federated learning training, and can also provide or transmit model segmentation information, provide model aggregation functions, and forward or transfer federated learning-related data or federated learning computation tasks.

[0303] Another embodiment of the present invention, such as Figure 9 As shown, a federated learning method is provided and applied to a third network element. The method includes:

[0304] S901, Obtain at least one trained model sent by the first network element.

[0305] It should be noted that the trained model sent by the first network element can be the model obtained by the first network element executing S201 to S204, which will not be elaborated here.

[0306] S902, aggregate the trained models sent by at least one of the first network elements to obtain an aggregated model.

[0307] In this embodiment of the invention, the third network element aggregates the models trained by all the first network elements to achieve multilateral model aggregation, obtain the aggregated model (i.e., the new model), and complete the federated learning task.

[0308] Optionally, after S902, the above method further includes:

[0309] The third network element can send a model update request to the model storage unit or the model location. The model update request includes at least one of the following: a model identifier and a new model. After the model is stored, the third network element can distribute the model to the model requester (i.e., the second network element that needs the model), including: a model identifier and a model URL. The model requester requests a new model from the model storage unit or the model location based on the model identifier and / or the model URL.

[0310] In one embodiment, optionally, before obtaining the trained model sent by at least one first network element in S901, the above method further includes:

[0311] Obtain resource information sent by multiple first network elements and link information sent by multiple second network elements;

[0312] Based on the resource information of multiple first network elements and the link information of multiple second network elements, at least one first network element and at least one second network element participating in federated learning are determined.

[0313] Send a first request to at least one of the first network elements.

[0314] In this embodiment of the invention, before S901, optionally, before sending the first request to the first network element, the third network element obtains resource information of multiple first network elements and link information of multiple second network elements. This allows it to select a suitable edge cluster (which includes at least one first network element and at least one corresponding second network element) for this federated learning process based on the resource information of the first network elements and the link information of the second network elements. For example, if a second network element under a certain first network element is in a high-speed moving state, posing a significant risk of network element switching for the second network element, the third network element will not select that first network element; or, if a certain first network element has a large amount of already occupied resources, making federated learning difficult, the third network element will also not select that first network element.

[0315] In addition, before obtaining the trained model sent by at least one first network element in S901, the above method further includes:

[0316] Based on the credit score of the second network element, at least one of the second network elements participating in federated learning is determined.

[0317] The third network element can prioritize selecting the second network element whose credit score is greater than the credit threshold and has a higher credit score.

[0318] It should be noted that the first network element needs to open its resource information, and the second network element needs to open its link information to the third network element, so that the third network element can quickly and selectively choose the participants in the federated learning, thereby reducing the additional signaling overhead caused by selection failure.

[0319] In one embodiment, optionally, before sending the first request to at least one of the first network elements, the method further includes:

[0320] Obtain the target request, which includes one of the following:

[0321] The federated learning request sent by the second network element;

[0322] The model request sent by the second network element;

[0323] The model request sent by the first network element;

[0324] Based on the target request, determine whether the model indicated by the target request needs to undergo federated learning.

[0325] It should be noted that the target request includes a request reason, and the third network element can determine whether the model needs to perform federated learning based on the request reason. The request reason includes at least one of the following: network degradation; decreased accuracy; outdated version; model change; data change. For example, if the network degradation reaches a degradation threshold, it is determined that federated learning is needed; if the accuracy decreases to an accuracy threshold, it is determined that federated learning is needed.

[0326] In one embodiment, optionally, the above method further includes:

[0327] Obtain the eighth request sent by the first network element, the eighth request being used to request the second network element to exit federated learning;

[0328] Update the credit score of the second network element according to the eighth request;

[0329] If the credit score of the second network element is less than the credit threshold, update the network element restriction information.

[0330] It is understandable that when a second network element withdraws from federated learning, it means that the second network element is participating in federated learning. Therefore, the second network element is a defaulter in federated learning. Based on the default behavior of the second network element, the third network element can update the credit score of the second network element. If the credit score is lower than the credit threshold, the network element can restrict the second network element from participating in or requesting federated learning through network element restriction information. This provides an exit mechanism and penalty measures for defaulters in federated learning, restricts the unjustified default of federated learning participants, and realizes credit management.

[0331] In this embodiment of the invention, the above-mentioned federated learning method can be executed by the first network function or the first network virtual function of the third network element. This function can be called the Federated Learning Management Function, which provides management services for federated learning training. This service enables consumers to request federated learning training or request models for federated learning training, and also provides model aggregation functions, as well as the ability to discover first and / or second network elements that meet the federated learning task. Furthermore, the second network function or the second network virtual function of the third network element can provide an exit mechanism and penalty measures for parties that default on federated learning; this function can be called a credit management function.

[0332] Another embodiment of the present invention, such as Figure 10 As shown, a federated learning method is provided and applied to a second network element. The method includes:

[0333] S1001, Obtain model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element.

[0334] It is understandable that since the first network element determines the model segmentation point based on the link information of the second network element and the resource information of the first network element, different second network elements correspond to different model segmentation points, thereby making full use of wireless resources.

[0335] S1002, the second sub-model is trained according to the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model.

[0336] S1003, send the trained second sub-model to the first network element.

[0337] In one embodiment, optionally, before obtaining the model segmentation information sent by the first network element in S1001, the method further includes:

[0338] Send the target request;

[0339] Obtain the second request sent by the first network element;

[0340] Send a first feedback message to the first network element, the first feedback message being used to instruct the second network element to agree to participate in federated learning;

[0341] The target request to be sent includes one of the following:

[0342] Send a first request to the first network element, the first request being used to request federated learning;

[0343] Send a federated learning request to the third network element;

[0344] Send a model request to the first network element;

[0345] Send a model request to the third network element.

[0346] Therefore, in this embodiment of the invention, the second network element can send a federated learning request or a model request to the first network element or the third network element.

[0347] In one implementation, optionally, in step S1002, the second sub-model is trained based on the first model parameters to obtain the trained second sub-model, including:

[0348] The second sub-model is trained by forward propagation based on network data to obtain the parameters of the second model.

[0349] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0350] In this embodiment of the invention, the second network element is trained using its own private data (i.e., network data) locally, ensuring the data privacy of the second network element. Furthermore, the interactive training mechanism with the first network element ensures the up-to-date state of the second sub-model, improving training efficiency.

[0351] In one implementation, optionally, in step S1002, the second sub-model is trained based on the first model parameters to obtain the trained second sub-model, including:

[0352] In the event that the first network element corresponding to the second network element is switched, a fourth request is sent to the first network element after the switch, the fourth request being used to request the second network element to continue federated learning;

[0353] Send the second model parameters to the first network element after the switch. The second model parameters are obtained by forward propagation training of the second sub-model based on the network data.

[0354] Obtain the first model parameters forwarded by the first target network element;

[0355] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0356] In one embodiment, optionally, before the handover of the first network element corresponding to the second network element occurs, the above method further includes one of the following:

[0357] Obtain the first notification sent by the first network element;

[0358] Send a first notification to the first network element;

[0359] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0360] In summary, the federated learning method described in this embodiment of the invention is a segmented federated learning method under a three-layer (cloud-side, edge-side, and terminal-side) federated learning architecture. By jointly managing the resource information of the three-layer devices, it dynamically selects model segmentation points for each second network element, so that the front-end federated learning training task is performed on the second network element, while the back-end federated learning training task is performed on the first network element. When the current first network element has temporary resource occupation, the segmented training task is offloaded to the cloud side or other first network elements, thereby making full use of wireless resources and improving computational and training efficiency without exposing terminal privacy.

[0361] like Figure 11 As shown, one embodiment of the present invention provides a federated learning device applied to a first network element, the device comprising:

[0362] The first acquisition module 1101 is used to segment the model according to the model segmentation point for each second network element, and acquire the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined according to the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element.

[0363] The first sending module 1102 is used to send corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model;

[0364] The second acquisition module 1103 is used to acquire the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model.

[0365] The second sending module 1104 is used to obtain the model trained by the first network element based on the trained second sub-model, and send the model trained by the first network element to the third network element.

[0366] Optionally, in the federated learning device, the second sending module 1104 is specifically used for:

[0367] The first sub-model and the second sub-model after training are merged to obtain the model after training the second network element. The first sub-model after training is obtained based on the parameters of the first model.

[0368] The models trained by at least one of the second network elements are aggregated to obtain the model trained by the first network element.

[0369] Optionally, the federated learning device further includes:

[0370] The fifth acquisition module is used to acquire the first request, which is used to request federated learning;

[0371] The fourth sending module is used to send a second request to at least one of the second network elements according to the first request, wherein the second request is used to request the second network element to participate in federated learning;

[0372] The sixth acquisition module is used to acquire at least one first feedback message sent by the second network element, wherein the first feedback message is used to instruct the second network element to agree to participate in federated learning.

[0373] Optionally, in the federated learning device, the fourth sending module is specifically used for:

[0374] Based on the network element restriction information in the first request, at least one second network element is determined to participate in federated learning, wherein the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning.

[0375] Send a second request to at least one of the second network elements.

[0376] Optionally, in the federated learning device, the first sending module 1102 is specifically used for:

[0377] Obtain the second model parameters sent by the second network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0378] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0379] Send the first model parameters to the second network element;

[0380] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element through backpropagation training of the second sub-model based on the parameters of the first model.

[0381] Optionally, in the federated learning device, the first sending module 1102 is specifically used for:

[0382] Obtain a third request sent by the first target network element, wherein the third request is sent based on a fourth request sent by the second network element, and both the third request and the fourth request are used to request the second network element to continue federated learning;

[0383] Obtain the second model parameters of the second network element forwarded by the first target network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0384] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0385] Send the first model parameters to the first target network element;

[0386] Obtain the second sub-model trained by the second network element forwarded by the first target network element.

[0387] Optionally, in the federated learning device, the third request is sent when the first network element corresponding to the second network element is switched;

[0388] Before the handover of the first network element corresponding to the second network element occurs, the device further includes one of the following:

[0389] The fifth sending module is used to send the first notification to the second network element;

[0390] The sixth acquisition module is used to acquire the first notification sent by the second network element;

[0391] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0392] Optionally, the federated learning apparatus further includes:

[0393] The sixth sending module is used to send a fifth request to the third network element when the first network element has temporary resource occupation. The fifth request is used to request the transfer of federated learning.

[0394] The seventh sending module is used to send a sixth request to the second target network element, the second target network element being determined by the third network element based on the resource information of at least one first network element other than the first network element, and the sixth request using federated learning, including the second model parameters of the second network element;

[0395] The seventh acquisition module is used to acquire the first model parameters sent by the second target network element. The first model parameters are obtained by the second target network element training the first sub-model based on the second model parameters.

[0396] Optionally, the federated learning apparatus further includes:

[0397] The eighth acquisition module is used to acquire the seventh request sent by the second network element, the seventh request being used to request to exit federated learning;

[0398] The eighth sending module is used to send an eighth request to the third network element. The eighth request is used to request the second network element to withdraw from federated learning. The third network element updates the credit score and network element restriction information of the second network element according to the eighth request.

[0399] The federated learning device provided in this embodiment of the invention can execute the federated learning method embodiment applied to the first network element described above. Its implementation principle and technical effect are similar, and will not be repeated here.

[0400] like Figure 12 As shown, another embodiment of the present invention provides a federated learning device applied to a third network element, the device comprising:

[0401] The third acquisition module 1201 is used to acquire at least one trained model sent by the first network element;

[0402] The first obtaining module 1202 is used to aggregate the trained models sent by at least one of the first network elements to obtain an aggregated model.

[0403] Optionally, the federated learning apparatus further includes, before acquiring the trained model sent by at least one first network element:

[0404] The ninth acquisition module is used to acquire resource information sent by multiple first network elements and link information sent by multiple second network elements;

[0405] The first determining module is used to determine at least one first network element and at least one second network element participating in federated learning based on the resource information of multiple first network elements and the link information of multiple second network elements.

[0406] The ninth sending module is used to send a first request to at least one of the first network elements.

[0407] Optionally, the federated learning apparatus further includes, before sending the first request to at least one of the first network elements:

[0408] The tenth acquisition module is used to acquire a target request, wherein the target request includes one of the following:

[0409] The federated learning request sent by the second network element;

[0410] The model request sent by the second network element;

[0411] The model request sent by the first network element;

[0412] The first judgment module is used to determine, based on the target request, whether the model indicated by the target request needs to undergo federated learning.

[0413] Optionally, the federated learning apparatus further includes:

[0414] The eleventh acquisition module is used to acquire the eighth request sent by the first network element, the eighth request being used to request the second network element to exit federated learning;

[0415] The first update module is used to update the credit score of the second network element according to the eighth request;

[0416] The second update module is used to update the network element restriction information when the credit score of the second network element is less than the credit threshold.

[0417] The federated learning device provided in this embodiment of the invention can execute the federated learning method embodiment applied to the third network element described above. Its implementation principle and technical effect are similar, and will not be repeated here.

[0418] like Figure 13As shown, another embodiment of the present invention provides a federated learning device applied to a second network element, the device comprising:

[0419] The fourth acquisition module 1301 is used to acquire model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element;

[0420] The second obtaining module 1302 is used to train the second sub-model according to the first model parameters to obtain the trained second sub-model, wherein the first model parameters are obtained by the first network element training the first sub-model;

[0421] The third sending module 1303 is used to send the trained second sub-model to the first network element.

[0422] Optionally, the federated learning apparatus further includes, before acquiring the model segmentation information sent by the first network element:

[0423] The tenth sending module is used to send the target request;

[0424] The twelfth acquisition module is used to acquire the second request sent by the first network element;

[0425] The eleventh sending module is used to send first feedback information to the first network element, and the first feedback information is used to instruct the second network element to agree to participate in federated learning.

[0426] Specifically, the tenth sending module is used for one of the following:

[0427] Send a first request to the first network element, the first request being used to request federated learning;

[0428] Send a federated learning request to the third network element;

[0429] Send a model request to the first network element;

[0430] Send a model request to the third network element.

[0431] Optionally, in the federated learning device, the second obtaining module 1302 is specifically used for:

[0432] The second sub-model is trained by forward propagation based on network data to obtain the parameters of the second model.

[0433] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0434] Optionally, in the federated learning device, the second obtaining module 1302 is specifically used for:

[0435] In the event of a switchover of the first network element corresponding to the second network element, a fourth request is sent to the first target network element, the fourth request being used to request the second network element to continue federated learning;

[0436] Send the second model parameters to the first target network element. The second model parameters are obtained by forward propagation training of the second sub-model based on the network data.

[0437] Obtain the first model parameters forwarded by the first target network element;

[0438] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0439] Optionally, the federated learning device further includes one of the following before the handover of the first network element corresponding to the second network element:

[0440] The thirteenth acquisition module is used to acquire the first notification sent by the first network element;

[0441] The twelfth sending module is used to send a first notification to the first network element;

[0442] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0443] The federated learning device provided in this embodiment of the invention can execute the federated learning method embodiment applied to the second network element described above. Its implementation principle and technical effect are similar, and will not be described again here.

[0444] like Figure 14 As shown, this embodiment of the invention also provides a network device, including: a processor 1401; and a memory 1402 connected to the processor 1401 via a bus interface, the memory 1402 being used to store programs and data used by the processor 1401 when performing operations, and the processor 1401 calling and executing the programs and data stored in the memory 1402.

[0445] The network device also includes a transceiver 1403, which is connected to a bus interface and is used to receive and send data under the control of the processor 1401.

[0446] The network device may be one of the first network element, the third network element, and the second network element mentioned above.

[0447] In this embodiment of the invention, the network device is taken as the first network element.

[0448] Specifically, the processor 1401 executes the following process:

[0449] For each second network element, the model is segmented according to the model segmentation point to obtain the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined based on the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element.

[0450] The transceiver 1403 is used to perform the following processes under the control of the processor 1401:

[0451] Send the corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model;

[0452] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model.

[0453] The trained model of the first network element is obtained based on the trained second sub-model, and the trained model of the first network element is sent to the third network element.

[0454] Among them, Figure 14 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 1401) and memory (memory 1402). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides a user interface 1404. A transceiver 1403 may be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over a transmission medium. Processor 1401 is responsible for managing the bus architecture and general processing, and memory 1402 may store data used by processor 1401 during operation.

[0455] The processor 1401 is responsible for managing the bus architecture and general processing, while the memory 1402 can store the data used by the processor 1401 when performing operations.

[0456] Optionally, in the network device, the processor 1401 is specifically configured to read the computer program and execute the following steps:

[0457] The trained second sub-model and the trained first sub-model are merged to obtain the model after training the second network element, wherein the trained first sub-model is obtained based on the parameters of the first model.

[0458] The models trained by at least one of the second network elements are aggregated to obtain the model trained by the first network element.

[0459] Optionally, in the network device, the transceiver 1403 is further configured to perform the following processes under the control of the processor 1401:

[0460] Obtain the first request, which is used to request federated learning;

[0461] Based on the first request, a second request is sent to at least one second network element, the second request being used to request the second network element to participate in federated learning;

[0462] Obtain at least one first feedback message sent by the second network element, wherein the first feedback message is used to instruct the second network element to agree to participate in federated learning.

[0463] Optionally, in the network device, the transceiver 1403 is specifically configured to perform the following processes under the control of the processor 1401:

[0464] Based on the network element restriction information in the first request, at least one second network element is determined to participate in federated learning, wherein the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning.

[0465] Send a second request to at least one of the second network elements.

[0466] Optionally, in the network device, the transceiver 1403 is specifically configured to perform the following processes under the control of the processor 1401:

[0467] Obtain the second model parameters sent by the second network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0468] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0469] Send the first model parameters to the second network element;

[0470] Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element through backpropagation training of the second sub-model based on the parameters of the first model.

[0471] Optionally, in the network device, the transceiver 1403 is specifically configured to perform the following processes under the control of the processor 1401:

[0472] Obtain a third request sent by the first target network element, wherein the third request is sent based on a fourth request sent by the second network element, and both the third request and the fourth request are used to request the second network element to continue federated learning;

[0473] Obtain the second model parameters of the second network element forwarded by the first target network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data.

[0474] The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters.

[0475] Send the first model parameters to the first target network element;

[0476] Obtain the second sub-model trained by the second network element forwarded by the first target network element.

[0477] Optionally, in the network device, the third request is sent when the first network element corresponding to the second network element undergoes a handover;

[0478] The transceiver 1403 is also used to perform one of the processes under the control of the processor 1401:

[0479] Send a first notification to the second network element;

[0480] Obtain the first notification sent by the second network element;

[0481] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0482] Optionally, in the network device, the transceiver 1403 is further configured to perform the following processes under the control of the processor 1401:

[0483] If the first network element has temporary resource occupation, a fifth request is sent to the third network element, the fifth request being used to request the transfer of federated learning;

[0484] A sixth request is sent to a second target network element, which is determined by the third network element based on resource information of at least one first network element other than the first network element. The sixth request uses federated learning to request the second model parameters of the second network element.

[0485] Obtain the first model parameters sent by the second target network element, wherein the first model parameters are obtained by the second target network element training the first sub-model based on the second model parameters.

[0486] Optionally, in the network device, the transceiver 1403 is further configured to perform the following processes under the control of the processor 1401:

[0487] Obtain the seventh request sent by the second network element, the seventh request being used to request to exit federated learning;

[0488] The third network element sends an eighth request to the second network element, which is used to request the second network element to withdraw from federated learning. The third network element updates the credit score and network element restriction information of the second network element according to the eighth request.

[0489] Next, in this embodiment of the invention, the network device is taken as a third network element.

[0490] Specifically, the transceiver is configured to perform the following processes under the control of the processor:

[0491] Obtain the trained model sent by at least one first network element;

[0492] The processor executes the following procedures:

[0493] The trained models sent by at least one of the first network elements are aggregated to obtain an aggregated model.

[0494] Optionally, in the network device, the transceiver is further configured to perform the following processes under the control of the processor:

[0495] Obtain resource information sent by multiple first network elements and link information sent by multiple second network elements;

[0496] The processor is also configured to read the computer program and perform the following steps:

[0497] Based on the resource information of multiple first network elements and the link information of multiple second network elements, at least one first network element and at least one second network element participating in federated learning are determined.

[0498] The transceiver is also configured to perform the following processes under the control of the processor:

[0499] Send a first request to at least one of the first network elements.

[0500] Optionally, in the network device, the transceiver is further configured to perform the following processes under the control of the processor:

[0501] Obtain the target request, which includes one of the following:

[0502] The federated learning request sent by the second network element;

[0503] The model request sent by the second network element;

[0504] The model request sent by the first network element;

[0505] The processor is also configured to read the computer program and perform the following steps:

[0506] Based on the target request, determine whether the model indicated by the target request needs to undergo federated learning.

[0507] Optionally, in the network device, the transceiver is further configured to perform the following processes under the control of the processor:

[0508] Obtain the eighth request sent by the first network element, the eighth request being used to request the second network element to exit federated learning;

[0509] The processor is also configured to read the computer program and perform the following steps:

[0510] Update the credit score of the second network element according to the eighth request;

[0511] If the credit score of the second network element is less than the credit threshold, update the network element restriction information.

[0512] Next, in this embodiment of the invention, the network device is taken as the second network element.

[0513] Specifically, the transceiver is configured to perform the following processes under the control of the processor:

[0514] Obtain model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element;

[0515] The processor executes the following procedures:

[0516] The second sub-model is trained based on the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model.

[0517] The transceiver is configured to perform the following processes under the control of the processor:

[0518] Send the trained second sub-model to the first network element.

[0519] Optionally, in the network device, the transceiver is further configured to perform the following processes under the control of the processor:

[0520] Send the target request;

[0521] Obtain the second request sent by the first network element;

[0522] Send a first feedback message to the first network element, the first feedback message being used to instruct the second network element to agree to participate in federated learning;

[0523] The target request to be sent includes one of the following:

[0524] Send a first request to the first network element, the first request being used to request federated learning;

[0525] Send a federated learning request to the third network element;

[0526] Send a model request to the first network element;

[0527] Send a model request to the third network element.

[0528] Optionally, in the network device, the processor is specifically configured to read the computer program and execute the following steps:

[0529] The second sub-model is trained by forward propagation based on network data to obtain the parameters of the second model.

[0530] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0531] Optionally, in the network device, the processor is specifically configured to read the computer program and execute the following steps:

[0532] In the event of a switchover of the first network element corresponding to the second network element, a fourth request is sent to the first target network element, the fourth request being used to request the second network element to continue federated learning;

[0533] Send the second model parameters to the first target network element. The second model parameters are obtained by forward propagation training of the second sub-model based on the network data.

[0534] Obtain the first model parameters forwarded by the first target network element;

[0535] The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

[0536] Optionally, in the network device, the transceiver is further configured to perform one of the following under the control of the processor:

[0537] Obtain the first notification sent by the first network element;

[0538] Send a first notification to the first network element;

[0539] The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

[0540] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps in the above-described federated learning method and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0541] In addition, embodiments of the present invention also provide a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 2 , Figure 9 or Figure 10 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0542] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0543] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can be physically comprised separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware or in the form of hardware plus software functional units.

[0544] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions that cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0545] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A federated learning method, characterized in that, Applied to the first network element, the method includes: For each second network element, the model is segmented according to the model segmentation point to obtain the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined based on the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element. Send the corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model; Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model. The trained model of the first network element is obtained based on the trained second sub-model, and the trained model of the first network element is sent to the third network element.

2. The federated learning method according to claim 1, characterized in that, The model trained on the first network element is obtained based on the trained second sub-model, including: The first sub-model and the second sub-model after training are merged to obtain the model after training the second network element, wherein the first sub-model after training is obtained based on the parameters of the first model. The models trained by at least one of the second network elements are aggregated to obtain the model trained by the first network element.

3. The federated learning method according to claim 1, characterized in that, Before segmenting the model according to the model segmentation points for each second network element and obtaining the first sub-model and second sub-model corresponding to the second network element, the method further includes: Obtain the first request, which is used to request federated learning; Based on the first request, a second request is sent to at least one second network element, the second request being used to request the second network element to participate in federated learning; Obtain at least one first feedback message sent by the second network element, wherein the first feedback message is used to instruct the second network element to agree to participate in federated learning.

4. The federated learning method according to claim 3, characterized in that, Based on the first request, a second request is sent to at least one of the second network elements, including: Based on the network element restriction information in the first request, at least one second network element is determined to participate in federated learning, wherein the network element restriction information is used to indicate the second network element that is restricted from participating in federated learning. Send a second request to at least one of the second network elements.

5. The federated learning method according to claim 1, characterized in that, Obtaining the trained second sub-model sent by the second network element includes: Obtain the second model parameters sent by the second network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data. The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters. Send the first model parameters to the second network element; Obtain the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element through backpropagation training of the second sub-model based on the parameters of the first model.

6. The federated learning method according to claim 1, characterized in that, Obtaining the trained second sub-model sent by the second network element includes: Obtain a third request sent by the first target network element, wherein the third request is sent based on a fourth request sent by the second network element, and both the third request and the fourth request are used to request the second network element to continue federated learning; Obtain the second model parameters of the second network element forwarded by the first target network element. The second model parameters are obtained by the second network element through forward propagation training of the second sub-model based on the network data. The first sub-model is trained by forward and backward propagation based on the second model parameters to obtain the first model parameters. Send the first model parameters to the first target network element; Obtain the second sub-model trained by the second network element forwarded by the first target network element.

7. The federated learning method according to claim 6, characterized in that, The third request is sent when the first network element corresponding to the second network element undergoes a handover; before the handover of the first network element corresponding to the second network element, the method further includes one of the following: Send a first notification to the second network element; Obtain the first notification sent by the second network element; The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

8. The federated learning method according to claim 1, characterized in that, The method further includes: If the first network element has temporary resource occupation, a fifth request is sent to the third network element, the fifth request being used to request the transfer of federated learning; A sixth request is sent to the second target network element, which is determined by the third network element based on the resource information of at least one first network element other than the first network element. The sixth request is used to request federated learning, including the second model parameters of the second network element. Obtain the first model parameters sent by the second target network element, wherein the first model parameters are obtained by the second target network element training the first sub-model based on the second model parameters.

9. The federated learning method according to claim 1, characterized in that, The method further includes: Obtain the seventh request sent by the second network element, the seventh request being used to request to exit federated learning; The third network element sends an eighth request to the second network element, which is used to request the second network element to withdraw from federated learning. The third network element updates the credit score and network element restriction information of the second network element according to the eighth request.

10. A federated learning method, characterized in that, Applied to a third network element, the method includes: Obtain the trained model sent by at least one first network element; The trained models sent by at least one of the first network elements are aggregated to obtain an aggregated model.

11. The federated learning method according to claim 10, characterized in that, Before acquiring the trained model sent by at least one first network element, the method further includes: Obtain resource information sent by multiple first network elements and link information sent by multiple second network elements; Based on the resource information of multiple first network elements and the link information of multiple second network elements, at least one first network element and at least one second network element participating in federated learning are determined. Send a first request to at least one of the first network elements.

12. The federated learning method according to claim 11, characterized in that, Before sending the first request to at least one of the first network elements, the method further includes: Obtain the target request, which includes one of the following: The federated learning request sent by the second network element; The first network element sends a federated learning request; The model request sent by the second network element; The model request sent by the first network element; Based on the target request, determine whether the model indicated by the target request needs to undergo federated learning.

13. The federated learning method according to claim 11, characterized in that, The method further includes: Obtain the eighth request sent by the first network element, the eighth request being used to request the second network element to exit federated learning; Update the credit score of the second network element according to the eighth request; If the credit score of the second network element is less than the credit threshold, update the network element restriction information.

14. A federated learning method, characterized in that, Applied to a second network element, the method includes: Obtain model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element; The second sub-model is trained based on the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model. The trained second sub-model is sent to the first network element.

15. The federated learning method according to claim 14, characterized in that, Before obtaining the model segmentation information sent by the first network element, the method further includes: Send the target request; Obtain the second request sent by the first network element; Send a first feedback message to the first network element, the first feedback message being used to instruct the second network element to agree to participate in federated learning; The target request to be sent includes one of the following: Send a first request to the first network element, the first request being used to request federated learning; Send a federated learning request to the third network element; Send a model request to the first network element; Send a model request to the third network element.

16. The federated learning method according to claim 14, characterized in that, The second sub-model is trained based on the parameters of the first model to obtain the trained second sub-model, including: The second sub-model is trained by forward propagation based on network data to obtain the parameters of the second model. The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

17. The federated learning method according to claim 16, characterized in that, The second sub-model is trained based on the parameters of the first model to obtain the trained second sub-model, including: In the event of a switchover of the first network element corresponding to the second network element, a fourth request is sent to the first target network element, the fourth request being used to request the second network element to continue federated learning; Send the second model parameters to the first target network element. The second model parameters are obtained by forward propagation training of the second sub-model based on the network data. Obtain the first model parameters forwarded by the first target network element; The second sub-model is trained by backpropagation based on the parameters of the first model to obtain the trained second sub-model.

18. The federated learning method according to claim 16, characterized in that, Before the handover of the first network element corresponding to the second network element occurs, the method further includes one of the following: Obtain the first notification sent by the first network element; Send a first notification to the first network element; The first notification is used to indicate that the first network element corresponding to the second network element is about to be switched.

19. A federated learning device, characterized in that, Applied to the first network element, the device includes: The first acquisition module is used to segment the model according to the model segmentation point for each second network element, and acquire the first sub-model and the second sub-model corresponding to the second network element. The model segmentation point is determined according to the resource information of the first network element and the link information of the second network element. The first sub-model is on the first network element and the second sub-model is on the second network element. The first sending module is used to send corresponding model segmentation information to the second network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model; The second acquisition module is used to acquire the trained second sub-model sent by the second network element. The trained second sub-model is obtained by the second network element training the second sub-model according to the first model parameters. The first model parameters are obtained by the first network element training the first sub-model. The second sending module is used to obtain the model trained by the first network element based on the trained second sub-model, and send the model trained by the first network element to the third network element.

20. A federated learning device, characterized in that, Applied to a third network element, the device includes: The third acquisition module is used to acquire at least one trained model sent by the first network element; The first obtaining module is used to aggregate the trained models sent by at least one of the first network elements to obtain an aggregated model.

21. A federated learning device, characterized in that, Applied to a second network element, the device includes: The fourth acquisition module is used to acquire model segmentation information sent by the first network element, wherein the model segmentation information is used to indicate the first sub-model and the second sub-model corresponding to the second network element; The second acquisition module is used to train the second sub-model according to the first model parameters to obtain the trained second sub-model. The first model parameters are obtained by the first network element training the first sub-model. The third sending module is used to send the trained second sub-model to the first network element.

22. A network device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the federated learning method as described in any one of claims 1 to 9, or implements the federated learning method as described in claims 10 to 13, or implements the federated learning method as described in claims 14 to 18.

23. A readable storage medium, characterized in that, The readable storage medium stores a program that, when executed by a processor, implements the federated learning method as described in any one of claims 1 to 9, or the federated learning method as described in claims 10 to 13, or the federated learning method as described in claims 14 to 18.

24. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the federated learning method as described in any one of claims 1 to 9, or the federated learning method as described in claims 10 to 13, or the federated learning method as described in claims 14 to 18.

Images