Electronic device for federated learning server, and electronic device for core network

Abstract

An electronic device for a federated learning server, and an electronic device for a core network The electronic device for a federated learning server comprises at least one processor and at least one memory, wherein the at least one memory comprises a computer program code, and the at least one memory and the computer program code are configured to enable, by means of the at least one processor, the electronic device to: send to a core network order-of-magnitude information of the order of magnitude of a parameter of a global model for federated learning, so as to assist the core network in identifying from among a plurality of federated learning clients a federated learning client, which cannot upload all of the parameters of a local model of the federated learning client, as a partial upload client, and aggregate, when a preset condition is met, some of the parameters of the local model that are received from the partial upload client, so as to obtain the global model in the current round of training of federated learning.

Description

Electronic devices for federated learning servers and electronic devices for the core network

[0001] This application claims priority to Chinese Patent Application No. 202311763059.1, filed on December 20, 2023, entitled "Electronic Device for Federated Learning Server and Electronic Device for Core Network", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure relates to the field of wireless communication technology, and more specifically to electronic devices for federated learning servers, electronic devices for core networks, and electronic devices for federated learning clients. More specifically, it relates to electronic devices for federated learning servers, electronic devices for core networks, and electronic devices for federated learning clients based on partial parameters of local models. Background Technology

[0003] With the continuous growth of intelligent terminal services, the requirements for wireless networks to support intelligent models are becoming increasingly stringent. Typically, user equipment (UE) datasets differ, and UEs are unwilling to share their datasets due to privacy concerns, creating data silos. Furthermore, UE-specific datasets are often small in size, making it difficult to train more accurate and generalized machine learning (ML) models. To address this, Federated Learning (FL) breaks down data silos between UEs through a training model where FL clients (e.g., UEs) train local models locally, the FL server aggregates the local models from multiple FL clients into a global model, and the global model is distributed to FL clients for training. This enables the training of more accurate and generalized ML models.

[0004] Summary of the Invention

[0005] A brief overview of the invention is given below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0006] According to one aspect of this disclosure, an electronic device for a federated learning server is provided, the electronic device including at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, via the at least one processor, to cause the electronic device to perform: sending to a core network magnitude information about the parameter magnitude of a global model for federated learning, to assist the core network in identifying federated learning clients from multiple federated learning clients that cannot upload all parameters of their local models as partial upload clients; and, if predetermined conditions are met, aggregating based on partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0007] According to one aspect of this disclosure, an electronic device for a core network is provided, the electronic device including at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, via the at least one processor, to cause the electronic device to: receive magnitude information about the magnitude of parameters of a global model for federated learning from a federated learning server, to identify federated learning clients from a plurality of federated learning clients that cannot upload all parameters of their local models as partial upload clients, so that, under predetermined conditions, the federated learning server aggregates the partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0008] According to one aspect of this disclosure, an electronic device for a federated learning client is provided, the electronic device including at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, through the at least one processor, to cause the electronic device to: send partial parameters of a local model for federated learning to a federated learning server, upon satisfying predetermined conditions, so that the federated learning server can aggregate the local model based on the partial parameters to obtain a global model in the current round of training of the federated learning, wherein the federated learning server sends information about the magnitude of the parameters of the global model to the core network to assist the core network in identifying the electronic device as a partial upload client that cannot upload all parameters of its local model.

[0009] According to one aspect of this disclosure, a method for a federated learning server is provided, comprising: sending to a core network magnitude information about the parameter magnitude of a global model for federated learning to assist the core network in identifying federated learning clients from multiple federated learning clients that cannot upload all parameters of their local models as partial upload clients; and, if predetermined conditions are met, aggregating based on partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0010] According to one aspect of this disclosure, a method for a core network is provided, comprising: receiving magnitude information about the magnitude of parameters of a global model for federated learning from a federated learning server, identifying federated learning clients from a plurality of federated learning clients that cannot upload all parameters of their local models as partial upload clients, so that, under predetermined conditions, the federated learning server aggregates the partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0011] According to one aspect of this disclosure, a method for a federated learning client is provided, comprising: sending partial parameters of a local model for federated learning to a federated learning server under predetermined conditions, so that the federated learning server can aggregate the local model based on the partial parameters to obtain a global model in the current round of training of the federated learning, wherein the federated learning server sends information about the magnitude of the parameters of the global model to the core network to assist the core network in identifying the federated learning client as a partial upload client that cannot upload all parameters of its local model.

[0012] According to other aspects of the present invention, computer program code and computer program product for implementing the above methods, as well as a computer-readable storage medium having the computer program code for implementing the above methods recorded thereon, are also provided. Attached Figure Description

[0013] To further illustrate the above and other advantages and features of the present invention, specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. The accompanying drawings, together with the following detailed description, are included in and form a part of this specification. Elements having the same function and structure are indicated by the same reference numerals. It should be understood that these drawings only depict typical examples of the invention and should not be construed as limiting the scope of the invention. In the drawings:

[0014] Figure 1 illustrates a schematic diagram of federated learning in a vehicle network scenario in the prior art;

[0015] Figure 2 shows a functional block diagram of an electronic device for a federated learning server according to an embodiment of the present disclosure;

[0016] Figure 3 is a diagram illustrating an example of a neural network model;

[0017] Figure 4 is a schematic diagram illustrating clustering using the K-Means algorithm according to an embodiment of the present disclosure;

[0018] Figure 5 is a schematic diagram illustrating the uploading of some parameters according to an embodiment of the present disclosure;

[0019] Figure 6 is a schematic diagram of the training process of a lightweight federated learning model according to an embodiment of the present disclosure;

[0020] Figure 7 is another schematic diagram of the training process of a lightweight federated learning model according to an embodiment of the present disclosure;

[0021] Figure 8 shows a functional block diagram of an electronic device for a core network according to another embodiment of the present disclosure;

[0022] Figure 9 shows a functional block diagram of an electronic device for a federated learning client according to another embodiment of the present disclosure;

[0023] Figure 10 shows a flowchart of a method for a federated learning server according to an embodiment of the present disclosure;

[0024] Figure 11 shows a flowchart of a method for a core network according to another embodiment of the present disclosure;

[0025] Figure 12 shows a flowchart of a method for a federated learning client according to yet another embodiment of the present disclosure;

[0026] Figure 13 is a block diagram illustrating a first example of a schematic configuration of an eNB or gNB to which the technologies of this disclosure can be applied;

[0027] Figure 14 is a block diagram illustrating a second example of a schematic configuration of an eNB or gNB to which the technologies of this disclosure can be applied;

[0028] Figure 15 is a block diagram illustrating an example of a schematic configuration of a smartphone to which the technologies of this disclosure can be applied;

[0029] Figure 16 is a block diagram illustrating an example of a schematic configuration of a car navigation device to which the technology of this disclosure can be applied; and

[0030] Figure 17 is a block diagram of an exemplary structure of a general-purpose personal computer in which methods and / or apparatus and / or systems according to embodiments of the present invention can be implemented. Detailed Implementation

[0031] Exemplary embodiments of the invention will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of actual implementations are described in the specification. However, it should be understood that many implementation-specific decisions must be made in the development of any such actual embodiment to achieve the developer's specific goals, such as complying with constraints related to the system and business, and these constraints may vary depending on the implementation. Furthermore, it should be understood that while development work can be very complex and time-consuming, such development work is merely a routine task for those skilled in the art who benefit from this disclosure.

[0032] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the device structure and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0033] Figure 1 illustrates a schematic diagram of federated learning in a vehicle network scenario in the prior art.

[0034] As shown in Figure 1, during the FL training process, within a certain communication range, multiple FL clients (e.g., FL client 1 to FL client 7 shown in Figure 1) train local models locally and upload the local models to the FL server. The FL server aggregates all local models and obtains the global FL model to complete the FL model training.

[0035] This application uses a vehicle-to-everything (V2X) scenario as an example for description, but it is not limited to V2X and can be applied to all scenarios where FL (Flexible Application) is used for model training. In the following description, for example, in a V2X scenario, the FL server is deployed on a base station and has AF (Application Function), and the FL client is a moving vehicle (UE).

[0036] Because FL models have a large number of parameters, uploading all parameters of a local model to the FL client would increase communication overhead. Therefore, this application proposes a lightweight FL model training method where the FL client only uploads a portion of the parameters of a local model. The lightweight FL model training is initiated by the FL server.

[0037] According to one embodiment of this disclosure, an electronic device 2000 for a federated learning server is provided. The electronic device 2000 includes at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, through the at least one processor, cause the electronic device 2000 to perform: sending magnitude information about the parameter magnitude of a global model for federated learning to a core network to assist the core network in identifying federated learning clients that cannot upload all parameters of their local models as partial upload clients from a plurality of federated learning clients; and, if predetermined conditions are met, aggregating based on partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0038] Figure 2 shows a functional block diagram of an electronic device 2000 for a federated learning server according to an embodiment of the present disclosure.

[0039] As shown in Figure 2, the electronic device 2000 used as the FL server includes: a server control unit 2001 for control; a processing unit 2003 configured to send information about the magnitude of parameters of the global model used for federated learning to the core network under the control of the server control unit 2001, so as to assist the core network in identifying federated learning clients that cannot upload all parameters of their local models as partial upload clients from multiple federated learning clients; and an aggregation unit 2005 configured to aggregate partial parameters of the local models received from the partial upload clients under the control of the server control unit 2001, provided that predetermined conditions are met, to obtain the global model in the current round of training for federated learning.

[0040] The server control unit 2001 can be implemented as one or more processing circuits and at least one memory. The processing circuits can be, for example, a processor or a chip, and the at least one memory can be RAM, ROM, etc., used to store computer program code and data required for the processing circuits to perform processing. The processing unit 2003 and the aggregation unit 2005 can operate under the control of the server control unit 2001. Furthermore, it should be understood that the various functional units in the electronic device 2000 shown in FIG. 2 are merely logical modules divided according to their specific functions, and are not intended to limit the specific implementation method.

[0041] Electronic device 2000 may be located at or communicatively connected to a base station. Electronic device 2000 may be located at or communicatively connected to a roadside unit (RSU). Electronic device 2000 may be located at or communicatively connected to a mobile edge computing (MEC). For example, electronic device 2000 may function as one of a base station, RSU, or MEC, and may also include external devices such as memory and transceivers (not shown). The memory may be used to store programs and related data information that electronic device 2000 needs to execute to perform various functions. The transceiver may include one or more communication interfaces to support communication with different devices (e.g., UE, base station, RSU, MEC, etc.), and the implementation of the transceiver is not specifically limited here.

[0042] As an example, the base station could be an eNB or a gNB.

[0043] As an example, the core network could be a Network Data Analysis Function (NWDAF) element.

[0044] The wireless communication system according to this disclosure can be a 5G NR (New Radio) communication system, a 5G+ communication system, or a 6G communication system. Further, the wireless communication system according to this disclosure can include a non-terrestrial network (NTN). Optionally, the wireless communication system according to this disclosure can also include a terrestrial network (TN). Additionally, those skilled in the art will understand that the wireless communication system according to this disclosure can also be a 4G or 3G communication system.

[0045] For example, some upload clients are FL clients that are unable to upload all the parameters of their local model due to poor channel quality.

[0046] For example, the aforementioned federated learning clients are all clients participating in federated learning.

[0047] As illustrated in Figure 1, in the context of federated learning within a vehicle network scenario, uploading all model parameters to the FL client in each training iteration incurs significant communication overhead. Furthermore, forcing some FL clients to upload all parameters of their local models when they cannot do so introduces transmission errors during the upload process. This results in errors in the local model received by the FL server, ultimately compromising the accuracy of the overall model.

[0048] According to an embodiment of this disclosure, the electronic device 2000 sends information about the magnitude of parameters of the global model to the core network to assist the core network in identifying some uploading clients. This allows some uploading clients to upload only a portion of the model parameters during the local model uploading process, thereby reducing the communication overhead of FL model training without affecting the accuracy of the global model.

[0049] In the following text, the electronic device 2000 will sometimes be referred to as the FL server.

[0050] For example, under the aforementioned predetermined conditions, the electronic device 2000 aggregates some parameters of the local model uploaded by some upload clients and all parameters of the local model uploaded by non-partial upload clients to obtain the global model in the current round of federated learning training.

[0051] For example, if the predetermined conditions are not met, the aforementioned multiple federated learning clients upload all parameters of the local model to the electronic device 2000. The electronic device 2000 then aggregates the received parameters of the local model to obtain the global model in the current round of federated learning training.

[0052] As an example, magnitude information (hereinafter sometimes referred to as model parameter magnitude) includes information about the size of the parameters of the global model (which may be simply referred to as model size). Other examples of magnitude information may be conceived by those skilled in the art, and will not be elaborated here.

[0053] For example, an untrusted FL server sends magnitude information to an NWDAF network element through a Network Open Function (NEF) network element. A trusted FL server, on the other hand, sends magnitude information directly to the NWDAF network element.

[0054] As an example, magnitude information is used as input to UE communication analysis in NWDAF, and the UE communication prediction output of UE communication analysis includes a list with some of the IDs of the uploading clients.

[0055] For UE Communication Analytics in NWDAF, please refer to section 6.7.3 of 3GPP TS23.288. For example, magnitude information can be placed in the service's input, and a list of FL clients that cannot upload all model parameters can be placed in the service's UE Communication Predictions output.

[0056] For example, adding parameters to an input can be done as follows:

[0057] For example, the output parameter can be added as follows:

[0058] As an example, the predetermined conditions include receiving a list of partial upload clients from the core network and / or the global model's accuracy in the previous training round reaching a preset accuracy, wherein the list includes the IDs of the partial upload clients, and multiple federated learning clients act as partial upload clients when the global model's accuracy in the previous training round reaches the preset accuracy.

[0059] For example, after the NWDAF network element determines that some FL clients cannot upload all the parameters of their local model due to poor channel quality, thus identifying the partial upload clients, it stores the IDs of the partial upload clients in a list and sends the list to the FL server to assist the FL server in subsequently notifying these partial upload clients to upload only some parameters.

[0060] For example, once the global model reaches a preset accuracy, uploading all parameters of the local model is no longer necessary; only the model parameters that most urgently need updating can be uploaded. As an example, those skilled in the art can pre-set the preset accuracy based on experience or application scenarios. For instance, the preset accuracy could be 95%, 80%, etc., which will not be elaborated upon here.

[0061] For example, if the global model's accuracy in the previous training round reaches a preset accuracy, all federated learning clients are considered partial upload clients. Similarly, if the list is empty and the global model's accuracy in the previous training round reaches a preset accuracy, all federated learning clients are considered partial upload clients even if there is no federated learning client unable to upload all parameters of its local model. Furthermore, when the global model's accuracy in the previous training round reaches a preset accuracy, the method for determining partial parameters is the same as when FL clients cannot upload all parameters of their local model due to poor channel quality. The method for determining partial parameters will be described in detail below.

[0062] As illustrated in Figure 1, in the context of federated learning within a vehicle network scenario, uploading all parameters of a local model by the client increases communication overhead due to the large number of model parameters. This is especially true when some FL clients have poor channel quality and / or the global model has already reached a preset accuracy; uploading all local model parameters further increases communication overhead. Furthermore, in cases where some FL clients have poor channel quality, forcibly uploading all parameters of their local models can easily introduce transmission errors during the upload process. This results in errors in the local models received by the FL server, ultimately impairing the accuracy of the global model. For example, in Figure 1, if the channel between FL client 2 and FL client 7 suddenly deteriorates, forcibly uploading all parameters of their local models would lead to large errors in the local models received by the FL server, thus weakening the performance of the global model.

[0063] In contrast, in the embodiments according to this disclosure, when the FL server receives a list including FL clients unable to upload all model parameters, the FL server triggers lightweight FL model training (i.e., some uploading clients only upload partial parameters of a local model). This not only reduces communication overhead but also reduces the error of the local model received by the FL server, thus not affecting the accuracy of the global model. When the global model accuracy of the FL server reaches a preset preset accuracy (model accuracy threshold), lightweight FL model training not only reduces communication overhead but also does not affect the accuracy of the global model.

[0064] When the global model's accuracy reaches the preset accuracy, since the accuracy difference in each iteration of FL training is not particularly large, lightweight FL model training, i.e., uploading partial parameters of the local model, can be triggered periodically. When the FL client communication quality is poor, since the client needs to upload its local model in each FL training iteration, it is necessary to predict the FL client's communication quality in each round of FL training iteration to determine whether the FL client can upload all parameters of its local model. For example, data transmission between the 5G core network (5GC) and the FL server does not occupy wireless communication resources and will not cause significant communication overhead. Therefore, in this case, lightweight FL model training can be triggered in real time.

[0065] If the list of FL clients returned by the NWDAF network element is empty, and the accuracy of the current global model has not reached the preset accuracy, the FL server sends a notification to the FL client to upload all parameters of the local model.

[0066] As an example, some of the upload client IDs included in the list above were determined by the core network based on magnitude information and predicted channel information between the electronic device 2000 and the federated learning client.

[0067] For example, the NWDAF network element has the function of predicting the channel information between the FL server and the FL client (e.g., base station and UE). The NWDAF network element determines whether the FL client can upload all the parameters of its local model to the FL server under the current communication conditions by using the predicted information and the magnitude information of the received model parameters.

[0068] As an example, the channel information includes at least one of the following: channel quality indicator (CQI), precoding matrix indicator (PMI), reference signal received power (RSRP), and signal-to-noise ratio (SNR) of the channel between the electronic device 2000 and the federated learning client.

[0069] For example, the NWDAF network element calculates the transmission rate using the SNR and bandwidth information between the FL server and the FL client, and calculates the amount of data that can be uploaded within the above latency limit based on the FL local model upload transmission latency limit, thereby determining whether the FL client can upload all parameters of its local model.

[0070] In expression (1), SNR is the SNR value of the channel between the FL server and the FL client, B is the bandwidth allocated by the FL server to the FL client, Delay is the upper limit of the FL local model upload transmission delay limited by the FL server, and S... l To ensure that the FL client can support the amount of data that can be uploaded, S g This refers to the amount of data for the parameters of the FL global model. By comparing S... l With S g It can determine whether the FL client can upload all parameters of a local model, and what proportion of model parameters it can support uploading when it cannot upload all parameters of a local model. From expression (1), it can be seen that in S... l Greater than S g In the current communication context, it is determined that the FL client can upload all parameters of its local model to the FL server; while in S... l Less than S g In the case of [unspecified event], it is determined that the FL client cannot upload all parameters of its local model to the FL server under the current communication conditions.

[0071] As an example, processing unit 2003 can be configured to, for the parameters of the global model in the previous training round and the parameters of the local models uploaded by multiple federated learning clients in the previous training round: calculate the local model parameter bias, representing the degree of dispersion among the parameters of the local models of the multiple federated learning clients, and calculate the global model parameter bias, representing the degree of deviation between the parameters of the local models of the multiple federated learning clients and the parameters of the global model; and determine some parameters based on the local model parameter bias and the global model parameter bias. In other words, electronic device 2000 can determine the parameters that need to be supplemented in the global model in the current training round (i.e., determine the parameters that some uploading clients need to upload in the current training round) based on the local model parameters received from the FL clients in the previous iteration.

[0072] As an example, the processing unit 2003 can be configured to calculate the average value of the i-th parameter based on the i-th parameter of the local model of multiple federated learning clients, and to calculate the degree of local model parameter deviation for the i-th parameter based on the difference between the i-th parameter of the local model of each federated learning client and the average value; and to calculate the degree of global model parameter deviation for the i-th parameter based on the difference between the i-th parameter of the local model of each federated learning client and the i-th parameter of the global model.

[0073] For example, the FL server calculates the degree of deviation of the local model parameters for the i-th parameter based on the i-th parameter of the local model uploaded by the FL client in the previous iteration:

[0074] In expression (2), N represents the number of multiple FL clients. This represents the i-th parameter of the n-th FL client (n is a positive integer greater than or equal to 1 and less than or equal to N). This represents the average value of the i-th parameter in all N FL client local models.

[0075] As mentioned above, σ l It measures the degree of dispersion of local model parameters among multiple FL clients. For example, the greater the dispersion of all FL clients for the i-th parameter, the greater the need for further updates to that parameter, meaning the greater the probability that the local model parameter will be uploaded to the FL server.

[0076] For example, the FL server uses the i-th parameter of the local model and the i-th parameter p of the global model uploaded by the FL client in the previous iteration. g The difference between (i) is used to calculate the degree of global model parameter bias for the i-th parameter:

[0077] As mentioned above, σ g It measures the average difference between all local model parameters of FL clients and the global model parameters. For example, for the i-th parameter in the model, if the average difference between the i-th parameter in all local models of FL clients and the i-th parameter in the global model is larger, then the parameter needs to be updated more, that is, the probability that the parameter is uploaded to the FL server is greater.

[0078] If among the N FL clients, there exists an FL client that did not upload the i-th parameter in the previous training round, then the FL client that did not upload the i-th parameter... It is 0.

[0079] As an example, the processing unit 2003 can be configured to cluster the parameters based on the degree of deviation of local model parameters and the degree of deviation of global model parameters, thereby obtaining multiple clusters after clustering. From the multiple clusters, the selected clusters that meet the predetermined cluster conditions are selected, and the parameters in the selected clusters are used as part of the parameters.

[0080] For example, global and local models are hierarchical neural network models. Typically, for neural network models, the parameters refer to the weight values ​​and bias values.

[0081] Figure 3 illustrates an example of a neural network model. As shown in Figure 3, the circles in the first column represent neurons in the input layer, the circles in the last column represent neurons in the output layer, the circles between the first and last columns represent neurons in the hidden layer, the lines connecting neurons represent weights, and each neuron has a bias value. Taking the first two layers of neurons in Figure 3 as an example, the index of each neuron is the value in its circle (neuron indices can also be defined in other ways, as long as they uniquely identify each neuron). For example, the weight value between neuron 1 and neuron 6 is named ω. 1，6 The weight value between neuron 5 and neuron 11 is named ω. 5，11 The weight values ​​between different neurons are named in the same way. Taking the neuron with index 6 as an example, the value of the neuron is calculated as shown in the following expression (4): x6=(ω 1，6 x1+ω 2，6 x2+ω 3，6 x3+ω 4，6 x4+ω 5，6 x5)+b6 expression (4)

[0082] In expression (4), ω is the weight value and b is the bias value.

[0083] As can be seen from the above description, the number of weights and biases in a neural network model is enormous. Calculating and clustering the local and global model parameter biases of these weights and biases would be computationally intensive. Therefore, in this application, the global model and all uploaded local models are processed with all 1s as input on the FL server. Specifically, all input values ​​are set to 1, for example, [1,1,1,1,…], and these values ​​are input into the global model and all uploaded local models to obtain the value of each neuron in the model (the neuron value calculation process is described in the expression (4) above for neuron 6). Therefore, the parameters of the model described in this application refer to the values ​​of the neurons in the FL model.

[0084] As an example, the predetermined cluster condition includes that the selection probability corresponding to the cluster is greater than a predetermined probability, and the processing unit 2003 can be configured to use the local model parameter deviation degree and the global model parameter deviation degree as coordinates in a two-dimensional coordinate system, and use the calculated local model parameter deviation degree and the global model parameter deviation degree corresponding to the parameter as coordinate values, thereby representing each parameter as a point in the coordinate system. The K-Means algorithm is used to cluster all points in the coordinate system to obtain multiple clusters. For each cluster in at least a portion of the multiple clusters, the selection probability of the cluster is calculated based on the distance between the cluster center and the origin of the coordinate system, and the angle between the line connecting the cluster center to the origin of the coordinate system and the dividing line of the first quadrant of the coordinate system. The dividing line is a straight line passing through the origin of the coordinate system and forming a predetermined angle with the horizontal axis of the first quadrant. The greater the distance, the greater the probability of selection, and the smaller the angle, the greater the probability of selection.

[0085] For example, those skilled in the art can determine the predetermined probability and predetermined angle in advance based on experience or application scenarios.

[0086] In the K-Means algorithm, samples are defined as parameters (i.e., points in the coordinate system above) expressed as the degree of bias in local and global model parameters. The specific steps for clustering all points in the coordinate system using the K-Means algorithm are as follows:

[0087] Step a: Randomly select K points from the sample as cluster centers (cluster centroids);

[0088] Step b: Calculate the distances of other samples in the sample to the K cluster centers, and assign these samples to the category of the nearest cluster center;

[0089] Step c: Calculate the average value for each category of the samples after the above clustering and solve for the new cluster centroids;

[0090] Step d: Compare the K cluster centroids obtained in the previous calculation. If the cluster centroids have changed, proceed to step b; otherwise, proceed to step e.

[0091] Step e: When the centroid no longer changes, stop the iteration and output the clustering results.

[0092] Step f: Select clusters from the clustering results and use the parameters of the selected clusters as partial parameters.

[0093] Figure 4 is a schematic diagram illustrating clustering using the K-Means algorithm according to an embodiment of the present disclosure.

[0094] In Figure 4, for example, the degree of deviation of local model parameters σ l The x-axis represents the degree of deviation of the global model parameters, σ. g Establish a coordinate system for the ordinate. Given that all parameters in the FL model have σ... l and σ g Therefore, each parameter in the FL model is a point in the established coordinate system. All parameters in the FL are represented in the established coordinate system, and the K-Means algorithm is used to cluster the model parameters. The clustering results are shown by circles in Figure 4 (illustrative rather than restrictive; Figure 4 shows the clustering results including cluster 1, cluster 2, and cluster 3), and each cluster has a cluster center.

[0095] In the following description, for simplicity, we assume the predetermined angle is 45 degrees. 。 If we describe it this way, then the dividing line above is a straight line at a 45-degree angle to the horizontal axis of the first quadrant.

[0096] For example, the expression for calculating the probability of a cluster being selected is as follows:

[0097] In expression (5), K represents the total number of clusters in the clustering result (e.g., K = 3 in Figure 4), k represents the kth cluster (e.g., k is 1 to 3 in Figure 4), P(k) represents the probability of selecting the kth cluster, d(k) represents the distance from the cluster center of the kth cluster to the origin o (d(1), d(2), d(3) are shown in Figure 4), θ(k) represents the deviation angle of the line connecting the cluster center of the kth cluster to the origin from the x-axis (θ(3) is shown in Figure 4), and g(·) represents the normalization function.

[0098] In the embodiments according to this disclosure, there are two cluster selection principles: one is to select clusters whose cluster centers are farthest from the coordinate origin (ensuring σ). l and σ gThe comprehensive value is large), so d(k) is considered in expression (5); secondly, clusters with small angles between the line connecting the cluster center to the origin and the dividing line of the first quadrant of the coordinate system (e.g., a straight line at a 45-degree angle to the x-axis, as shown by the dashed line in Figure 4) are selected (to ensure σ l and σ g The values ​​are all very large, and there will be no extreme cases of bias to one side (e.g., avoiding cluster centers on the x-axis and far from the origin). Therefore, |θ(k)-45°| is considered in the cluster selection expression (5). By calculating the P(k) values ​​of all clusters, clusters are selected based on the P(k) values. For example, all parameters in the cluster with the largest P(k) value can be selected as the partial parameters to be uploaded, that is, the cluster with the largest deviation (local model parameter deviation σ) is selected. l The degree of deviation σ from the global model parameters g A cluster of parameters (all as large as possible) is used as the part of the parameters that need to be supplemented.

[0099] The above text uses the K-Means algorithm as an example to describe clustering. However, those skilled in the art can conceive of other clustering methods, which will not be elaborated here.

[0100] As an example, processing unit 2003 can be configured to send notification information about partial parameters to partial upload clients so that the partial upload clients can upload only partial parameters in the current round of training.

[0101] As mentioned above, the global model is a hierarchical neural network model. As an example, the notification information includes the location information of some parameters within the neural network model. This location information could be, for example, the index of a neuron mentioned above.

[0102] As an example, aggregation unit 2005 can be configured to align partial parameters of a local model received from a partial upload client and all parameters of a local model received from other clients among multiple federated learning clients based on location information, and then aggregate the local models.

[0103] Figure 5 is a schematic diagram illustrating the uploading of some parameters according to an embodiment of the present disclosure.

[0104] After receiving a notification message indicating that some parameters need to be uploaded (e.g., a list of indices of neurons to be uploaded), the FL client uploads a portion of the model's parameters. For example, Figure 5 shows the neurons to be uploaded (referred to as uploaded neurons). The FL client uploads the weights of all upper-layer neurons connected to the uploaded neuron, along with the bias values ​​of the uploaded neuron, to the FL server. After receiving all the local models uploaded by the FL clients, the FL server aligns them according to the neuron indices to aggregate them and obtain the global model.

[0105] In scenarios where FL clients move (e.g., connected vehicle scenarios, drone scenarios, etc.), some FL clients may leave the FL server's coverage area before completing the training of their local models. It may be necessary to perform computational unloading of FL clients that are about to leave during each round of FL training.

[0106] For example, drone scenarios can include the following two scenarios. In scenario 1, the drone acts as the FL server, and multiple vehicles act as FL clients; in scenario 2, the base station acts as the FL server, and multiple drones performing emergency missions act as FL clients.

[0107] The above problem can be solved by receiving the compute offload indication triggered by the electronic device 2000 from the core network. This disclosure adds a compute offload analysis function to the NWDAF network element in the 5G core network. That is, when the FL server (e.g., gNB, RSU, MEC, drone, etc.) and the FL client (e.g., UE, drone, etc.) are conducting each round of FL training iteration, the NWDAF assists in analyzing the dwell time of all FL clients within the coverage area of ​​the FL server to determine whether a compute offload indication is triggered.

[0108] As an example, the processing unit 2003 can be configured to receive a computation offload instruction from the core network that instructs the federated learning client to perform computation offload, and to notify all federated learning clients within the coverage area of ​​the electronic device 2000 of the computation offload instruction so that the federated learning clients can determine whether to perform computation offload, wherein the computation offload instruction is sent by the core network based on the dwell time of all federated learning clients within the coverage area of ​​the electronic device 2000.

[0109] For example, the aforementioned dwell time could be the average dwell time, median dwell time, or other values ​​of all federated learning clients within the coverage area of ​​Electronic Device 2000.

[0110] For example, in computational offloading, a federated learning client transmits the parameters of its local model to other federated learning clients so that the other clients can perform federated learning based on the parameters, while the original federated learning client no longer participates in the federated learning process. For example, other federated learning clients may fuse their local models based on the parameters and the parameters of their local models (e.g., by taking a weighted average of the parameters and the parameters of their local models) to perform federated learning based on the fused local model.

[0111] Other federated learning clients can improve the accuracy of federated learning by performing federated learning based on the fused local models.

[0112] In FL learning, computational unloading can also be called model unloading. In the following text, computational unloading will sometimes be written as model / computational unloading.

[0113] For example, the FL client decides whether to perform computational offloading based on whether it has completed training of the local model.

[0114] As an example, the computational offload instruction is sent by the core network when it determines that the dwell time is less than the preset training duration of the local model.

[0115] For example, those skilled in the art can set the training time of a preset local model based on experience or application scenarios.

[0116] For example, if the core network determines that the dwell time is greater than or equal to the preset training duration of the local model, the computation unloading instruction will not be triggered.

[0117] Those skilled in the art may also conceive of other examples of the core network sending computation offload instructions based on dwell time, which will not be elaborated here.

[0118] As an example, the dwell time is obtained by the core network based on the ID of the electronic device 2000, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the electronic device 2000.

[0119] For example, the ID of electronic device 2000 can be a unique identifier for electronic device 2000, and the ID of electronic device 2000 can indicate the location information of electronic device 2000. For example, the ID of electronic device 2000 can be the IP address of electronic device 2000.

[0120] For example, a time ID can be a timestamp ID, a time slot ID, etc. For example, a time ID can represent morning peak hours, evening peak hours, and off-peak hours, etc.

[0121] The density mentioned above refers to the distribution density of federated learning clients within the coverage area of ​​the electronic device 2000.

[0122] As an example, dwell time is also obtained by the core network using the movement speed of the federated learning client.

[0123] For example, the movement speed of the federated learning client could be the average movement speed of the federated learning clients within the coverage area of ​​the electronic device 2000, or it could be a list of the movement speeds of all federated learning clients within the coverage area of ​​the electronic device 2000.

[0124] For example, the core network can use a neural network model to obtain the predicted average moving speed of all federated learning clients within the coverage area of ​​electronic device 2000, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​electronic device 2000, and obtain the average dwell time based on the predicted average moving speed.

[0125] For example, the neural network model used to obtain the residence time could be a Long Short-Term Memory (LSTM) network, etc.

[0126] For example, the input to the compute unloading analysis function (dataset information from AF) is:

[0127]

[0128] For example, the output of the calculation unloading analysis function (output analysis) is:

[0129] Figure 6 is a schematic diagram of a lightweight FL model training process according to an embodiment of the present disclosure. In Figure 6, it is assumed that the FL server (electronic device 2000) is not trusted, and the multiple federated learning clients mentioned above include FL client 1 to FL client N.

[0130] In step 1, the FL server selects FL clients. For example, the FL server selects FL clients that need to participate in FL training based on requirements. For example, the dataset value of each FL client may differ (i.e., the datasets possessed by each FL client are heterogeneous), and the FL server selects FL clients based on the value of their datasets to fully train the FL model.

[0131] In step 2, the FL server performs initial global model distribution. After selecting an FL client, the FL server distributes the initial global model to the selected FL client for subsequent local model training on the FL client.

[0132] In step 3 (including steps 3a and 3b), the FL server sends a model parameter magnitude notification. As shown in steps 3a and 3b, the FL server sends the model parameter magnitude notification to the NWDAF network element through the NEF network element to assist the NWDAF network element in determining which FL clients cannot upload all parameters of their local model under the current communication conditions.

[0133] In step 4, the NWDAF network element performs a comparative analysis of the model parameter magnitude and the FL client channel quality to determine whether the FL client can upload all parameters of its local model to the FL server under the current communication conditions.

[0134] In step 5 (including steps 5a and 5b), the NWDAF network element sends a list of FL clients that are unable to upload all parameters of the model to the FL server.

[0135] In step 6, the FL server determines the missing parameters of the partial model. Under the predetermined conditions mentioned above, the FL server determines the parameters of the partial model that needs to be uploaded.

[0136] In step 7, the FL server sends a notification to upload a partial model (all model parameters or some model parameters). For example, after the FL server determines which model parameters need to be uploaded (or supplemented), it notifies the FL client which parameters should be uploaded in the current iteration by sending a notification message that includes the location information of the model parameters that need to be supplemented. For example, this notification message is a list containing the indices of the locations of all model parameters that need to be supplemented. Furthermore, for FL clients that have not been notified of model parameter supplementation, uploading all parameters of their partial model is sufficient.

[0137] Optionally, in step a, the FL client performs model / computation unloading. For a description of model / computation unloading, please refer to the description above regarding the computation unloading analysis function.

[0138] In step 8, the FL client uploads a partial model (all or some model parameters). For example, after receiving a notification to upload a partial model (all or some parameters), if some FL clients have poor channel quality, those clients upload the corresponding model parameters from their partial model based on the location information of the parameters to be uploaded, while other FL clients upload all parameters of their partial model. If the global model accuracy reaches a preset model accuracy threshold, all FL clients upload the corresponding model parameters from their partial model based on the location information of the parameters. In these cases, when uploading partial parameters of the partial model, the FL client uploads the parameter values ​​at the corresponding locations based on the parameter location information.

[0139] In step 9, the FL server performs model aggregation. For example, after the FL server receives all the local models uploaded by FL clients, it aligns the local model parameters of the clients based on the location and value information of the parameters in the local models and completes global model aggregation.

[0140] In step 10, the FL server updates / terminates the process. For example, based on the requirements of the aggregation task initiator, the FL server decides whether to continue updating or terminate the FL training process.

[0141] In step 11, the FL server distributes the aggregated global model. The FL server distributes the aggregated global model from this iteration to the FL client for use in the next round of local model training by the FL client.

[0142] In step 12, the FL client performs a local model update. After receiving the new global model, the FL client begins a new round of training and updates the local model.

[0143] In each round of training, steps 4 to 12 are executed iteratively.

[0144] Figure 7 is another schematic diagram of the lightweight FL model training process according to an embodiment of the present disclosure. In Figure 7, it is assumed that the FL server (electronic device 2000) is trusted.

[0145] As described above, untrusted FL servers send model parameter magnitude notifications to NWDAF network elements through NEF network elements, while trusted FL servers send model parameter magnitude notifications directly to NWDAF network elements. Therefore, there are no NEF network elements in Figure 7, and except for steps 3 and 5, the other steps in Figure 7 are the same as in Figure 6, and will not be repeated here.

[0146] This disclosure also provides an electronic device 8000 for a core network according to another embodiment of this disclosure. The electronic device 8000 includes at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, via the at least one processor, to cause the electronic device 8000 to: receive magnitude information about the magnitude of parameters of a global model for federated learning from a federated learning server; identify federated learning clients from a plurality of federated learning clients that cannot upload all parameters of their local models as partial upload clients; and, upon meeting predetermined conditions, aggregate the partial parameters of the local models received from the partial upload clients to obtain the global model for the current round of training in federated learning.

[0147] Figure 8 shows a functional block diagram of an electronic device 8000 for a core network according to another embodiment of the present disclosure.

[0148] As shown in Figure 8, the electronic device 8000 includes: a core network control unit 8001, which performs control; and an identification unit 8003, which, under the control of the core network control unit 8001, receives information about the magnitude of parameters of the global model used for federated learning from the federated learning server, so as to identify federated learning clients that cannot upload all parameters of their local models from multiple federated learning clients as partial upload clients, so that, under predetermined conditions, the federated learning server can aggregate the partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training of federated learning.

[0149] The core network control unit 8001 can be implemented as one or more processing circuits and at least one memory. The processing circuits can be, for example, a processor or a chip, and the at least one memory can be RAM, ROM, etc. The at least one memory is used to store computer program code and data required for the processing circuits to perform processing. The identification unit 8003 can perform operations under the control of the core network control unit 8001. Furthermore, it should be understood that the various functional units in the electronic device 8000 shown in Figure 8 are merely logical modules divided according to their specific functions, and are not used to limit the specific implementation method.

[0150] For example, electronic device 8000 can function as the core network device itself, and may also include external devices such as memory and transceivers (not shown). The memory can be used to store programs and related data information that electronic device 8000 needs to execute to perform various functions. The transceiver may include one or more communication interfaces to support communication with different devices (e.g., UE, base station, etc.), and there is no specific limitation on the implementation of the transceiver.

[0151] As an example, the federated learning server in the embodiment of electronic device 8000 can be the electronic device 2000 mentioned above. As an example, electronic device 8000 can be the core network involved in the embodiment of electronic device 2000 mentioned above.

[0152] For example, some upload clients are FL clients that are unable to upload all the parameters of their local model due to poor channel quality.

[0153] For example, the aforementioned federated learning clients are all clients participating in federated learning.

[0154] The electronic device 8000 according to an embodiment of the present disclosure can identify some uploading clients based on the magnitude information of the parameters of the global model received from the FL server, so that some uploading clients only need to upload some model parameters during the local model uploading process, thereby reducing the communication overhead of FL model training without affecting the accuracy of the global model.

[0155] As an example, magnitude information is used as input to UE communication analysis in NWDAF, and the UE communication prediction output of UE communication analysis includes a list with some of the IDs of the uploading clients.

[0156] As an example, magnitude information includes information about the size of the parameters of the global model. Other examples of magnitude information will be conceived by those skilled in the art, and will not be elaborated here.

[0157] For example, electronic device 8000 can be an NWDAF network element. For instance, an untrusted FL server sends magnitude information to an NWDAF network element through a NEF network element. A trusted FL server, on the other hand, sends magnitude information directly to the NWDAF network element.

[0158] As an example, the predetermined conditions include the federated learning server receiving a list of some upload clients from the electronic device 8000 and / or the global model's accuracy in the previous training round reaching a preset accuracy, wherein the list includes the IDs of some upload clients.

[0159] For a description of the predetermined conditions, please refer to the relevant section in the embodiments of the electronic device 2000, which will not be repeated here.

[0160] As an example, the identification unit 8003 can be configured to determine the IDs of a subset of upload clients included in the list based on magnitude information and the predicted channel information between the federated learning server and the federated learning client.

[0161] For example, electronic device 8000 has the function of predicting channel information between FL server and FL client (e.g., base station and UE). Electronic device 8000 determines whether FL client can upload all parameters of its local model to FL server under the current communication conditions by using the predicted information and the magnitude information of the received model parameters.

[0162] As an example, the channel information includes at least one of the following: channel quality indicator (CQI), precoding matrix indicator (PMI), reference signal received power (RSRP), and signal-to-noise ratio (SNR) of the channel between the electronic device 8000 and the federated learning client.

[0163] For a description of how the electronic device 8000 determines a portion of the upload clients based on magnitude information and SNR, please refer to the description of expression (1) in the embodiment of the electronic device 2000, which will not be repeated here.

[0164] As an example, the identification unit 8003 can be configured to send a computation unloading instruction to the federated learning server based on the dwell time of all federated learning clients within the coverage area of ​​the federated learning server, instructing the federated learning clients to perform computation unloading, so that the federated learning server can notify all federated learning clients within its coverage area of ​​the computation unloading instruction, so that the federated learning clients can determine whether to perform computation unloading.

[0165] For example, the aforementioned dwell time could be the average dwell time, median dwell time, or other values ​​of all federated learning clients within the coverage area of ​​the federated learning server.

[0166] As an example, the recognition unit 8003 can be configured to send a computation unloading instruction when it is determined that the dwell time is less than the preset training time of the local model.

[0167] As an example, the identification unit 8003 can be configured to obtain the dwell time based on the ID of the federated learning server, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the federated learning server.

[0168] As an example, the identification unit 8003 can be configured to also obtain the dwell time based on the movement speed of the federated learning client.

[0169] For a description of the model / computation offloading, please refer to the description of the computation offloading analysis function in the above embodiment of the electronic device 2000, which will not be repeated here.

[0170] This disclosure also provides an electronic device 9000 for a federated learning client according to yet another embodiment of this disclosure. The electronic device 9000 includes at least one processor and at least one memory, the at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, via the at least one processor, to cause the electronic device 9000 to: send partial parameters of a local model for federated learning to a federated learning server, upon satisfying predetermined conditions, so that the federated learning server can aggregate the local model based on the partial parameters to obtain a global model in the current round of training for federated learning, wherein the federated learning server sends magnitude information about the parameter magnitude of the global model to the core network to assist the core network in identifying the electronic device 9000 as a partial upload client that cannot upload all parameters of its local model.

[0171] Figure 9 shows a functional block diagram of an electronic device 9000 for a federated learning client according to another embodiment of the present disclosure.

[0172] As shown in Figure 9, the electronic device 9000 includes: a client control unit 9001, which performs control; and a communication unit 9003, which, under the control of the client control unit 9001, sends partial parameters of the local model for federated learning to the federated learning server when predetermined conditions are met. The federated learning server then aggregates the local model based on the partial parameters to obtain the global model in the current round of training. The federated learning server sends information about the magnitude of the parameters of the global model to the core network to help the core network identify the electronic device 9000 as a partial upload client that cannot upload all the parameters of its local model.

[0173] The client control unit 9001 can be implemented as one or more processing circuits and at least one memory. The processing circuits can be, for example, a processor or a chip, and the at least one memory can be RAM, ROM, etc. The at least one memory is used to store, for example, computer program code and data required for the processing circuits to perform processing. The communication unit 9003 can operate under the control of the client control unit 9001. Furthermore, it should be understood that the various functional units in the electronic device 9000 shown in FIG. 9 are merely logical modules divided according to their specific functions, and are not intended to limit the specific implementation method.

[0174] For example, the electronic device 9000 can function as the user equipment itself, and may also include external devices such as memory and transceivers (not shown in the figure). The memory can be used to store programs and related data information that the user equipment needs to execute to perform various functions. The transceiver may include one or more communication interfaces to support communication with different devices (e.g., base stations, other user equipment, etc.), and there is no specific limitation on the implementation of the transceiver.

[0175] As an example, the federated learning server in the embodiment of electronic device 9000 can be the electronic device 2000 mentioned above, and the core network in the embodiment of electronic device 9000 can be the electronic device 8000 mentioned above. As an example, electronic device 9000 can be some of the upload clients involved in the embodiments of electronic device 2000 and electronic device 8000 mentioned above.

[0176] For example, Electronic Device 9000 is an FL client that cannot upload all parameters of its local model due to poor channel quality.

[0177] In the process of uploading a local model, the electronic device 9000 according to the embodiments of this disclosure only needs to upload some model parameters, thereby reducing the communication overhead of FL model training and not affecting the global model accuracy.

[0178] As an example, magnitude information is used as input to UE communication analysis in NWDAF, and the UE communication prediction output of UE communication analysis includes a list with some of the IDs of the uploading clients.

[0179] As an example, magnitude information includes information about the size of the parameters of the global model. Other examples of magnitude information will be conceived by those skilled in the art, and will not be elaborated here.

[0180] As an example, the predetermined conditions include the federated learning server receiving from the core network a list including the IDs of some of the uploading clients and / or the global model's accuracy in the previous training round reaching a preset accuracy.

[0181] For a description of the predetermined conditions, please refer to the relevant section in the embodiments of the electronic device 2000, which will not be repeated here.

[0182] As an example, the communication unit 9003 can be configured to receive notification information about certain parameters from the federated learning server.

[0183] As an example, the global model is a hierarchical neural network model, and the notification information includes the location information of some parameters in the neural network model.

[0184] For a description of the location information of certain parameters in the neural network model, please refer to the corresponding part in the embodiment of the electronic device 2000, which will not be repeated here.

[0185] As an example, the communication unit 9003 can be configured to receive a computation offload instruction from the federated learning server, which instructs the federated learning client to perform computation offload, in order to determine whether to perform computation offload. The computation offload instruction is sent by the core network to the federated learning server based on the dwell time of all federated learning clients within the coverage area of ​​the federated learning server.

[0186] For example, the aforementioned dwell time could be the average dwell time, median dwell time, or other values ​​of all federated learning clients within the coverage area of ​​the federated learning server.

[0187] As an example, the computational offload instruction is sent by the core network when it determines that the dwell time is less than the preset training duration of the local model.

[0188] As an example, the dwell time is obtained by the core network based on the ID of the federated learning server, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the federated learning server.

[0189] As an example, dwell time is also obtained by the core network based on the movement speed of the federated learning client.

[0190] For a description of the model / computation offloading, please refer to the description of the computation offloading analysis function in the above embodiment of the electronic device 2000, which will not be repeated here.

[0191] In the process of describing electronic devices 2000 and 9000 in the embodiments described above, some processes or methods have obviously been disclosed. Hereinafter, without repeating some details already discussed above, a summary of these methods is given. However, it should be noted that although these methods are disclosed in the description of the above electronic devices, these methods do not necessarily employ or are performed by the components described. For example, the embodiments of the above electronic devices can be implemented partially or entirely using hardware and / or firmware, while the methods discussed below can be implemented entirely by computer-executable programs, although these methods can also employ the hardware and / or firmware of the electronic device.

[0192] Figure 10 shows a flowchart of a method S1000 for a federated learning server according to an embodiment of the present disclosure. Method S1000 begins at step S1002. In step S1004, magnitude information about the parameter magnitudes of the global model used for federated learning is sent to the core network to assist the core network in identifying federated learning clients from multiple federated learning clients that cannot upload all parameters of their local models as partial upload clients. In step S1006, if predetermined conditions are met, partial parameters of the local models received from the partial upload clients are aggregated to obtain the global model for the current round of training in federated learning. Method S1000 ends at step S1008.

[0193] This method can be executed, for example, by the electronic device 2000 described above. For details, please refer to the above description of the relevant processing of the electronic device 2000, which will not be repeated here.

[0194] Figure 11 shows a flowchart of method S1100 for a core network according to another embodiment of the present disclosure. Method S1100 begins at step S1102. In step S1104, information about the magnitude of parameters of the global model used for federated learning is received from the federated learning server to identify federated learning clients that cannot upload all parameters of their local models as partial upload clients from among multiple federated learning clients. Under predetermined conditions, the federated learning server aggregates the partial parameters of the local models received from the partial upload clients to obtain the global model for the current round of training in federated learning. Method S1100 ends at step S1106.

[0195] This method can be executed, for example, by the electronic device 8000 described above. For details, please refer to the above description of the relevant processing of the electronic device 8000, which will not be repeated here.

[0196] Figure 12 shows a flowchart of method S1200 for a federated learning client according to another embodiment of the present disclosure. Method S1200 begins at step S1202. In step S1204, under predetermined conditions, partial parameters of a local model for federated learning are sent to the federated learning server so that the federated learning server can aggregate the local model based on the partial parameters to obtain the global model in the current round of training for federated learning. The federated learning server sends information about the magnitude of the parameters of the global model to the core network to assist the core network in identifying the federated learning client as a partial upload client that cannot upload all parameters of its local model. Method S1200 ends at step S1206.

[0197] This method can be executed, for example, by the electronic device 9000 described above. For details, please refer to the description of the relevant processing of the electronic device 9000 above, which will not be repeated here.

[0198] The technology disclosed herein can be applied to a variety of products.

[0199] Electronic device 2000 can be located on the base station side or connected to the base station. The base station can be implemented as any type of evolved Node B (eNB) or gNB (5G base station). eNBs include, for example, macro eNBs and small eNBs. Small eNBs can be eNBs covering cells smaller than macro cells, such as pico eNBs, micro eNBs, and femtocell eNBs. A similar situation can occur with gNBs. Alternatively, the base station can be implemented as any other type of base station, such as a NodeB and a Base Transceiver Station (BTS). The base station can include: a main body configured to control wireless communication (also called base station equipment); and one or more remote radio heads (RRHs) located in a different location from the main body. Furthermore, various types of electronic devices can operate as base stations by temporarily or semi-persistently performing base station functions.

[0200] Electronic device 9000 can be implemented as various user devices. User devices can be implemented as mobile terminals (such as smartphones, tablet PCs, laptop PCs, portable gaming terminals, portable / dongle-type mobile routers, and digital camera devices) or in-vehicle terminals (such as car navigation devices). User devices can also be implemented as terminals performing machine-to-machine (M2M) communication (also known as machine-type communication (MTC) terminals). Furthermore, user devices can be wireless communication modules (such as integrated circuit modules comprising a single chip) installed on each of the aforementioned terminals.

[0201] [Application examples of base stations]

[0202] (First application example)

[0203] Figure 13 is a block diagram illustrating a first example of a schematic configuration of an eNB or gNB to which the technologies of this disclosure can be applied. Note that the following description uses an eNB as an example, but it can also be applied to a gNB. The eNB 800 includes one or more antennas 810 and a base station device 820. The base station device 820 and each antenna 810 can be connected to each other via RF cables.

[0204] Each of the antennas 810 includes one or more antenna elements (such as multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna) and is used by the base station device 820 to transmit and receive wireless signals. As shown in Figure 13, the eNB 800 may include multiple antennas 810. For example, multiple antennas 810 may be compatible with multiple frequency bands used by the eNB 800. Although Figure 13 shows an example in which the eNB 800 includes multiple antennas 810, the eNB 800 may also include a single antenna 810.

[0205] The base station equipment 820 includes a controller 821, a memory 822, a network interface 823, and a wireless communication interface 825.

[0206] The controller 821 can be, for example, a CPU or a DSP, and operates various higher-level functions of the base station equipment 820. For example, the controller 821 generates data packets based on data in signals processed by the wireless communication interface 825, and transmits the generated packets via the network interface 823. The controller 821 can bundle data from multiple baseband processors to generate bundled packets and transmit the generated bundled packets. The controller 821 may have logical functions that perform controls such as radio resource control, radio bearer control, mobility management, admission control, and scheduling. This control can be performed in conjunction with nearby eNBs or core network nodes. The memory 822 includes RAM and ROM, and stores programs executed by the controller 821 and various types of control data (such as terminal lists, transmission power data, and scheduling data).

[0207] Network interface 823 is a communication interface used to connect base station equipment 820 to core network 824. Controller 821 can communicate with core network nodes or other eNBs via network interface 823. In this case, eNB 800 and core network nodes or other eNBs can be connected to each other through logical interfaces (such as S1 and X2 interfaces). Network interface 823 can also be a wired communication interface or a wireless communication interface for wireless backhaul. If network interface 823 is a wireless communication interface, it can use a higher frequency band for wireless communication compared to the frequency band used by wireless communication interface 825.

[0208] The wireless communication interface 825 supports any cellular communication scheme (such as LTE and LTE-Advanced) and provides wireless connectivity to terminals located in the cell of eNB 800 via antenna 810. The wireless communication interface 825 typically includes, for example, a baseband (BB) processor 826 and RF circuitry 827. The BB processor 826 can perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and performs various types of signal processing at layers (e.g., Layer 1, Medium Access Control (MAC), Radio Link Control (RLC), and Packet Data Convergence Protocol (PDCP)). Instead of controller 821, the BB processor 826 can have some or all of the above-described logical functions. The BB processor 826 can be a memory storing communication control programs, or a module including a processor and associated circuitry configured to execute programs. Update programs can change the functionality of the BB processor 826. The module can be a card or blade inserted into a slot in base station equipment 820. Alternatively, the module can also be a chip mounted on a card or blade. Meanwhile, the RF circuit 827 may include, for example, a mixer, a filter, and an amplifier, and transmits and receives wireless signals via the antenna 810.

[0209] As shown in Figure 13, the wireless communication interface 825 may include multiple BB processors 826. For example, the multiple BB processors 826 may be compatible with multiple frequency bands used by the eNB 800. As shown in Figure 13, the wireless communication interface 825 may include multiple RF circuits 827. For example, the multiple RF circuits 827 may be compatible with multiple antenna elements. Although Figure 13 shows an example in which the wireless communication interface 825 includes multiple BB processors 826 and multiple RF circuits 827, the wireless communication interface 825 may also include a single BB processor 826 or a single RF circuit 827.

[0210] When the electronic device 2000 is implemented as the eNB 800 shown in Figure 13, its transceiver can be implemented by the wireless communication interface 825. At least a portion of the functionality can also be implemented by the controller 821. For example, the controller 821 can reduce the communication overhead of FL model training without affecting the global model accuracy by executing the functions of the units in the electronic device 2000.

[0211] (Second application example)

[0212] Figure 14 is a block diagram illustrating a second example of a schematic configuration of an eNB or gNB to which the technologies of this disclosure can be applied. Note that, similarly, the following description uses an eNB as an example, but it can also be applied to a gNB. The eNB 830 includes one or more antennas 840, a base station device 850, and an RRH 860. The RRH 860 and each antenna 840 can be connected to each other via RF cables. The base station device 850 and the RRH 860 can be connected to each other via high-speed lines such as fiber optic cables.

[0213] Each of the antennas 840 includes one or more antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used by the RRH 860 to transmit and receive wireless signals. As shown in Figure 14, the eNB 830 may include multiple antennas 840. For example, multiple antennas 840 may be compatible with multiple frequency bands used by the eNB 830. Although Figure 14 shows an example in which the eNB 830 includes multiple antennas 840, the eNB 830 may also include a single antenna 840.

[0214] The base station device 850 includes a controller 851, a memory 852, a network interface 853, a wireless communication interface 855, and a connection interface 857. The controller 851, memory 852, and network interface 853 are the same as the controller 821, memory 822, and network interface 823 described with reference to FIG14.

[0215] The wireless communication interface 855 supports any cellular communication scheme (such as LTE and LTE-Advanced) and provides wireless communication to terminals located in the sector corresponding to the RRH 860 via the RRH 860 and antenna 840. The wireless communication interface 855 may typically include, for example, a BB processor 856. The BB processor 856 is identical to the BB processor 826 described with reference to FIG14, except that it is connected to the RF circuitry 864 of the RRH 860 via a connection interface 857. As shown in FIG14, the wireless communication interface 855 may include multiple BB processors 856. For example, multiple BB processors 856 may be compatible with multiple frequency bands used by the eNB 830. Although FIG14 shows an example in which the wireless communication interface 855 includes multiple BB processors 856, the wireless communication interface 855 may also include a single BB processor 856.

[0216] Connection interface 857 is an interface for connecting base station device 850 (wireless communication interface 855) to RRH 860. Connection interface 857 can also be a communication module for connecting base station device 850 (wireless communication interface 855) to the aforementioned high-speed line of RRH 860.

[0217] The RRH 860 includes a connectivity interface 861 and a wireless communication interface 863.

[0218] Connection interface 861 is an interface for connecting RRH 860 (wireless communication interface 863) to base station equipment 850. Connection interface 861 can also be a communication module for communication in the aforementioned high-speed line.

[0219] Wireless communication interface 863 transmits and receives wireless signals via antenna 840. Wireless communication interface 863 typically includes, for example, RF circuitry 864. RF circuitry 864 may include, for example, mixers, filters, and amplifiers, and transmits and receives wireless signals via antenna 840. As shown in FIG14, wireless communication interface 863 may include multiple RF circuits 864. For example, multiple RF circuits 864 may support multiple antenna elements. Although FIG14 shows an example in which wireless communication interface 863 includes multiple RF circuits 864, wireless communication interface 863 may also include a single RF circuit 864.

[0220] When the electronic device 2000 is implemented as the eNB 830 shown in Figure 14, its transceiver can be implemented by the wireless communication interface 855. At least a portion of the functionality can also be implemented by the controller 851. For example, the controller 851 can reduce the communication overhead of FL model training without affecting the global model accuracy by performing the functions of the units in the electronic device 2000.

[0221] [Application examples related to user equipment]

[0222] (First application example)

[0223] Figure 15 is a block diagram illustrating an example of a schematic configuration of a smartphone 900 to which the technology of this disclosure can be applied. The smartphone 900 includes a processor 901, a memory 902, a storage device 903, an external connection interface 904, a camera device 906, a sensor 907, a microphone 908, an input device 909, a display device 910, a speaker 911, a wireless communication interface 912, one or more antenna switches 915, one or more antennas 916, a bus 917, a battery 918, and an auxiliary controller 919.

[0224] The processor 901 can be, for example, a CPU or a system-on-a-chip (SoC), and controls the application layer and other functions of the smartphone 900. The memory 902 includes RAM and ROM, and stores data and programs executed by the processor 901. The storage device 903 can include storage media such as semiconductor memory and hard disks. The external connectivity interface 904 is an interface for connecting external devices, such as memory cards and Universal Serial Bus (USB) devices, to the smartphone 900.

[0225] The camera device 906 includes an image sensor (such as a charge-coupled device (CCD) and complementary metal-oxide-semiconductor (CMOS)) and generates captured images. The sensor 907 may include a set of sensors, such as a measurement sensor, a gyroscope sensor, a magnetometer sensor, and an accelerometer sensor. The microphone 908 converts sound input to the smartphone 900 into an audio signal. The input device 909 includes, for example, a touch sensor, keypad, keyboard, buttons, or switches configured to detect touches on the screen of the display device 910 and receives operations or information input from the user. The display device 910 includes a screen (such as a liquid crystal display (LCD) and an organic light-emitting diode (OLED) display) and displays the output image of the smartphone 900. The speaker 911 converts the audio signal output from the smartphone 900 into sound.

[0226] The wireless communication interface 912 supports any cellular communication scheme (such as LTE and LTE-Advanced) and performs wireless communication. The wireless communication interface 912 typically includes, for example, a BB processor 913 and RF circuitry 914. The BB processor 913 can perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and performs various types of signal processing for wireless communication. Meanwhile, the RF circuitry 914 can include, for example, mixers, filters, and amplifiers, and transmits and receives wireless signals via antenna 916. Note that although the figure shows a scenario where one RF link is connected to one antenna, this is only illustrative; scenarios where one RF link is connected to multiple antennas via multiple phase shifters also exist. The wireless communication interface 912 can be a single chip module on which the BB processor 913 and RF circuitry 914 are integrated. As shown in Figure 15, the wireless communication interface 912 can include multiple BB processors 913 and multiple RF circuits 914. Although Figure 15 shows an example where the wireless communication interface 912 includes multiple BB processors 913 and multiple RF circuits 914, the wireless communication interface 912 can also include a single BB processor 913 or a single RF circuitry 914.

[0227] In addition to cellular communication schemes, the wireless communication interface 912 can support other types of wireless communication schemes, such as short-range wireless communication schemes, near-field communication schemes, and wireless local area network (LAN) schemes. In this case, the wireless communication interface 912 may include a BB processor 913 and RF circuitry 914 for each wireless communication scheme.

[0228] Each of the antenna switches 915 switches the connection destination of the antenna 916 among multiple circuits (e.g., circuits for different wireless communication schemes) included in the wireless communication interface 912.

[0229] Each of the antennas 916 includes one or more antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used for transmitting and receiving wireless signals through the wireless communication interface 912. As shown in Figure 15, the smartphone 900 may include multiple antennas 916. Although Figure 15 shows an example in which the smartphone 900 includes multiple antennas 916, the smartphone 900 may also include a single antenna 916.

[0230] Furthermore, the smartphone 900 may include an antenna 916 for each wireless communication scheme. In this case, the antenna switch 915 can be omitted from the configuration of the smartphone 900.

[0231] Bus 917 connects processor 901, memory 902, storage device 903, external connection interface 904, camera device 906, sensor 907, microphone 908, input device 909, display device 910, speaker 911, wireless communication interface 912, and auxiliary controller 919 to each other. Battery 918 supplies power to the various blocks of smartphone 900 shown in FIG. 15 via feeders, which are partially shown as dashed lines in the figure. Auxiliary controller 919 operates the minimum necessary functions of smartphone 900, for example, in sleep mode.

[0232] When the electronic device 9000 is implemented, for example, as a smartphone on the user equipment side, such as the smartphone 900 shown in FIG. 15, the transceiver of the electronic device 9000 can be implemented by the wireless communication interface 912. At least a portion of the functionality can also be implemented by the processor 901 or the auxiliary controller 919. For example, by executing the functions of the units in the electronic device 9000 described above, the processor 901 or the auxiliary controller 919 enables the reduction of communication overhead for FL model training without affecting the global model accuracy.

[0233] (Second application example)

[0234] Figure 16 is a block diagram illustrating an example of a schematic configuration of a car navigation device 920 to which the technology of this disclosure can be applied. The car navigation device 920 includes a processor 921, a memory 922, a Global Positioning System (GPS) module 924, a sensor 925, a data interface 926, a content player 927, a storage medium interface 928, an input device 929, a display device 930, a speaker 931, a wireless communication interface 933, one or more antenna switches 936, one or more antennas 937, and a battery 938.

[0235] The processor 921 can be, for example, a CPU or a SoC, and controls the navigation functions and other functions of the car navigation device 920. The memory 922 includes RAM and ROM, and stores data and programs executed by the processor 921.

[0236] GPS module 924 uses GPS signals received from GPS satellites to measure the location (such as latitude, longitude, and altitude) of car navigation device 920. Sensor 925 may include a set of sensors, such as a gyroscope sensor, a geomagnetic sensor, and an air pressure sensor. Data interface 926 is connected to, for example, an in-vehicle network 941 via a terminal not shown, and acquires data generated by the vehicle (such as vehicle speed data).

[0237] Content player 927 reproduces content stored on storage media (such as CDs and DVDs), which is inserted into storage media interface 928. Input device 929 includes, for example, a touch sensor, button, or switch configured to detect touch on the screen of display device 930, and receives operations or information input from the user. Display device 930 includes a screen such as an LCD or OLED display and displays images or reproduced content for navigation functions. Speaker 931 outputs sound for navigation functions or reproduced content.

[0238] The wireless communication interface 933 supports any cellular communication scheme (such as LTE and LTE-Advanced) and performs wireless communication. The wireless communication interface 933 typically includes, for example, a BB processor 934 and RF circuitry 935. The BB processor 934 can perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and performs various types of signal processing for wireless communication. Meanwhile, the RF circuitry 935 can include, for example, mixers, filters, and amplifiers, and transmits and receives wireless signals via antenna 937. The wireless communication interface 933 can also be a chip module on which the BB processor 934 and RF circuitry 935 are integrated. As shown in Figure 16, the wireless communication interface 933 can include multiple BB processors 934 and multiple RF circuits 935. Although Figure 16 shows an example where the wireless communication interface 933 includes multiple BB processors 934 and multiple RF circuits 935, the wireless communication interface 933 can also include a single BB processor 934 or a single RF circuitry 935.

[0239] In addition to cellular communication schemes, the wireless communication interface 933 can support other types of wireless communication schemes, such as short-range wireless communication schemes, near-field communication schemes, and wireless LAN schemes. In this case, for each wireless communication scheme, the wireless communication interface 933 may include a BB processor 934 and an RF circuit 935.

[0240] Each of the antenna switches 936 switches the connection destination of the antenna 937 among multiple circuits (such as circuits for different wireless communication schemes) included in the wireless communication interface 933.

[0241] Each of the antennas 937 includes one or more antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used for transmitting and receiving wireless signals through the wireless communication interface 933. As shown in Figure 16, the car navigation device 920 may include multiple antennas 937. Although Figure 16 shows an example in which the car navigation device 920 includes multiple antennas 937, the car navigation device 920 may also include a single antenna 937.

[0242] Furthermore, the car navigation device 920 may include an antenna 937 for each wireless communication scheme. In this case, the antenna switch 936 can be omitted from the configuration of the car navigation device 920.

[0243] Battery 938 supplies power to the various blocks of the car navigation device 920 shown in Figure 16 via feeders, which are partially shown as dashed lines in the figure. Battery 938 accumulates the power supplied from the vehicle.

[0244] When the electronic device 9000 is implemented, for example, as a car navigation device on the user equipment side, such as the car navigation device 920 shown in FIG16, the transceiver of the electronic device 9000 can be implemented by the wireless communication interface 933. At least a portion of the functionality can also be implemented by the processor 921. For example, by executing the functions of the units in the electronic device 9000 described above, the processor 921 enables the reduction of communication overhead for FL model training without affecting the global model accuracy.

[0245] The technology disclosed herein can also be implemented as an in-vehicle system (or vehicle) 940 comprising one or more of the following blocks: a car navigation device 920, an in-vehicle network 941, and a vehicle module 942. The vehicle module 942 generates vehicle data (such as vehicle speed, engine speed, and fault information) and outputs the generated data to the in-vehicle network 941.

[0246] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in the form of hardware, firmware, software or a combination thereof. This can be achieved by those skilled in the art using their basic circuit design knowledge or basic programming skills after reading the description of the present invention.

[0247] Furthermore, this invention also proposes a program product storing machine-readable instruction code. When the instruction code is read and executed by a machine, the method described above according to embodiments of the present invention can be performed.

[0248] Accordingly, the storage medium used to carry the program product storing the machine-readable instruction code is also included in the disclosure of this invention. Storage media include, but are not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, etc.

[0249] When the present invention is implemented by software or firmware, the program constituting the software is installed from a storage medium or network onto a computer with a dedicated hardware structure (e.g., the general-purpose computer 1700 shown in FIG17), which is capable of performing various functions when various programs are installed.

[0250] In Figure 17, the Central Processing Unit (CPU) 1701 executes various processes based on programs stored in Read-Only Memory (ROM) 1702 or programs loaded into Random Access Memory (RAM) 1703 from Storage Section 1708. The RAM 1703 also stores data required as needed when the CPU 1701 executes various processes, etc. The CPU 1701, ROM 1702, and RAM 1703 are connected to each other via bus 1704. An input / output interface 1705 is also connected to bus 1704.

[0251] The following components are connected to the input / output interface 1705: input section 1706 (including keyboard, mouse, etc.), output section 1707 (including monitor, such as cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.), storage section 1708 (including hard disk, etc.), and communication section 1709 (including network interface card, such as LAN card, modem, etc.). The communication section 1709 performs communication processing via a network, such as the Internet. If necessary, a drive 1710 may also be connected to the input / output interface 1705. Removable media 1711, such as disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on the drive 1710 as needed, so that computer programs read from them can be installed into the storage section 1708 as needed.

[0252] When the above series of processes are implemented by software, the program constituting the software is installed from a network such as the Internet or a storage medium such as removable media 1711.

[0253] Those skilled in the art will understand that such storage media are not limited to the removable medium 1711 shown in FIG. 17, which stores programs and is distributed separately from the device to provide programs to users. Examples of removable media 1711 include magnetic disks (including floppy disks (registered trademark)), optical disks (including optical disc read-only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including mini-discs (MD) (registered trademark)), and semiconductor memories. Alternatively, the storage medium may be ROM 1702, a hard disk included in storage section 1708, etc., which stores programs and is distributed to users along with the device containing them.

[0254] It should also be noted that in the apparatus, method, and system of the present invention, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Furthermore, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order. Some steps can be performed in parallel or independently of each other.

[0255] Finally, it should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Furthermore, unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0256] While embodiments of the present invention have been described in detail above with reference to the accompanying drawings, it should be understood that the embodiments described above are merely illustrative and do not constitute a limitation thereof. Those skilled in the art can make various modifications and alterations to the above embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims and their equivalents.

[0257] This technology can also be implemented as follows.

[0258] Option 1. An electronic device for a federated learning server, comprising:

[0259] At least one processor; and

[0260] At least one memory, including computer program code, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute via the at least one processor:

[0261] The core network sends information about the magnitude of parameters in the global model used for federated learning to assist it in identifying federated learning clients from multiple clients that cannot upload all parameters of their local models as partial upload clients.

[0262] Under predetermined conditions, the partial parameters of the local model received from the partial upload client are aggregated to obtain the global model in the current round of federated learning training.

[0263] Option 2. The electronic device according to Option 1, wherein the magnitude information includes information about the magnitude of the parameters of the global model.

[0264] Option 3. The electronic device according to Option 1 or 2, wherein,

[0265] The predetermined conditions include receiving a list of the partial upload clients from the core network and / or the global model's accuracy in the previous training round reaching a preset accuracy, wherein the list includes the IDs of the partial upload clients, and

[0266] If the accuracy of the global model in the previous training round reaches the preset accuracy, all of the multiple federated learning clients will serve as the partial upload clients.

[0267] Option 4. The electronic device according to Option 3, wherein,

[0268] The IDs of some of the upload clients included in the list are determined by the core network based on the magnitude information and the predicted channel information between the electronic device and the federated learning client.

[0269] Option 5. The electronic device according to Option 4, wherein,

[0270] The channel information includes at least one of the following: channel quality indicator (CQI), precoding matrix indicator (PMI), reference signal received power (RSRP), and signal-to-noise ratio (SNR) of the channel between the electronic device and the federated learning client.

[0271] Option 6. An electronic device according to any one of Options 1 to 5, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0272] Regarding the parameters of the global model from the previous training round and the parameters of the local models uploaded by the multiple federated learning clients in the previous training round:

[0273] Calculate the local model parameter bias, representing the degree of dispersion among the parameters of the local models of the multiple federated learning clients, and calculate the global model parameter bias, representing the degree of deviation between the parameters of the local models of the multiple federated learning clients and the parameters of the global model.

[0274] The partial parameters are determined based on the degree of deviation of the local model parameters and the degree of deviation of the global model parameters.

[0275] Option 7. The electronic device according to Option 6, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0276] The average value of the i-th parameter is calculated based on the local model of the multiple federated learning clients, and the degree of local model parameter bias for the i-th parameter is calculated based on the difference between the i-th parameter of the local model of each federated learning client and the average value; and

[0277] Based on the difference between the i-th parameter of the local model and the i-th parameter of the global model in each federated learning client, the degree of deviation of the global model parameter for the i-th parameter is calculated.

[0278] Option 8. The electronic device according to Option 6 or 7, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0279] Based on the degree of deviation of the local model parameters and the degree of deviation of the global model parameters, a clustering method is used to cluster the parameters, thereby obtaining multiple clusters.

[0280] From the plurality of clusters, select the clusters that meet the predetermined cluster conditions, and

[0281] The parameters in the selected cluster are used as the partial parameters.

[0282] Option 9. The electronic device according to Option 8, wherein,

[0283] The predetermined cluster conditions include the probability of the cluster being selected being greater than a predetermined probability, and

[0284] The at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0285] Using the local model parameter deviation and the global model parameter deviation as coordinates in a two-dimensional coordinate system, and taking the calculated parameter's corresponding local and global model parameter deviation as coordinate values, each parameter is represented as a point in the coordinate system.

[0286] The K-Means algorithm is used to cluster all points in the coordinate system to obtain the multiple clusters.

[0287] For each of at least a subset of the plurality of clusters, the probability of selection of the cluster is calculated based on the distance between the cluster center and the origin, and the angle between the line connecting the cluster center to the origin and the dividing line of the first quadrant of the coordinate system, wherein the dividing line is a straight line passing through the origin and forming a predetermined angle with the horizontal axis of the first quadrant.

[0288] Specifically, the greater the distance, the greater the probability of being selected; and the smaller the included angle, the greater the probability of being selected.

[0289] Option 10. An electronic device according to any one of Options 1 to 9, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0290] Send notification information about the partial parameters to the partial upload client so that the partial upload client can upload only the partial parameters in the current round of training.

[0291] Option 11. The electronic device according to Option 10, wherein,

[0292] The global model is a hierarchical neural network model.

[0293] The notification information includes the location information of the aforementioned parameters within the neural network model.

[0294] Option 12. The electronic device according to Option 11, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0295] Based on the location information, partial parameters of the local model received from the partial upload client and all parameters of the local model received from other clients among the plurality of federated learning clients are aligned, and the local models are aggregated.

[0296] Scheme 13. An electronic device according to any one of Schemes 1 to 12, wherein the magnitude information is used as input to UE communication analysis in the Network Data Analysis Function (NWDAF), and the UE communication prediction output of the UE communication analysis includes a list having the IDs of the partially uploaded clients.

[0297] Option 14. An electronic device according to any one of Options 1 to 13, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0298] Receives from the core network a computation offload instruction that instructs the federated learning client to perform computation offload, and

[0299] The computation offload instruction is sent to all federated learning clients within the coverage area of ​​the electronic device, so that the federated learning clients can determine whether to perform computation offload.

[0300] The computation offload instruction is sent by the core network based on the dwell time of all federated learning clients within the coverage area of ​​the electronic device.

[0301] Option 15. The electronic device according to Option 14, wherein,

[0302] The computational offload instruction is sent by the core network when it determines that the dwell time is less than the preset training duration of the local model.

[0303] Option 16. The electronic device according to Option 14 or 15, wherein,

[0304] The dwell time is obtained by the core network based on the ID of the electronic device, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the electronic device.

[0305] Option 17. The electronic device according to Option 16, wherein,

[0306] The dwell time is also obtained by the core network based on the movement speed of the federated learning client.

[0307] Option 18. An electronic device for a core network, comprising:

[0308] At least one processor; and

[0309] At least one memory, including computer program code, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute via the at least one processor:

[0310] The system receives information about the magnitude of parameters of the global model used for federated learning from the federated learning server. It identifies federated learning clients that cannot upload all parameters of their local models from among multiple federated learning clients as partial upload clients. Under predetermined conditions, the federated learning server aggregates the partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0311] Option 19. The electronic device according to Option 18, wherein the magnitude information includes information about the magnitude of the parameters of the global model.

[0312] Option 20. The electronic device according to Option 18 or 19, wherein,

[0313] The predetermined conditions include the federated learning server receiving a list of the partial upload clients from the electronic device and / or the global model's accuracy in the previous training round reaching a preset accuracy, wherein the list includes the IDs of the partial upload clients.

[0314] Option 21. The electronic device according to Option 20, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0315] The IDs of some upload clients included in the list are determined based on the magnitude information and the predicted channel information between the federated learning server and the federated learning client.

[0316] Option 22. The electronic device according to Option 21, wherein,

[0317] The channel information includes at least one of the following: Channel Quality Indicator (CQI), Precoding Matrix Indicator (PMI), Reference Signal Received Power (RSRP), and Signal-to-Noise Ratio (SNR) of the channel between the federated learning server and the federated learning client.

[0318] Scheme 23. An electronic device according to any one of Schemes 18 to 22, wherein the magnitude information is used as input to UE communication analysis in the Network Data Analysis Function (NWDAF), and the UE communication prediction output of the UE communication analysis includes a list having the IDs of the partially uploaded clients.

[0319] Option 24. An electronic device according to any one of Options 18 to 23, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0320] Based on the dwell time of all federated learning clients within the coverage area of ​​the federated learning server, a computation unloading instruction is sent to the federated learning server, instructing the federated learning clients to perform computation unloading. The federated learning server then notifies all federated learning clients within its coverage area of ​​the computation unloading instruction, allowing the federated learning clients to determine whether to perform computation unloading.

[0321] Option 25. The electronic device according to Option 24, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0322] If the dwell time is determined to be less than the preset training time of the local model, the computation unloading instruction is sent.

[0323] Option 26. The electronic device according to Option 24 or 25, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0324] The dwell time is obtained based on the ID of the federated learning server, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the federated learning server.

[0325] Option 27. The electronic device according to Option 26, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0326] The dwell time is also obtained based on the movement speed of the federated learning client.

[0327] Option 28. An electronic device for a federated learning client, comprising:

[0328] At least one processor; and

[0329] At least one memory, including computer program code, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute via the at least one processor:

[0330] Under predetermined conditions, partial parameters of the local models used for federated learning are sent to the federated learning server, so that the federated learning server can aggregate the local models based on the partial parameters to obtain the global model in the current round of federated learning training.

[0331] The federated learning server sends information about the magnitude of parameters in the global model to the core network to help the core network identify the electronic device as a partial upload client that cannot upload all parameters of its local model.

[0332] Option 29. The electronic device according to Option 28, wherein the magnitude information includes information about the magnitude of the parameters of the global model.

[0333] Option 30. The electronic device according to Option 28 or 29, wherein,

[0334] The predetermined conditions include the federated learning server receiving from the core network a list including the IDs of some of the uploaded clients and / or the global model in the previous training round reaching a preset accuracy.

[0335] Option 31. An electronic device according to any one of Options 28 to 30, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0336] Receive notification information about the aforementioned parameters from the federated learning server.

[0337] Option 32. The electronic device according to Option 31, wherein,

[0338] The global model is a hierarchical neural network model.

[0339] The notification information includes the location information of the aforementioned parameters within the neural network model.

[0340] Scheme 33. An electronic device according to any one of Schemes 28 to 32, wherein the magnitude information is used as input to UE communication analysis in the Network Data Analysis Function (NWDAF), and the UE communication prediction output of the UE communication analysis includes a list with some of the IDs of the uploading clients.

[0341] Option 34. An electronic device according to any one of Options 28 to 33, wherein the at least one memory and the computer program code are configured to cause the electronic device to execute, via the at least one processor:

[0342] Receive a computation unloading instruction from the federated learning server instructing the federated learning client to perform computation unloading, in order to determine whether to perform the computation unloading.

[0343] The computation offload instruction is sent by the core network to the federated learning server based on the dwell time of all federated learning clients within the coverage area of ​​the federated learning server.

[0344] Option 35. The electronic device according to Option 34, wherein,

[0345] The computational offload instruction is sent by the core network when it determines that the dwell time is less than the preset training duration of the local model.

[0346] Option 36. The electronic device according to Option 34 or 35, wherein,

[0347] The dwell time is obtained by the core network based on the ID of the federated learning server, the time ID used to represent time, and the density of federated learning clients within the coverage area of ​​the federated learning server.

[0348] Option 37. The electronic device according to Option 36, wherein,

[0349] The dwell time is also obtained by the core network based on the movement speed of the federated learning client.

[0350] Option 38. A method for a federated learning server, comprising:

[0351] The core network sends information about the magnitude of parameters in the global model used for federated learning to assist it in identifying federated learning clients from multiple clients that cannot upload all parameters of their local models as partial upload clients.

[0352] Under predetermined conditions, the partial parameters of the local model received from the partial upload client are aggregated to obtain the global model in the current round of federated learning training.

[0353] Option 39. A method for a core network, comprising:

[0354] The system receives information about the magnitude of parameters of the global model used for federated learning from the federated learning server. It identifies federated learning clients that cannot upload all parameters of their local models from among multiple federated learning clients as partial upload clients. Under predetermined conditions, the federated learning server aggregates the partial parameters of the local models received from the partial upload clients to obtain the global model in the current round of training for federated learning.

[0355] Option 40. A method for a federated learning client, comprising:

[0356] Under predetermined conditions, partial parameters of the local models used for federated learning are sent to the federated learning server, so that the federated learning server can aggregate the local models based on the partial parameters to obtain the global model in the current round of federated learning training.

[0357] The federated learning server sends information about the magnitude of parameters in the global model to the core network to help the core network identify the federated learning client as a partial upload client that cannot upload all parameters of its local model.

[0358] Scheme 41. A computer-readable storage medium having stored thereon computer-executable instructions that, when executed, perform the method according to any one of Schemes 38 to 40.

Claims

1. An electronic device for a federated learning server, comprising: at least one processor; and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: sending magnitude information about a parameter magnitude of a global model for federated learning to a core network to assist the core network to identify, from a plurality of federated learning clients, a federated learning client that is unable to upload all parameters of its local model as a partial upload client, and based on partial parameters of a local model received from the partial upload client, performing aggregation to obtain the global model in a current round training of federated learning, if a predetermined condition is met. 2.The electronic device of claim 1, wherein, The magnitude information comprises information about a size of a parameter of the global model.

3. The electronic device according to claim 1 or 2, wherein the predetermined condition comprises receiving, from the core network, a list about the partial upload client and / or an accuracy of the global model in a last round training reaches a preset accuracy, wherein the list comprises an ID of the partial upload client, and in a case that the accuracy of the global model in the last round training reaches the preset accuracy, all of the plurality of federated learning clients are the partial upload client.

4. The electronic device according to claim 3, wherein the ID of the partial upload client included in the list is determined by the core network based on the magnitude information and channel information between the electronic device and a federated learning client predicted.

5. The electronic device according to claim 4, wherein the channel information comprises at least one of a channel quality indication (CQI), a precoding matrix indication (PMI), a reference signal received power (RSRP) and a signal to noise ratio (SNR) of a channel between the electronic device and a federated learning client.

6. The electronic device of any of claims 1-5, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: for parameters of the global model in a last round training and parameters of local models uploaded by the plurality of federated learning clients in the last round training: calculating a local model parameter deviation degree representing a degree of dispersion among the parameters of the local models of the plurality of federated learning clients, and calculating a global model parameter deviation degree representing a degree of deviation between the parameters of the local models of the plurality of federated learning clients and the parameters of the global model, and based on the local model parameter deviation degree and the global model parameter deviation degree, determining the partial parameters.

7. The electronic device of claim 6, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: based on an i-th parameter of a local model of the plurality of federated learning clients, calculating an average value of the i-th parameter, and based on a difference between the i-th parameter of the local model of each federated learning client and the average value, calculating a local model parameter deviation degree for the i-th parameter; and and The global model parameter deviation degree for the i-th parameter is calculated based on a difference between the i-th parameter of the local model of each federated learning client and the i-th parameter of the global model.

8. The electronic device of claim 6 or 7, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: based on the local model parameter deviation degree and the global model parameter deviation degree, clustering parameters by using a clustering method, thereby obtaining a plurality of clusters after clustering, from the plurality of clusters, selecting a selected cluster that satisfies a predetermined cluster condition, and taking the parameters in the selected cluster as the partial parameters.

9. The electronic device according to claim 8, wherein the predetermined cluster condition comprises that a selected probability corresponding to a cluster is greater than a predetermined probability, and the at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: respectively taking the local model parameter deviation degree and the global model parameter deviation degree as coordinates of a two-dimensional coordinate system, and taking the calculated local model parameter deviation degree and global model parameter deviation degree corresponding to the parameters as coordinate values, thereby representing each parameter as a point in the coordinate system, clustering all points in the coordinate system by using a K-Means algorithm, thereby obtaining the plurality of clusters, and for each cluster in at least part of the plurality of clusters, based on a distance between a cluster center of the cluster and a coordinate origin, and an included angle between a connecting line from the cluster center to the coordinate origin and a dividing line of a first quadrant of the coordinate system, calculating a selected probability of the cluster, wherein the dividing line is a straight line passing through the coordinate origin and forming a predetermined angle with a horizontal axis of the first quadrant, wherein the greater the distance is, the greater the selected probability is, and the smaller the included angle is, the greater the selected probability is.

10. The electronic device of any of claims 1-9, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: sending notification information about the partial parameters to the partial upload clients, for the partial upload clients to upload only the partial parameters in the current round of training.

11. The electronic device according to claim 10, wherein the global model is a hierarchical neural network model, the notification information comprises location information of the partial parameters in the neural network model.

12. The electronic device of claim 11, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: based on the location information, aligning partial parameters of a local model received from the partial upload clients and all parameters of a local model received from other clients in the plurality of federated learning clients, and performing aggregation of the local model.

13. The electronic device of any of claims 1-12, wherein, The magnitude information is input to UE communication analysis in a Network Data Analytics Function (NWDAF), and a UE communication prediction output of the UE communication analysis comprises a list with IDs of the partial upload clients. The magnitude information is input to UE communication analysis in a Network Data Analytics Function (NWDAF), and a UE communication prediction output of the UE communication analysis comprises a list with IDs of the partial upload clients.

14. The electronic device of any of claims 1-13, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: receiving, from the core network, a computation offloading indication indicating that a federated learning client is to perform computation offloading, and informing all federated learning clients within a coverage range of the electronic device of the computation offloading indication for the federated learning clients to determine whether to perform computation offloading, wherein the computation offloading indication is sent by the core network based on a residence time of all federated learning clients within the coverage range of the electronic device.

15. The electronic device of claim 14, wherein the computation offloading indication is sent by the core network in a case where the residence time is determined to be less than a preset training duration of a local model.

16. The electronic device of claim 14 or 15, wherein the residence time is obtained by the core network based on an ID of the electronic device, a time ID for characterizing time, and a density of federated learning clients within the coverage range of the electronic device.

17. The electronic device of claim 16, wherein the residence time is obtained by the core network further based on a moving speed of a federated learning client.

18. An electronic device for a core network, comprising: at least one processor; and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: receiving, from a federated learning server, magnitude information about a parameter magnitude of a global model for federated learning, to identify, from among a plurality of federated learning clients, a federated learning client that is unable to upload all parameters of a local model thereof as a partial upload client, for the federated learning server to aggregate, based on partial parameters of a local model received from the partial upload client, a global model in a current round training of federated learning in a case where a predetermined condition is met. The magnitude information includes information about a size of a parameter of the global model.

19. The electronic device of claim 18, wherein, 20. The electronic device of claim 18 or 19, wherein the predetermined condition includes the federated learning server receiving, from the electronic device, a list of the partial upload client and / or an accuracy of a global model in a last round training reaching a preset accuracy, wherein the list includes an ID of the partial upload client. The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform:

21. The electronic device of claim 20, wherein, determining, based on the magnitude information and predicted channel information between the federated learning server and a federated learning client, an ID of a partial upload client included in the list.

22. The electronic device of claim 21, wherein the channel information includes at least one of a channel quality indication (CQI), a precoding matrix indication (PMI), a reference signal received power (RSRP), and a signal-to-noise ratio (SNR) of a channel between the federated learning server and a federated learning client. ​ 23. The electronic device of any of claims 18-22, wherein, The magnitude information is taken as an input of UE communication analysis in a network data analytics function (NWDAF), and a UE communication prediction output of the UE communication analysis includes a list with IDs of the partial upload clients.

24. The electronic device of any of claims 18-23, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: based on a residence time of all federated learning clients within a coverage of the federated learning server, sending, to the federated learning server, a computation offloading indication indicating that a federated learning client is to perform computation offloading, for the federated learning server to notify all federated learning clients within its coverage of the computation offloading indication, for a federated learning client to determine whether to perform computation offloading.

25. The electronic device of claim 24, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: in a case where it is determined that the residence time is less than a preset training duration of a local model, sending the computation offloading indication.

26. The electronic device of claim 24 or 25, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: based on an ID of the federated learning server, a time ID for representing time, and a density of federated learning clients within a coverage of the federated learning server, obtaining the residence time.

27. The electronic device of claim 26, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: further based on a moving speed of a federated learning client, obtaining the residence time. 28.An electronic device for a federated learning client, comprising: at least one processor; and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: in a case where a predetermined condition is met, sending, to a federated learning server, partial parameters of a local model for federated learning, for the federated learning server to perform aggregation of the local model based on the partial parameters to obtain a global model in a current round training of the federated learning, wherein the federated learning server sends, to a core network, magnitude information about a magnitude of parameters of the global model, to assist the core network to identify the electronic device as a partial upload client that is unable to upload all parameters of its local model.

29. The electronic device of claim 28, wherein, The magnitude information includes information about a size of the parameters of the global model. 30.The electronic device of claim 28 or 29, wherein the predetermined condition includes that the federated learning server receives, from the core network, a list including IDs of the partial upload clients and / or that an accuracy of the global model in a last round training reaches a preset accuracy.

31. The electronic device of any of claims 28-30, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: receiving, from the federated learning server, notification information about the partial parameters. 32.The electronic device of claim 31, wherein The global model is a hierarchical neural network model, The notification information includes position information of the partial parameters in the neural network model.

33. The electronic device of any of claims 28-32, wherein, The magnitude information is input to UE communication analysis in a network data analysis function (NWDAF), and a UE communication prediction output of the UE communication analysis includes a list with IDs of partial upload clients.

34. The electronic device of any of claims 28-33, wherein, The at least one memory and the computer program code are configured to, with the at least one processor, cause the electronic device to perform: receiving, from the federated learning server, a computation offloading indication indicating that a federated learning client is to perform computation offloading, for determining whether to perform the computation offloading, The computation offloading indication is sent by the core network to the federated learning server based on a residence time of all federated learning clients within a coverage range of the federated learning server.

35. The electronic device of claim 34, wherein The computation offloading indication is sent by the core network in a case where the residence time is determined to be less than a preset training duration of a local model.

36. The electronic device of claim 34 or 35, wherein The residence time is obtained by the core network based on an ID of the federated learning server, a time ID for characterizing time, and a density of federated learning clients within a coverage range of the federated learning server.

37. The electronic device of claim 36, wherein The residence time is obtained by the core network further based on a moving speed of a federated learning client.

38. A method for a federated learning server, comprising: sending, to a core network, magnitude information about a parameter magnitude of a global model for federated learning, to assist the core network in identifying, from a plurality of federated learning clients, a federated learning client that is unable to upload all parameters of a local model thereof as a partial upload client, and based on partial parameters of a local model received from the partial upload client, performing aggregation to obtain the global model in a current round training of federated learning in a case where a predetermined condition is met.

39. A method for a core network, comprising: receiving, from a federated learning server, magnitude information about a parameter magnitude of a global model for federated learning, to identify, from a plurality of federated learning clients, a federated learning client that is unable to upload all parameters of a local model thereof as a partial upload client, for the federated learning server to perform aggregation based on partial parameters of a local model received from the partial upload client to obtain the global model in a current round training of federated learning in a case where a predetermined condition is met.

40. A method for a federated learning client, comprising: sending, to a federated learning server, partial parameters of a local model for federated learning in a case where a predetermined condition is met, for the federated learning server to perform aggregation of the local model based on the partial parameters to obtain the global model in a current round training of federated learning, ​ In the method, the federated learning server sends magnitude information about a parameter magnitude of a global model to a core network to assist the core network in identifying the federated learning client as a partial upload client that cannot upload all parameters of a local model of the federated learning client.

41. A computer-readable storage medium having stored thereon computer- executable instructions which, when executed, perform a method according to any one of claims 38 to 40.

Images