Model training method and communication device

Abstract

The present application discloses a model training method, an apparatus, and a communication device, which belong to the field of communication technology. The model training method of the embodiment of the present application includes the following steps: when a first network element performs a federated model training process corresponding to a model training task, the first network element sends a federated model training request message to at least one second network element, where the at least one second network element is a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between each second network element are different but have the same sample features; the first network element receives first information sent by the at least one second network element, where the first information at least includes a first training result, and the first training result corresponds to the training data used in the federated model training process in the second network element; and the first network element performs model training based on the first model and the first training result reported by the at least one second network element, to obtain a target model and / or a second training result.

Description

[Technical field]

[0001] The present application relates to the field of communication technology, and in particular to a model training method, apparatus and communication device. [Background technology]

[0002] With the rapid development of artificial intelligence (AI), it has been widely applied in various fields.

[0003] Here, taking the communications field as an example, with the introduction of AI functions, there is inevitably a need for machine learning. For example, for a certain smart network service business, in order to realize more accurate services, it is necessary to train models with a large amount of data generated in different domains, different network elements, etc. However, as data security and privacy issues are increasingly emphasized, there exists a problem of data isolation between different domains, different network elements, etc. in the network, so how to realize model training based on data on different domains and different network elements is currently a problem to be solved. Summary of the Invention [Problem to be solved by the invention]

[0004] The embodiments of the present application provide a model training method, an apparatus, and a communication device that can realize model training by linking data located in different domains and different network elements without sharing the data. [Means for solving the problem]

[0005] A first aspect provides a model training method, the model training method including: when a first network element performs a federated model training process corresponding to a model training task, a step of sending a federated model training request message to at least one second network element, where the at least one second network element is a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between each of the second network elements are different but have the same sample features; a step of the first network element receiving first information sent by the at least one second network element, where the first information includes at least a first training result, and the first training result corresponds to training data used in the federated model training process in the second network element; and a step of the first network element performing model training based on a first model and the first training result reported by the at least one second network element, to obtain a target model and / or a second training result.

[0006] A second aspect provides a model training method, the method including: a step of receiving, by a second network element, a federated model training request message sent by a first network element to request the second network element to participate in a federated model training process corresponding to a model training task; a step of the second network element performing model training based on the federated model training request message to obtain a first training result; and a step of the second network element transmitting first information including at least the first training result to the first network element, where the first training result corresponds to training data used in the federated model training process in the second network element, samples used in the federated model training in the second network element are different from samples used in the federated model training in a fifth network element but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federated model training process.

[0007] A third aspect provides a model training method, the method including a step of receiving, by a fourth network element, target model related information transmitted by a first network element, where the target model related information is used at least to characterize the target model as a horizontal federation model.

[0008] A fourth aspect provides a model training apparatus, the model training apparatus including: a first sending module for sending a federated model training request message to at least one second network element when performing a federated model training process corresponding to a model training task, the at least one second network element being a network element participating in the federated model training process, where samples corresponding to training data used in the federated model training process between each of the second network elements are different but have the same sample features; a first receiving module for receiving first information sent by the at least one second network element, the first information including at least a first training result, the first training result corresponding to training data used in the federated model training process in the second network element; and a first model training module for performing model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0009] A fifth aspect provides a model training apparatus, the model training apparatus including: a second receiving module for receiving a federated model training request message sent by a first network element by a second network element to request the second network element to participate in a federated model training process corresponding to a model training task; a second model training module for performing model training based on the federated model training request message to obtain a first training result; and a second transmitting module for transmitting first information including at least the first training result to the first network element, where the first training result corresponds to training data used in the federated model training process in the second network element, samples used in the federated model training in the second network element are different from samples used in the federated model training in a fifth network element but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federated model training process.

[0010] A sixth aspect provides a model training apparatus, the model training apparatus including a third receiving module for receiving target model related information transmitted by a first network element, where the target model related information is used at least to characterize the target model as a horizontal federation model.

[0011] A seventh aspect provides a communications device, the communications device comprising a processor and a memory, the memory storing a program or instructions operable to run on the processor, the program or instructions, when executed by the processor, implementing steps of a method according to the first aspect, the second aspect or the third aspect.

[0012] An eighth aspect provides a communications device, the communications device comprising a processor and a communications interface, wherein the communications interface is coupled to the processor, the processor being adapted to run a program or instructions to perform steps of the method of the first aspect, or to perform steps of the method of the second aspect, or to perform steps of the method of the third aspect.

[0013] A ninth aspect provides a model training system, the model training system comprising a first network element, a second network element and a third network element, the first network element may be used to perform steps of the model training method described in the first aspect, the second network element may be used to perform steps of the model training method described in the second aspect, and the third network element may be used to perform steps of the model training method described in the third aspect.

[0014] A tenth aspect provides a readable storage medium having a program or instructions stored thereon, the program or instructions, when executed by a processor, performing steps of the method of the first aspect, or performing steps of the method of the second aspect, or performing steps of the method of the third aspect.

[0015] An eleventh aspect provides a chip, the chip including a processor and a communication interface, the communication interface coupled to the processor, the processor being adapted to run a program or instructions to perform steps of the method of the first aspect, or to perform steps of the method of the second aspect, or to perform steps of the method of the third aspect.

[0016] A twelfth aspect provides a computer program / program product, the computer program / program product being stored on a storage medium, the computer program / program product being executed by at least one processor to perform the steps of the method of the first aspect, or to perform the steps of the method of the second aspect, or to perform the steps of the method of the third aspect. Effect of the Invention

[0017] In the embodiment of the present application, the first network element uses a horizontal federation model training method to cooperate with each second network element to realize local distributed model training, thereby ensuring data privacy and data security in each second network element without sharing data on each second network element in the communication network, and also ensuring the model training effect. [Brief description of the drawings]

[0018] [Figure 1] FIG. 1 is a structural schematic diagram of a wireless communication system according to an exemplary embodiment of the present application; [Diagram 2] FIG. 1 is a structural schematic diagram of a model training system according to an exemplary embodiment of the present application; [Diagram 3] 1 is a flowchart of a model training method according to an embodiment of the present application. [Figure 4] 2 is a second flowchart of a model training method according to an embodiment of the present application. [Diagram 5] 3 is a flowchart of a model training method according to an embodiment of the present application. [Figure 6a] 1 is an interaction flowchart of a model training method according to an embodiment of the present application; [Figure 6b] 2 is a second interaction flowchart of a model training method according to an embodiment of the present application. [Figure 7]4 is a fourth flowchart of a model training method according to an embodiment of the present application. [Figure 8] 5 is a fifth flowchart of a model training method according to an embodiment of the present application. [Figure 9] FIG. 2 is a structural schematic diagram of a model training device according to an embodiment of the present application; [Figure 10] FIG. 2 is a second structural schematic diagram of a model training device according to an embodiment of the present application. [Figure 11] 3 is a third structural schematic diagram of a model training device according to an embodiment of the present application. [Figure 12] FIG. 2 is a structural schematic diagram of a communication device according to an exemplary embodiment of the present application; [Figure 13] FIG. 2 is a structural schematic diagram of a terminal according to an exemplary embodiment of the present application; [Figure 14] FIG. 2 is a schematic structural diagram of a network side device according to an embodiment of the present application; [Figure 15] FIG. 2 is a second schematic structural diagram of a network side device according to an embodiment of the present application; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] The following clearly describes the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only some of the embodiments of the present application, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application are all within the scope of protection of the present application.

[0020] The terms "first," "second," etc. in the specification and claims of the present application are intended to distinguish between similar objects and are not intended to describe a particular order or sequence. It is to be understood that the terms so used are interchangeable where appropriate, such that the embodiments of the present application may be performed in an order other than that shown or described herein, and that the objects distinguished by "first" and "second" are generally of the same type and do not limit the number of objects, e.g., the first object may be one or more. It is to be noted that "and / or" in the specification and claims represents at least one of the objects connected, and the character " / " generally represents an "or" relationship between the related objects.

[0021] It should be noted that the techniques described in the embodiments of the present application are not limited to Long Term Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be applied to other wireless communication systems, such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA), and other systems. The terms "system" and "network" in the embodiments of the present application are always used interchangeably, and the techniques described may be used in the above-mentioned systems and radio technologies, or in other systems and radio technologies. Although the following description describes a New Radio (NR) system for illustrative purposes and uses NR terminology in most of the description below, these techniques may be applied to systems other than NR systems, such as 6th Generation (6G) communication systems.

[0022] 1 is a block diagram of a wireless communication system to which the embodiment of the present application can be applied. The wireless communication system includes a terminal 11 and a network side device 12. Here, the terminal 11 may be a terminal device such as a mobile phone, a tablet personal computer, a laptop computer (also called a notebook computer), a personal digital assistant (PDA), a palmtop computer, a netbook, an ultra-mobile personal computer (UMPC), a mobile Internet device (MID), an augmented reality (AR) / virtual reality (VR) device, a robot, a wearable device, a vehicle-mounted equipment (VUE), a pedestrian terminal (PUE), a smart home (home equipment having a wireless communication function, such as a refrigerator, a television, a washing machine, or furniture), a game console, a personal computer (PC), a deposit machine, or a self-service machine, and the wearable device may include a smart watch, a smart bracelet, a smart earphone, a smart glasses, a smart accessory (smart bracelet, a smart hand chain, a smart ring, a smart necklace, a smart ankle bracelet, a smart anklet, etc.), a smart band, a smart garment, etc. It should be noted that the terminal 11 in the embodiment of the present application is not limited to a specific type. The network side equipment 12 may include an access network equipment or a core network equipment, where the access network equipment 12 may be called a radio access network equipment, a radio access network (Radio Access Network, RAN), a radio access network function or a radio access network unit.The access network device 12 may include a base station, a Wireless Local Area Network (WLAN) access point, or a Wireless Fidelity (WiFi) node, etc., and the base station may be called a Node B, an evolved Node B (eNB), an access point, a Base Transceiver Station (BTS), a radio base station, a radio transceiver, a Basic Service Set (BSS), an Extended Service Set (ESS), a home B node, a home evolved B node, a Transmitting Receiving Point (TRP), or any other suitable term in the art, and as long as the same technical effect is achieved, the base station is not limited to a specific technical term, and it should be explained that the embodiment of this application takes a base station in an NR system as an example, and does not limit the specific type of the base station. In the following, the technical solution according to the embodiment of this application will be described in detail through several embodiments and their application scenarios in conjunction with the drawings.

[0023] Based on the above-mentioned wireless communication system, as shown in FIG. 2, an embodiment of the present application further provides a structural schematic diagram of a model training system (which may be referred to as a horizontal federation learning system or a horizontal federation model training system), which may include a first network element, a second network element, and a fourth network element.

[0024] The first network element may be a network element or device that can provide machine learning functions in a communication network as a coordinator in the model training system, such as a network element or device dedicated to providing network intelligence services, such as a network data analytics function (NWDAF), a management data analytics service (MDAS), and a management data analytic function (MDAF). Or, the first network element may be a network element or device that provides other communication-related services (such as a mobility management (MM) service and a session management (SM) service) and simultaneously has an intelligence function, such as an access and mobility management function (AMF), a session management function (SMF), an application function (AF), etc., and the AF may be an AF deployed by a telecommunications carrier or a third-party AF. In this application, the first network element may further accept model training applications proposed by other network elements (eg, a fourth network element) or devices.

[0025] The second network element may be a network element or device that has local data but does not want to share data as a training entity in the model training system, and the network element or device has local machine learning capabilities. In this case, the second network element may be, but is not limited to, an AI function network element in a Radio Access Network (RAN) domain, an AI function network element in a Core Network (CN) domain, a third-party AI application, an AI client in a User Equipment (UE), a local communication service device, etc. It should be noted that, for ease of understanding, only two second network elements are illustrated in FIG. 2, but in practice, the second network element may be one or more.

[0026] The fourth network element may be a network element or device that requires model consumption in a wireless communication system, such as a third-party AI application, a UE, a network side device, etc., as a model consumer in the model training system. As can be understood, the model training system may include the fourth network element (e.g., the federated model training process may be triggered by the fourth network element) as shown in FIG. 2, or may not include a fourth network element (e.g., the federated model training process may be triggered by the first network element according to its own need, etc.), etc. That is, the model training system may include more or less network elements or devices than those shown in FIG. 2, and is not limited here.

[0027] The following describes in detail the technical solutions according to the embodiments of the present application through several embodiments and their application scenarios in conjunction with the drawings.

[0028] As shown in Figure 3, a flowchart of a model training method 300 according to an exemplary embodiment of the present application, the method 300 may be performed by a first network element (such as a terminal or a network side device), but is not limited thereto, and may be specifically performed by hardware and / or software installed in the first network element. In this embodiment, the method 300 may at least include the following steps:

[0029] At S310, when the first network element performs a federation model training process corresponding to the model training task, it sends a federation model training request message to at least one second network element.

[0030] Here, the model training task may be an AI model training task corresponding to data analysis, an AI model training task corresponding to communication robustness learning, an AI model training task for model training on the movement trajectory of a UE, etc., and is not limited thereto.

[0031] The federated model training process may be triggered by the first network element based on a model training task required by it, or may be triggered by a fourth network element, for example, the fourth network element may send a model request message corresponding to a model training task required by it to the first network element to trigger the federated model training process.

[0032] It should be noted that the federation model training process may be understood as a horizontal federation learning process, the essence of which is the cooperation of samples, which is applied to a scenario in which participants have the same service type but serve different customers, i.e., features (i.e., sample features) overlap a lot, but users (i.e., samples) overlap less, such as the same service (e.g., MM service or SM service, i.e., sample features are the same) for different users or different user groups (e.g., UE or UE group, i.e., samples are different) in base station equipment with different RAN domains. Based on this, the present application increases the sample dimension corresponding to the training data by cooperation of the same sample features (which may also be referred to as data features) of different samples of model training participants (i.e., second network elements), thereby training to obtain one better model.

[0033] In this case, the federated model training request message sent by the first network element to each of the second network elements requests at least one of the second network elements to participate in the federated model training process (i.e., the at least one second network element is a network element participating in the federated model training process, and the samples used in the federated model training between each of the second network elements are different but have the same sample features), for example, to perform local distributed model training, so that each of the second network elements can cooperate to realize local distributed model training without sharing raw data (or training data) on each of the second network elements, thereby ensuring data privacy and data security in each of the second network elements, and also ensuring the effect of model training.

[0034] Of course, the manner of determining the second network element participating in the federation model training process may be various. In one implementation manner, the second network element may be a network element obtained by the first network element from a Network Repository Function (NRF) or a Unified Data Management Entity (UDM) based on the model training task, and can support the federation model training process. For example, the first network element queries the NRF to obtain AMF instances (instance(s)) participating in the current federation model training process. Also, for example, the first network element queries the UDM to obtain UE instances (instance(s)) participating in the current federation model training process, etc., and is not limited here.

[0035] At S320, the first network element receives first information including at least a first training result transmitted by the at least one second network element.

[0036] As can be understood, after receiving the federation model training request message, each of the second network elements may perform model training based on the federation model training request message to obtain a first training result, and send the first training result to the first network element through first information (i.e., the first information includes at least the first training result). Here, when the second network element performs model training based on the federation model training request message, the second network element may perform model training based on a locally pre-configured initial model, may perform model training based on a local training model that has already completed training before receiving the federation model training request, or may perform model training based on a first model indicated by the federation model training request message, and is not limited thereto.

[0037] Furthermore, the first training result corresponds to training data for training a federation model in the second network element, and the samples corresponding to the training data used in the federation model training process between each of the second network elements are different, but the sample features are the same; for example, the training data in the second network element A corresponds to sample S1, the training data in the second network element B corresponds to sample S2, and the training data in the third network element C corresponds to sample S3, and the sample features of sample S1, sample S2, and sample S3 used in this federation model training process are all the same.

[0038] At S330, the first network element performs model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0039] Here, the first model may be an initial model (which may also be called an initialized global model) generated by the first network element for the model training task, or the first model may be an intermediate model obtained by the first network element performing one or more model training / updates on the initial model, and is not limited thereto.

[0040] The process of the first network element training a model based on the first training result is not limited here, for example, the first network element may update the first model based on the first training result reported by the at least one second network element. It should be noted that the first training results reported by each of the second network elements all correspond to the first model, i.e., each of the second network elements trains on the first model, reports the first training result obtained by training to the first network element, and the first network element performs model training / update, etc. on the first model based on each of the first training results to obtain a target model. As can be understood, when the first network element trains on the first model to obtain the target model, it may undergo one or more federated model training processes in between, and for example, when the target model does not converge, it may continue to execute the federated model training process with the target model as the intermediate model (i.e., as the first model again).

[0041] The second training result may include at least gradient information of the loss function of the target model with respect to the target model parameters. Here, the loss function of the target model may be configured for the target model in advance by the first network element. Since the target model may be only an intermediate model that has not converged, the second training result may correspond to a target model with a different training degree (or number of trainings), and the corresponding second training result may also be different, and is not limited here.

[0042] In this embodiment, the first network element uses a horizontal federation learning method to cooperate with each second network element to realize local distributed model training, thereby ensuring data privacy and data security in each second network element without sharing data on each second network element in the communication network, and also ensuring the model training effect.

[0043] As shown in Figure 4, a flowchart of a model training method 400 according to an exemplary embodiment of the present application, the method 400 may be performed by a first network element (such as a terminal or a network side device), but is not limited thereto, and may be specifically performed by hardware and / or software installed in the first network element. In this embodiment, the method 400 may at least include the following steps:

[0044] At S410, when the first network element performs a federation model training process corresponding to the model training task, it sends a federation model training request message to at least one second network element.

[0045] Here, the at least one second network element is a network element participating in the federation model training process, and samples corresponding to training data used in the federation model training process between each of the second network elements are different, but the sample features are the same.

[0046] As can be understood, the implementation process of S410 can refer to the relevant description in method embodiment 300, and in one implementation manner, the federation model training request message may include at least one of the following (11)-(19):

[0047] (11) Model instance identifier information (model instance ID), which is used to uniquely identify the first model, intermediate model, or target model generated in the first network element (i.e., the model instance identifier information corresponds to the first model, intermediate model, or target model), so as to facilitate identification, invocation, consumption, etc. of the first model, intermediate model, or target model.

[0048] In this embodiment, the model instance identifier information may be assigned by the first network element, for example, when the first network element decides to perform a federated model training process corresponding to a model training task, the first network element correspondingly assigns model instance identifier information for the model training task to indicate a model entity obtained by training through the federated model training process, i.e., for example, the first model, or an intermediate model, or the target model.

[0049] Based on this, the model instance identifier information corresponds to or includes at least one of the following (111)-(114):

[0050] (111) Related information of the first network element, such as the identity (ID) information of the first network element, fully qualified domain name (FQDN) information, name information, address information, etc.

[0051] (112) A first time for indicating that the model training task is performed based on training data generated within the first time.

[0052] Alternatively, the first time and the subsequent second and third times may be a single time period or point in time, for example, the first time may be October 11, 2021.

[0053] It should be noted that the first time, the second time, and the third time may be relative times or absolute times, and there is no limitation thereto here.

[0054] (113) A second time for indicating the completion time of the federated model training process.

[0055] (114) Related information of the second network element, such as name information, FQDN information, identifier information, address information, etc. of the second network element.

[0056] (12) Type information of the model training task, which may include an analytics ID, a model ID, etc., to describe which machine learning task (i.e., a model training task) the federated model training process is training a model for.

[0057] For example, the string "analytics ID / model ID=UE mobility" may be used to indicate that the UE movement trajectory AI model needs to be obtained by performing model training on the UE movement trajectory. It should be noted that the type information may be expressed using a number representation method or other coding representation method in addition to the above-mentioned string type representation method, and is not limited thereto.

[0058] (13) Identifier information of the model training task. Similar to the type information of the model training task, the identifier information of the model training task may include an analytics ID, a model ID, etc., to describe which machine learning task (i.e., model training task) the federated model training process is to train a model for.

[0059] For example, the string "analytics ID / model ID=UE mobility" may be used to indicate that the UE mobility trajectory AI model needs to be obtained by performing model training on the UE mobility trajectory. It should be noted that the identifier information may be expressed using a number representation method or other coding representation method in addition to the above-mentioned string type representation method, and is not limited thereto.

[0060] As can be understood, in order to avoid information redundancy, the federation model training request message may include any one of identifier information of the model training task and type information of the model training task.

[0061] (14) First instruction information for indicating that the federation model training process is a single horizontal federation learning process. Here, the first instruction information may be indicated in an implicit or explicit manner. For example, when indicated in an implicit manner, the second network element may determine that the federation model training process is a single horizontal federation learning process, etc., based on information that the number of network elements participating in the federation model training process included in the federation model training request message is multiple, and is not limited thereto.

[0062] (15) First filter related information, which may be understood as model training filtering information, is used to define at least one of a target object (e.g., target UE(s)), a target time (e.g., target time period), and a target area (e.g., target area of ​​interest (target AOI)) corresponding to the model training task, so that each second network element can collect appropriate training data based on the first filter related information to perform local model training.

[0063] For example, the target object may be a serving object or a correspondingly managed object of the second network element, and the target area may be a service area or a correspondingly managed area of ​​the second network element, without any limitation thereto.

[0064] (16) Model training configuration information. This model training configuration information is used to arrange a model training method and / or training parameters for the horizontal federation learning process. That is, the model training configuration information is used to instruct the second network element how to perform local model training, the model algorithm or the local training model to be adopted, etc. For example, the second network element may train a local training model based on the model training configuration information.

[0065] In one implementation, the model training configuration information is determined by the first network element for the model training task or by the federated model training process. Optionally, the model training configuration information may include at least one of the following (161)-(166):

[0066] (161) Model structure information. The model structure information is used to indicate the specific model structure (e.g., neural network, deep neural network (DNN), convolutional neural network (CNN), cycle neural network (RNN), linear structure, etc.) used in the horizontal federation learning training process, so that each training entity uses the model structure information to train the model.

[0067] In addition, for the neural network model structure, the model structure information may sub-indicate the layers of the neural network, the neuron nodes, and the relationship between the inputs and outputs of each layer, etc. Of course, each of the second network elements may determine the model structure that needs to be used in this federation model training process according to the model structure information, and there is no limitation here.

[0068] (162) Model hyperparameter information. The model hyperparameter information is a configuration variable external to the target model, generally used in the training process of model parameters, generally may be directly specified by an implementer, and is set, for example, before starting the federated model training process, and is a parameter for defining a higher-level concept related to the model, such as complexity or learning ability.

[0069] Optionally, the model hyperparameter information may include an algorithm employed in model training, the number of trees or tree depth, the number of latent factors in matrix decomposition, a learning rate, the number of hidden layers in a deep neural network, the number of clusters in a k-means cluster, a loss function corresponding to the target model, a predetermined value (corresponding to the loss function) for determining whether the target model converges, etc., but are not limited thereto.

[0070] It should be noted that in addition to the first network element specifying the model hyperparameters, the second network element may determine the model hyperparameter information by itself, that is, if the federated model training request message does not include the model hyperparameter information, each of the second network elements may determine the model hyperparameter information by itself.

[0071] (163) Type information of training data in the federation model training process. Here, the training data may be understood as input data and / or output data in the model training process, so as to recommend to the second network element the type of input data and / or the type of output data used in the current federation model training process. Here, the type of input data may be sample features to be trained, etc., and is not limited here.

[0072] Exemplarily, when the second network element is two AMFs (AMF1 and AMF2), the training data type information is used to recommend two input data such as UE location and AMF load to be used for the current federation model training to the two AMFs to train the AMF local training model.

[0073] (164) Model training condition information for instructing the second network element of conditions under which the first network element starts model training (or model updating, calculation, etc.) based on the first training result reported by the second network element.

[0074] In one implementation, the model training condition information may include at least one of the following (a) and (b):

[0075] (a) a third time, where the third time is used to instruct the first network element to start model training when a waiting time for feeding back the first training result to the second network element reaches the third time.

[0076] For example, assuming that there are five second network elements participating in the federation model training process, and the waiting time for the first network element to wait for feedback of the first training results from the five second network elements reaches the third time, the first network element may attempt to start updating the first model based on the currently received first training results sent by the three second network elements.

[0077] (b) a first threshold, which is used to instruct the first network element to start model training when the number of the first training results received reaches the first threshold.

[0078] For example, assuming that there are five second network elements participating in the federated model training process and the first threshold is four, when the number of the first training results received by the first network element is four, the first network element may attempt to start updating the first model based on the currently received four first training results.

[0079] Also for example, assuming that there are five second network elements participating in the federation model training process and a first threshold is 80%, the number of the first training results received by the first network element is four, and when it reaches 80% (4 / 5), the first network element can attempt to start updating the first model based on the currently received four first training results.

[0080] (165) is model training times information for indicating the number of times of local model training that the second network element needs to perform before sending the first information to the first network element, that is, the model training times information is used to indicate to the second network element the number of times information that the second network element needs to perform local model training before returning one first training result to the first network element.

[0081] It should be noted that the model training count information may further include information such as the total number of training iterations (or total number) planned in this federation model training process.

[0082] (166) Model training time length information for indicating a time length of local model training that the second network element needs to perform before transmitting first information to the first network element, that is, the model training time length information is used to indicate to the second network element available time length information for which the second network element can perform local model training before returning a first training result to the first network element once.

[0083] As can be understood, the model training number information and model training length information described in (165)-(166) above may be determined by the second network element based on the training capability information (e.g., computing capability, data storage capability, and time length information required for each round of training iteration) of each of the second network elements. That is, for different second network elements, the model training number information and / or the model training length information may be different.

[0084] For example, taking the model training times as an example, the first network element determines the model training times information of the second network element according to the training ability information of the second network element, thereby ensuring that the second network element with weak training ability is trained relatively less times, but the second network element with strong training ability is trained relatively more times, and further synchronizes the pace at which each second network element participating in the federation model training process returns the first training result, that is, the first network element obtains the first training result reported by each second network element almost simultaneously, and performs subsequent model training and update operations.

[0085] In addition, the training capability information of the second network element may be obtained from the NRF by the first network element, which is not limited here.

[0086] (17) Report information of the first training result for limiting information such as a report format and report conditions when each second network element feeds back the first training result.

[0087] Here, the report format may be understood as model framework information, for example, a representation method based on TensorFlow or Pytorch model framework, or a model representation method based on a framework (e.g., ONNX), or the report format may be used to indicate at least one of the entire model after the second network element returns to local training, all model parameters after local training, or model parameters / gradient information changed after local training, etc.

[0088] The reporting conditions may include an event trigger and / or a periodic trigger, where the event trigger may include any one of (171)-(173).

[0089] (171) When the number of training rounds (or times) that the second network element trains the local model reaches a predetermined value, the second network element reports a first training result, where the number of training rounds specified by the first network element for different second network elements may be different, so that the first network element can align the pace of the first training results reported by each training entity (i.e., the second network element) in the model training iteration process, and further ensure that the reporting pace of each training entity is consistent, thereby preventing the appearance of the dropout problem and ensuring the model training effect.

[0090] (172) The second network element reports the first training result before the reporting time reaches the longest waiting time (or feedback deadline). As can be understood, the longest waiting time may be understood as the time length information of the model training, i.e., the second network element may report the first training result before the reporting time reaches the time length information of the model training.

[0091] (173) When the local model converges (the loss function corresponding to the local model reaches a preset value), the second network element reports the first training result.

[0092] The periodic trigger means that the second network element can report the first training result periodically, such as once every 5 minutes.

[0093] (18) Related information of each network element participating in the current federation model training process, for indicating to a second network element other second network elements and / or first network elements participating in the current federation model training process, etc. Optionally, the related information of each network element may include network element type information, identifier information, or address information, etc. of each of the second network elements, and network element type information, identifier information, or address information, etc. of the first network element.

[0094] (19) Related information of the first model for each of the second network elements to perform local model training, thereby avoiding the process of the second network elements training a model from scratch, and further improving the efficiency of model training.

[0095] Optionally, the related information of the first model may be model identifier information, model structure information, model parameter information, etc. of the first model, and is not limited thereto. In addition, in the federated model training process, when the first federated model training is performed, the related information of the first model is the related information of the initial model, and when the subsequent second, third, etc. federated model training is performed, the related information of the first model is the related information of the intermediate model.

[0096] It can be understood that which information among the above (11)-(19) is specifically included in the federation model training request message may be determined by a protocol, configured by an upper layer, or configured by a network side, and is not limited herein. In addition, when there are multiple second network elements participating in the federation model training process, the contents in the federation model training request message for each of the second network elements may be partially the same or partially different, and is not limited herein.

[0097] In addition to the above-mentioned implementation manner, in this embodiment, when the first network element determines whether to perform a federated model training process corresponding to a model training task, it may be triggered by the first network element itself or the fourth network element, and may also need to first determine whether a first condition is met or not, and then determine whether to perform a federated model training process according to whether the first condition is met or not, for example, when determining that a first condition is met, the first network element determines to perform a federated model training process corresponding to a model training task. Of course, the first condition may be understood as an element or trigger condition for determining that the first network element needs to perform a federated model training process, etc.

[0098] Based on this, the first condition may include at least one of the following (21)-(23).

[0099] (21) The first network element does not store or cannot obtain all or part of the training data corresponding to the model training task. For example, due to data security or data privacy issues, the first network element may lack some or all of the training data of each training entity (i.e., the second network element), so the first network element needs to use a federated model training process to cooperate with each second network element to perform local distributed model training.

[0100] (22) The at least one second network element may provide all or a portion of the training data corresponding to the model training task.

[0101] (23) The samples of training data between the second network elements corresponding to the model training task are different, but the sample features are the same. For example, in order to analyze the behavior trajectory rules of all users in the A region (which has multiple base stations in charge), it is necessary to obtain the trajectory model of users in the A region, and the training data required for training the trajectory model may be location data (i.e., sample features) generated by different users (i.e., samples) under different base stations.

[0102] As can be understood, which of the above (21)-(23) the first condition includes may be agreed upon by a protocol, may be configured by a higher layer, or may be configured by the network side, and is not limited herein.

[0103] At S420, the first network element receives first information sent by the at least one second network element.

[0104] Here, the first information includes at least a first training result, and the first training result corresponds to training data used in the federation model training process in the second network element, i.e., the first training result is obtained based on training data in the second network element.

[0105] It can be understood that the implementation process of S420 can refer to the relevant description in method embodiment 300. In one possible implementation manner, for each of the second network elements, after receiving the federation model training request message sent by the first network element, the second network element may perform model training to obtain a local training model, and then perform calculations based on the local training model to generate the first training result, and then send it to the first network element.

[0106] Exemplarily, when the second network element performs model training based on a federation model training request message to obtain a local training model, the process of each of the second network elements obtaining a first training result may include the following (31)-(35), and the content is as follows:

[0107] (31) The second network element determines that a federation model training process needs to be performed based on the first instruction information. As can be understood, the first instruction information may be obtained from a federation model training request message sent by the first network element, or the first instruction information may be obtained directly from the first network element, and is not limited thereto.

[0108] (32) The second network element determines a model training task corresponding to the federated model training process based on the type information of the model training task or the identifier information (e.g., analytics ID or model ID) of the model training task. For example, the second network element can determine that model training needs to be performed for an application of a user to obtain a service experience model based on a string "analytics ID / model ID=service experience".

[0109] (33) The second network element obtains training data based on the model training task and the association information of the first filter.

[0110] For example, the second network element may determine to collect location data of all UEs in cell 1 of the service corresponding to the second network element during the 07:00-09:00 time slot every Monday as training data based on the relevant information of the model training task and the first filter, and the collected data type may be configured within a training entity or determined based on the type information of the training data included in the federation model training request message, and is not limited thereto.

[0111] (34) The second network element determines a local training model and a model training scheme to be used for training the local training model based on the model training configuration information and / or the related information of the first model.

[0112] For example, the second network element may determine a training manner of a local training model, such as a model structure, an algorithm, a hyperparameter, etc., based on the model training configuration information. Also for example, the second network element determines that the first model is the local training model based on related information of the first model.

[0113] Of course, in addition to determining the local training model based on the related information of the first model, the local training model may be a model pre-configured in the second network element, or a model obtained by the second network element from other network elements, or a model obtained by the second network element through local model training, etc., and is not limited thereto herein.

[0114] (35) The second network element trains the local training model based on the model training scheme, and calculates the training data based on the trained local training model to obtain a first training result.

[0115] For example, assuming that the second network element is a base station, the i-th base station will obtain a cell model after t rounds of model training. JPEG2025504349000002.jpg54 is obtained from this model JPEG2025504349000003.jpg54 is reported to the first network element as the first training result, and similarly, the i+1th base station reports the cell model after performing t rounds of model training. JPEG2025504349000004.jpg56, model JPEG2025504349000005.jpg56 is reported to the first network element as a first training result.

[0116] Of course, it should be noted that the first training result may be a model (i.e., model information of the second model) obtained by training the second network element, or may be first gradient information obtained by training the second network element based on a local training model, and the first gradient information may be a gradient calculation value of the loss function of the second model obtained by training the second network element with respect to the second model parameters.

[0117] Alternatively, in one implementation, the first information may further include model instance identifier information, which is used by the first network element to perform model association, for example, the first network element associates first training results obtained by the same first model reported by different second network elements based on the model instance identifier information, thereby performing model training / updating for the first model.

[0118] As can be appreciated, the model instance identifier information reported by the second network element may be determined based on the model instance identifier information in the federation model training request message.

[0119] At S430, the first network element performs model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0120] It can be understood that the implementation process of S430 can refer to the relevant description in embodiment 300 of the above-mentioned method. In one possible implementation manner, based on the difference of the above-mentioned model training condition information, the process of the first network element performing model training based on the first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result may be different. The following describes the model training process of the first network element by combining method 1 and method 2, and the content is as follows:

[0121] Method 1: Assuming that all second network elements have reported first training results or the waiting time for the first network element to feed back the first training results to the second network element has reached the third time or the number of first training results of the second network element feedback received by the first network element has reached the first threshold, the first network element updates the first model based on the first training results reported by each of the second network elements.

[0122] For example, when the first training result is first gradient information corresponding to a second model, the first network element performs a weighted average on the received first gradient information to obtain an average model gradient calculation result, and uses the average model gradient calculation result to update the first model.

[0123] Also, for example, when the first training result is model information of a second model, the model information of the second model reported by the i-th second network element received by the first network element is JPEG2025504349000006.jpg54, model information of the second model reported by the i+1th second network element Assuming that the image is JPEG2025504349000007.jpg56, the first network element calculates the following equation (1): JPEG2025504349000008.jpg54 and The first model can be updated based on JPEG2025504349000009.jpg56.

[0124]

number

[0125] Method 2: Each time the first network element receives a first training result reported by one second network element, it updates the first model based on the first training result.

[0126] For example, the first training result reported by the i-th second network element received by the first network element is JPEG2025504349000011.jpg54, the first network element is The first model can be updated based on JPEG2025504349000012.jpg54.

[0127]

number

[0128] As can be understood, after the first network element trains to obtain the target model, the second training result may be a gradient value of the loss function of the target model with respect to the parameters of the target model, i.e., the second training result includes at least second gradient information of the loss function of the target model with respect to the parameters of the target model.

[0129] Of course, if the calculation result of the loss function of the target model does not meet the predetermined requirement (for example, the accuracy value of the loss function does not reach a predetermined value), the first network element can perform the federation model training process again until the calculation result of the loss function of the target model trained by the first network element meets the predetermined requirement. As can be understood, when the calculation result of the loss function of the target model does not meet the predetermined requirement, the target model may be understood as an intermediate model, that is, the second training result is the gradient value of the loss function of the intermediate model with respect to the current parameters of the intermediate model.

[0130] In one implementation, the first network element can resend second information to the at least one second network element when performing a federation model training process again, the second information including at least the second training result, so that each second network element retrains a local training model based on the second training result to re-acquire the first training result.

[0131] For example, the first network element issues the same second training result corresponding to a different second network element. For example, the first network element sends second gradient information of the loss function of the target model with respect to the current parameters of the target model to RAN1 (i.e., the second network element) and RAN2, respectively. Correspondingly, after receiving the second training result, the RAN1 and the RAN2 can respectively adjust local model parameters according to the second training result to obtain a parameter-adjusted local model, and then re-obtain the first training result according to the parameter-adjusted local model.

[0132] In another implementation manner, when the first network element performs the federated model training process again, the data of different training entities may be different, or different training entities may have different data for model learning, or a training entity may go offline, have too high an extension, and so on, and thus is no longer suitable for the current federated model training process, it may be replaced with another training entity in the next round iteration, that is, the training entity participating in the federated model training process may be at least partially different from the training entity participating in the federated model training process in the previous round.

[0133] That is, the first network element may re-determine the network elements, such as at least one third network element, participating in the federated model training process when performing the next round of federated model training, and after the first network element determines the at least one third network element participating in the current round of federated model training, it sends second information to the at least one third network element, where the second information at least includes the second training result, and the second training result is used by the third network element to perform local model training to obtain a third training result.

[0134] As can be understood, for the selection of the third network element, please refer to the above-mentioned relevant description of the selection process of the second network element, and for the realization of the third training result, please refer to the above-mentioned relevant description of the first training result, and in order to avoid repetition, no further description will be given here.

[0135] It should be noted that the at least one third network element may be at least partially the same as or different from the at least one second network element, but considering that the third network element participates in a federated model training process for the first time, in order to ensure that the third network element can smoothly complete the training of a local model, the second information sent to the third network element may include the target model information or second gradient information, and may further include some or all of the information in the federated model training request other than the related information of the first model.

[0136] Therefore, in a scenario in which the first network element transmits the second information to the second network element, or in a scenario in which the third network element is the same as the second network element and the first network element transmits the second information to the third network element, the first network element can adopt incremental transmission or full transmission when transmitting the second information.

[0137] Here, the so-called incremental transmission method refers to sending, for each of the second network elements, specific information belonging to the second information and different from the information included in the federation model training request message to the second network element.

[0138] For example, if the training entity has not changed compared with the iterative process of the previous round, the second information may not include other information repeated with the federation model training request message, but may include updated target model information and / or second training results, that is, the first network element only needs to issue the updated target model information and / or second training results without repeatedly issuing the other information in the federation model training request message, thereby avoiding the problem of resource waste or poor transmission reliability caused by too large data.

[0139] The so-called full transmission method means that, for each of the second network elements, all the information in the second information is transmitted to the second network element.

[0140] For example, if the training entity has changed compared to the previous round of the iterative process, the second information may include updated target model information and / or second training results, and may also issue the other information in the federated model training request message, such as model instance identifier information, first training result report information, model training configuration information, related information of each network element participating in the federated model training process, etc., to the newly selected training entity.

[0141] That is, in addition to the second training result, the second information may further include at least one of the following (51)-(55).

[0142] (51) Model instance identifier information for the second network element to perform model association.

[0143] (52) Information of the target model for instructing the second network element or the third network element to perform local model training based on the target model. Optionally, the information of the target model may include model structure information, model hyperparameter information, etc., but is not limited thereto.

[0144] (53) This is report information on the results of the first training exercise.

[0145] (54) Related information of each network element participating in the federation model training process.

[0146] (55) Model training configuration information corresponding to the target model.

[0147] Here, for the description of (51)-(55), please refer to the above related description in the federation model training request message, and in order to avoid repetition, no further description will be given here.

[0148] It should be noted that in the above-mentioned federated model training process, the first network element can determine whether to repeat the above-mentioned federated model training process until the target model obtained in the first network element converges according to whether the target model converges, where the so-called target model converges means that the accuracy value of the loss function corresponding to the target model meets the budget requirement or reaches a predetermined value.

[0149] As shown in Figure 5, a flowchart of a model training method 500 according to an exemplary embodiment of the present application, the method 500 may be performed by a first network element (such as a terminal or a network side device), but is not limited thereto, and may be specifically performed by hardware and / or software installed in the first network element. In this embodiment, the method 500 may at least include the following steps:

[0150] At S510, the first network element receives a model request message sent by a fourth network element.

[0151] Here, the model request message includes at least one of the following (61)-(64).

[0152] (61) Type information of the model training task.

[0153] (62) Identifier information of the model training task.

[0154] Here, the type information of the model training task and the identifier information of the model training task can be referred to the above-mentioned related description of the type information of the model training task and the identifier information of the model training task in the federation model training request message, and in order to avoid repetition, they will not be described further here.

[0155] (63) Related information of a second filter for limiting at least one of a target object, a target time, and a target area corresponding to the model training task.

[0156] (64) Model feedback related information, wherein the model feedback related information may include at least one of a model feedback format and a feedback condition.

[0157] Here, the model feedback format may be understood as model framework information, for example, a representation method based on TensorFlow or Pytorch model framework, or a model representation method based on a framework (for example, ONNX), etc.

[0158] The feedback conditions may include an event trigger and / or a periodic trigger, where the event trigger includes at least one of (641)-(643).

[0159] (641) When the number of training rounds (or times) that the first network element trains the target model reaches a predetermined value, the first network element feeds back the target model.

[0160] (642) Before the training time reaches the longest latency time, the first network element feeds back a target model.

[0161] (643) When the target model converges (the loss function corresponding to the target model reaches a preset value and a corresponding preset value), the first network element feeds back the target model, and so on.

[0162] The periodic trigger means that the first network element can feed back the target model to the fourth network element periodically, such as feeding back the target model once every 5 minutes.

[0163] Of course, in addition to the above feedback conditions, the model feedback information may instruct the first network element to perform feedback of the target model according to methods such as batch feedback, multiple continuous update feedback, feedback deadline, periodic feedback, etc., and are not limited thereto.

[0164] It should be noted that the federation model training request message sent by the first network element to each of the second network elements may be determined by the first network element based on the model request message, for example, the type information of the model training task and the identifier information of the model training task in the federation model training request message may be the same as the type information of the model training task and the identifier information of the model training task in the model request message, and for example, the related information of the first filter in the federation model training request message may be the same as or different from the related information of the second filter in the model request message, etc., which is not limited here.

[0165] At S520, when the first network element performs a federation model training process corresponding to the model training task, it sends a federation model training request message to at least one second network element.

[0166] Here, the at least one second network element is a network element participating in the federation model training process, and samples corresponding to training data used in the federation model training process between each of the second network elements are different, but the sample features are the same.

[0167] At S530, the first network element receives first information sent by the at least one second network element, the first information including at least a first training result, the first training result corresponding to training data used in the federation model training process in the second network element.

[0168] At S540, the first network element performs model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0169] As can be understood, the implementation process of S520-S540 can be referred to the relevant descriptions in the above method embodiments 300 and / or 400, and will not be further described here to avoid repetition.

[0170] Of course, in one implementation manner, when the calculation result of the loss function of the target model meets a predetermined requirement (for example, the accuracy value of the loss function reaches a predetermined value, i.e., the target model converges), the first network element can send related information of the target model to the fourth network element, so that the fourth network element performs corresponding tasks based on the target model.

[0171] Here, the relevant information of the target model includes at least one of the following (71)-(74).

[0172] (71) The model instance identifier information, which can be referred to the above description of the model instance identifier information, and will not be further described here to avoid repetition.

[0173] (72) The target model information, where the target model information includes at least one of the following (721)-(726).

[0174] (721) Model structure information. Here, the model structure information is used to indicate what kind of model structure of the target model is specifically (for example, neural network, deep neural network (DNN), convolutional neural network (CNN), cycle neural network (RNN), linear structure, etc.). For the neural network model structure, the model structure information may indicate the layers of the neural network, neuron nodes, and the relationship between the input and output of each layer in a more detailed manner.

[0175] (722) Model parameter information. Here, the model parameter information is a configuration variable inside the target model to define the function of the target model, and may be obtained by a manner of data estimation or model training. For example, the model parameters may include weights in an artificial neural network, support vectors in a support vector machine, coefficients in linear regression or logistic regression, etc.

[0176] (723) Model algorithm information. For example, the model algorithm information may include, but is not limited to, a policy tree, a Bayesian classifier, a K-nearest neighbor, a support vector machine, etc.

[0177] (724) Model hyperparameter information, where the model hyperparameter information is a configuration variable external to the target model, generally used in the training process of model parameters, and generally may be directly specified by an implementer. For example, the model hyperparameter information may include a learning rate of a training neural network, C and sigma hyperparameters of a support vector machine, a parameter k in K-nearest neighbors, etc.

[0178] (725) Type information of model input data.

[0179] (726) Type information of model output data.

[0180] (73) Second indication information for indicating that the target model is a horizontal federation learning model or for indicating that the model is generated by a horizontal federation learning process. Optionally, the second indication information may be an explicit indication or an implicit indication. For example, when the second indication information is implicit indication information, the implicit indication may be realized by model instance identifier information or related information of a second network element.

[0181] (74) Related information of the second network element and / or the third network element, such as network element type information, identifier information, FQDN information, address information, etc. of the second network element and the third network element.

[0182] In this embodiment, when the first network element receives a model training request sent by the fourth network element, it triggers a federated model training process, and further cooperates with each of the second network elements to perform local distributed model training to obtain a target model, thereby ensuring the security and privacy of data on the second network element and meeting the model training demand on the fourth network element.

[0183] Based on the description of the above method embodiments 200-500, the following further illustratively describes the model training method provided by the present application in conjunction with Figures 6a and 6b, taking the federation model training process as an example that is triggered by the fourth network element, and the content is as follows: Here, it is assumed that there are two second network elements participating in the federation model training process.

[0184] Example 1 As shown in FIG. 6a, the model training process given by Example 1 is as follows.

[0185] In S601, a fourth network element sends a model request message to the first network element.

[0186] At S602, the first network element determines to perform a federation model training process based on the model request message.

[0187] At S603, the first network element obtains two second network elements, such as training entity A and training entity B, that can participate in the federation model training process from the NRF.

[0188] At S604, the first network element sends a federation model training request message to training entity A and training entity B.

[0189] In S605, after receiving the federation model training request message, training entity A trains a local training model based on the federation model training request message, and further obtains a first training result based on the local training model obtained by training.

[0190] Correspondingly, after receiving the federation model training request message, training entity B trains a local training model based on the federation model training request message, and further obtains a first training result based on the local training model obtained by training.

[0191] At S606, training entity A and training entity B each report first information including at least the first training result to the first network element.

[0192] In S607, the first network element performs model training based on the received first training results reported by training entity A and training entity B, to obtain a target model and / or a second training result.

[0193] In S608, the first network element determines whether a calculation result of a loss function corresponding to the target model meets a predetermined requirement.

[0194] In S609, the first network element sends second information including at least the second training result to the training entity A and training entity B if the calculation result of the loss function of the target model does not satisfy a predetermined requirement.

[0195] In S610, training entity A and training entity B re-train the local model based on the received second training result, and re-acquire the first training result.

[0196] As can be understood, the first network element and the training entity A and training entity B repeatedly execute S605-S610, and when the calculation result of the loss function corresponding to the target model meets a predetermined requirement, i.e., the target model converges, the federated model training process is stopped.

[0197] In S611, if the calculation result of the loss function of the target model satisfies a predetermined requirement, the first network element sends related information of the target model to the fourth network element.

[0198] It should be noted that the implementation process of the above S601-S611 can refer to the relevant descriptions in the above method embodiments 300-500 to achieve the same or corresponding technical effects, and will not be further described here to avoid repetition.

[0199] It should be noted that in this embodiment, the federation model training process may include, but is not limited to, the above-mentioned S601-S611, for example, may include more or less steps than the above-mentioned S601-S611, and is not limited thereto.

[0200] Example 2 As shown in FIG. 6b, the model training process given by Example 2 is as follows.

[0201] At S621, the fourth network element sends a model request message to the first network element.

[0202] At S622, the first network element determines to perform a federation model training process based on the model request message.

[0203] At S623, the first network element obtains two second network elements, such as training entity A and training entity B, that can participate in the federation model training process from the NRF.

[0204] At S624, the first network element sends a federation model training request message to training entity A and training entity B.

[0205] In S625, after receiving the federation model training request message, training entity A trains a local training model based on the federation model training request message, and further obtains a first training result based on the local training model obtained by training.

[0206] At S626, training entity A reports first information including at least the first training result to the first network element.

[0207] In S627, the first network element performs model training based on the received first training result sent by the training entity A, to obtain a target model and / or a second training result.

[0208] At S628, the first network element determines whether a calculation result of a loss function corresponding to the target model satisfies a predetermined requirement.

[0209] In S629, the first network element sends second information including at least the second training result to the training entity A if the calculation result of the loss function of the target model does not satisfy a predetermined requirement.

[0210] In S630, training entity A re-acquires and reports the first training result by performing local model training again based on the received second training result, thereby realizing cyclic iteration for the federated model training process.

[0211] In S631, after receiving a federation model training request message, training entity B trains a local training model based on the federation model training request message, and further obtains a first training result based on the local training model obtained by training.

[0212] At S632, each training entity B reports first information including at least the first training result to the first network element.

[0213] In S633, the first network element performs model training based on the received first training result sent by the training entity B, to obtain a target model and / or a second training result.

[0214] At S634, the first network element determines whether a calculation result of a loss function corresponding to the target model satisfies a predetermined requirement.

[0215] At S635, the first network element sends second information including at least the second training result to training entity B if the calculation result of the loss function of the target model does not satisfy a predetermined requirement.

[0216] In S636, training entity B re-acquires and reports the first training result by performing local model training again based on the received second training result, thereby achieving cyclic iteration for the federated model training process.

[0217] As can be understood, the first network element and the training entity A and training entity B repeatedly execute S605-S610, and when the calculation result of the loss function corresponding to the target model meets a predetermined requirement, i.e., the target model converges, the federated model training process is stopped.

[0218] In S637, if the calculation result of the loss function of the target model satisfies a predetermined requirement, the first network element sends related information of the target model to the fourth network element.

[0219] It should be noted that the implementation process of the above S621-S637 can refer to the relevant descriptions in the above method embodiments 300-500 to achieve the same or corresponding technical effects, and will not be further described here to avoid repetition.

[0220] It should be noted that in this embodiment, the federation model training process may include, but is not limited to, the above-mentioned S621-S637, for example, may include more or less steps than the above-mentioned S621-S637, and is not limited thereto.

[0221] As shown in Figure 7, a flowchart of a model training method 700 according to an exemplary embodiment of the present application, the method 700 may be performed by a second network element (such as a terminal or a network side device), but is not limited thereto, and may be specifically performed by hardware and / or software installed in the second network element. In this embodiment, the method 700 may at least include the following steps:

[0222] At S710, a second network element receives a federated model training request message sent by a first network element to request the second network element to participate in a federated model training process corresponding to a model training task.

[0223] At S720, the second network element performs model training based on the federation model training request message to obtain a first training result.

[0224] At S730, the second network element transmits first information including at least the first training result to the first network element.

[0225] Wherein, the first training result corresponds to the training data used in the federation model training process in the second network element, the samples used in the federation model training in the second network element are different from the samples used in the federation model training in the fifth network element, but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federation model training process. For example, assuming that the second network elements participating in model training include network element 1, network element 2, network element 3, and network element 4, and the second network element performing the embodiment 700 of the present method is network element 1, the fifth network element is one or more of network element 2, network element 3, and network element 4, and is not limited here.

[0226] Optionally, the federated model training request message includes at least one of: model instance identifier information corresponding to a target model and assigned by the first network element; type information of the model training task; identifier information of the model training task; first instruction information for indicating that the federated model training process is a single horizontal federation learning process; related information of a first filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task; related information of a first model for each of the second network elements to perform local model training; model training configuration information; report information of the first training result; and related information of each network element participating in the federated model training process.

[0227] Optionally, the model instance identifier information corresponds to at least one of: associated information of the first network element, a first time for indicating that the model training task is performed based on training data generated within the first time, a second time for indicating a completion time of the federated model training process, and associated information of each network element participating in the federated model training process.

[0228] Optionally, the model training configuration information includes at least one of model structure information, model hyperparameter information, type information of training data in the federated model training process, model training condition information for instructing the second network element on a condition under which the first network element starts model training based on the first training result reported by the second network element, model training number information for instructing a number of times of local model training that the second network element needs to perform before sending first information to the first network element, and model training time length information for instructing a time length of local model training that the second network element performs before sending first information to the first network element.

[0229] Optionally, the model training condition information includes at least one of a third time for instructing the first network element to start model training when a waiting time for the first network element to feed back the first training result to the second network element reaches the third time, and a first threshold for instructing the first network element to start model training when a number of received first training results reaches the first threshold.

[0230] Optionally, the reporting information of the first training result includes at least one of a reporting format of the first training result and a reporting condition of the first training result.

[0231] Optionally, the first training result includes model information of a second model obtained by training the second network element based on a local training model and / or first gradient information corresponding to the second model.

[0232] Optionally, the second network element is a network element obtained from a network storage function NRF by the first network element based on the model training task and capable of supporting the federated model training process.

[0233] Optionally, the first information further includes at least one of model instance identifier information for the first network element to perform model association.

[0234] Optionally, the step of the second network element performing model training based on the model training request message and obtaining a first training result includes the steps of: the second network element determining, based on the first instruction information, that a federated model training process needs to be performed; the second network element determining a model training task corresponding to the federated model training based on type information or identifier information of the model training task; the second network element obtaining training data based on the model training task and related information of the first filter; the second network element determining a local training model and a model training scheme used for training the local training model based on the model training configuration information and / or related information of the first model; the second network element training the local training model based on the model training scheme, and calculating the training data based on the trained local training model to obtain a first training result.

[0235] Optionally, the second network element receives second information transmitted by the first network element, the second information including at least a second training result.

[0236] Optionally, the step of the second network element receiving the second information transmitted by the first network element includes any one of the steps of the second network element receiving, by the first network element, specific information belonging to the second information transmitted by the first network element and different from the information included in the federation model training request message, and the second network element receiving all information in the second information transmitted by the first network element.

[0237] Optionally, the second information is transmitted by the first network element when a calculation result of a loss function of a target model does not satisfy a predetermined requirement.

[0238] Optionally, the second training result includes at least second gradient information of a loss function of a target model with respect to parameters of the target model or information of the target model.

[0239] Optionally, the second information further includes at least one of model instance identifier information for the second network element to perform model association, association information of each network element participating in the federated model training process, and model training configuration information corresponding to the target model.

[0240] As can be understood, the implementation process of each implementation method in method embodiment 700 can refer to the relevant descriptions in the above method embodiments 300-600 to achieve the same or corresponding technical effects, and will not be further described here to avoid repetition of description.

[0241] As shown in Figure 8, a flowchart of a model training method 800 according to an exemplary embodiment of the present application, the method 800 may be performed by a fourth network element (such as a terminal or a network side device), but is not limited thereto, and may be specifically performed by hardware and / or software installed in the fourth network element. In this embodiment, the method 800 may at least include the following steps:

[0242] At S810, the fourth network element receives the relevant information of the target model sent by the first network element.

[0243] Here, the related information of the target model is at least used to characterize the target model as a horizontal federation model.

[0244] Optionally, the related information of the target model includes at least one of: model instance identifier information corresponding to a target model and assigned by the first network element; third model information including network structure information and / or model parameter information of the target model; second instruction information for indicating that the target model is a horizontal federation learning model; and related information of a second network element and / or a third network element which are network elements participating in the target model training.

[0245] Optionally, the relevant information of the target model is sent by the first network element when a calculation result of a loss function meets a predetermined requirement.

[0246] Optionally, the method further includes a step of the fourth network element sending a model request message to the first network element, where the model request message includes at least one of type information of the model training task, identifier information of the model training task, related information of a second filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task, and model feedback related information including at least one of a model description scheme and a feedback time.

[0247] As can be understood, the implementation process of each implementation method in method embodiment 800 can refer to the relevant descriptions in the above method embodiments 300-600 to achieve the same or corresponding technical effects, and will not be further described here to avoid repetition of description.

[0248] The execution body of the model training method 300-800 according to the embodiment of the present application may be a model training device. In the embodiment of the present application, the model training device executes the model training method as an example to describe the model training device according to the embodiment of the present application.

[0249] As shown in FIG. 9 , it is a structural schematic diagram of a model training apparatus 900 according to an exemplary embodiment of the present application. When performing a federated model training process corresponding to a model training task, the apparatus 900 includes: a first sending module 910 for sending a federated model training request message to at least one second network element, where the at least one second network element is a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between each of the second network elements are different but have the same sample features; a first receiving module 920 for receiving first information sent by the at least one second network element, where the first information includes at least a first training result, and the first training result corresponds to the training data used in the federated model training process in the second network element; and a first model training module 930 for performing model training according to a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0250] Optionally, the federated model training request message includes at least one of: model instance identifier information corresponding to the target model and assigned by the first network element; type information of the model training task; identifier information of the model training task; first instruction information for indicating that the federated model training process is a single horizontal federation learning process; related information of a first filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task; related information of a first model for each of the second network elements to perform local model training; model training configuration information; report information of the first training result; and related information of each network element participating in the federated model training process.

[0251] Optionally, the model instance identifier information corresponds to at least one of: associated information of the first network element, a first time for indicating that the model training task is performed based on training data generated within the first time, a second time for indicating a completion time of the federated model training process, and associated information of the second network element.

[0252] Optionally, the model training configuration information includes at least one of model structure information, model hyperparameter information, type information of training data in the federated model training process, model training condition information for instructing the second network element on a condition under which the first network element starts model training based on the first training result reported by the second network element, model training number information for instructing a number of times of local model training that the second network element needs to perform before sending first information to the first network element, and model training time length information for instructing a time length of local model training that the second network element needs to perform before sending first information to the first network element.

[0253] Optionally, the model training condition information includes at least one of a third time for instructing the first network element to start model training when a waiting time for the first network element to feed back the first training result to the second network element reaches the third time, and a first threshold for instructing the first network element to start model training when a number of received first training results reaches the first threshold.

[0254] Optionally, the reporting information of the first training result includes at least one of a reporting format of the first training result and a reporting condition of the first training result.

[0255] Optionally, the first training result includes model information of a second model obtained by training the second network element based on a local training model and / or first gradient information corresponding to the second model.

[0256] Optionally, the first information further includes at least one of model instance identifier information for the first network element to perform model association.

[0257] Optionally, the second network element is a network element obtained from a network storage function NRF by the first network element based on the model training task and capable of supporting the federated model training process.

[0258] Optionally, the first model training module 930 is further used to determine that a first condition is met, where the first condition includes at least one of: all or a portion of the training data corresponding to the model training task is not stored or cannot be obtained in the first network element; the at least one second network element is able to provide all or a portion of the training data corresponding to the model training task; and samples of training data between each of the second network elements corresponding to the model training task are different but have the same sample characteristics.

[0259] Optionally, the first sending module 910 is further used for sending second information to the at least one second network element when the calculation result of the loss function of the target model does not meet a predetermined requirement, where the second information at least includes the second training result, and the second training result is used by the second network element to re-train a local model and re-acquire the first training result.

[0260] Optionally, the step of the first sending module 910 sending the second information to the at least one second network element includes one of the following steps: for each of the second network elements, sending to the second network element specific information that belongs to the second information and is different from the information included in the federation model training request message; and for each of the second network elements, sending to the second network element all of the information in the second information.

[0261] Optionally, the first sending module 910 is further used for sending second information to at least one third network element when the calculation result of the loss function of the target model does not meet a predetermined requirement, where the second information at least includes the second training result, and the second training result is used by the third network element to perform local model training to obtain a third training result, and the third network element is a network element that is re-determined by the first network element and participates in the federated model training process.

[0262] Optionally, the second training result includes at least second gradient information of the loss function with respect to parameters of the target model or information of the target model.

[0263] Optionally, the second information further includes at least one of model instance identifier information for the second network element to perform model association, report information of the first training result, association information of each network element participating in the federated model training process, and model training configuration information corresponding to the target model.

[0264] Optionally, the first receiving module 920 is further used for receiving a model request message sent by a fourth network element, where the model request message includes at least one of type information of the model training task, identifier information of the model training task, related information of a second filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task, and model feedback related information including at least one of a model feedback format and a feedback condition.

[0265] Optionally, the first sending module 910 is further used for sending related information of the target model to the fourth network element when the calculation result of the loss function of the target model meets a predetermined requirement, where the related information of the target model includes at least one of the model instance identifier information, the target model information, second indication information for indicating that the target model is a horizontal federation learning model, and related information of the second network element and / or a third network element.

[0266] As shown in FIG. 10 , it is a structural schematic diagram of a model training apparatus 1000 according to an exemplary embodiment of the present application, in which the apparatus 1000 includes: a second receiving module 1010 for receiving a federated model training request message sent by a first network element, for a second network element to request the second network element to participate in a federated model training process corresponding to a model training task; a second model training module 1020 for performing model training based on the federated model training request message to obtain a first training result; and a second sending module 1030 for sending first information including at least the first training result to the first network element, where the first training result corresponds to training data used in the federated model training process in the second network element, the samples used in the federated model training in the second network element are different from the samples used in the federated model training in a fifth network element but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federated model training process.

[0267] Optionally, the federated model training request message includes at least one of: model instance identifier information corresponding to a target model and assigned by the first network element; type information of the model training task; identifier information of the model training task; first instruction information for indicating that the federated model training process is a single horizontal federation learning process; related information of a first filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task; related information of a first model for each of the second network elements to perform local model training; model training configuration information; report information of the first training result; and related information of each network element participating in the federated model training process.

[0268] Optionally, the model instance identifier information corresponds to at least one of: associated information of the first network element, a first time for indicating that the model training task is performed based on training data generated within the first time, a second time for indicating a completion time of the federated model training process, and associated information of each network element participating in the federated model training process.

[0269] Optionally, the model training configuration information includes at least one of model structure information, model hyperparameter information, type information of training data in the federated model training process, model training condition information for instructing the second network element a condition under which the first network element starts model training based on the first training result reported by the second network element, model training number information for instructing a number of times of local model training that the second network element needs to perform before sending first information to the first network element, and model training time length information for instructing a time length of local model training that the second network element needs to perform before sending first information to the first network element.

[0270] Optionally, the model training condition information includes at least one of a third time for instructing the first network element to start model training when a waiting time for the first network element to feed back the first training result to the second network element reaches the third time, and a first threshold for instructing the first network element to start model training when a number of received first training results reaches the first threshold.

[0271] Optionally, the reporting information of the first training result includes at least one of a reporting format of the first training result and a reporting condition of the first training result.

[0272] Optionally, the first training result includes model information of a second model obtained by training the second network element based on a local training model and / or first gradient information corresponding to the second model.

[0273] Optionally, the second network element is a network element obtained from a network storage function NRF by the first network element based on the model training task and capable of supporting the federated model training process.

[0274] Optionally, the first information further includes at least one of model instance identifier information for the first network element to perform model association.

[0275] Optionally, the step of the second model training module 1020 performing model training based on the model training request message and obtaining a first training result includes the steps of: determining, based on the first instruction information, that a federated model training process needs to be performed; determining a model training task corresponding to the federated model training based on type information or identifier information of the model training task; obtaining training data based on the model training task and related information of the first filter; determining a local training model and a model training scheme used for training the local training model based on the model training configuration information and / or related information of the first model; training the local training model based on the model training scheme, and calculating the training data based on the local training model after training, to obtain a first training result.

[0276] Optionally, the second receiving module 1010 is further used for receiving second information sent by the first network element, the second information including at least a second training result.

[0277] Optionally, the step of the second receiving module 1010 receiving the second information transmitted by the first network element includes any one of the steps of receiving specific information transmitted by the first network element that belongs to the second information and is different from the information included in the federation model training request message, and receiving all information in the second information transmitted by the first network element.

[0278] Optionally, the second information is transmitted by the first network element when a calculation result of a loss function of a target model does not satisfy a predetermined requirement.

[0279] Optionally, the second training result includes at least second gradient information of a loss function of a target model with respect to parameters of the target model or information of the target model.

[0280] Optionally, the second information further includes at least one of model instance identifier information for the second network element to perform model association, association information of each network element participating in the federated model training process, and model training configuration information corresponding to the target model.

[0281] As shown in FIG. 11, it is a structural schematic diagram of a model training apparatus 1100 according to an exemplary embodiment of the present application, where the apparatus 1100 includes a third receiving module 1110 for receiving target model related information sent by a first network element, where the target model related information is at least used to characterize the target model as a horizontal federation model.

[0282] Optionally, the related information of the target model includes at least one of: model instance identifier information corresponding to a target model and assigned by the first network element; third model information including network structure information and / or model parameter information of the target model; second instruction information for indicating that the target model is a horizontal federation learning model; and related information of a second network element and / or a third network element which are network elements participating in the target model training.

[0283] Optionally, the relevant information of the target model is sent by the first network element when a calculation result of a loss function meets a predetermined requirement.

[0284] Optionally, the apparatus 1100 further includes a third sending module for sending a model request message to the first network element, where the model request message includes at least one of type information of the model training task, identifier information of the model training task, related information of a second filter for limiting at least one of a target object, a target time and a target area corresponding to the model training task, and model feedback related information including at least one of a model description scheme and a feedback time.

[0285] The model training device 900, 1000, 1100 in the embodiment of the present application may be a communication device, such as a communication device having an operating system, or may be a component in the communication device, such as an integrated circuit or chip. The communication device may be a terminal or a network side device. Exemplarily, the terminal may include, but is not limited to, the types of terminal 11 listed above, and the network side device may include, but is not limited to, the types of network side device 12 listed above, and the embodiment of the present application is not specifically limited.

[0286] The model training apparatuses 900, 1000, 1100 according to the embodiments of the present application can realize each process realized by the embodiments of the methods of Figures 3 to 8 and achieve the same technical effects, and will not be further described here to avoid repetition of description.

[0287] Optionally, as shown in Fig. 12, the embodiment of the present application further provides a communication device 1200, which includes a processor 1201 and a memory 1202, and the memory 1202 stores a program or instruction that can run on the processor 1201, for example, when the communication device 1200 is a terminal, the program or instruction can be executed by the processor 1201 to realize each step of the above-mentioned model training method embodiment 300-800, and achieve the same technical effect. When the communication device 1200 is a network side device, the program or instruction can be executed by the processor 1201 to realize each step of the above-mentioned 300-800 method embodiment, and achieve the same technical effect, which will not be described further here in order to avoid repetition.

[0288] In one implementation, the communication device 1200 may be a terminal, which may include a processor and a communication interface, and the communication interface is coupled to the processor, and the processor is used to run programs or instructions to realize the steps of the method in the method embodiments 300-800. This terminal embodiment corresponds to the above-mentioned terminal-side method embodiment, and each implementation process and implementation manner of the above-mentioned method embodiment can be applied to this terminal embodiment, and the same technical effects can be achieved. Specifically, Figure 13 is a schematic diagram of the hardware structure for realizing the terminal of the embodiment of the present application.

[0289] The terminal 1300 includes at least some components such as, but not limited to, a radio frequency unit 1301, a network module 1302, an audio output unit 1303, an input unit 1304, a sensor 1305, a display unit 1306, a user input unit 1307, an interface unit 1308, a memory 1309, and a processor 1310.

[0290] As can be understood by those skilled in the art, the terminal 1300 may further include a power source (e.g., a battery) for powering each component, and the power source may be logically connected to the processor 1310 by a power management system, so that the power management system can realize functions such as charge / discharge management and power consumption management. The terminal structure shown in FIG. 13 does not constitute a limitation on the terminal, and the terminal may include more or less components than the number of components shown, or a combination of some components, or a different arrangement of components, and will not be further described here.

[0291] It should be understood that in the embodiment of the present application, the input unit 1304 may include a graphics processing unit (GPU) 13041 and a microphone 13042, and the graphics processor 13041 processes image data of still or video images obtained by an image capture device (e.g., a camera) in a video capture mode or an image capture mode. The display unit 1306 may include a display panel 13061, and the display panel 13061 may be configured in the form of a liquid crystal display, an organic light emitting diode, etc. The user input unit 1307 includes at least one of a touch panel 13071 and other input devices 13072. The touch panel 13071 is also called a touch screen. The touch panel 13071 may include two parts: a touch detection device and a touch controller. The other input devices 13072 may include, but are not limited to, a physical keyboard, function keys (e.g., volume control buttons, switch buttons, etc.), a trackball, a mouse, and an operation lever, which will not be described further herein.

[0292] In the embodiment of the present application, the radio frequency unit 1301 can receive downlink data from the network side device, and then transmit the data to the processor 1310 for processing, and can transmit uplink data to the network side device. In general, the radio frequency unit 1301 includes, but is not limited to, an antenna, an amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, etc.

[0293] The memory 1309 may be used to store software programs or instructions and various data. The memory 1309 may mainly include a first storage area in which programs or instructions are stored and a second storage area for stored data, where the first storage area can store an operating system, an application program or instructions required for at least one function (e.g., a sound playback function, an image playback function, etc.), etc. The memory 1309 may include a volatile memory or a non-volatile memory, or the memory 1309 may include both volatile and non-volatile memory. Here, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be Random Access Memory (RAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), Synch link Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). Memory 1309 in embodiments of the present application may include, but is not limited to, these and any other suitable types of memory.

[0294] The processor 1310 may include one or more processing units. Optionally, the processor 1310 may be integrated with an application processor and a modem processor, where the application processor mainly processes operations related to the operating system, user interface, and application programs, and the like, and the modem processor mainly processes wireless communication signals, e.g., a baseband processor. As can be understood, the modem processor may not be integrated into the processor 1310.

[0295] Here, when performing a federated model training process corresponding to a model training task, the radio frequency unit 1301 is used to send a federated model training request message to at least one second network element, where the at least one second network element is a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between each of the second network elements are different but sample features are the same; and receive first information sent by the at least one second network element, where the first information at least includes a first training result, and the first training result corresponds to the training data used in the federated model training process in the second network element; and the processor 1310 is used to perform model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

[0296] In this embodiment, a horizontal federation learning method is used to link each second network element to realize local distributed model training, so that data privacy and data security in each second network element can be ensured without sharing data on each second network element in the communication network, and the model training effect can also be ensured.

[0297] In another implementation manner, the communication device 1200 may be a network side device, which includes a processor and a communication interface, and the communication interface is coupled to the processor, and the processor runs a program or instruction to realize the steps of the method in the embodiments 300 to 800. This network side device embodiment corresponds to the above network side device method embodiment, and each implementation process and implementation manner of the above method embodiment can be applied to this network side device embodiment, and the same technical effects can be achieved.

[0298] Illustratively, as shown in FIG. 14, is a structural schematic diagram of a network side device according to an embodiment of the present application. The network side device 1400 includes an antenna 1401, a radio frequency device 1402, a baseband device 1403, a processor 1404, and a memory 1405. The antenna 1401 and the radio frequency device 1402 are connected. In the uplink direction, the radio frequency device 1402 receives information through the antenna 1401, and transmits the received information to the baseband device 1403 for processing. In the downlink direction, the baseband device 1403 processes the information to be transmitted and transmits it to the radio frequency device 1402, and the radio frequency device 1402 sends out the received information through the antenna 1401 after processing it.

[0299] The methods performed by the network side equipment in the above embodiments may be implemented in the baseband device 1403, which includes a baseband processor.

[0300] The baseband device 1403 may include, for example, at least one baseband board, on which multiple chips are installed, and as shown in FIG. 14, one of the chips is, for example, a baseband processor, which is connected to the memory 1405 via a bus interface, calls the program in the memory 1405, and performs the operation of the network side equipment shown in the embodiment of the above method.

[0301] The network side device may further include a network interface 1406, which may be, for example, a common public radio interface (CPRI).

[0302] Specifically, the network side device 1400 of the embodiment of the present invention further includes instructions or programs stored in the memory 1405 and capable of running on the processor 1404, and the processor 1404 can call the instructions or programs in the memory 1405 to execute the methods performed by each module shown in Figures 9-11, and achieve the same technical effects, which will not be described further here in order to avoid repetition.

[0303] Illustratively, as shown in Figure 15, which is a structural schematic diagram of a network side device 1500 according to another example embodiment of the present application, the network side device 1500 includes a processor 1501, a network interface 1502, and a memory 1503. Here, the network interface 1502 is, for example, a common public radio interface (CPRI).

[0304] Specifically, the network side device 1500 of the embodiment of the present invention further includes instructions or programs stored in the memory 1503 and capable of running on the processor 1501, and the processor 1501 can call the instructions or programs in the memory 1503 to execute the methods performed by each module shown in Figures 9-11, and achieve the same technical effects, which will not be described further here in order to avoid repetition.

[0305] An embodiment of the present application further provides a readable storage medium, in which a program or instruction is stored, which, when executed by a processor, can realize each process of the embodiment of the above model training method and achieve the same technical effect, and will not be further described here in order to avoid repetition.

[0306] Wherein, the processor is the processor in the terminal according to the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read only memory ROM, a random access memory RAM, a magnetic disk, or an optical disk.

[0307] An embodiment of the present application further provides a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run a program or instruction of a network side equipment and realize each process of the embodiment of the above model training method, and can achieve the same technical effect, and in order to avoid repetition of description, no further description will be given here.

[0308] It should be understood that the chips referred to in the embodiments of the present application may be referred to as system level chips, system chips, chip systems, or systems on chips.

[0309] An embodiment of the present application further provides a computer program product, which includes a processor, a memory, and a program or instruction stored in the memory and capable of running on the processor, which, when executed by the processor, can realize each process of the embodiment of the above model training method and achieve the same technical effect, and will not be further described here in order to avoid repetition.

[0310] An embodiment of the present application further provides a model training system, which includes a first network element, a second network element, and a third network element, where the first network element may be used to perform the method steps in method embodiments 300-600, the second network element may be used to perform the method steps in method embodiment 700, and the third network element may be used to perform the method steps in method embodiment 800.

[0311] It should be explained that in this specification, the terms "comprise", "include", or any other variants thereof are intended to cover the non-exclusive "comprise", whereby a process, method, article, or apparatus that includes a set of elements includes not only those elements, but also other elements not expressly listed or inherent to such process, method, article, or apparatus. Absent any further limitations, an element limited by the phrase "comprises a" does not preclude the presence of other identical elements in the process, method, article, or apparatus that includes this element. It should be pointed out that the scope of the method and apparatus in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may include performing functions in an essentially simultaneous manner or in reverse order based on the functions involved, for example, the described method can be performed in a different order than described, and various steps can be added, omitted, or combined. Also, features described with reference to some examples can be combined in other examples.

[0312] From the above description of the embodiments, it is clear to those skilled in the art that the methods of the above embodiments can be realized in the form of software and a necessary general-purpose hardware platform. Of course, they can also be realized in hardware, but in many cases the former is a more preferred embodiment. In light of this understanding, the technical proposal of the present application may be substantially or the part that contributes to the prior art may be embodied in the form of a computer software product, which is stored in a storage medium (e.g., ROM / RAM, magnetic disk, optical disk) and includes some instructions for causing a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0313] The above describes the embodiments of the present application in conjunction with the drawings, but the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not limiting. Those skilled in the art can take the suggestions of this application and make many forms without departing from the spirit and scope of the claims of this application, all of which belong to the protection scope of this application.

[0314] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to Chinese Patent Application No. 202111662860.8 filed on December 30, 2021, the entire contents of which are hereby incorporated by reference.

Claims

1. 1. A method for training a model, comprising: When a first network element performs a federated model training process corresponding to a model training task, the first network element sends a federated model training request message to at least one second network element, where the at least one second network element is a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between the second network elements are different but have the same sample features; receiving, by the first network element, first information transmitted by the at least one second network element, the first information including at least a first training result, the first training result corresponding to training data used in the federation model training process in the second network element; A model training method comprising a step of the first network element performing model training based on a first model and the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

2. The federated model training request message includes: model instance identifier information corresponding to the target model and assigned by the first network element; type information of the model training task; Identifier information for the model training task; First indication information for indicating that the federation model training process is a one-time horizontal federation learning process; Related information of a first filter for limiting at least one of a target object, a target time, and a target area corresponding to the model training task; Related information of a first model for each of the second network elements to perform local model training; Model training configuration information; and Report information of the first training result; and related information of each network element participating in the federated model training process.

3. The model instance identifier information related information of the first network element; a first time period for indicating that the model training task is to be performed based on training data generated within the first time period; a second time for indicating a completion time of the federated model training process; The method of claim 2 , wherein the second network element is associated with at least one of the first and second network elements.

4. The model training configuration information: Model structure information; Model hyperparameter information, type information of training data in the federation model training process; Model training condition information for instructing the second network element on a condition for the first network element to start model training based on the first training result reported by the second network element; model training times information for indicating a number of local model trainings that the second network element needs to perform before transmitting the first information to the first network element; and model training duration information for indicating a duration of local model training that the second network element needs to perform before transmitting the first information to the first network element.

5. The model training condition information is a third time for the first network element to instruct the second network element to start model training when a waiting time for feeding back the first training result reaches the third time; and a first threshold for instructing the first network element to start model training when a number of the first training results received reaches the first threshold.

6. The first training result report information includes: a report format of the first training result; and a reporting condition for the first training result.

7. The method of claim 1 , wherein the first training result includes model information of a second model obtained by training the second network element based on a local training model and / or first gradient information corresponding to the second model.

8. The first information is The method of claim 1 , wherein the first network element further comprises at least one of model instance identifier information for performing model association.

9. 2. The method of claim 1, wherein the second network element is obtained from a network storage function NRF by the first network element based on the model training task and is a network element capable of supporting the federated model training process.

10. The method comprises: The method further includes the first network element determining that a first condition exists, where the first condition includes: the first network element does not store or cannot obtain all or a portion of the training data corresponding to the model training task; the at least one second network element is capable of providing all or a portion of training data corresponding to the model training task; 10. The method of claim 1, further comprising at least one of: samples of training data between the second network elements corresponding to the model training tasks are different, but sample features are the same.

11. The method comprises: The method further includes the step of: the first network element sending second information to the at least one second network element when the calculation result of the loss function of the target model does not satisfy a predetermined requirement; 10. The method of claim 1, wherein the second information includes at least the second training result, and the second training result is used by the second network element to re-perform local model training and re-obtain the first training result.

12. The step of transmitting second information to the at least one second network element includes: For each of the second network elements, sending to the second network element specific information that belongs to the second information and is different from the information included in the federation model training request message; and for each said second network element, transmitting all information in said second information to said second network element.

13. The method comprises: The method further includes the step of: when the calculation result of the loss function of the target model does not satisfy a predetermined requirement, the first network element sends second information to at least one third network element; 10. The method of claim 1, wherein the second information includes at least the second training result, the second training result is used by the third network element to perform local model training to obtain a third training result, and the third network element is a network element that is re-determined by the first network element and participates in the federated model training process.

14. The method of claim 11 or 13, wherein the second training result includes at least second gradient information of the loss function with respect to parameters of the target model or information of the target model.

15. The second information is model instance identifier information for the second network element to perform model association; Report information on the results of the first training exercise; and Related information of each network element participating in the federated model training process; and model training configuration information corresponding to the target model.

16. The method comprises: The method further includes the step of receiving, by the first network element, a model request message sent by a fourth network element, where the model request message comprises: type information of the model training task; Identifier information for the model training task; Related information of a second filter for limiting at least one of a target object, a target time, and a target area corresponding to the model training task; 16. The method of claim 1, further comprising at least one of: a model feedback format; and model feedback related information including at least one of a feedback condition.

17. The method comprises: The first network element further includes, when the calculation result of the loss function of the target model satisfies a predetermined requirement, sending related information of the target model to the fourth network element; Here, the relevant information of the target model is: the model instance identifier information; and The target model information; second indication information for indicating that the target model is a horizontal federation learning model; and related information of the second network element and / or a third network element.

18. 1. A method for training a model, comprising: receiving, by a second network element, a federated model training request message sent by a first network element to request the second network element to participate in a federated model training process corresponding to a model training task; The second network element performs model training according to the federation model training request message to obtain a first training result; The second network element transmits first information including at least the first training result to the first network element; wherein the first training result corresponds to training data used in the federation model training process in the second network element, the samples used in the federation model training in the second network element are different from the samples used in the federation model training in a fifth network element but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federation model training process.

19. The federated model training request message includes: model instance identifier information corresponding to a target model and assigned by the first network element; type information of the model training task; Identifier information for the model training task; First indication information for indicating that the federation model training process is a one-time horizontal federation learning process; Related information of a first filter for limiting at least one of a target object, a target time, and a target area corresponding to the model training task; Related information of a first model for each of the second network elements to perform local model training; Model training configuration information; and Report information of the first training result; and associated information of each network element participating in the federated model training process.

20. The model instance identifier information related information of the first network element; a first time period for indicating that the model training task is to be performed based on training data generated within the first time period; a second time for indicating a completion time of the federated model training process; and associated information of each network element participating in the federated model training process.

21. The model training configuration information: Model structure information; Model hyperparameter information, type information of training data in the federation model training process; Model training condition information for instructing the second network element on a condition for the first network element to start model training based on the first training result reported by the second network element; model training times information for indicating a number of local model trainings that the second network element needs to perform before transmitting the first information to the first network element; and model training duration information for indicating a duration of local model training that the second network element needs to perform before transmitting the first information to the first network element.

22. The model training condition information is a third time for the first network element to instruct the second network element to start model training when a waiting time for feeding back the first training result reaches the third time; and a first threshold for instructing the first network element to start model training when a number of the first training results received reaches the first threshold.

23. The first training result report information includes: a report format of the first training result; and a reporting condition for the first training result.

24. The method of claim 18 , wherein the first training result includes model information of a second model obtained by training the second network element based on a local training model and / or first gradient information corresponding to the second model.

25. 20. The method of claim 18, wherein the second network element is a network element obtained by the first network element from a network storage function NRF based on the model training task and capable of supporting the federated model training process.

26. The first information is 20. The method of claim 18, wherein the first network element further comprises at least one of model instance identifier information for performing model association.

27. The step of the second network element performing model training according to the model training request message to obtain a first training result includes: determining that the second network element needs to perform a federation model training process based on the first indication information; The second network element determines a model training task corresponding to the federated model training based on type information or identifier information of the model training task; The second network element obtains training data based on the model training task and the related information of the first filter; The second network element determines a local training model and a model training scheme used for training the local training model based on the model training configuration information and / or the related information of the first model; 20. The method of claim 19, further comprising: the second network element training the local training model based on the model training scheme; and calculating the training data based on the trained local training model to obtain a first training result.

28. The method comprises:

20. The method of claim 18, further comprising the step of the second network element receiving second information transmitted by the first network element, the second information including at least a second training result.

29. The step of receiving, by the second network element, second information transmitted by the first network element, comprises: receiving, by the second network element, specification information belonging to the second information and different from information included in the federation model training request message sent by the first network element; and the second network element receiving all information in the second information transmitted by the first network element.

30. 29. The method of claim 28, wherein the second information is transmitted by the first network element when a calculation result of a loss function of a target model does not satisfy a predetermined requirement.

31. 29. The method of claim 28, wherein the second training result includes at least second gradient information of a loss function of a target model with respect to parameters of the target model or information of the target model.

32. The second information is model instance identifier information for the second network element to perform model association; Information on the target model, Related information of each network element participating in the federated model training process; and model training configuration information corresponding to the target model.

33. 1. A method for training a model, comprising: A fourth network element receives the target model related information sent by the first network element; wherein the relevant information of the target model is at least used to characterize the target model as a horizontal federation model.

34. The relevant information of the target model is model instance identifier information corresponding to a target model and assigned by the first network element; Third model information including network structure information and / or model parameter information of the target model; second indication information for indicating that the target model is a horizontal federation learning model; and related information of a second network element and / or a third network element that are network elements participating in the target model training.

35. The method of claim 33, wherein the target model related information is sent by the first network element when a calculation result of a loss function satisfies a predetermined requirement.

36. The method comprises: The method further includes the step of the fourth network element sending a model request message to the first network element, where the model request message comprises: type information of the model training task; Identifier information for the model training task; Related information of a second filter for limiting at least one of a target object, a target time, and a target area corresponding to the model training task; 35. The method of claim 34, comprising at least one of a model description scheme and model feedback related information including at least one of a feedback time.

37. A model training device, When performing a federated model training process corresponding to a model training task, a first sending module is used to send a federated model training request message to at least one second network element, the at least one second network element being a network element participating in the federated model training process, and samples corresponding to training data used in the federated model training process between the second network elements are different but have the same sample features; a first receiving module for receiving first information transmitted by the at least one second network element, the first information including at least a first training result, the first training result corresponding to training data used in the federation model training process in the second network element; A model training device comprising: a first model and a first model training module for performing model training based on the first training result reported by the at least one second network element to obtain a target model and / or a second training result.

38. A model training device, a second receiving module for receiving, by a second network element, a federated model training request message sent by the first network element to request the second network element to participate in a federated model training process corresponding to a model training task; a second model training module for performing model training according to the federation model training request message to obtain a first training result; a second transmitting module for transmitting first information including at least the first training result to the first network element; wherein the first training result corresponds to training data used in the federation model training process in the second network element, the samples used in the federation model training in the second network element are different from the samples used in the federation model training in a fifth network element but have the same sample features, and the fifth network element is another network element other than the second network element among a plurality of network elements participating in the federation model training process.

39. A model training device, a third receiving module for receiving the target model related information sent by the first network element; wherein the relevant information of the target model is at least used to characterize the target model as a horizontal federation model.

40. 32. A communications device comprising a processor and a memory, the memory storing a program or instructions operable to run on the processor, the program or instructions, when executed by the processor, implementing the steps of a model training method according to any one of claims 1 to 17, or implementing the steps of a model training method according to any one of claims 18 to 32, or implementing the steps of a model training method according to any one of claims 33 to 36.

41. 32. A readable storage medium having a program or instructions stored therein, the program or instructions being, when executed by a processor, to implement steps of a model training method according to any one of claims 1 to 17, or to implement steps of a model training method according to any one of claims 18 to 32, or to implement steps of a model training method according to any one of claims 33 to 36.

Images