WO2023208043A1 - 用于无线通信系统的电子设备、方法和存储介质 - Google Patents

用于无线通信系统的电子设备、方法和存储介质 Download PDF

Info

Publication number
WO2023208043A1
WO2023208043A1 PCT/CN2023/090868 CN2023090868W WO2023208043A1 WO 2023208043 A1 WO2023208043 A1 WO 2023208043A1 CN 2023090868 W CN2023090868 W CN 2023090868W WO 2023208043 A1 WO2023208043 A1 WO 2023208043A1
Authority
WO
WIPO (PCT)
Prior art keywords
entity
server entity
group
participant
electronic device
Prior art date
Application number
PCT/CN2023/090868
Other languages
English (en)
French (fr)
Inventor
崔焘
孙晨
Original Assignee
索尼集团公司
崔焘
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 索尼集团公司, 崔焘 filed Critical 索尼集团公司
Publication of WO2023208043A1 publication Critical patent/WO2023208043A1/zh

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network
    • H04L67/104—Peer-to-peer [P2P] networks
    • H04L67/1044—Group management mechanisms 
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/098—Distributed learning, e.g. federated learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network
    • H04L67/104—Peer-to-peer [P2P] networks
    • H04L67/1042—Peer-to-peer [P2P] networks using topology management mechanisms
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network
    • H04L67/104—Peer-to-peer [P2P] networks
    • H04L67/1059—Inter-group management mechanisms, e.g. splitting, merging or interconnection of groups
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/52—Network services specially adapted for the location of the user terminal
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H04W4/023—Services making use of location information using mutual or relative location information between multiple location based services [LBS] targets or of distance thresholds
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H04W4/029—Location-based management or tracking services
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/06—Selective distribution of broadcast services, e.g. multimedia broadcast multicast service [MBMS]; Services to user groups; One-way selective calling services
    • H04W4/08—User group management

Definitions

  • the present disclosure relates to the field of wireless communications, and more particularly, to electronic devices, methods and storage media for federated learning in the field of wireless communications.
  • Federated Learning has become a very important technology due to its unique advantages in ensuring data privacy security and legal compliance, and its ability to improve the effectiveness of machine learning models through joint modeling of multiple devices.
  • Distributed artificial intelligence framework or distributed machine learning framework has become a very important technology due to its unique advantages in ensuring data privacy security and legal compliance, and its ability to improve the effectiveness of machine learning models through joint modeling of multiple devices.
  • Distributed artificial intelligence framework or distributed machine learning framework has become a very important technology due to its unique advantages in ensuring data privacy security and legal compliance, and its ability to improve the effectiveness of machine learning models through joint modeling of multiple devices.
  • a FL server (such as a cloud server) can generate a global model by aggregating local models trained locally on multiple terminal devices. Specifically, the FL server first initializes the local model in each terminal device. In each iterative operation, each terminal device relies on its local training data to train and learn the local model. Then, the terminal device may report the local model (specifically, the parameters of the local model, which parameters may be characterized in a gradient manner) to the FL server through the uplink of the wireless communication system (eg, 5G network). The FL server aggregates local model parameters from multiple terminal devices (e.g., averages the local model parameters of the terminal devices) to obtain an updated global model.
  • the wireless communication system eg, 5G network
  • the updated global model (specifically, the parameters of the global model) can be distributed to each terminal device through the downlink of the wireless communication system. Therefore, each terminal device updates the local model according to the received global model, and then continues to perform the next round of local training based on the updated local model.
  • federated learning needs to be adjusted to adapt to time-varying wireless channel environments, unstable training data, and device heterogeneity.
  • the base station acts as the server of federated learning, and each terminal device acts as a participant in federated learning.
  • the base station can obtain the global model by aggregating local models from each terminal device. , and sends the global model to each terminal device to update its local model.
  • the first layer is to distribute the global model and upload the local model.
  • the first layer is the terminal device that is a participant in federated learning.
  • Such a two-layer structure requires each FL participant to communicate directly with the FL server. Considering the time-varying characteristics and instability of the wireless channel, the communication delay between the server and participants may be large, making aggregation based on the local model It takes a long time to obtain the global model, so the participants cannot obtain a more accurate model quickly, which may affect the prediction process based on the local model.
  • the electronic device may include processing circuitry that may be configured to: determine at least one level one federated learning (FL) server entity and a plurality of FL participations corresponding to each level one FL server entity.
  • Party entity wherein a first-level FL server entity and its corresponding multiple FL participant entities together form a group, and the at least one first-level FL server entity serves as the second-level FL server entity included in the electronic device.
  • the FL participant is capable of federated learning with the secondary FL server entity; and sending the information of the formed group to the at least one primary FL server entity and the FL participant entity corresponding to each primary FL server entity, to Federated learning can be performed within each group.
  • the electronic device may include a processing circuit system, and the processing circuit system may be configured to: receive from the electronic device on the network device side information about the group in which the electronic device on the user equipment side is located, wherein the group includes a A primary FL server entity and a plurality of FL participant entities corresponding to it, and at least one primary FL server entity determined by the electronic device on the network device side as a secondary FL included in the electronic device on the network device side
  • the FL participants of the server entity can perform federated learning with the secondary FL server entity; and perform federated learning within the group based on the information of the group.
  • the method may include: determining at least one level one federated learning (FL) server entity and a plurality of FL participant entities corresponding to each level one FL server entity, wherein one level one FL server entity and Its corresponding multiple FL participant entities together form a group, and the at least one first-level FL server entity serves as a network side device package FL participants containing a secondary FL server entity are capable of performing federated learning with the secondary FL server entity; and sending the information of the formed group to the at least one primary FL server entity and with each primary FL server entity Corresponding FL participant entities to enable federated learning within each group.
  • FL federated learning
  • the method may include: receiving information from a network device about a group in which the terminal device is located, wherein the group includes a first-level FL server entity and multiple FL participant entities corresponding to it, and the network device includes: At least one primary FL server entity determined by the device as a FL participant of the secondary FL server entity included in the network device can perform federated learning with the secondary FL server entity; and based on the information of the group, perform federation within the group study.
  • Yet another aspect of the present disclosure relates to a computer-readable storage medium storing one or more instructions.
  • the one or more instructions when executed by one or more processors of the electronic device, may cause the electronic device to perform the above-described method.
  • Figure 1 is a schematic diagram of the process of federated learning in a wireless communication network in the related art
  • Figure 2 is a schematic diagram of a federated learning architecture according to an embodiment of the present disclosure
  • Figure 3 is a flowchart of a method performed by a network device in the federated learning architecture proposed herein according to an embodiment of the present disclosure
  • Figure 4 is a diagram of a method performed by a terminal device in the federated learning architecture proposed herein according to an embodiment of the present disclosure. flow chart;
  • Figure 5 is a flow chart of a method adopted in the federated learning process in the federated learning architecture proposed herein according to an embodiment of the present disclosure
  • Figure 6 is a sequence diagram of an example of information exchange in the federated learning architecture proposed herein according to an embodiment of the present disclosure
  • FIG. 7 is a block diagram of an example structure of a personal computer as an information processing device employable in the embodiment of the present disclosure
  • FIG. 8 is a block diagram illustrating a first example of a schematic configuration of a gNB to which the technology of the present disclosure may be applied;
  • FIG. 9 is a block diagram illustrating a second example of a schematic configuration of a gNB to which the technology of the present disclosure may be applied;
  • FIG. 10 is a block diagram showing an example of a schematic configuration of a smartphone to which the technology of the present disclosure can be applied.
  • FIG. 11 is a block diagram showing an example of a schematic configuration of a car navigation device to which the technology of the present disclosure can be applied.
  • the FL server 110 can also be called a FL training server, cloud server, etc.
  • multiple terminal devices A to E jointly perform federated learning to aggregate the local data of the terminal devices A to E through the FL server 110
  • the trained local model is used to obtain the global model, and the local model in the terminal device is updated by issuing the global model.
  • the FL server 110 can be set in the base station or other network equipment. middle.
  • the terminal devices A to E are terminal devices that can wirelessly communicate with it, including any form of user equipment such as a smartphone, a desktop computer, and the like.
  • the terminal devices A to E send a training resource report to the FL server 110, including their respective available computing resources, wireless channel environment, geographical location, etc.
  • the FL server 110 performs device selection among the terminal devices A to E according to the received reports to determine the terminal devices participating in model training in the local iteration.
  • the terminal devices A to E in the wireless communication system may also have other data services that require uplink transmission.
  • the FL server 110 selects terminal devices A, C, and D for federated learning in the N-th iteration.
  • the FL server 110 distributes the global model and training parameter configuration to the selected terminal devices A, C, and D.
  • terminal devices A, C, and D update the local model to be consistent with the global model, and train the local model based on local data.
  • the terminal devices A, C, and D report the training results (eg, parameters of the local model) to the FL server 110 .
  • the FL server 110 aggregates the received training results to obtain a new global model.
  • the Nth iteration process ends. In the N+1th iteration, a similar process to the Nth iteration is repeated.
  • the FL server 110 selects terminal devices B, C, and E to perform federated learning based on the re-reported training resource report.
  • the terminal device participating in federated learning needs to be selected at the beginning of each iteration, if the conditions of the terminal device (such as the computing resources of the terminal device, wireless channel environment, etc.) do not change substantially, then FL The operation of the server selecting the terminal device and issuing the training parameter configuration do not need to be performed in each iteration.
  • a terminal device skips the model aggregation iteration of federated learning one or more times, it may affect the accuracy of the federated learning model. Therefore, the terminal devices participating in federated learning can be alternately arranged over time to maximize Ensure independent and identically distributed sampling effects and give all terminal devices equal opportunities to contribute to the global model.
  • FIG. 2 A schematic diagram of a federated learning architecture 200 according to an embodiment of the present disclosure is shown in FIG. 2 .
  • the structure is divided into three layers, the first layer as the lower layer contains FL participant entities, the second layer as the middle layer (can also be considered as the layer that performs preliminary aggregation) contains the first-level FL server entity, and the third layer as the upper layer
  • the tier (which can also be thought of as the tier that performs fine-grained aggregation) contains secondary FL server entities.
  • the "entity” referred to herein may be a hardware circuit or circuit system that physically exists in a tangible device such as a base station, user equipment, terminal equipment, server equipment, or the tangible device itself (i.e., in the form of implemented in hardware), or may be computer-executable instructions or program code stored in the memory of a device capable of performing wireless or network communications and executed on a processor of the device (i.e., implemented in software).
  • entity can be a physical entity or a logical entity.
  • an "entity contained in a device” it may mean that the entity is a hardware component of the device or the device itself, or it may mean that the entity is computer-executable instructions or program code that runs in the device.
  • a FL participant entity may correspond to or be included in a terminal device
  • a primary FL server entity may correspond to or be included in a terminal device
  • a secondary FL server entity may correspond to
  • a network-side device for example, a base station, a cloud server, or a control device in a core network
  • a network-side device may be included in the network-side device.
  • an entity When an entity is located on a certain layer, its corresponding device will also be located on that layer.
  • it may contain only one FL participant entity, it may contain only one first-level FL server entity, or it may contain both one FL participant entity and one first-level FL server entity.
  • the device on the network side can be a base station, a server device or a control device in the core network, for the simplicity of description, the following description takes the base station located at the third layer as an example.
  • the device on the network side is Other forms of equipment can be used instead of base stations to implement the solutions of the embodiments of the present disclosure.
  • a first-level FL server entity and its corresponding multiple FL participant entities form a group in which federated learning can be performed.
  • the first-level FL server entity acts as the server of traditional federated learning
  • the FL participant entity acts as the participant of traditional federated learning to jointly perform federated learning.
  • Figure 2 shows three groupings, each containing a primary FL server entity indicated by a triangle in the figure and a plurality of FL participant entities indicated by circles in the figure.
  • the number of three groupings and the number of FL party entities in the grouping This is given only as an example, and those skilled in the art will understand that there may be more or fewer groups within the coverage area of the base station, and there may be more or fewer FL participant entities in each group.
  • all terminal devices are grouped in Figure 2, there may be situations where terminal devices are not grouped. In this case, the ungrouped end devices contain primary FL server entities, located at the second level of the architecture.
  • the primary FL server entities in the second layer and the secondary FL server entities in the third layer can also perform federated learning.
  • the secondary FL server entity acts as a server of traditional federated learning and the primary FL server entity acts as a participant of traditional federated learning to jointly perform federated learning.
  • the first layer and the second layer can jointly perform federated learning
  • the second layer and the third layer can jointly perform federated learning, allowing the federated learning architecture to be extended to multiple layers.
  • the terminal devices located on the first layer do not need to interact with the base station located on the third layer every time to update the local model. They can quickly update the local model by communicating with the terminal devices located on the second layer, which delays the model update. improvements.
  • the base station located on the third layer can receive more comprehensive model-related information from the second-layer terminal device that collects the local model-related information of the first-layer terminal device, so that the global model can aggregate more comprehensive model-related information, thereby Global models can be more accurate.
  • devices at the third layer are required to group terminal devices.
  • the base station receives any one or more of the processing capabilities, machine learning capabilities, geographical location, wireless channel quality and movement trajectory information of the terminal device from the terminal device within its coverage, for example, it can be received through the Uu link .
  • the link used to receive the above items is called a Uu link here, those skilled in the art can understand that the link can also have other representations in different standards to describe base stations and terminals with such links. Communication between devices.
  • the base station may receive only a portion of these items.
  • the base station may receive only the processing capability of the terminal device, only the machine learning capability of the terminal device, only the geographical location of the terminal device, only the wireless channel quality of the terminal device, only the Movement trajectory information, or receiving different combinations of these items.
  • the base station can select a manager device from the terminal devices within its coverage based on the received items.
  • Each manager device contains a first-level FL server entity.
  • the base station may determine, based on the CPU processing speed reported by the terminal device, the terminal device whose CPU processing speed is ranked in a predetermined position (for example, first, top two, top three, etc.) before all terminal devices as the manager device.
  • the base station may also determine the terminal device with more computing resources allocated to machine learning (for example, located at the front predetermined position) as the manager device based on the computing resources reported by the terminal device for machine learning.
  • the base station can also determine the distance from the base station or other terminal equipment reported by the terminal equipment.
  • the distance between terminal devices determines at least one terminal device that has the most adjacent terminal devices with a distance smaller than a predetermined value as the manager device.
  • the base station can also rank at least one terminal device or channel quality with good channel quality (for example, signal-to-noise ratio (SNR) less than a predetermined value, reference signal received power (RSRP) greater than a predetermined value, etc.) based on the channel quality report reported by the terminal device.
  • the terminal devices with predetermined positions (such as the first, first two, first three, etc.) in front of all terminal devices are determined to be the manager devices.
  • the base station can also select a terminal device that is expected to pass through a specific location as the manager device based on the future path planning information reported by the terminal device.
  • the base station can determine the terminal device within a predetermined distance from the manager device as being in the first layer and containing the FL participant entity, and form a group together with the manager device.
  • a terminal device may be located in only one group, or a terminal device may be located in two or more groups. For example, the distance between the terminal device and different manager devices is within a predetermined distance.
  • the manager device may contain both a first-level FL server entity and a FL participant entity. This means that the manager device may actually be a terminal device with powerful computing capabilities. It not only trains local models, but also serves as a group Manager to aggregate local models within a group. At this time, the group in which the manager device is located contains its own FL participant entities.
  • terminal devices that are not grouped, for example, these terminal devices are more than a predetermined distance from each manager device, then these terminal devices are determined to contain the first-level FL server entity and will be connected to other devices that are also located at the second layer.
  • the primary FL server entities perform federated learning together as FL participants of the secondary FL server entities.
  • the base station can send the group information to the terminal equipment within its coverage.
  • the base station may send to each terminal device in a group any one of the identifier ID of the terminal device where the primary FL server entity in the group is located and the group ID of the group, or Multiple items. This means that the base station can send to each terminal device in the group only the identifier ID of the terminal device located on the second layer, only the group ID of the group, or both.
  • the group ID of the group may be an ID assigned by the base station, or may be related to the identifier ID of the manager device, so that the terminal device can determine the manager device of the group by analyzing these IDs.
  • the base station may also send the group ID of the group in adjacent geographical locations to each terminal device in the group. For example, the base station may determine one or more other manager devices that are closest to the manager device in the group, determine the group in which these manager devices are located as an adjacent group, and then notify the group ID of the adjacent group to Terminals in a group equipment, thus facilitating communication between adjacent groups.
  • the FL participant entity of the group uploads local model-related information to the first-level FL server entity of the group, so that the first-level FL server entity updates by aggregating the received local model-related information.
  • the local model of the FL party entity for this group can use direct communication to upload local model related information to the first-level FL server entity through a link between the terminal device where the FL participant entity is located and the electronic device where the first-level FL server entity is located.
  • the local model related information may be information related to the parameters of the local model (eg it may be an encrypted version of the parameters of the local model) from which the primary FL server entity is able to obtain the parameters of the local model.
  • the local model may be updated by transmitting the output results of the local model (which may also be called prediction results, prediction scores, etc.).
  • the local model related information may be the output result calculated by the FL participant entity based on the local model based on the common data issued by the first-level FL server entity.
  • a first-level FL server entity may first select a common set of data (e.g., a subset of a common data set) as input to the FL party entity's local model, and then send these to all FL party entities within its group. common data.
  • the FL participant entities that receive the common data use the common data as the input of the local model, and input it into the local model trained through local data in turn to obtain a series of output results, and return these output results to the first-level FL server. entity.
  • the primary FL server entity After the primary FL server entity receives the output results sent from each participant entity in the group, by averaging the output results, the primary FL server entity can determine the aggregated output results corresponding to the common data.
  • the first-level FL server entity can deliver the obtained aggregated output results to each FL participant entity in the group, so that each FL participant entity can again based on the previously received common data and the corresponding aggregated output results.
  • Train the local model to update the parameters of the local model to obtain an updated local model. Since the data used for retraining the local model contains information from other FL participant entities, the training result is more accurate than if the FL participant entity only uses local data for training.
  • the first-level FL server entity can also use a weighted average to obtain the aggregated output results.
  • the first-level FL server entity may receive from each FL participant entity in the group some relevant information about the terminal device where the FL participant entity is located, including the first quantity related to the local model prediction accuracy of the terminal device, the terminal A second quantity related to the channel quality of the device and the history and/or future of the end device Any two or more of the trajectory-related third quantities, ie at least two of these terms may be received in different embodiments.
  • the terminal device including the FL participant entity may calculate a first quantity to characterize the local model prediction accuracy determined by comparing the local model with the updated new local model, and may also calculate a second quantity to characterize the local model prediction accuracy based on CSI, SNR, RSRP, the bit error rate of the received signal, etc. represent the channel quality.
  • the second quantity may be equal to the CSI divided by the received signal RSRP.
  • the terminal device may also calculate a third quantity to reflect the historical and/or future trajectory. For example, if the distance moved within the predetermined time period is longer, the third quantity may be larger.
  • Weight alpha*first quantity+beta*second quantity+gamma*third quantity
  • alpha, beta and gamma are predetermined constants, which can be determined according to the importance of the first quantity, the second quantity and the third quantity. For example, when the first quantity is more important, the weight assigned to it can be greater.
  • the above weight formula accordingly omits the corresponding item.
  • the first-level FL server entity can multiply the output results received from each FL participant entity by the corresponding weight and add them up to weight the output results from each FL participant entity.
  • the output results of the square entities are weighted and summed to obtain the aggregated output results.
  • the above weight may also be received from the base station.
  • the base station needs to calculate the weight corresponding to the FL participant entity.
  • the FL participant entity may send the above-mentioned first quantity, second quantity and/or third quantity directly to the base station (for example, a secondary FL server entity included in the base station), and the base station adopts the communication with the primary FL server entity.
  • the primary FL server entity that receives the weight may weight the output results uploaded by the corresponding FL participant entities with the received weights as described above to obtain an aggregated output result.
  • the aggregated output results can be made more consistent with the actual situation of each FL participant entity, and thus more accurate.
  • transmitting output results and aggregated output results is more secure due to reduced exposure to the internal state of the model, and reduces the amount of data transferred and the communication load will also be significantly reduced.
  • the output result calculated by the FL participant entity based on the common data delivered by the first-level FL server entity can be sent to other FL participant entities located in the same group in a unicast or multicast manner.
  • a FL participant entity can send a request message to other FL participant entities via Synchronization Channel Information (SCI), and then via the SCI Carry information used to demodulate and decode the physical side link control channel (PSSCH) and carry the output results through the PSSCH to send the output results to other FL participant entities.
  • SCI Synchronization Channel Information
  • PSSCH physical side link control channel
  • the operation of sending SCI can be completed in stage-1 (stage 1), and the operation of sending SCI and PSSSCH can be completed in stage-2 (stage 2).
  • the FL participant entity can also send the output results to the FL participant entities of different groups in a unicast or multicast manner. Such transmission can be performed through the PC5 link. As a result, information related to the local model of the FL participant entity can be shared more quickly, thereby promoting other FL participant entities to learn new knowledge.
  • the operation of the FL participant entity uploading the output results to the first-level FL server entity can be performed in a D2D manner.
  • the FL participant entities can also upload the output results to the first-level FL server entities of different groups in a D2D manner to share model-related information faster.
  • the first-level FL server entity since it exists as the FL server side of the FL participant entity to jointly perform federated learning, the first-level FL server entity has a global model that is valid within the group, which is the group where the first-level FL server entity is located.
  • the aggregation result of the local model of FL participant entities can also be called a local global model.
  • the first-level FL server entity can also train the local global model based on the above common data and aggregated output results, thereby obtaining the parameters of the local global model.
  • the parameters of the local global model may be issued to the FL participant entity to update its local model instead of the above aggregated output results issued to the FL participant entity.
  • the first-level FL server entity may issue an aggregation to the FL participant entity
  • the output results can also deliver the parameters of the local global model to the FL participant entity to update the local model of the FL participant entity.
  • the local global model of the first level FL server entity it contains is not passed It is obtained by aggregating the local models of FL participant entities, but is obtained through interaction with the secondary FL server entity located in the third layer and local training.
  • the secondary FL server entity can send the global model to ungrouped terminal devices to initialize the local global model therein, and the ungrouped terminal devices can train the local global model based on local data to update the local global model for uploading. to the secondary FL server entity.
  • the first-level FL server entity When the first-level FL server entity receives a request from a FL participant entity of another group, the first-level FL server entity can deliver the parameters of the local global model to the FL participant entity so that it can quickly update its local model. For example, when a FL participant entity is moving from one group to another, in order to get new points faster For model-related information of the group, the FL participant entity can request the first-level FL server entity of the adjacent group to issue a local global model based on the group ID of the adjacent group issued by the base station.
  • the secondary FL server entity contained in the base station receives local and global model related information from the primary FL server entity contained in the terminal device, and aggregates the received
  • the local global model related information is used to update the local global model of the first-level FL server entity.
  • the local global model related information may be information related to the parameters of the local global model (eg it may be an encrypted version of the parameters of the local global model) from which the secondary FL server entity can obtain the parameters of the local global model.
  • the local global model can be updated by transmitting the output results of the local global model.
  • the information related to the local global model may be the output result calculated by the first-level FL server entity based on the local global model based on the common data issued by the second-level FL server entity.
  • the secondary FL server entity may first select a set of common data (such as a subset of a common data set) as input to the local global model of the primary FL server entity, and then provide The primary FL server entity sends these common data.
  • the first-level FL server entity that receives the common data uses the common data as the input of the local global model and inputs it into the local global model in turn, thereby obtaining a series of output results, and returns these output results to the second-level FL server entity.
  • the secondary FL server entity After the secondary FL server entity receives these output results, by averaging the output results, the secondary FL server entity can determine an aggregated output result corresponding to the common data.
  • the secondary FL server entity can deliver the obtained aggregated output results to each primary FL server entity, so that each primary FL server entity can train again based on the previously received common data and the corresponding aggregated output results.
  • the local global model is used to update the parameters of the local global model to obtain the updated local global model.
  • the data used for retraining the local global model contains information from other first-level FL server entities (and thus information from other FL participant entities), so compared to the first-level FL participant entities, only the local model within the group is used.
  • the resulting local and global model may be more accurate.
  • the secondary FL server entity can also use a weighted average to obtain the aggregated output results.
  • the secondary FL server entity may receive from each primary FL server entity some relevant information about the terminal device where the primary FL server entity is located, including the first quantity related to the local global model prediction accuracy of the terminal device, the terminal device Any two or more of the second quantity related to the channel quality and the third quantity related to the history and/or future trajectory of the terminal device, that is, in different embodiments, at least one of these items can be received Two items.
  • the terminal device including the primary FL server entity may calculate a first quantity to characterize the prediction accuracy of the local global model determined by comparing the local global model with a new local global model based on data updates issued by the secondary FL server entity.
  • a second quantity may also be calculated to characterize the channel quality based on CSI, SNR, RSRP, bit error rate of the received signal, etc., for example, the second quantity may be equal to the CSI divided by the received signal RSRP.
  • the terminal device including the primary FL server entity can also calculate a third quantity to reflect the historical and/or future trajectory. For example, the third quantity can be larger if the distance moved within the predetermined time period is longer.
  • the weight corresponding to the first-level FL participant entity included in the terminal device can be determined based on at least two of the first quantity, the second quantity, and the third quantity:
  • Weight alpha’*first quantity+beta’*second quantity+gamma’*third quantity
  • alpha', beta' and gamma' are predetermined constants, which can be the same as or different from the above-mentioned alpha, beta and gamma. They can be determined according to the importance of the first quantity, the second quantity and the third quantity. For example, when The weight assigned to the first quantity can be greater when it is more important. When the primary FL server entity does not send any of the first amount, the second amount, and the third amount to the secondary FL server entity, the above weight formula accordingly omits the corresponding item.
  • the second-level FL server entity can multiply the output results received from each first-level FL server entity by the corresponding weight and add them together to weight the output results from each first-level FL server entity.
  • the output results of the first-level FL server entities are weighted and summed to obtain the aggregated output results.
  • the aggregate output result can be made more consistent with the actual situation of each first-level FL server entity, and thus may be more accurate.
  • transmitting output results and aggregated output results is more secure due to reduced exposure to the internal state of the model, and reduces the amount of data transferred and the communication load will also be significantly reduced.
  • the secondary FL server entity contained in the base station since it exists as the FL server side of the primary FL server entity to jointly perform federated learning, it has a global model regarding the local global model. Since the secondary FL server entity is also the highest-level processing entity of the entire federated learning architecture, such a global model is valid within the coverage of the base station.
  • the secondary FL server entity can train the global model based on the above common data and aggregated output results, thereby obtaining the parameters of the global model.
  • the parameters of the global model may be sent to the first-level FL server entity to update its local global model instead of the above-mentioned aggregate output results sent to the first-level FL server entity.
  • the secondary FL server entity may Deliver the aggregated output results, and also deliver global model parameters to the first-level FL server entity to update A local global model of the first-level FL server entity.
  • the secondary FL server entity can also directly deliver the global model to the FL participant entity.
  • the secondary FL server entity can deliver the parameters of the global model to each FL participant entity to directly update the local model of the FL participant entity.
  • the local global model uploaded to the secondary FL server entity can indicate a specific event (such as a traffic accident, traffic control, etc.)
  • the secondary FL server entity can immediately deliver the parameters of the global model after aggregating the global model.
  • the corresponding terminal equipment such as vehicles, etc.
  • the above describes the multi-layer federated learning architecture in detail, and specifically describes the interaction of entities between different layers. Due to the introduction of first-level federated learning server entities and groups, the local model can be quickly updated within the group, allowing the terminal device to obtain a more accurate model in a shorter time than the local model based only on local data. Compared with the traditional way of direct interaction between terminal equipment and base stations, the delay is reduced. In addition, it helps model updates by interacting with output results based on common data, improving security and reducing the burden of data transmission.
  • the network equipment here includes secondary FL server entities, such as base stations, cloud servers, core network equipment, etc.
  • the network device determines at least one first-level FL server entity and multiple FL participant entities corresponding to each first-level FL server entity, where one first-level FL server entity and multiple FL participant entities corresponding to it The entities together form a group, and the at least one primary FL server entity, as a FL participant of the secondary FL server entity included in the network side device, can perform federated learning with the secondary FL server entity.
  • the network device sends the information of the formed group to the at least one first-level FL server entity and the FL participant entity corresponding to each first-level FL server entity, so that federation can be performed within each group. study. For S310 and S320 and other operations, please refer to the above description of FIG. 2 and will not be described again.
  • FIG. 4 shows a flowchart of a method that can be executed by a terminal device in the federated learning architecture proposed herein according to an embodiment of the present disclosure.
  • the terminal device here includes a first-level FL server entity and/or a FL participant entity, which can be a user equipment (UE), for example, or other devices with information processing capabilities, such as tablet computers, desktop computers, Computers, supercomputers, etc.
  • UE user equipment
  • the terminal device receives information about the group in which the terminal device is located from the network device, where the group includes a first-level FL server entity and multiple FL participant entities corresponding to it, and at least one entity determined by the network device.
  • the primary FL server entity as a FL participant of the secondary FL server entity included in the network device, can perform federated learning with the secondary FL server entity.
  • the terminal device performs federated learning within the group based on the information of the group. S410 and S420 and other operations may refer to the above description of FIG. 2 and will not be described again.
  • Figure 5 is a flowchart of a method adopted in the federated learning process in the federated learning architecture proposed herein according to an embodiment of the present disclosure.
  • a reporting operation is performed. All terminal devices report their computing resources, wireless channel environment, location information, path planning, etc. to the cloud server (for example, located in a base station).
  • the cloud server for example, located in a base station.
  • a grouping operation is performed.
  • the cloud server selects the manager device based on predetermined criteria, which contains the first-level FL server entity, and notifies the federated learning group information to all terminal devices within its coverage that participate in federated learning, so that each terminal device knows what group it is in. and/or at what level of the architecture.
  • a model learning operation is performed.
  • the FL participant entity located in the terminal device of the first layer and the first-level FL server entity located in the terminal device of the second layer without lower layer can respectively perform local data processing according to the local data.
  • model and local global model (the initialization of these two models is completed by the global model issued by the cloud server, etc.) for training.
  • the FL participant entity located in the first-level terminal device reports the prediction score of the intermediate result of local model learning (that is, the output result of the local model based on the common data issued by the first-level FL server entity) to the belonging group
  • the first-level FL server entity located in the second-layer terminal device (which may also be called a manager device) reports, for example, in a D2D manner. It should be noted that when the same terminal device includes both FL participant entities and first-level FL participant entities, the terminal device can be located on the first layer and the second layer at the same time. For a terminal device without a lower layer, it reports the prediction score of the intermediate result of local global model learning (that is, the output result of the local model based on the common data issued by the secondary FL server entity) to the cloud server.
  • a model update operation is performed.
  • the first-level FL server entity in the manager device aggregates the prediction scores from different terminal devices in the group (or corresponding area) it is responsible for, and delivers the average of the prediction scores to this terminal devices so that they update their local models.
  • the first-level FL server entity also performs federated learning with the second-level FL server entity in the cloud server to update the local global model of the first-level FL server entity.
  • the secondary FL server entity obtains the global model, it can deliver the global model to different terminal devices in a point-to-point manner according to the requests of different terminal devices.
  • the above S510 to S540 may be performed in each iteration.
  • the administrator devices can also interact with the local parts of the group (or corresponding area) they are responsible for in a D2D manner.
  • the prediction score of the global model is performed in each iteration.
  • Figure 6 is a sequence diagram of an example of information exchange in the federated learning architecture proposed herein according to an embodiment of the present disclosure.
  • each terminal device reports their respective capabilities to the cloud server.
  • the cloud server groups these terminal devices according to the reported capabilities, thereby determining, as an example, the manager devices A and A that both contain the first-level FL server entity.
  • FIG 6 Although only two manager devices and one terminal device are shown in Figure 6, there may be more or less manager devices, and each manager device has one or more FL participants in its group. Physical terminal equipment.
  • Terminal device C uploads the prediction score to manager device A.
  • Manager device A aggregates the prediction scores uploaded by all FL participant entities in the group to obtain an aggregated prediction score, and helps terminal device C update the local model by sending the aggregated prediction score to it.
  • Manager devices A and B can also send the prediction scores of the local global model to the cloud server, so that the cloud server can aggregate the prediction scores uploaded by all first-level FL participant entities to obtain an aggregated prediction score.
  • the cloud server can obtain the global model based on the common data previously issued to obtain the prediction score and the aggregated prediction score, and deliver the global model to manager devices A and B and terminal device C to update their respective models.
  • the above federated learning architecture can be utilized in the Internet of Vehicles.
  • a base station can serve multiple intersections, with different traffic flows at different intersections. According to the location of the intersection, the base station issues information that the vehicles that can be used for federated learning belong to different areas.
  • the base station selects the multi-access edge computing equipment (MEC) near each intersection as the manager device of the area. Vehicles traveling at different intersections send the model prediction results of local learning results to the corresponding MEC, and the MEC fuses the model prediction results to update the local models of all vehicles participating in federated learning in the area. After the vehicle updates the local model, it then performs data training iterations based on local data.
  • MEC multi-access edge computing equipment
  • MEC uploads the model prediction results of the local global model to the cloud server.
  • the cloud server integrates the results of multiple MECs and can generate a global model.
  • the global model In response to the needs of vehicle A at a certain intersection, the global model directly It is sent to vehicle A because vehicle A is preparing to drive to another intersection within the coverage area and needs this information as a priori model information.
  • the prediction results of MEC's local global model include traffic accidents in the corresponding area
  • the local global model of the area is quickly calculated and fused by the cloud server, and then the resulting global model is distributed to each vehicle participating in federated learning in the coverage area. , to ensure that any vehicle can adopt a universal avoidance strategy when it reaches this location.
  • machine-executable instructions in the machine-readable storage medium or program product may be configured to perform operations corresponding to the above-described apparatus and method embodiments.
  • the embodiments of the machine-readable storage medium or program product will be clear to those skilled in the art, and therefore will not be described again.
  • Machine-readable storage media and program products for carrying or including the above-described machine-executable instructions are also within the scope of the present disclosure.
  • Such storage media may include, but are not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, and the like.
  • the above series of processes and devices can also be implemented through software and/or firmware.
  • the program constituting the software is installed from a storage medium or a network to a computer having a dedicated hardware structure, such as the general-purpose personal computer 1300 shown in FIG. 7 , and the computer is installed with various programs. , can perform various functions and so on.
  • 7 is a block diagram showing an example structure of a personal computer as an information processing device employable in the embodiment of the present disclosure.
  • the personal computer may correspond to the above-described exemplary terminal device according to the present disclosure.
  • a central processing unit (CPU) 1301 executes various processes according to a program stored in a read-only memory (ROM) 1302 or a program loaded from a storage section 1308 into a random access memory (RAM) 1303 .
  • ROM read-only memory
  • RAM random access memory
  • data required when the CPU 1301 performs various processes and the like is also stored as necessary.
  • the CPU 1301, ROM 1302 and RAM 1303 are connected to each other via a bus 1304.
  • Input/output interface 1305 is also connected to bus 1304.
  • the following components are connected to the input/output interface 1305: an input part 1306, including a keyboard, a mouse, etc.; an output part 1307, including a display, such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage part 1308 , including hard disk, etc.; and communication part 1309, including network interface card such as LAN card, modem Devices etc.
  • the communication section 1309 performs communication processing via a network such as the Internet.
  • Driver 1310 is also connected to input/output interface 1305 as needed.
  • Removable media 1311 such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc. are installed on the drive 1310 as necessary, so that computer programs read therefrom are installed into the storage section 1308 as needed.
  • the program constituting the software is installed from a network such as the Internet or a storage medium such as the removable medium 1311.
  • this storage medium is not limited to the removable medium 1311 shown in FIG. 7 in which the program is stored and distributed separately from the device to provide the program to the user.
  • the removable media 1311 include magnetic disks (including floppy disks (registered trademark)), optical disks (including compact disk read-only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including minidiscs (MD) (registered trademark) )) and semiconductor memory.
  • the storage medium may be a ROM 1302, a hard disk contained in the storage section 1308, or the like, in which programs are stored and distributed to users together with the device containing them.
  • the base stations mentioned in this disclosure may be implemented as any type of evolved Node B (gNB), such as macro gNB and small gNB.
  • a small gNB may be a gNB covering a smaller cell than a macro cell, such as pico gNB, micro gNB, and home (femto) gNB.
  • the base station may be implemented as any other type of base station, such as a NodeB and a Base Transceiver Station (BTS).
  • the base station may include: a main body (also called a base station device) configured to control wireless communication; and one or more remote radio heads (RRH) disposed in a different place from the main body.
  • RRH remote radio heads
  • various types of terminals to be described below may operate as base stations by temporarily or semi-persistently performing base station functions.
  • the terminal device mentioned in this disclosure is also called a user device in some examples, and may be implemented as a mobile terminal (such as a smartphone, a tablet personal computer (PC), a notebook PC, a portable game terminal, a portable/dongle mobile routers and digital cameras) or vehicle-mounted terminals (such as car navigation equipment).
  • the user equipment may also be implemented as a terminal performing machine-to-machine (M2M) communication (also known as a machine type communication (MTC) terminal).
  • M2M machine-to-machine
  • MTC machine type communication
  • the user equipment may be a wireless communication module (such as an integrated circuit module including a single die) installed on each of the above-mentioned terminals.
  • the term base station in this disclosure has the full breadth of its ordinary meaning and at least includes those used for A wireless communication station that is part of a wireless communication system or radio system to facilitate communications.
  • a base station may be, for example but not limited to, the following: the base station may be one or both of a base transceiver station (BTS) and a base station controller (BSC) in the GSM system, and may be a radio network controller in the WCDMA system.
  • BTS base transceiver station
  • BSC base station controller
  • One or both of (RNC) and Node B can be the eNB in LTE and LTE-Advanced systems, or can be the corresponding network node in the future communication system (such as gNB, eLTE that may appear in the 5G communication system eNB etc.).
  • Some functions in the base station of the present disclosure can also be implemented as entities with communication control functions in D2D, M2M and V2V communication scenarios, or as entities that play a spectrum coordination role in cognitive radio communication
  • gNB 1400 includes multiple antennas 1410 and base station equipment 1420.
  • the base station device 1420 and each antenna 1410 may be connected to each other via an RF cable.
  • the gNB 1400 (or base station device 1420) here may correspond to the above-mentioned electronic devices 300A, 1300A and/or 1500B.
  • Antennas 1410 each include a single or multiple antenna elements, such as multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna, and are used by base station device 1420 to transmit and receive wireless signals.
  • gNB 1400 may include multiple antennas 1410.
  • multiple antennas 1410 may be compatible with multiple frequency bands used by gNB 1400.
  • the base station device 1420 includes a controller 1421, a memory 1422, a network interface 1423, and a wireless communication interface 1425.
  • the controller 1421 may be, for example, a CPU or a DSP, and operates various functions of higher layers of the base station device 1420 . For example, the controller 1421 generates data packets based on the data in the signal processed by the wireless communication interface 1425 and delivers the generated packets via the network interface 1423 . The controller 1421 may bundle data from multiple baseband processors to generate bundled packets, and deliver the generated bundled packets. The controller 1421 may have logical functions to perform controls such as radio resource control, radio bearer control, mobility management, admission control, and scheduling. This control can be performed in conjunction with nearby gNB or core network nodes.
  • the memory 1422 includes RAM and ROM, and stores programs executed by the controller 1421 and various types of control data such as terminal lists, transmission power data, and scheduling data.
  • the network interface 1423 is a communication interface used to connect the base station device 1420 to the core network 1424. Controller 1421 may communicate with core network nodes or additional gNBs via network interface 1423. In this case, gNB 1400 Core network nodes or other gNBs may be connected to each other through logical interfaces such as S1 interface and X2 interface.
  • the network interface 1423 may also be a wired communication interface or a wireless communication interface for a wireless backhaul line. If the network interface 1423 is a wireless communication interface, the network interface 1423 may use a higher frequency band for wireless communication than the frequency band used by the wireless communication interface 1425.
  • the wireless communication interface 1425 supports any cellular communication scheme, such as Long Term Evolution (LTE) and LTE-Advanced, and provides wireless connectivity to terminals located in the cell of the gNB 1400 via the antenna 1410.
  • Wireless communication interface 1425 may generally include, for example, a baseband (BB) processor 1426 and RF circuitry 1427.
  • the BB processor 1426 may perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform layers such as L1, Medium Access Control (MAC), Radio Link Control (RLC), and Packet Data Convergence Protocol ( Various types of signal processing for PDCP)).
  • MAC Medium Access Control
  • RLC Radio Link Control
  • Packet Data Convergence Protocol Various types of signal processing for PDCP
  • the BB processor 1426 may have part or all of the above-mentioned logical functions.
  • the BB processor 1426 may be a memory that stores a communication control program, or a module including a processor and related circuitry configured to execute the program.
  • the update program can cause the functionality of the BB processor 1426 to change.
  • the module may be a card or blade that plugs into a slot of the base station device 1420. Alternatively, the module may be a chip mounted on a card or blade.
  • the RF circuit 1427 may include, for example, a mixer, filter, and amplifier, and transmit and receive wireless signals via the antenna 1410.
  • FIG. 8 shows an example in which one RF circuit 1427 is connected to one antenna 1410, the present disclosure is not limited to this illustration, and one RF circuit 1427 can be connected to multiple antennas 1410 at the same time.
  • the wireless communication interface 1425 may include multiple BB processors 1426 .
  • multiple BB processors 1426 may be compatible with multiple frequency bands used by gNB 1400.
  • wireless communication interface 1425 may include a plurality of RF circuits 1427.
  • multiple RF circuits 1427 may be compatible with multiple antenna elements.
  • FIG. 8 shows an example in which the wireless communication interface 1425 includes multiple BB processors 1426 and multiple RF circuits 1427, the wireless communication interface 1425 may also include a single BB processor 1426 or a single RF circuit 1427.
  • gNB 1530 includes multiple antennas 1540, base station equipment 1550 and RRH 1560. RRH 1560 and each antenna 1540 may be connected to each other via RF cables. The base station equipment 1550 and the RRH 1560 may be connected to each other via high-speed lines such as fiber optic cables.
  • the gNB 1530 (or base station device 1550) here may correspond to the above-mentioned electronic devices 300A, 1300A and/or 1500B.
  • Antennas 1540 each include single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by RRH 1560 to transmit and receive wireless signals.
  • gNB 1530 can include Includes multiple antennas 1540.
  • multiple antennas 1540 may be compatible with multiple frequency bands used by gNB 1530.
  • the base station device 1550 includes a controller 1551, a memory 1552, a network interface 1553, a wireless communication interface 1555, and a connection interface 1557.
  • the controller 1551, the memory 1552, and the network interface 1553 are the same as the controller 1421, the memory 1422, and the network interface 1423 described with reference to FIG. 8 .
  • the wireless communication interface 1555 supports any cellular communication scheme (such as LTE and LTE-Advanced) and provides wireless communication to terminals located in the sector corresponding to the RRH 1560 via the RRH 1560 and the antenna 1540.
  • the wireless communication interface 1555 may generally include a BB processor 1556, for example.
  • the BB processor 1556 is the same as the BB processor 1426 described with reference to FIG. 8 except that the BB processor 1556 is connected to the RF circuit 1564 of the RRH 1560 via the connection interface 1557.
  • the wireless communication interface 1555 may include multiple BB processors 1556.
  • multiple BB processors 1556 may be compatible with multiple frequency bands used by gNB 1530.
  • FIG. 9 shows an example in which the wireless communication interface 1555 includes multiple BB processors 1556, the wireless communication interface 1555 may also include a single BB processor 1556.
  • connection interface 1557 is an interface for connecting the base station device 1550 (wireless communication interface 1555) to the RRH 1560.
  • the connection interface 1557 may also be a communication module used to connect the base station device 1550 (wireless communication interface 1555) to the communication in the above-mentioned high-speed line of the RRH 1560.
  • RRH 1560 includes a connection interface 1561 and a wireless communication interface 1563.
  • connection interface 1561 is an interface for connecting the RRH 1560 (wireless communication interface 1563) to the base station device 1550.
  • the connection interface 1561 may also be a communication module used for communication in the above-mentioned high-speed line.
  • Wireless communication interface 1563 transmits and receives wireless signals via antenna 1540.
  • Wireless communication interface 1563 may generally include RF circuitry 1564, for example.
  • RF circuitry 1564 may include, for example, mixers, filters, and amplifiers, and transmit and receive wireless signals via antenna 1540 .
  • FIG. 9 shows an example in which one RF circuit 1564 is connected to one antenna 1540, the present disclosure is not limited to this illustration, and one RF circuit 1564 can be connected to multiple antennas 1540 at the same time.
  • wireless communication interface 1563 may include a plurality of RF circuits 1564.
  • multiple RF circuits 1564 may support multiple antenna elements.
  • FIG. 9 shows an example in which the wireless communication interface 1563 includes a plurality of RF circuits 1564, the wireless communication interface 1563 may also include a single RF circuit 1564.
  • the smart phone 1600 includes a processor 1601, a memory 1602, a storage device 1603, an external connection interface 1604, a camera 1606, a sensor 1607, a microphone 1608, an input device 1609, a display device 1610, a speaker 1611, a wireless communication interface 1612, one or more Antenna switch 1615, one or more antennas 1616, bus 1617, battery 1618, and auxiliary controller 1619.
  • the smart phone 1600 (or processor 1601) here may correspond to the above-mentioned terminal device 300B and/or 1500A.
  • the processor 1601 may be, for example, a CPU or a system on a chip (SoC), and controls functions of the application layer and other layers of the smartphone 1600 .
  • the memory 1602 includes RAM and ROM, and stores data and programs executed by the processor 1601 .
  • the storage device 1603 may include storage media such as semiconductor memory and hard disk.
  • the external connection interface 1604 is an interface for connecting external devices, such as memory cards and Universal Serial Bus (USB) devices, to the smartphone 1600 .
  • the camera 1606 includes an image sensor such as a charge coupled device (CCD) and a complementary metal oxide semiconductor (CMOS) and generates a captured image.
  • Sensors 1607 may include a group of sensors such as measurement sensors, gyroscope sensors, geomagnetic sensors, and acceleration sensors.
  • the microphone 1608 converts the sound input to the smartphone 1600 into an audio signal.
  • the input device 1609 includes, for example, a touch sensor, a keypad, a keyboard, a button, or a switch configured to detect a touch on the screen of the display device 1610, and receives an operation or information input from a user.
  • the display device 1610 includes a screen such as a liquid crystal display (LCD) and an organic light emitting diode (OLED) display, and displays an output image of the smartphone 1600 .
  • the speaker 1611 converts the audio signal output from the smartphone 1600 into sound.
  • the wireless communication interface 1612 supports any cellular communication scheme such as LTE and LTE-Advanced, and performs wireless communication.
  • Wireless communication interface 1612 may generally include, for example, BB processor 1613 and RF circuitry 1614.
  • the BB processor 1613 can perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform various types of signal processing for wireless communication.
  • RF circuitry 1614 may include, for example, mixers, filters, and amplifiers, and transmit and receive wireless signals via antenna 1616.
  • the wireless communication interface 1612 may be a chip module on which the BB processor 1613 and the RF circuit 1614 are integrated.
  • the wireless communication interface 1612 may include multiple BB processors 1613 and multiple RF circuits 1614.
  • FIG. 10 shows an example in which the wireless communication interface 1612 includes multiple BB processors 1613 and multiple RF circuits 1614, the wireless communication interface 1612 may also include a single BB processor 1613 or a single RF circuit 1614.
  • the wireless communication interface 1612 may support additional types of wireless communication methods. solutions, such as short-range wireless communication solutions, near-field communication solutions, and wireless local area network (LAN) solutions.
  • the wireless communication interface 1612 may include a BB processor 1613 and an RF circuit 1614 for each wireless communication scheme.
  • Each of the antenna switches 1615 switches the connection destination of the antenna 1616 between a plurality of circuits included in the wireless communication interface 1612 (for example, circuits for different wireless communication schemes).
  • Antennas 1616 each include a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by wireless communication interface 1612 to transmit and receive wireless signals.
  • smartphone 1600 may include multiple antennas 1616.
  • FIG. 10 shows an example in which smartphone 1600 includes multiple antennas 1616
  • smartphone 1600 may also include a single antenna 1616 .
  • smartphone 1600 may include an antenna 1616 for each wireless communication scheme.
  • the antenna switch 1615 may be omitted from the configuration of the smartphone 1600.
  • the bus 1617 connects the processor 1601, the memory 1602, the storage device 1603, the external connection interface 1604, the camera 1606, the sensor 1607, the microphone 1608, the input device 1609, the display device 1610, the speaker 1611, the wireless communication interface 1612, and the auxiliary controller 1619 to each other. connect.
  • the battery 1618 provides power to the various blocks of the smartphone 1600 shown in Figure 10 via feeders, which are partially shown as dashed lines in the figure.
  • the auxiliary controller 1619 operates the minimum necessary functions of the smartphone 1600 in the sleep mode, for example.
  • the car navigation device 1720 includes a processor 1721, a memory 1722, a global positioning system (GPS) module 1724, a sensor 1725, a data interface 1726, a content player 1727, a storage media interface 1728, an input device 1729, a display device 1730, a speaker 1731, a wireless Communication interface 1733, one or more antenna switches 1736, one or more antennas 1737, and battery 1738.
  • the car navigation device 1720 (or processor 1721) here may correspond to the above-mentioned terminal device 300B and/or 1500A.
  • the processor 1721 may be, for example, a CPU or an SoC, and controls the navigation function and other functions of the car navigation device 1720 .
  • the memory 1722 includes RAM and ROM, and stores data and programs executed by the processor 1721 .
  • the GPS module 1724 measures the location (such as latitude, longitude, and altitude) of the car navigation device 1720 using GPS signals received from GPS satellites.
  • Sensors 1725 may include a group of sensors such as gyroscope sensors, geomagnetic sensors, and air pressure sensors.
  • the data interface 1726 is connected to, for example, the vehicle-mounted network 1741 via a terminal not shown, and acquires data generated by the vehicle (such as vehicle speed data).
  • the content player 1727 reproduces content stored in storage media, such as CDs and DVDs, which are inserted into the storage media interface 1728 .
  • the input device 1729 includes, for example, a touch sensor, a button, or a switch configured to detect a touch on the screen of the display device 1730, and receives an operation or information input from a user.
  • the display device 1730 includes a screen such as an LCD or an OLED display, and displays an image of a navigation function or reproduced content.
  • the speaker 1731 outputs the sound of the navigation function or the reproduced content.
  • the wireless communication interface 1733 supports any cellular communication scheme such as LTE and LTE-Advanced, and performs wireless communication.
  • Wireless communication interface 1733 may generally include, for example, BB processor 1734 and RF circuitry 1735.
  • the BB processor 1734 can perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform various types of signal processing for wireless communications.
  • the RF circuit 1735 may include, for example, a mixer, filter, and amplifier, and transmit and receive wireless signals via the antenna 1737.
  • the wireless communication interface 1733 may also be a chip module on which the BB processor 1734 and the RF circuit 1735 are integrated. As shown in FIG.
  • the wireless communication interface 1733 may include a plurality of BB processors 1734 and a plurality of RF circuits 1735.
  • FIG. 11 shows an example in which the wireless communication interface 1733 includes multiple BB processors 1734 and multiple RF circuits 1735, the wireless communication interface 1733 may also include a single BB processor 1734 or a single RF circuit 1735.
  • the wireless communication interface 1733 may support other types of wireless communication schemes, such as short-range wireless communication schemes, near field communication schemes, and wireless LAN schemes.
  • the wireless communication interface 1733 may include a BB processor 1734 and an RF circuit 1735 for each wireless communication scheme.
  • Each of the antenna switches 1736 switches the connection destination of the antenna 1737 between a plurality of circuits included in the wireless communication interface 1733, such as circuits for different wireless communication schemes.
  • Antennas 1737 each include a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by wireless communication interface 1733 to transmit and receive wireless signals.
  • car navigation device 1720 may include multiple antennas 1737 .
  • FIG. 11 shows an example in which the car navigation device 1720 includes multiple antennas 1737, the car navigation device 1720 may also include a single antenna 1737.
  • the car navigation device 1720 may include an antenna 1737 for each wireless communication scheme.
  • the antenna switch 1736 may be omitted from the configuration of the car navigation device 1720.
  • the battery 1738 provides power to the various blocks of the car navigation device 1720 shown in FIG. 11 via feeders, which are partially shown as dashed lines in the figure. Battery 1738 accumulates power provided from the vehicle.
  • the technology of the present disclosure may also be implemented to include a car navigation device 1720, a vehicle network 1741, and a vehicle Onboard system (or vehicle) 1740 of one or more blocks in vehicle module 1742.
  • the vehicle module 1742 generates vehicle data such as vehicle speed, engine speed, and fault information, and outputs the generated data to the in-vehicle network 1741 .
  • a plurality of functions included in one unit in the above embodiments may be implemented by separate devices.
  • multiple functions implemented by multiple units in the above embodiments may be implemented by separate devices respectively.
  • one of the above functions may be implemented by multiple units. Needless to say, such a configuration is included in the technical scope of the present disclosure.
  • steps described in the flowchart include not only processing performed in time series in the stated order but also processing performed in parallel or individually and not necessarily in time series. Furthermore, even in steps processed in time series, it goes without saying that the order can be appropriately changed.
  • An electronic device for a network device side in a wireless communication system comprising a processing circuit system, the processing circuit system being configured as:
  • the information of the formed groups is sent to the at least one primary FL server entity and the FL participant entity corresponding to each primary FL server entity to enable federated learning within each group.
  • processing circuitry is further configured to:
  • each manager device contains a primary FL server entity
  • a terminal device within a predetermined distance from the manager device is determined to contain a FL participant entity corresponding to the primary FL server entity included in the manager device.
  • the FL participant entity of the group uses direct communication to the group through the link between the terminal device and the electronic device.
  • the primary FL server entity uploads the local model related information, so that the primary FL server entity updates the local model of the group's FL participant entities by aggregating the received local model related information.
  • the local model related information is an output result calculated by the FL participant entity based on the local model based on common data issued by the first-level FL server entity.
  • the local global model of the first-level FL server entity is updated by aggregating the received local global model related information.
  • the local global model related information is an output result calculated by the primary FL server entity based on the local global model based on common data issued by the secondary FL server entity.
  • processing circuitry is further configured to:
  • each primary FL server entity Receives from each primary FL server entity a first quantity related to prediction accuracy of the local global model of the terminal device where the entity is located, a second quantity related to channel quality, and a third quantity related to historical and/or future trajectories. any two or more;
  • the local global model of the first-level FL server entity is updated by weighting and aggregating the output results received from the first-level FL server entity with the weight corresponding to the first-level FL server entity.
  • processing circuitry is further configured to:
  • the global model is sent to multiple FL participant entities corresponding to each first-level FL server entity.
  • the information of the formed group includes any one or more of the following: the identifier ID of the terminal device where the primary FL server entity in the group is located and the group group ID.
  • An electronic device used on the user equipment side in a wireless communication system including a processing circuit system, said processing circuit system
  • the management circuit system is configured as:
  • federated learning is performed within the group.
  • the local model of the FL party entity of the group is updated by aggregating the received local model related information.
  • the local model related information is an output result calculated by the FL participant entity based on the local model based on common data issued by the first-level FL server entity.
  • processing circuitry is further configured to:
  • the local model of the FL participant entity of the group is updated by weighting and aggregating the output results received from the FL participant entity with the weight corresponding to the FL participant entity.
  • processing circuitry is further configured to:
  • a weight corresponding to each FL participant entity of the group is received from the electronic device on the network device side, the weight is based on the weight received by the electronic device on the network device side from each FL participant entity of the group. Any two or more of the first quantity related to the local model prediction accuracy of the terminal equipment where the FL participant entity is located, the second quantity related to the channel quality, and the third quantity related to the historical and/or future trajectory are determined. of,
  • the local model of the FL participant entity of the group is updated by weighting and aggregating the output results received from the FL participant entity with the weight corresponding to the FL participant entity.
  • processing circuitry is further configured to:
  • the local global model is obtained by aggregating the received local model related information, where the local global model of the first-level FL server entity is the aggregation result of the local models of the FL participant entities of the group where the first-level FL server entity is located;
  • the local global model is sent to a FL participant entity of another group upon request from that FL participant entity.
  • the local global model is the aggregation result of the local models of the FL participant entities in the group where the first-level FL server entity is located.
  • the output results are sent to other FL participant entities by carrying information for demodulating and decoding the physical side link control channel PSSCH through the SCI and carrying the output results through the PSSCH.
  • the information of the group includes any one or more of the following: the identifier ID of the terminal device where the primary FL server entity in the group is located and the group ID of the group .
  • the electronic device of clause 29, wherein the grouped information further includes: a group ID of the group in an adjacent geographical location.
  • a method for use in a wireless communication system comprising:
  • the information of the formed groups is sent to the at least one primary FL server entity and the FL participant entity corresponding to each primary FL server entity to enable federated learning within each group.
  • a method for use in a wireless communication system comprising:
  • the group includes a first-level FL server entity and multiple FL participant entities corresponding to it, and at least one first-level FL server entity determined by the network device as The FL participants of the secondary FL server entity included in the network device can perform federated learning with the secondary FL server entity;
  • federated learning is performed within the group.
  • a computer-readable storage medium storing one or more instructions which, when executed by one or more processors of an electronic device, cause the electronic device to perform the performance of clause 31 or 32 Methods.

Abstract

本公开涉及用于无线通信系统的电子设备、方法和存储介质。一种包括处理电路系统的网络设备侧的电子设备，所述处理电路系统被配置为：确定至少一个一级联邦学习(FL)服务器实体以及对应的多个FL参与方实体，其中，一个一级FL服务器实体和对应的多个FL参与方实体共同形成一个分组，并且一级FL服务器实体作为电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及将所形成的分组的信息发送到所述至少一个一级FL服务器实体和对应的FL参与方实体，以在每个分组内能够进行联邦学习。由于一级联邦学习服务器实体和分组的引入，可以在分组内快速进行本地模型的更新，使得终端设备能够在更短的时间内得到更准确的模型。

Description

用于无线通信系统的电子设备、方法和存储介质 技术领域
本公开涉及无线通信领域，更具体地，涉及用于在无线通信领域中进行联邦学习的电子设备、方法和存储介质。
背景技术
联邦学习(Federated Learning，FL)由于其在保证数据隐私安全性和合法合规性方面具有独特的优势，并且可以通过多个设备共同建模来提升机器学习模型的效果，已经成为当前非常重要的分布式人工智能框架或分布式机器学习框架。
在传统的联邦学习中，FL服务器(例如云端服务器)可以通过聚合多个终端设备本地训练得到的本地模型来生成一个全局模型。具体而言，FL服务器首先初始化各个终端设备中的本地模型，在每次的迭代运算中，每个终端设备依靠它本地的训练数据对本地模型进行训练学习。然后，终端设备可以通过无线通信系统(例如5G网络)的上行链路将本地模型(具体而言，本地模型的参数，所述参数例如可以以梯度的方式表征)上报给FL服务器。FL服务器聚合来自多个终端设备的本地模型参数(例如对终端设备的本地模型参数求平均)，以得到更新的全局模型。接着，更新后的全局模型(具体而言，全局模型的参数)可以通过无线通信系统的下行链路分发给每个终端设备。于是，每个终端设备根据接收到的全局模型更新本地模型，然后基于更新后的本地模型继续执行下一轮的本地训练。
联邦学习与无线通信网络相结合正逐渐成为未来网络智能化的趋势之一。针对无线通信系统，联邦学习需要进行调整以适应时变的无线信道环境、不稳定的训练数据以及设备的异质性。目前，在无线通信系统下实现联邦学习的过程中，由基站充当联邦学习的服务器，由各终端设备充当联邦学习的参与方，这样基站通过聚合来自各终端设备的本地模型，可以聚合得到全局模型，并将全局模型发送给各终端设备以更新其本地模型。
然而，现有的联邦学习仅仅涉及两层架构，一层是分发全局模型并对上传的本地模 型进行聚合的服务器，一层是作为联邦学习参与方的终端设备。这样的两层结构要求各FL参与方与FL服务器直接进行通信，考虑到无线信道的时变特性和不稳定性，服务器和参与方之间的通信的延时可能较大，使得根据本地模型聚合得到全局模型的时间较长，从而参与方不能较快地得到更准确的模型，由此可能影响根据本地模型的预测处理。
因此，希望能够提供一种全新的联邦学习架构，使得终端设备能够在更短的时间内得到比仅基于本地数据得到的本地模型更准确的模型。
发明内容
本公开的一个方面涉及一种用于无线通信系统中的网络设备侧的电子设备。根据一个实施例，该电子设备可以包括处理电路系统，该处理电路系统可以被配置为：确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为所述电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以在每个分组内能够进行联邦学习。
本公开的另一个方面涉及一种用于无线通信系统中的用户设备侧的电子设备。根据一个实施例，该电子设备可以包括处理电路系统，该处理电路系统可以被配置为：从网络设备侧的电子设备接收所述用户设备侧的电子设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备侧的电子设备确定的至少一个一级FL服务器实体作为所述网络设备侧的电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及基于该分组的信息，在该分组内进行联邦学习。
本公开的再一个方面涉及一种在无线通信系统中使用的方法。根据一个实施例，该方法可以包括：确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为网络侧设备包 含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以使在每个分组内能够进行联邦学习。
本公开的又一个方面涉及一种在无线通信系统中使用的方法。在一个实施例中，该方法可以包括：从网络设备接收终端设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备确定的至少一个一级FL服务器实体作为所述网络设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及基于该分组的信息，在该分组内进行联邦学习。
本公开的又一个方面涉及存储有一个或多个指令的计算机可读存储介质。在一些实施例中，该一个或多个指令可以在由电子设备的一个或多个处理器执行时使该电子设备执行上述方法。
提供上述概述是为了总结一些示例性的实施例，以提供对本文所描述的主题的各方面的基本理解。因此，上述特征仅仅是例子并且不应该被解释为以任何方式缩小本文所描述的主题的范围或精神。本文所描述的主题的其他特征、方面和优点将从以下结合附图描述的具体实施方式而变得明晰。
附图说明
当结合附图考虑实施例的以下具体描述时，可以获得对本公开内容更好的理解。在各附图中使用了相同或相似的附图标记来表示相同或者相似的部件。各附图连同下面的具体描述一起包含在本说明书中并形成说明书的一部分，用来例示说明本公开的实施例和解释本公开的原理和优点。其中：
图1是相关技术中的无线通信网络中的联邦学习的过程的示意图；
图2是根据本公开实施例的联邦学习架构的示意图；
图3是根据本公开实施例的在本文提出的联邦学习架构中由网络设备执行的方法的流程图；
图4是根据本公开实施例的在本文提出的联邦学习架构中由终端设备执行的方法的 流程图；
图5是根据本公开实施例的在本文提出的联邦学习架构中在联邦学习过程中采用的方法的流程图；
图6是根据本公开实施例的在本文提出的联邦学习架构中的信息交换示例的时序图；
图7是作为本公开的实施例中可采用的信息处理设备的个人计算机的示例结构的框图；
图8是示出可以应用本公开的技术的gNB的示意性配置的第一示例的框图；
图9是示出可以应用本公开的技术的gNB的示意性配置的第二示例的框图；
图10是示出可以应用本公开的技术的智能电话的示意性配置的示例的框图；以及
图11是示出可以应用本公开的技术的汽车导航设备的示意性配置的示例的框图。
虽然在本公开内容中所描述的实施例可能易于有各种修改和另选形式，但是其具体实施例在附图中作为例子示出并且在本文中被详细描述。但是，应当理解，附图以及对其的详细描述不是要将实施例限定到所公开的特定形式，而是相反，目的是要涵盖属于权利要求的精神和范围内的所有修改、等同和另选方案。
具体实施方式
以下描述根据本公开的设备和方法等各方面的代表性应用。这些例子的描述仅是为了增加上下文并帮助理解所描述的实施例。因此，对本领域技术人员而言明晰的是，以下所描述的实施例可以在没有具体细节当中的一些或全部的情况下被实施。在其他情况下，众所周知的过程步骤没有详细描述，以避免不必要地模糊所描述的实施例。其他应用也是可能的，本公开的方案并不限制于这些示例。
首先参考图1描述在相关技术中如何在无线通信网络中实现联邦学习。
如图1所示，FL服务器(也可以被成为FL训练服务器、云端服务器等)110和多个终端设备A至E共同进行联邦学习，以通过FL服务器110聚合终端设备A至E在本地数据上训练得到的本地模型而得到全局模型，并通过下发全局模型来对终端设备中的本地模型进行更新。FL服务器110可以被设置在基站中，也可以被设置在其它网络设备 中。当FL服务器110被设置在基站中时，终端设备A至E是可以与它无线通信的终端设备，包括诸如智能电话、台式计算机之类的任何形式的用户设备。
在联邦学习的每次迭代中，终端设备A至E向FL服务器110发送训练资源报告，包括它们各自可用的计算资源、无线信道环境、地理位置等。FL服务器110根据接收到的报告在终端设备A至E中进行设备选择，以确定本地迭代中参与模型训练的终端设备。由于无线通信系统中的终端设备A至E除了执行联邦学习任务之外，还可能有其他的数据业务需要进行上行链路传输，因此，当这些数据业务有比较高的优先级同时对于时延的忍受度很低时，这些数据的传输会影响终端设备上传本地训练模型的能力，使得该终端设备此时不是很适合参与联邦学习，导致FL服务器根据终端设备的报告不会选择该终端设备。所以，在终端设备选择的时候需要考虑终端设备的本地模型上传和终端设备的业务数据上传之间的平衡。
在图1所示的示例中，FL服务器110在第N次迭代中选择终端设备A、C和D进行联邦学习。接下来，FL服务器110向选择出的终端设备A、C和D分发全局模型和训练参数配置。然后，终端设备A、C和D将本地模型更新为与全局模型一致，并基于本地数据对本地模型进行训练。接着，终端设备A、C和D将训练结果(例如本地模型的参数)报告给FL服务器110。FL服务器110根据接收到的训练结果进行聚合，以得到新的全局模型。至此，第N次迭代过程结束。在第N+1次迭代中，重复与第N次迭代相似的过程。例如，在图1所示的示例中，FL服务器110根据重新上报的训练资源报告，选择终端设备B、C和E进行联邦学习。
尽管在图1的示例中，在每次迭代开始时都需要选择参与联邦学习的终端设备，但是如果终端设备的条件(例如终端设备的计算资源、无线信道环境等)没有发生实质变化，那么FL服务器选择终端设备的操作和下发训练参数配置的操作不需要在每次迭代时都进行。此外，如果一个终端设备一次或多次跳过联邦学习的模型聚合迭代，那么可能会影响联邦学习模型的准确性，因此，可以随着时间的流逝交替安排参与联邦学习的终端设备，使得能够尽量保证独立同分布的采样效果，给予所有终端设备贡献全局模型的平等机会。
由于在相关技术的联邦学习中仅仅涉及两层架构，使得在每次模型更新过程中都需要基站和终端设备经由时变无线信道进行直接通信，因此在本地模型的更新速度方面可 能存在一定的改进空间；并且由于每次选择参与联邦学习的终端设备可能不同且不全面，存在全局模型没有考虑到所有终端设备基于本地数据得到的本地模型的情况，因此模型的准确性方面也存在一定的改进空间。在下文中，将详细描述本公开实施例提供的全新的联邦学习架构。
在图2中示出了根据本公开实施例的联邦学习架构200的示意图。该结构分为三层，作为低层的第一层包含FL参与方实体，作为中间层的第二层(也可以被认为是执行初步聚合的层)包含一级FL服务器实体，作为高层的第三层(也可以被认为是执行精细聚合的层)包含二级FL服务器实体。应该理解的是，本文中所称的“实体”可以是物理上存在于诸如基站、用户设备、终端设备、服务器设备之类的有形设备中的硬件电路或电路系统或者有形设备本身(即，以硬件形式实现)，也可以是存储在可执行无线或网络通信的设备的存储器中并在设备的处理器上执行的计算机可执行指令或程序代码(即，以软件形式实现)。换句话说，“实体”可以是物理实体，也可以是逻辑实体。当提到“某设备包含的实体”时，可以指该实体是该设备的硬件组成部分或该设备本身，也可以指该实体是在该设备中运行的计算机可执行指令或程序代码。具体而言，一个FL参与方实体可以对应于一个终端设备或包括在该终端设备中，一级FL服务器实体可以对应于一个终端设备或包括在该终端设备中，二级FL服务器实体可以对应于一个网络侧的设备(例如，基站、云端服务器或核心网中的控制设备)或包括在该网络侧的设备中。当某实体位于某一层时，其对应的设备也将位于该层。对于一个终端设备而言，它可能仅包含一个FL参与方实体，也可能仅包含一个一级FL服务器实体，还可能同时包含一个FL参与方实体和一个一级FL服务器实体。尽管网络侧的设备可以是基站、服务器设备或核心网中的控制设备，但是为了描述的简便，下文中以基站位于第三层为例进行描述，本领域技术人员可以理解当网络侧的设备为其它形式的设备时，可以用其替代基站来实施本公开实施例的方案。
在图2的架构中，一个一级FL服务器实体和与其对应的多个FL参与方实体构成一个分组，在该分组中可以进行联邦学习。此时，一级FL服务器实体作为传统联邦学习的服务器而FL参与方实体作为传统联邦学习的参与方来共同进行联邦学习。图2示出了三个分组，在每个分组中包含一个由图中的三角形指示的一级FL服务器实体和多个由图中的圆形指示的FL参与方实体。这里的三个分组的数量和分组中的FL参与方实体的数量 仅仅是作为例子给出的，本领域技术人员可以理解，在基站的覆盖范围内也可以存在更多或更少的分组，每个分组中可以存在更多或更少的FL参与方实体。尽管在图2中所有的终端设备都被分组，但是可能存在终端设备未被分组的情况。在这种情况下，未被分组的终端设备包含一级FL服务器实体，位于架构的第二层。
第二层的一级FL服务器实体和第三层中的二级FL服务器实体也可以进行联邦学习。此时，二级FL服务器实体作为传统联邦学习的服务器而一级FL服务器实体作为传统联邦学习的参与方来共同进行联邦学习。
由此，第一层和第二层可以共同进行联邦学习，第二层和第三层可以共同进行联邦学习，使得联邦学习的架构能够扩展到多层。这样，位于第一层的终端设备无需每次与位于第三层的基站进行交互来更新本地模型，它们通过与位于第二层的终端设备通信就可以快速地更新本地模型，使得模型更新的延时得到的改进。此外，位于第三层的基站能够从收集到第一层终端设备的本地模型相关信息的第二层终端设备接收到更全面的模型相关信息，使得全局模型能聚合的模型相关信息更全面，从而全局模型可以更加准确。
为了形成图2所示的架构，需要位于第三层的设备(例如基站)对终端设备进行分组。具体而言，基站从其覆盖范围内的终端设备接收终端设备的处理能力、机器学习能力、地理位置、无线信道质量和移动轨迹信息中的任意一项或多项，例如可以通过Uu链路接收。尽管用于接收上述项的链路在这里被称为Uu链路，但是本领域技术人员可以理解在不同的标准中该链路也可以具有另外的表示，以用这样的链路描述基站和终端设备之间进行的通信。在不同的实施例，基站可以仅接收这些项中的一部分。例如，在不同的实施例中，基站可以仅接收终端设备的处理能力，仅接收终端设备的机器学习能力，仅接收终端设备的地理位置，仅接收终端设备的无线信道质量，仅接收终端设备的移动轨迹信息，或者接收这些项的不同组合。
接着，基站可以根据接收到的项，从它覆盖范围内的终端设备中选择管理者设备，每个管理者设备包含一个一级FL服务器实体。例如，基站可以根据终端设备上报的CPU处理速度，将CPU处理速度排在所有终端设备前预定位(例如第一位、前两位、前三位等)的终端设备确定为管理者设备。基站也可以根据终端设备上报的用于机器学习的计算资源，将分配给机器学习的计算资源较多(例如位于前预定位)的终端设备确定为管理者设备。基站还可以根据终端设备上报的距离基站或其它终端设备的距离，确定各个 终端设备之间的距离，将具有与其距离小于预定值的相邻终端设备最多的至少一个终端设备确定为管理者设备。基站也可以根据终端设备上报的信道质量报告，将信道质量良好(例如信噪比(SNR)小于预定值、参考信号接收功率(RSRP)大于预定值等)的至少一个终端设备或者信道质量排在所有终端设备前预定位(例如第一位、前两位、前三位等)的终端设备确定为管理者设备。此外，基站还可以根据终端设备上报的未来路径规划信息，选取预期要通过某特定地点的终端设备作为管理者设备。本领域技术人员可以理解，上述各示例可以灵活组合，也可以存在其它利用上报的信息选择管理者设备的方式。
选择了管理者设备之后，基站可以将离管理者设备预定距离内的终端设备确定为处于第一层并包含FL参与方实体，并且与该管理者设备共同构成一个分组。在本公开的不同实施例中，一个终端设备可能仅位于一个分组中，也可能一个终端设备位于两个或两个以上的分组中，例如该终端设备与不同的管理者设备之间的距离都在预定距离内。
上文提到了管理者设备可能同时包含一级FL服务器实体和FL参与方实体，这意味着管理者设备实际上可能是一个运算能力强大的终端设备，不仅进行本地模型的训练，而且还作为分组的管理者来聚合分组内的本地模型。此时，该管理者设备所在的分组包含它自己所包含的FL参与方实体。
如果有未被分组的终端设备，例如这些终端设备距离每个管理者设备距离都超过了预定距离，那么这些终端设备被确定为包含一级FL服务器实体，并将与同样位于第二层的其它一级FL服务器实体一起作为二级FL服务器实体的FL参与方来进行联邦学习。
确定了各分组之后，基站可以将分组的信息发送给其覆盖范围内的终端设备。例如，在不同的实施例中，基站可以向一个分组中的每个终端设备发送该分组中的一级FL服务器实体所在的终端设备的标识符ID和该分组的组ID中的任意一项或多项。这意味着，基站可以向分组中的每个终端设备仅发送位于第二层的终端设备的标识符ID，仅发送分组的组ID，或者发送这两者。分组的组ID可以是基站分配的ID，也可以与管理者设备的标识符ID有关，以使得终端设备通过分析这些ID就可以确定所在分组的管理者设备。此外，基站还可以向分组中的每个终端设备发送相邻地理位置中的分组的组ID。例如，基站可以确定与分组中的管理者设备最近的一个或多个其它管理者设备，并将这些管理者设备所在的分组确定为相邻分组，接着可以将相邻分组的组ID也通知给分组中的终端 设备，从而有助于相邻分组之间的通信。
接下来将描述发生在图2所示的第一层和第二层之间的联邦学习，即，在分组内的联邦学习。
在分组内进行联邦学习的过程中，该分组的FL参与方实体向该分组的一级FL服务器实体上传本地模型相关信息，以使一级FL服务器实体通过聚合所接收的本地模型相关信息来更新该分组的FL参与方实体的本地模型。例如，FL参与方实体可以通过FL参与方实体所在的终端设备与一级FL服务器实体所在的电子设备之间的链路使用直连通信将本地模型相关信息上传到一级FL服务器实体。本地模型相关信息可以是与本地模型的参数相关的信息(例如它可以是本地模型的参数的加密版本)，一级FL服务器实体能够从其得到本地模型的参数。
在本发明的另外的实施例中，可以通过传输本地模型的输出结果(也可以被称为预测结果、预测分数等)来进行本地模型的更新。此时，本地模型相关信息可以是FL参与方实体根据一级FL服务器实体下发的共同数据基于本地模型计算得到的输出结果。具体而言，一级FL服务器实体首先可以选择作为FL参与方实体的本地模型输入的一系列共同数据(例如公共数据集的子集)，然后向它所在分组内的所有FL参与方实体发送这些共同数据。接收到共同数据的FL参与方实体将共同数据作为本地模型的输入，依次输入到通过本地数据训练得到的本地模型中，以得到一系列的输出结果，并将这些输出结果返回给一级FL服务器实体。一级FL服务器实体接收到来自分组内的每个参与方实体发送的输出结果之后，通过对这些输出结果进行平均处理，一级FL服务器实体可以确定对应于共同数据的聚合的输出结果。一级FL服务器实体可以将得到的聚合输出结果下发给分组内的每个FL参与方实体，这样每个FL参与方实体可以基于之前接收到的共同数据和与之对应的聚合输出结果，再次训练本地模型以更新本地模型的参数，从而得到更新后的本地模型。该本地模型再次训练所用的数据由于包含来自其它FL参与方实体的信息，因此相比于FL参与方实体仅利用本地数据进行训练的结果更加准确。
一级FL服务器实体得到聚合输出结果的方式除了简单的算术平均的方式之外，还可以采用加权平均的方式。例如，一级FL服务器实体可以从该分组内的每个FL参与方实体接收该FL参与方实体所在的终端设备的一些相关信息，包括终端设备的本地模型预测准确度相关的第一量、终端设备的信道质量相关的第二量和终端设备的历史和/或未来 轨迹相关的第三量中的任意两项或多项，即在不同实施例中可以接收这些项中的至少两项。包含FL参与方实体的终端设备可以计算第一量来表征通过将本地模型与更新后的新的本地模型进行比较而确定的本地模型预测准确度，还可以计算第二量来基于CSI、SNR、RSRP、接收信号的误码率等表征信道质量，例如，第二量可以等于CSI除以接收信号RSRP。此外，该终端设备还可以计算第三量来反映历史和/或未来轨迹，例如如果在预定时间段内移动的距离越长的话，那么第三量可以越大。通过采用以下公式，可以根据第一量、第二量和第三量中的至少两项确定终端设备中包含的FL参与方实体对应的权重：
权重＝alpha*第一量+beta*第二量+gamma*第三量
其中，alpha、beta和gamma是预先确定的常数，其可以根据第一量、第二量和第三量的重要性确定，例如当第一量更重要时分配给它的权重可以更大。当FL参与方实体未向一级FL服务器实体发送第一量、第二量、第三量中的任一项时，上述权重公式相应地省去对应项。计算出每个FL参与方实体对应的权重之后，一级FL服务器实体可以将从每个FL参与方实体接收到的输出结果乘以对应的权重并相加，以加权的方式对来自各FL参与方实体的输出结果进行加权求和来得到聚合的输出结果。
在本公开的实施例中，上述权重除了通过一级FL服务器实体根据FL参与方实体上报的信息确定之外，还可以从基站接收。在这种情况下，需要基站计算FL参与方实体对应的权重。例如，FL参与方实体可以将上述的第一量、第二量和/或第三量直接发送给基站(例如，基站中包含的二级FL服务器实体)，由基站采用与一级FL服务器实体计算权重的方式相同的方式来计算各FL参与方实体的权重，并将计算得到的权重发送到FL参与方实体所在分组的一级FL服务器实体。接收到权重的一级FL服务器实体可以如上所述那样用接收到的权重加权对应FL参与方实体上传的输出结果，以得到聚合的输出结果。
通过加权求和的方式，可以使聚合输出结果更加符合各个FL参与方实体的实际情况，从而更加准确。此外，相比于在联邦学习各方之间传输模型参数的传统方式而言，传输输出结果和聚合的输出结果由于减少了对模型内部状态的暴露而更加安全，并且传输的数据量和通信负载也会明显减小。
另外，FL参与方实体根据一级FL服务器实体下发的共同数据计算得到的输出结果可以以单播或组播的方式发送给位于相同分组的其它FL参与方实体。例如，FL参与方实体可以通过同步信道信息(SCI)向其它FL参与方实体发送请求消息，然后通过SCI 携带用于解调解码物理侧链路控制信道(PSSCH)的信息并通过PSSCH携带输出结果，来向其它FL参与方实体发送输出结果。发送SCI的操作可以在stage-1(阶段1)中完成，发送SCI和PSSSCH的操作可以在stage-2(阶段2)中完成。
在不同分组具有相同的共同数据的情况下，FL参与方实体还可以将输出结果以单播或组播的方式发送到不同分组的FL参与方实体，这样的发送可以通过PC5链路进行。由此，可以更快地共享与FL参与方实体的本地模型有关的信息，促进其它FL参与方实体学习到新的知识。另一方面，FL参与方实体将输出结果上传给一级FL服务器实体的操作可以通过D2D的方式进行。在不同分组具有相同的共同数据的情况下，FL参与方实体还可以将输出结果以D2D的方式上传给不同分组的一级FL服务器实体，以更快地共享模型相关信息。
对于一级FL服务器实体，由于它作为FL参与方实体的FL服务器方存在以共同进行联邦学习，因此一级FL服务器实体具有在分组内有效的全局模型，其是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果，也可以被称为局部全局模型。一级FL服务器实体根据上述的共同数据和聚合的输出结果，也可以对局部全局模型进行训练，由此得到局部全局模型的参数。局部全局模型的参数可以取代于上述下发给FL参与方实体的聚合输出结果而被下发给FL参与方实体以更新其本地模型。换句话说，当FL参与方实体向一级FL服务器实体上传本地模型基于共同数据得到的输出结果时，在本公开的不同实施例中，一级FL服务器实体可以向FL参与方实体下发聚合的输出结果，也可以向FL参与方实体下发局部全局模型的参数，以更新FL参与方实体的本地模型。
要注意的是，由于可能存在未被分组的终端设备，如上所述它位于第二层，并且没有位于其下的其它终端设备，因此它包含的一级FL服务器实体的局部全局模型并不是通过聚合FL参与方实体的本地模型得到的，而是通过与位于第三层的二级FL服务器实体的交互以及本地训练而得到的。例如。二级FL服务器实体可以将全局模型发送给未被分组的终端设备以初始化其中的局部全局模型，并且未被分组的终端设备可以基于本地数据对局部全局模型进行训练来更新局部全局模型以供上传到二级FL服务器实体。
当一级FL服务器实体接收到来自其它分组的FL参与方实体的请求时，一级FL服务器实体可以将局部全局模型的参数下发给该FL参与方实体，以便于其快速更新其本地模型。例如，FL参与方实体正在从一个分组移动到另一个分组时，为了更快地得到新分 组的模型相关信息，FL参与方实体可以根据基站下发的相邻分组的组ID来请求相邻分组的一级FL服务器实体下发局部全局模型。
接下来将描述发生在图2所示的第二层和第三层之间的联邦学习。
在第二层和第三层之间的联邦学习的过程中，基站中包含的二级FL服务器实体从终端设备中包含的一级FL服务器实体接收局部全局模型相关信息，并通过聚合接收到的局部全局模型相关信息来更新一级FL服务器实体的局部全局模型。例如，局部全局模型相关信息可以是与局部全局模型的参数相关的信息(例如它可以是局部全局模型的参数的加密版本)，二级FL服务器实体能够从其得到局部全局模型的参数。
在本发明的另外的实施例中，可以通过传输局部全局模型的输出结果来进行局部全局模型的更新。此时，局部全局模型相关信息可以是一级FL服务器实体根据二级FL服务器实体下发的共同数据基于局部全局模型计算得到的输出结果。具体而言，二级FL服务器实体首先可以选择作为一级FL服务器实体的局部全局模型输入的一系列共同数据(例如公共数据集的子集)，然后向二级FL服务器实体的覆盖范围内的一级FL服务器实体发送这些共同数据。接收到共同数据的一级FL服务器实体将共同数据作为局部全局模型的输入，依次输入到局部全局模型中，由此得到一系列的输出结果，并将这些输出结果返回给二级FL服务器实体。二级FL服务器实体接收到这些输出结果之后，通过对这些输出结果进行平均处理，二级FL服务器实体可以确定对应于共同数据的聚合的输出结果。二级FL服务器实体可以将得到的聚合输出结果下发给每个一级FL服务器实体，这样每个一级FL服务器实体可以基于之前接收到的共同数据和与之对应的聚合输出结果，再次训练局部全局模型以更新局部全局模型的参数，从而得到更新后的局部全局模型。局部全局模型再次训练所用的数据由于包含来自其它一级FL服务器实体的信息(以及由此的其它FL参与方实体的信息)，因此相比于一级FL参与方实体仅利用分组内的本地模型得到的局部全局模型而言可能更加准确。
二级FL服务器实体得到聚合输出结果的方式除了简单的算术平均的方式之外，还可以采用加权平均的方式。例如，二级FL服务器实体可以从每个一级FL服务器实体接收该一级FL服务器实体所在的终端设备的一些相关信息，包括终端设备的局部全局模型预测准确度相关的第一量、终端设备的信道质量相关的第二量和终端设备的历史和/或未来轨迹相关的第三量中的任意两项或多项，即在不同实施例中可以接收这些项中的至少 两项。包含一级FL服务器实体的终端设备可以计算第一量来表征通过将局部全局模型与基于二级FL服务器实体下发的数据更新得到的新的局部全局模型进行比较而确定的局部全局模型预测准确度，还可以计算第二量来基于CSI、SNR、RSRP、接收信号的误码率等表征信道质量，例如，第二量可以等于CSI除以接收信号RSRP。此外，包含一级FL服务器实体的终端设备还可以计算第三量来反映历史和/或未来轨迹，例如如果在预定时间段内移动的距离越长的话，那么第三量可以越大。通过采用以下公式，可以根据第一量、第二量和第三量中的至少两项确定终端设备中包含的一级FL参与方实体对应的权重：
权重＝alpha’*第一量+beta’*第二量+gamma’*第三量
其中，alpha’、beta’和gamma’是预先确定的常数，可以与上述的alpha、beta和gamma相同或不同，其可以根据第一量、第二量和第三量的重要性确定，例如当第一量更重要时分配给它的权重可以更大。当一级FL服务器实体未向二级FL服务器实体发送第一量、第二量、第三量中的任一项时，上述权重公式相应地省去对应项。计算出每个一级FL服务器实体对应的权重之后，二级FL服务器实体可以将从每个一级FL服务器实体接收到的输出结果乘以对应的权重并相加，以加权的方式对来自各一级FL服务器实体的输出结果进行加权求和来得到聚合的输出结果。
通过加权求和的方式，可以使聚合输出结果更加符合各个一级FL服务器实体的实际情况，从而可能更加准确。此外，相比于在联邦学习各方之间传输模型参数的传统方式而言，传输输出结果和聚合的输出结果由于减少了对模型内部状态的暴露而更加安全，并且传输的数据量和通信负载也会明显减小。
对于基站包含的二级FL服务器实体，由于它作为一级FL服务器实体的FL服务器方存在以共同进行联邦学习，因此它具有关于局部全局模型的全局模型。由于二级FL服务器实体也是整个联邦学习架构的最高层的处理实体，因此这样的全局模型在基站的覆盖范围内有效。二级FL服务器实体根据上述的共同数据和聚合的输出结果，可以对全局模型进行训练，由此得到全局模型的参数。全局模型的参数可以取代于上述下发给一级FL服务器实体的聚合输出结果而被下发给一级FL服务器实体以更新其局部全局模型。换句话说，当一级FL服务器实体向二级FL服务器实体上传局部全局模型基于共同数据得到的输出结果时，在本公开的不同实施例中，二级FL服务器实体可以向一级FL服务器实体下发聚合的输出结果，也可以向一级FL服务器实体下发全局模型的参数，以更新 一级FL服务器实体的局部全局模型。
在本公开的实施例中，二级FL服务器实体也可以将全局模型直接下发给FL参与方实体。此时，二级FL服务器实体可以将全局模型的参数分别下发给各个FL参与方实体以直接更新FL参与方实体的本地模型。例如，当上传到二级FL服务器实体的局部全局模型能够指示特定事件(例如交通事故、交通管制等)时，二级FL服务器实体在聚合得到全局模型之后，可以立即将全局模型的参数下发到每个FL参与方实体和每个一级FL服务器实体，以帮助相应的终端设备(例如车辆等)采取避让策略。
以上具体描述了多层的联邦学习架构，并具体描述了不同层之间各实体的交互。由于一级联邦学习服务器实体和分组的引入，可以在分组内快速进行本地模型的更新，使得终端设备能够在更短的时间内得到比仅基于本地数据得到的本地模型更准确的模型，相比于传统终端设备直接与基站交互的方式减小了时延。此外，通过交互基于共同数据得到的输出结果来帮助模型的更新，提高了安全性并减小了数据传输的负担。并且，联邦学习参与方实体、一级联邦学习服务器实体和二级联邦学习服务器实体之间存在灵活的交互方式，使得增加了信息共享的灵活性和自由性，有利于各终端设备掌握更多的关于模型的信息。接下来将描述在本文提出的全新的联邦学习架构中使用的方法。
图3示出了根据本公开实施例的在本文提出的联邦学习架构中可由网络设备执行的方法的流程图。这里的网络设备包含二级FL服务器实体，例如可以是基站、云端服务器、核心网设备等。
在S310中，网络设备确定至少一个一级FL服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为网络侧设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习。在S320中，网络设备将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以使在每个分组内能够进行联邦学习。S310和S320以及其它操作可以参考上述关于图2的描述，再次不再赘述。
图4示出了根据本公开实施例的在本文提出的联邦学习架构中可由终端设备执行的方法的流程图。这里的终端设备包含一级FL服务器实体和/或FL参与方实体，例如可以是用户设备(UE)，也可以是其它具有信息处理能力的设备，例如平板电脑、台式电脑、 计算机、超级计算机等。
在S410中，终端设备从网络设备接收终端设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备确定的至少一个一级FL服务器实体作为所述网络设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习。在S420中，终端设备基于该分组的信息，在该分组内进行联邦学习。S410和S420以及其它操作可以参考上述关于图2的描述，再次不再赘述。
图5是根据本公开实施例的在本文提出的联邦学习架构中在联邦学习过程中采用的方法的流程图。
在S510中，进行上报操作。所有终端设备向云端服务器(例如位于基站中)上报自己的计算资源、无线信道环境、位置信息、路径规划等。
在S520中，进行分组操作。云端服务器基于预定准则选择管理者设备，其包含一级FL服务器实体，并将联邦学习分组的信息告知参与联邦学习的其覆盖范围内的全部终端设备，以使每个终端设备知道自己位于什么分组和/或处于架构的什么层级中。
在S530中，进行模型学习操作。位于第一层的终端设备中的FL参与方实体和没有下层的位于第二层的终端设备中的一级FL服务器实体(如果存在这样的一级FL服务器实体的话)可以根据本地数据分别对本地模型和局部全局模型(其中，这两个模型的初始化都由云端服务器等下发全局模型完成)进行训练。位于第一层的终端设备中的FL参与方实体将本地模型学习的中间结果的预测分数(即，上述基于一级FL服务器实体下发的共同数据得到的本地模型的输出结果)上报给所属分组的位于第二层的终端设备(也可以称为管理者设备)中的一级FL服务器实体，例如以D2D的方式上报。要注意的是，当同一个终端设备既包括FL参与方实体、又包括一级FL参与方实体时，该终端设备可以同时位于第一层和第二层。对于没有下层的终端设备，它将局部全局模型学习的中间结果的预测分数(即，上述基于二级FL服务器实体下发的共同数据得到的本地模型的输出结果)上报给云端服务器。
在S540中，进行模型更新操作。管理者设备中的一级FL服务器实体聚合来自它所负责分组(或相应区域)的不同终端设备的预测分数，并将预测分数的平均值下发给这 些终端设备，以使它们更新各自的本地模型。另外，一级FL服务器实体还与云端服务器中的二级FL服务器实体进行联邦学习，以更新一级FL服务器实体的局部全局模型。并且，二级FL服务器实体在得到全局模型之后，可以根据不同终端设备的请求以点对点的方式向其下发全局模型。
在每次迭代中可以进行上述S510至S540。当需要管理员设备之间的模型迭代时，在每次迭代中，除了进行上述S510至S540之外，还可以在管理者设备之间通过D2D的方式交互各自负责分组(或相应区域)的局部全局模型的预测分数。
图6是根据本公开实施例的在本文提出的联邦学习架构中的信息交换示例的时序图。
如图6所示，各终端设备向云端服务器报告它们各自的能力，云端服务器根据上报的能力对这些终端设备进行分组，从而作为例子，确定出均包含一级FL服务器实体的管理者设备A和B以及包含位于管理者设备A的下层的FL参与方实体的终端设备C。尽管图6中仅示出了两个管理者设备和一个终端设备，但可以存在更多或更少的管理者设备，并且每个管理者设备在其分组中具有一个或多个包含FL参与方实体的终端设备。
终端设备C将预测分数上传给管理者设备A。管理者设备A根据分组内的所有FL参与方实体上传的预测分数进行聚合以得到聚合的预测分数，并通过将聚合的预测分数发送给终端设备C来帮助其更新本地模型。管理者设备A和B还可以向云端服务器发送局部全局模型的预测分数，以使得云端服务器可以根据所有一级FL参与方实体上传的预测分数进行聚合以得到聚合的预测分数。云端服务器可以根据之前为得到预测分数而下发的共同数据以及聚合的预测分数，来得到全局模型，并将全局模型下发给管理者设备A和B以及终端设备C以更新它们各自的模型。
上述的联邦学习架构可以利用在车联网中。例如，一个基站可以服务于多个十字路口，不同十字路口的交通流量不同。根据十字路口所在的位置，基站下发可用于联邦学习的车辆归属于不同区域的信息，同时基站选择每个路口就近的多接入边缘计算设备(MEC)作为区域的管理者设备。行驶在不同路口的车辆将本地学习结果的模型预测结果发送给相应的MEC，MEC融合模型预测结果以更新该区域的所有参与联邦学习的车辆的本地模型。车辆更新本地模型后，接着根据本地数据进行数据训练的迭代。另一方面，MEC将局部全局模型的模型预测结果上传给云端服务器，云端服务器融合多个MEC的结果，并可以生成全局模型。响应于某一路口的车辆A的需求，该全局模型直接 下发给车辆A，因为车辆A正准备驶向另一覆盖范围内的路口，需要该信息作为先验模型信息。当MEC的局部全局模型的预测结果包含对应区域的交通事故时，该区域的局部全局模型快速地被云端服务器计算融合，然后将得到的全局模型下发到覆盖区域的每个参与联邦学习的车辆，以保证任何车辆到该位置可以采取通用的避让策略。
以上分别描述了根据本公开实施例的各示例性电子设备和方法。应当理解，这些电子设备的操作或功能可以相互组合，从而实现比所描述的更多或更少的操作或功能。各方法的操作步骤也可以以任何适当的顺序相互组合，从而类似地实现比所描述的更多或更少的操作。
应当理解，根据本公开实施例的机器可读存储介质或程序产品中的机器可执行指令可以被配置为执行与上述设备和方法实施例相应的操作。当参考上述设备和方法实施例时，机器可读存储介质或程序产品的实施例对于本领域技术人员而言是明晰的，因此不再重复描述。用于承载或包括上述机器可执行指令的机器可读存储介质和程序产品也落在本公开的范围内。这样的存储介质可以包括但不限于软盘、光盘、磁光盘、存储卡、存储棒等等。
另外，应当理解，上述系列处理和设备也可以通过软件和/或固件实现。在通过软件和/或固件实现的情况下，从存储介质或网络向具有专用硬件结构的计算机，例如图7所示的通用个人计算机1300安装构成该软件的程序，该计算机在安装有各种程序时，能够执行各种功能等等。图7是示出作为本公开的实施例中可采用的信息处理设备的个人计算机的示例结构的框图。在一个例子中，该个人计算机可以对应于根据本公开的上述示例性终端设备。
在图7中，中央处理单元(CPU)1301根据只读存储器(ROM)1302中存储的程序或从存储部分1308加载到随机存取存储器(RAM)1303的程序执行各种处理。在RAM 1303中，也根据需要存储当CPU 1301执行各种处理等时所需的数据。
CPU 1301、ROM 1302和RAM 1303经由总线1304彼此连接。输入/输出接口1305也连接到总线1304。
下述部件连接到输入/输出接口1305：输入部分1306，包括键盘、鼠标等；输出部分1307，包括显示器，比如阴极射线管(CRT)、液晶显示器(LCD)等，和扬声器等；存储部分1308，包括硬盘等；和通信部分1309，包括网络接口卡比如LAN卡、调制解调 器等。通信部分1309经由网络比如因特网执行通信处理。
根据需要，驱动器1310也连接到输入/输出接口1305。可拆卸介质1311比如磁盘、光盘、磁光盘、半导体存储器等等根据需要被安装在驱动器1310上，使得从中读出的计算机程序根据需要被安装到存储部分1308中。
在通过软件实现上述系列处理的情况下，从网络比如因特网或存储介质比如可拆卸介质1311安装构成软件的程序。
本领域技术人员应当理解，这种存储介质不局限于图7所示的其中存储有程序、与设备相分离地分发以向用户提供程序的可拆卸介质1311。可拆卸介质1311的例子包含磁盘(包含软盘(注册商标))、光盘(包含光盘只读存储器(CD-ROM)和数字通用盘(DVD))、磁光盘(包含迷你盘(MD)(注册商标))和半导体存储器。或者，存储介质可以是ROM 1302、存储部分1308中包含的硬盘等等，其中存有程序，并且与包含它们的设备一起被分发给用户。
本公开的技术能够应用于各种产品。例如，本公开中提到的基站可以被实现为任何类型的演进型节点B(gNB)，诸如宏gNB和小gNB。小gNB可以为覆盖比宏小区小的小区的gNB，诸如微微gNB、微gNB和家庭(毫微微)gNB。代替地，基站可以被实现为任何其他类型的基站，诸如NodeB和基站收发台(Base Transceiver Station，BTS)。基站可以包括：被配置为控制无线通信的主体(也称为基站设备)；以及设置在与主体不同的地方的一个或多个远程无线头端(Remote Radio Head，RRH)。另外，下面将描述的各种类型的终端均可以通过暂时地或半持久性地执行基站功能而作为基站工作。
例如，本公开中提到的终端设备在一些示例中也称为用户设备，可以被实现为移动终端(诸如智能电话、平板个人计算机(PC)、笔记本式PC、便携式游戏终端、便携式/加密狗型移动路由器和数字摄像装置)或者车载终端(诸如汽车导航设备)。用户设备还可以被实现为执行机器对机器(M2M)通信的终端(也称为机器类型通信(MTC)终端)。此外，用户设备可以为安装在上述终端中的每个终端上的无线通信模块(诸如包括单个晶片的集成电路模块)。
以下将参照图8至图11描述根据本公开的应用示例。
[关于基站的应用示例]
应当理解，本公开中的基站一词具有其通常含义的全部广度，并且至少包括被用于 作为无线通信系统或无线电系统的一部分以便于通信的无线通信站。基站的例子可以例如是但不限于以下：基站可以是GSM系统中的基站收发信机(BTS)和基站控制器(BSC)中的一者或两者，可以是WCDMA系统中的无线电网络控制器(RNC)和Node B中的一者或两者，可以是LTE和LTE-Advanced系统中的eNB，或者可以是未来通信系统中对应的网络节点(例如可能在5G通信系统中出现的gNB，eLTE eNB等等)。本公开的基站中的部分功能也可以实现为在D2D、M2M以及V2V通信场景下对通信具有控制功能的实体，或者实现为在认知无线电通信场景下起频谱协调作用的实体。
第一应用示例
图8是示出可以应用本公开内容的技术的gNB的示意性配置的第一示例的框图。gNB 1400包括多个天线1410以及基站设备1420。基站设备1420和每个天线1410可以经由RF线缆彼此连接。在一种实现方式中，此处的gNB 1400(或基站设备1420)可以对应于上述电子设备300A、1300A和/或1500B。
天线1410中的每一个均包括单个或多个天线元件(诸如包括在多输入多输出(MIMO)天线中的多个天线元件)，并且用于基站设备1420发送和接收无线信号。如图8所示，gNB 1400可以包括多个天线1410。例如，多个天线1410可以与gNB 1400使用的多个频段兼容。
基站设备1420包括控制器1421、存储器1422、网络接口1423以及无线通信接口1425。
控制器1421可以为例如CPU或DSP，并且操作基站设备1420的较高层的各种功能。例如，控制器1421根据由无线通信接口1425处理的信号中的数据来生成数据分组，并经由网络接口1423来传递所生成的分组。控制器1421可以对来自多个基带处理器的数据进行捆绑以生成捆绑分组，并传递所生成的捆绑分组。控制器1421可以具有执行如下控制的逻辑功能：该控制诸如为无线资源控制、无线承载控制、移动性管理、接纳控制和调度。该控制可以结合附近的gNB或核心网节点来执行。存储器1422包括RAM和ROM，并且存储由控制器1421执行的程序和各种类型的控制数据(诸如终端列表、传输功率数据以及调度数据)。
网络接口1423为用于将基站设备1420连接至核心网1424的通信接口。控制器1421可以经由网络接口1423而与核心网节点或另外的gNB进行通信。在此情况下，gNB 1400 与核心网节点或其他gNB可以通过逻辑接口(诸如S1接口和X2接口)而彼此连接。网络接口1423还可以为有线通信接口或用于无线回程线路的无线通信接口。如果网络接口1423为无线通信接口，则与由无线通信接口1425使用的频段相比，网络接口1423可以使用较高频段用于无线通信。
无线通信接口1425支持任何蜂窝通信方案(诸如长期演进(LTE)和LTE-先进)，并且经由天线1410来提供到位于gNB 1400的小区中的终端的无线连接。无线通信接口1425通常可以包括例如基带(BB)处理器1426和RF电路1427。BB处理器1426可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行层(例如L1、介质访问控制(MAC)、无线链路控制(RLC)和分组数据汇聚协议(PDCP))的各种类型的信号处理。代替控制器1421，BB处理器1426可以具有上述逻辑功能的一部分或全部。BB处理器1426可以为存储通信控制程序的存储器，或者为包括被配置为执行程序的处理器和相关电路的模块。更新程序可以使BB处理器1426的功能改变。该模块可以为插入到基站设备1420的槽中的卡或刀片。可替代地，该模块也可以为安装在卡或刀片上的芯片。同时，RF电路1427可以包括例如混频器、滤波器和放大器，并且经由天线1410来传送和接收无线信号。虽然图8示出一个RF电路1427与一根天线1410连接的示例，但是本公开并不限于该图示，而是一个RF电路1427可以同时连接多根天线1410。
如图8所示，无线通信接口1425可以包括多个BB处理器1426。例如，多个BB处理器1426可以与gNB 1400使用的多个频段兼容。如图8所示，无线通信接口1425可以包括多个RF电路1427。例如，多个RF电路1427可以与多个天线元件兼容。虽然图8示出其中无线通信接口1425包括多个BB处理器1426和多个RF电路1427的示例，但是无线通信接口1425也可以包括单个BB处理器1426或单个RF电路1427。
第二应用示例
图9是示出可以应用本公开内容的技术的gNB的示意性配置的第二示例的框图。gNB 1530包括多个天线1540、基站设备1550和RRH 1560。RRH 1560和每个天线1540可以经由RF线缆而彼此连接。基站设备1550和RRH 1560可以经由诸如光纤线缆的高速线路而彼此连接。在一种实现方式中，此处的gNB 1530(或基站设备1550)可以对应于上述电子设备300A、1300A和/或1500B。
天线1540中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)并且用于RRH 1560发送和接收无线信号。如图9所示，gNB 1530可以包 括多个天线1540。例如，多个天线1540可以与gNB 1530使用的多个频段兼容。
基站设备1550包括控制器1551、存储器1552、网络接口1553、无线通信接口1555以及连接接口1557。控制器1551、存储器1552和网络接口1553与参照图8描述的控制器1421、存储器1422和网络接口1423相同。
无线通信接口1555支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且经由RRH 1560和天线1540来提供到位于与RRH 1560对应的扇区中的终端的无线通信。无线通信接口1555通常可以包括例如BB处理器1556。除了BB处理器1556经由连接接口1557连接到RRH 1560的RF电路1564之外，BB处理器1556与参照图8描述的BB处理器1426相同。如图9所示，无线通信接口1555可以包括多个BB处理器1556。例如，多个BB处理器1556可以与gNB 1530使用的多个频段兼容。虽然图9示出其中无线通信接口1555包括多个BB处理器1556的示例，但是无线通信接口1555也可以包括单个BB处理器1556。
连接接口1557为用于将基站设备1550(无线通信接口1555)连接至RRH 1560的接口。连接接口1557还可以为用于将基站设备1550(无线通信接口1555)连接至RRH 1560的上述高速线路中的通信的通信模块。
RRH 1560包括连接接口1561和无线通信接口1563。
连接接口1561为用于将RRH 1560(无线通信接口1563)连接至基站设备1550的接口。连接接口1561还可以为用于上述高速线路中的通信的通信模块。
无线通信接口1563经由天线1540来传送和接收无线信号。无线通信接口1563通常可以包括例如RF电路1564。RF电路1564可以包括例如混频器、滤波器和放大器，并且经由天线1540来传送和接收无线信号。虽然图9示出一个RF电路1564与一根天线1540连接的示例，但是本公开并不限于该图示，而是一个RF电路1564可以同时连接多根天线1540。
如图9所示，无线通信接口1563可以包括多个RF电路1564。例如，多个RF电路1564可以支持多个天线元件。虽然图9示出其中无线通信接口1563包括多个RF电路1564的示例，但是无线通信接口1563也可以包括单个RF电路1564。
[关于用户设备的应用示例]
第一应用示例
图10是示出可以应用本公开内容的技术的智能电话1600的示意性配置的示例的框图。智能电话1600包括处理器1601、存储器1602、存储装置1603、外部连接接口1604、摄像装置1606、传感器1607、麦克风1608、输入装置1609、显示装置1610、扬声器1611、无线通信接口1612、一个或多个天线开关1615、一个或多个天线1616、总线1617、电池1618以及辅助控制器1619。在一种实现方式中，此处的智能电话1600(或处理器1601)可以对应于上述终端设备300B和/或1500A。
处理器1601可以为例如CPU或片上系统(SoC)，并且控制智能电话1600的应用层和另外层的功能。存储器1602包括RAM和ROM，并且存储数据和由处理器1601执行的程序。存储装置1603可以包括存储介质，诸如半导体存储器和硬盘。外部连接接口1604为用于将外部装置(诸如存储卡和通用串行总线(USB)装置)连接至智能电话1600的接口。
摄像装置1606包括图像传感器(诸如电荷耦合器件(CCD)和互补金属氧化物半导体(CMOS))，并且生成捕获图像。传感器1607可以包括一组传感器，诸如测量传感器、陀螺仪传感器、地磁传感器和加速度传感器。麦克风1608将输入到智能电话1600的声音转换为音频信号。输入装置1609包括例如被配置为检测显示装置1610的屏幕上的触摸的触摸传感器、小键盘、键盘、按钮或开关，并且接收从用户输入的操作或信息。显示装置1610包括屏幕(诸如液晶显示器(LCD)和有机发光二极管(OLED)显示器)，并且显示智能电话1600的输出图像。扬声器1611将从智能电话1600输出的音频信号转换为声音。
无线通信接口1612支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且执行无线通信。无线通信接口1612通常可以包括例如BB处理器1613和RF电路1614。BB处理器1613可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行用于无线通信的各种类型的信号处理。同时，RF电路1614可以包括例如混频器、滤波器和放大器，并且经由天线1616来传送和接收无线信号。无线通信接口1612可以为其上集成有BB处理器1613和RF电路1614的一个芯片模块。如图10所示，无线通信接口1612可以包括多个BB处理器1613和多个RF电路1614。虽然图10示出其中无线通信接口1612包括多个BB处理器1613和多个RF电路1614的示例，但是无线通信接口1612也可以包括单个BB处理器1613或单个RF电路1614。
此外，除了蜂窝通信方案之外，无线通信接口1612可以支持另外类型的无线通信方 案，诸如短距离无线通信方案、近场通信方案和无线局域网(LAN)方案。在此情况下，无线通信接口1612可以包括针对每种无线通信方案的BB处理器1613和RF电路1614。
天线开关1615中的每一个在包括在无线通信接口1612中的多个电路(例如用于不同的无线通信方案的电路)之间切换天线1616的连接目的地。
天线1616中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)，并且用于无线通信接口1612传送和接收无线信号。如图10所示，智能电话1600可以包括多个天线1616。虽然图10示出其中智能电话1600包括多个天线1616的示例，但是智能电话1600也可以包括单个天线1616。
此外，智能电话1600可以包括针对每种无线通信方案的天线1616。在此情况下，天线开关1615可以从智能电话1600的配置中省略。
总线1617将处理器1601、存储器1602、存储装置1603、外部连接接口1604、摄像装置1606、传感器1607、麦克风1608、输入装置1609、显示装置1610、扬声器1611、无线通信接口1612以及辅助控制器1619彼此连接。电池1618经由馈线向图10所示的智能电话1600的各个块提供电力，馈线在图中被部分地示为虚线。辅助控制器1619例如在睡眠模式下操作智能电话1600的最小必需功能。
第二应用示例
图11是示出可以应用本公开内容的技术的汽车导航设备1720的示意性配置的示例的框图。汽车导航设备1720包括处理器1721、存储器1722、全球定位系统(GPS)模块1724、传感器1725、数据接口1726、内容播放器1727、存储介质接口1728、输入装置1729、显示装置1730、扬声器1731、无线通信接口1733、一个或多个天线开关1736、一个或多个天线1737以及电池1738。在一种实现方式中，此处的汽车导航设备1720(或处理器1721)可以对应于上述终端设备300B和/或1500A。
处理器1721可以为例如CPU或SoC，并且控制汽车导航设备1720的导航功能和另外的功能。存储器1722包括RAM和ROM，并且存储数据和由处理器1721执行的程序。
GPS模块1724使用从GPS卫星接收的GPS信号来测量汽车导航设备1720的位置(诸如纬度、经度和高度)。传感器1725可以包括一组传感器，诸如陀螺仪传感器、地磁传感器和空气压力传感器。数据接口1726经由未示出的终端而连接到例如车载网络1741，并且获取由车辆生成的数据(诸如车速数据)。
内容播放器1727再现存储在存储介质(诸如CD和DVD)中的内容，该存储介质被插入到存储介质接口1728中。输入装置1729包括例如被配置为检测显示装置1730的屏幕上的触摸的触摸传感器、按钮或开关，并且接收从用户输入的操作或信息。显示装置1730包括诸如LCD或OLED显示器的屏幕，并且显示导航功能的图像或再现的内容。扬声器1731输出导航功能的声音或再现的内容。
无线通信接口1733支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且执行无线通信。无线通信接口1733通常可以包括例如BB处理器1734和RF电路1735。BB处理器1734可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行用于无线通信的各种类型的信号处理。同时，RF电路1735可以包括例如混频器、滤波器和放大器，并且经由天线1737来传送和接收无线信号。无线通信接口1733还可以为其上集成有BB处理器1734和RF电路1735的一个芯片模块。如图11所示，无线通信接口1733可以包括多个BB处理器1734和多个RF电路1735。虽然图11示出其中无线通信接口1733包括多个BB处理器1734和多个RF电路1735的示例，但是无线通信接口1733也可以包括单个BB处理器1734或单个RF电路1735。
此外，除了蜂窝通信方案之外，无线通信接口1733可以支持另外类型的无线通信方案，诸如短距离无线通信方案、近场通信方案和无线LAN方案。在此情况下，针对每种无线通信方案，无线通信接口1733可以包括BB处理器1734和RF电路1735。
天线开关1736中的每一个在包括在无线通信接口1733中的多个电路(诸如用于不同的无线通信方案的电路)之间切换天线1737的连接目的地。
天线1737中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)，并且用于无线通信接口1733传送和接收无线信号。如图11所示，汽车导航设备1720可以包括多个天线1737。虽然图11示出其中汽车导航设备1720包括多个天线1737的示例，但是汽车导航设备1720也可以包括单个天线1737。
此外，汽车导航设备1720可以包括针对每种无线通信方案的天线1737。在此情况下，天线开关1736可以从汽车导航设备1720的配置中省略。
电池1738经由馈线向图11所示的汽车导航设备1720的各个块提供电力，馈线在图中被部分地示为虚线。电池1738累积从车辆提供的电力。
本公开内容的技术也可以被实现为包括汽车导航设备1720、车载网络1741以及车 辆模块1742中的一个或多个块的车载系统(或车辆)1740。车辆模块1742生成车辆数据(诸如车速、发动机速度和故障信息)，并且将所生成的数据输出至车载网络1741。
以上参照附图描述了本公开的示例性实施例，但是本公开当然不限于以上示例。本领域技术人员可在所附权利要求的范围内得到各种变更和修改，并且应理解这些变更和修改自然将落入本公开的技术范围内。
例如，在以上实施例中包括在一个单元中的多个功能可以由分开的装置来实现。替选地，在以上实施例中由多个单元实现的多个功能可分别由分开的装置来实现。另外，以上功能之一可由多个单元来实现。无需说，这样的配置包括在本公开的技术范围内。
在该说明书中，流程图中所描述的步骤不仅包括以所述顺序按时间序列执行的处理，而且包括并行地或单独地而不是必须按时间序列执行的处理。此外，甚至在按时间序列处理的步骤中，无需说，也可以适当地改变该顺序。
虽然已经详细说明了本公开及其优点，但是应当理解在不脱离由所附的权利要求所限定的本公开的精神和范围的情况下可以进行各种改变、替代和变换。而且，本公开实施例的术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
从本文中的描述将认识到的是，本公开实施例可以被配置如下：
1.一种用于无线通信系统中的网络设备侧的电子设备，包括处理电路系统，所述处理电路系统被配置为：
确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为所述电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以在每个分组内能够进行联邦学习。
2.根据条款1所述的电子设备，其中，所述处理电路系统进一步被配置为：
从所述电子设备的覆盖范围内的终端设备接收终端设备的处理能力、机器学习能力、地理位置、无线信道质量和移动轨迹信息中的任意一项或多项；
根据所接收的终端设备的处理能力、地理位置、无线信道质量和移动轨迹信息中的所述任意一项或多项，在所述终端设备中选择至少一个管理者设备，其中，每个管理者设备包含一级FL服务器实体；以及
对于每个管理者设备，将距该管理者设备预定距离内的终端设备确定为包含与该管理者设备包含的一级FL服务器实体对应的FL参与方实体。
3.根据条款2所述的电子设备，其中，至少一个管理者设备还包含FL参与方实体，以及管理者设备包含的一级FL服务器实体和FL参与方实体两者处于同一分组内。
4.根据条款2所述的电子设备，其中，将距每个管理者设备预定距离之外的终端设备确定为包含与一级FL服务器实体一起作为FL参与方与二级服务器实体进行联邦学习的FL实体。
5.根据条款1所述的电子设备，其中，在分组内进行联邦学习的过程中，该分组的FL参与方实体通过终端设备与所述电子设备之间的链路使用直连通信向该分组的一级FL服务器实体上传本地模型相关信息，以使一级FL服务器实体通过聚合所接收的本地模型相关信息来更新该分组的FL参与方实体的本地模型。
6.根据条款5所述的电子设备，其中，所述本地模型相关信息是FL参与方实体根据一级FL服务器实体下发的共同数据基于本地模型计算得到的输出结果。
7.根据条款1所述的电子设备，其中，在所述至少一个一级FL服务器实体和二级FL服务器实体进行联邦学习的过程中，所述处理电路系统进一步被配置为：
从所述至少一个一级FL服务器实体中的每个一级FL服务器实体接收该一级FL服务器实体的局部全局模型相关信息，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果；以及
通过聚合接收到的局部全局模型相关信息来更新一级FL服务器实体的局部全局模型。
8.根据条款7所述的电子设备，其中，所述局部全局模型相关信息是一级FL服务器实体根据二级FL服务器实体下发的共同数据基于局部全局模型计算得到的输出结果。
9.根据条款8所述的电子设备，其中，所述处理电路系统进一步被配置为：
从每个一级FL服务器实体接收该实体所在的终端设备的局部全局模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相关的第三量中的任意两项或多项；以及
基于第一量、第二量和第三量中的所述任意两项或多项，确定该一级FL服务器实体对应的权重，
其中，通过用一级FL服务器实体对应的权重加权聚合从一级FL服务器实体接收到的输出结果来更新一级FL服务器实体的局部全局模型。
10.根据条款9所述的电子设备，其中，所述权重被计算为第一量、第二量和第三量中的至少两项的线性求和。
11.根据条款7所述的电子设备，其中，所述处理电路系统进一步被配置为：
通过聚合接收到的局部全局模型相关信息来得到二级FL服务器实体的全局模型；以及
将全局模型发送给与每个一级FL服务器实体对应的多个FL参与方实体。
12.根据条款1所述的电子设备，其中，所形成的分组的信息包括如下中的任意一项或多项：该分组中的一级FL服务器实体所在的终端设备的标识符ID和该分组的组ID。
13.根据条款12所述的电子设备，其中，所形成的分组的信息进一步包括：相邻地理位置中的分组的组ID。
14.一种用于无线通信系统中的用户设备侧的电子设备，包括处理电路系统，所述处 理电路系统被配置为：
从网络设备侧的电子设备接收所述用户设备侧的电子设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备侧的电子设备确定的至少一个一级FL服务器实体作为所述网络设备侧的电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
基于该分组的信息，在该分组内进行联邦学习。
15.根据条款14所述的电子设备，其中，在该电子设备包含一级FL服务器实体的情况下，所述处理电路系统被配置为：
从该电子设备所在分组的FL参与方实体接收本地模型相关信息；
通过聚合所接收的本地模型相关信息来更新该分组的FL参与方实体的本地模型。
16.根据条款15所述的电子设备，其中，所述本地模型相关信息是FL参与方实体根据一级FL服务器实体下发的共同数据基于本地模型计算得到的输出结果。
17.根据条款16所述的电子设备，其中，所述处理电路系统进一步被配置为：
从该电子设备所在分组的每个FL参与方实体接收该FL参与方实体所在的终端设备的本地模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相关的第三量中的任意两项或多项；以及
基于第一量、第二量和第三量中的所述任意两项或多项，确定该FL参与方实体对应的权重，
其中，通过用FL参与方实体对应的权重加权聚合从FL参与方实体接收到的输出结果来更新该分组的FL参与方实体的本地模型。
18.根据条款16所述的电子设备，其中，所述处理电路系统进一步被配置为：
从所述网络设备侧的电子设备接收与该分组的每个FL参与方实体对应的权重，所述权重是所述网络设备侧的电子设备根据从该分组的每个FL参与方实体接收到的该FL参与方实体所在的终端设备的本地模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相关的第三量中的任意两项或多项确定的，
其中，通过用FL参与方实体对应的权重加权聚合从FL参与方实体接收到的输出结果来更新该分组的FL参与方实体的本地模型。
19.根据条款17或18所述的电子设备，其中，所述权重被计算为第一量、第二量和第三量中的至少两项的线性求和。
20.根据条款15所述的电子设备，其中，所述处理电路系统进一步被配置为：
通过聚合所接收的本地模型相关信息来得到局部全局模型，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果；以及
根据来自其它分组的FL参与方实体的请求，向该FL参与方实体发送局部全局模型。
21.根据条款14所述的电子设备，其中，在该电子设备包含一级FL服务器实体的情况下，所述处理电路系统被配置为：
向二级FL服务器实体发送局部全局模型相关信息，以使二级FL服务器实体通过聚合接收到的局部全局模型相关信息来更新一级FL服务器实体的局部全局模型，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果。
22.根据条款21所述的电子设备，其中，所述局部全局模型相关信息是一级FL服务器实体根据二级FL服务器实体下发的共同数据基于局部全局模型计算得到的输出结果。
23.根据条款22所述的电子设备，其中，所述处理电路系统进一步被配置为：
与其它一级FL服务器实体交换局部全局模型相关信息。
24.根据条款14所述的电子设备，其中，在该电子设备包含FL参与方实体的情况下，所述处理电路系统进一步被配置为：
向该电子设备所在分组的一级FL服务器实体发送本地模型相关信息，以使一级FL服务器通过聚合从该分组的FL参与方实体接收的本地模型相关信息来更新该分组的FL 参与方实体的本地模型。
25.根据条款24所述的电子设备，其中，所述处理电路系统进一步被配置为：
从一级FL服务器实体接收共同数据；
根据共同数据基于本地模型计算输出结果；以及
将输出结果发送给相同分组内的其它FL参与方实体。
26.根据条款25所述的电子设备，其中，所述处理电路系统进一步被配置为：
通过同步信道信息SCI向其它FL参与方实体发送请求消息；以及
通过SCI携带用于解调解码物理侧链路控制信道PSSCH的信息并通过PSSCH携带输出结果，来向其它FL参与方实体发送输出结果。
27.根据条款25所述的电子设备，其中，所述处理电路系统进一步被配置为：
将输出结果通过PC5链路发送给不同分组内的FL参与方实体。
28.根据条款25所述的电子设备，其中，所述处理电路系统进一步被配置为：
将输出结果通过D2D的方式发送给相同分组和/或不同分组中的一级FL服务器实体。
29.根据条款14所述的电子设备，其中，分组的信息包括如下中的任意一项或多项：该分组中的一级FL服务器实体所在的终端设备的标识符ID和该分组的组ID。
30.根据条款29所述的电子设备，其中，分组的信息进一步包括：相邻地理位置中的分组的组ID。
31.一种在无线通信系统中使用的方法，包括：
确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为网络侧设备包含的二级 FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以使在每个分组内能够进行联邦学习。
32.一种在无线通信系统中使用的方法，包括：
从网络设备接收终端设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备确定的至少一个一级FL服务器实体作为所述网络设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
基于该分组的信息，在该分组内进行联邦学习。
33.一种存储有一个或多个指令的计算机可读存储介质，所述一个或多个指令在由电子设备的一个或多个处理器执行时使该电子设备执行根据条款31或32所述的方法。

Claims (33)

  1. 一种用于无线通信系统中的网络设备侧的电子设备，包括处理电路系统，所述处理电路系统被配置为：
    确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为所述电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
    将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以在每个分组内能够进行联邦学习。
  2. 根据权利要求1所述的电子设备，其中，所述处理电路系统进一步被配置为：
    从所述电子设备的覆盖范围内的终端设备接收终端设备的处理能力、机器学习能力、地理位置、无线信道质量和移动轨迹信息中的任意一项或多项；
    根据所接收的终端设备的处理能力、地理位置、无线信道质量和移动轨迹信息中的所述任意一项或多项，在所述终端设备中选择至少一个管理者设备，其中，每个管理者设备包含一级FL服务器实体；以及
    对于每个管理者设备，将距该管理者设备预定距离内的终端设备确定为包含与该管理者设备包含的一级FL服务器实体对应的FL参与方实体。
  3. 根据权利要求2所述的电子设备，其中，至少一个管理者设备还包含FL参与方实体，以及管理者设备包含的一级FL服务器实体和FL参与方实体两者处于同一分组内。
  4. 根据权利要求2所述的电子设备，其中，将距每个管理者设备预定距离之外的终端设备确定为包含与一级FL服务器实体一起作为FL参与方与二级服务器实体进行联邦学习的FL实体。
  5. 根据权利要求1所述的电子设备，其中，在分组内进行联邦学习的过程中，该分组的FL参与方实体通过终端设备与所述电子设备之间的链路使用直连通信向该分组的一级FL服务器实体上传本地模型相关信息，以使一级FL服务器实体通过聚合所接收的本地 模型相关信息来更新该分组的FL参与方实体的本地模型。
  6. 根据权利要求5所述的电子设备，其中，所述本地模型相关信息是FL参与方实体根据一级FL服务器实体下发的共同数据基于本地模型计算得到的输出结果。
  7. 根据权利要求1所述的电子设备，其中，在所述至少一个一级FL服务器实体和二级FL服务器实体进行联邦学习的过程中，所述处理电路系统进一步被配置为：
    从所述至少一个一级FL服务器实体中的每个一级FL服务器实体接收该一级FL服务器实体的局部全局模型相关信息，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果；以及
    通过聚合接收到的局部全局模型相关信息来更新一级FL服务器实体的局部全局模型。
  8. 根据权利要求7所述的电子设备，其中，所述局部全局模型相关信息是一级FL服务器实体根据二级FL服务器实体下发的共同数据基于局部全局模型计算得到的输出结果。
  9. 根据权利要求8所述的电子设备，其中，所述处理电路系统进一步被配置为：
    从每个一级FL服务器实体接收该实体所在的终端设备的局部全局模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相关的第三量中的任意两项或多项；以及
    基于第一量、第二量和第三量中的所述任意两项或多项，确定该一级FL服务器实体对应的权重，
    其中，通过用一级FL服务器实体对应的权重加权聚合从一级FL服务器实体接收到的输出结果来更新一级FL服务器实体的局部全局模型。
  10. 根据权利要求9所述的电子设备，其中，所述权重被计算为第一量、第二量和第三量中的至少两项的线性求和。
  11. 根据权利要求7所述的电子设备，其中，所述处理电路系统进一步被配置为：
    通过聚合接收到的局部全局模型相关信息来得到二级FL服务器实体的全局模型；以 及
    将全局模型发送给与每个一级FL服务器实体对应的多个FL参与方实体。
  12. 根据权利要求1所述的电子设备，其中，所形成的分组的信息包括如下中的任意一项或多项：该分组中的一级FL服务器实体所在的终端设备的标识符ID和该分组的组ID。
  13. 根据权利要求12所述的电子设备，其中，所形成的分组的信息进一步包括：相邻地理位置中的分组的组ID。
  14. 一种用于无线通信系统中的用户设备侧的电子设备，包括处理电路系统，所述处理电路系统被配置为：
    从网络设备侧的电子设备接收所述用户设备侧的电子设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备侧的电子设备确定的至少一个一级FL服务器实体作为所述网络设备侧的电子设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
    基于该分组的信息，在该分组内进行联邦学习。
  15. 根据权利要求14所述的电子设备，其中，在该电子设备包含一级FL服务器实体的情况下，所述处理电路系统被配置为：
    从该电子设备所在分组的FL参与方实体接收本地模型相关信息；
    通过聚合所接收的本地模型相关信息来更新该分组的FL参与方实体的本地模型。
  16. 根据权利要求15所述的电子设备，其中，所述本地模型相关信息是FL参与方实体根据一级FL服务器实体下发的共同数据基于本地模型计算得到的输出结果。
  17. 根据权利要求16所述的电子设备，其中，所述处理电路系统进一步被配置为：
    从该电子设备所在分组的每个FL参与方实体接收该FL参与方实体所在的终端设备的本地模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相 关的第三量中的任意两项或多项；以及
    基于第一量、第二量和第三量中的所述任意两项或多项，确定该FL参与方实体对应的权重，
    其中，通过用FL参与方实体对应的权重加权聚合从FL参与方实体接收到的输出结果来更新该分组的FL参与方实体的本地模型。
  18. 根据权利要求16所述的电子设备，其中，所述处理电路系统进一步被配置为：
    从所述网络设备侧的电子设备接收与该分组的每个FL参与方实体对应的权重，所述权重是所述网络设备侧的电子设备根据从该分组的每个FL参与方实体接收到的该FL参与方实体所在的终端设备的本地模型预测准确度相关的第一量、信道质量相关的第二量、和历史和/或未来轨迹相关的第三量中的任意两项或多项确定的，
    其中，通过用FL参与方实体对应的权重加权聚合从FL参与方实体接收到的输出结果来更新该分组的FL参与方实体的本地模型。
  19. 根据权利要求17或18所述的电子设备，其中，所述权重被计算为第一量、第二量和第三量中的至少两项的线性求和。
  20. 根据权利要求15所述的电子设备，其中，所述处理电路系统进一步被配置为：
    通过聚合所接收的本地模型相关信息来得到局部全局模型，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果；以及
    根据来自其它分组的FL参与方实体的请求，向该FL参与方实体发送局部全局模型。
  21. 根据权利要求14所述的电子设备，其中，在该电子设备包含一级FL服务器实体的情况下，所述处理电路系统被配置为：
    向二级FL服务器实体发送局部全局模型相关信息，以使二级FL服务器实体通过聚合接收到的局部全局模型相关信息来更新一级FL服务器实体的局部全局模型，其中，一级FL服务器实体的局部全局模型是一级FL服务器实体所在分组的FL参与方实体的本地模型的聚合结果。
  22. 根据权利要求21所述的电子设备，其中，所述局部全局模型相关信息是一级FL服务器实体根据二级FL服务器实体下发的共同数据基于局部全局模型计算得到的输出结果。
  23. 根据权利要求22所述的电子设备，其中，所述处理电路系统进一步被配置为：
    与其它一级FL服务器实体交换局部全局模型相关信息。
  24. 根据权利要求14所述的电子设备，其中，在该电子设备包含FL参与方实体的情况下，所述处理电路系统进一步被配置为：
    向该电子设备所在分组的一级FL服务器实体发送本地模型相关信息，以使一级FL服务器通过聚合从该分组的FL参与方实体接收的本地模型相关信息来更新该分组的FL参与方实体的本地模型。
  25. 根据权利要求24所述的电子设备，其中，所述处理电路系统进一步被配置为：
    从一级FL服务器实体接收共同数据；
    根据共同数据基于本地模型计算输出结果；以及
    将输出结果发送给相同分组内的其它FL参与方实体。
  26. 根据权利要求25所述的电子设备，其中，所述处理电路系统进一步被配置为：
    通过同步信道信息SCI向其它FL参与方实体发送请求消息；以及
    通过SCI携带用于解调解码物理侧链路控制信道PSSCH的信息并通过PSSCH携带输出结果，来向其它FL参与方实体发送输出结果。
  27. 根据权利要求25所述的电子设备，其中，所述处理电路系统进一步被配置为：
    将输出结果通过PC5链路发送给不同分组内的FL参与方实体。
  28. 根据权利要求25所述的电子设备，其中，所述处理电路系统进一步被配置为：
    将输出结果通过D2D的方式发送给相同分组和/或不同分组中的一级FL服务器实体。
  29. 根据权利要求14所述的电子设备，其中，分组的信息包括如下中的任意一项或多项：该分组中的一级FL服务器实体所在的终端设备的标识符ID和该分组的组ID。
  30. 根据权利要求29所述的电子设备，其中，分组的信息进一步包括：相邻地理位置中的分组的组ID。
  31. 一种在无线通信系统中使用的方法，包括：
    确定至少一个一级联邦学习(FL)服务器实体以及与每个一级FL服务器实体对应的多个FL参与方实体，其中，一个一级FL服务器实体和与它对应的多个FL参与方实体共同形成一个分组，并且所述至少一个一级FL服务器实体作为网络侧设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
    将所形成的分组的信息发送到所述至少一个一级FL服务器实体和与每个一级FL服务器实体对应的FL参与方实体，以使在每个分组内能够进行联邦学习。
  32. 一种在无线通信系统中使用的方法，包括：
    从网络设备接收终端设备所在的分组的信息，其中，分组包括一个一级FL服务器实体和与它对应的多个FL参与方实体，以及由所述网络设备确定的至少一个一级FL服务器实体作为所述网络设备包含的二级FL服务器实体的FL参与方能够与二级FL服务器实体进行联邦学习；以及
    基于该分组的信息，在该分组内进行联邦学习。
  33. 一种存储有一个或多个指令的计算机可读存储介质，所述一个或多个指令在由电子设备的一个或多个处理器执行时使该电子设备执行根据权利要求31或32所述的方法。
PCT/CN2023/090868 2022-04-29 2023-04-26 用于无线通信系统的电子设备、方法和存储介质 WO2023208043A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210471942.2 2022-04-29
CN202210471942.2A CN117014449A (zh) 2022-04-29 2022-04-29 用于无线通信系统的电子设备、方法和存储介质

Publications (1)

Publication Number Publication Date
WO2023208043A1 true WO2023208043A1 (zh) 2023-11-02

Family

ID=88517821

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/090868 WO2023208043A1 (zh) 2022-04-29 2023-04-26 用于无线通信系统的电子设备、方法和存储介质

Country Status (2)

Country Link
CN (1) CN117014449A (zh)
WO (1) WO2023208043A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110380917A (zh) * 2019-08-26 2019-10-25 深圳前海微众银行股份有限公司 联邦学习系统的控制方法、装置、终端设备及存储介质
CN110490738A (zh) * 2019-08-06 2019-11-22 深圳前海微众银行股份有限公司 一种混合联邦学习方法及架构
US20200285980A1 (en) * 2019-03-08 2020-09-10 NEC Laboratories Europe GmbH System for secure federated learning
CN112636989A (zh) * 2020-12-31 2021-04-09 中国农业银行股份有限公司 一种联邦学习通信方法及装置
CN114116198A (zh) * 2021-10-21 2022-03-01 西安电子科技大学 一种移动车辆的异步联邦学习方法、系统、设备及终端
CN114363911A (zh) * 2021-12-31 2022-04-15 哈尔滨工业大学（深圳） 一种部署分层联邦学习的无线通信系统及资源优化方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200285980A1 (en) * 2019-03-08 2020-09-10 NEC Laboratories Europe GmbH System for secure federated learning
CN110490738A (zh) * 2019-08-06 2019-11-22 深圳前海微众银行股份有限公司 一种混合联邦学习方法及架构
CN110380917A (zh) * 2019-08-26 2019-10-25 深圳前海微众银行股份有限公司 联邦学习系统的控制方法、装置、终端设备及存储介质
CN112636989A (zh) * 2020-12-31 2021-04-09 中国农业银行股份有限公司 一种联邦学习通信方法及装置
CN114116198A (zh) * 2021-10-21 2022-03-01 西安电子科技大学 一种移动车辆的异步联邦学习方法、系统、设备及终端
CN114363911A (zh) * 2021-12-31 2022-04-15 哈尔滨工业大学（深圳） 一种部署分层联邦学习的无线通信系统及资源优化方法

Also Published As

Publication number Publication date
CN117014449A (zh) 2023-11-07

Similar Documents

Publication Publication Date Title
Zhao et al. A novel cost optimization strategy for SDN-enabled UAV-assisted vehicular computation offloading
US20210345141A1 (en) Electronic device and method for wireless communication, and computer-readable storage medium
JP2019024248A (ja) 移動通信ネットワークにおける近隣サービスを実現する方法及び装置
CN110401918A (zh) 一种通信方法及设备
WO2014063568A1 (zh) 缓存状态报告发送与接收方法、用户设备和基站
US20190124581A1 (en) Device, method and user equipment in a wireless communication system
US10849129B2 (en) Communication control apparatus, communication control method and terminal apparatus
CN110366112A (zh) 一种定位方法及相关设备
WO2021047359A1 (zh) 路径规划方法及通信装置
US20230107308A1 (en) Apparatus, method, and storage medium for federated learning
WO2018223983A1 (zh) 频谱管理装置和方法、频谱协调装置和方法以及电子设备
WO2019206061A1 (zh) 用于无线通信系统的电子设备、方法和存储介质
US10951289B2 (en) Electronic device, method applied to electronic device, and data processing device for reusing idle CSI-RS ports of an adjacent cell
WO2023208043A1 (zh) 用于无线通信系统的电子设备、方法和存储介质
WO2023280032A1 (zh) 用于无线通信系统的电子设备、方法和存储介质
US10862561B2 (en) Electronic device and method for network control terminal and network node
WO2024027676A1 (zh) 用于分层联邦学习网络中的切换的装置、方法和介质
Lee et al. Vehicles clustering for low-latency message dissemination in VANET
WO2023185562A1 (zh) 用于无线通信的电子设备和方法、计算机可读存储介质
WO2021185129A1 (zh) 用于无线通信的电子设备和方法、计算机可读存储介质
WO2024012319A1 (zh) 用于无线通信的电子设备和方法、计算机可读存储介质
WO2023179262A1 (zh) 小区信息的配置方法、装置、可读存储介质及芯片系统
WO2023185566A1 (zh) 用于无线通信的方法和电子设备以及计算机可读存储介质
CN108141859A (zh) 信息处理设备和通信系统
CN109906646A (zh) 信息传输方法、基站和终端设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23795475

Country of ref document: EP

Kind code of ref document: A1