Intelligent selection training method and device for vertical federal learning client

Abstract

The invention discloses an intelligent selection training method and device for a vertical federated learning client, and the method comprises the steps: initiating a federated learning task, determining a task target and a task demand of the federated learning task, and generating a corresponding task data package; state information of the FL clients is obtained, and the FL clients meeting task requirements are screened out according to the state information; scoring each type of state information in the FL clients meeting the task requirements, performing weighted calculation on each type of state information to obtain a total evaluation score of each FL client, sorting the FL clients according to the total evaluation scores, and selecting the corresponding FL clients to participate in intelligent training; and after the FL client carries out model training, receiving a training result returned by the FL client, updating the global model by using the training result, and carrying out corresponding iterative training until a preset ending condition is reached. Compared with the prior art, the optimal client can be efficiently and accurately screened to participate in federal learning.

Description

Technical Field

[0001] The present invention relates to the fields of communication and artificial intelligence model training, and in particular to intelligent selection training of vertical federated learning clients. Background Art

[0002] With the rapid development of 5G communication technology and artificial intelligence model training, cross-device and cross-network intelligent applications are increasing. In particular, in scenarios where user privacy and data security need to be protected, traditional centralized model training methods face increasing challenges. Federated learning has emerged as the times require. It allows each participating node (such as user equipment, network function modules, etc.) to jointly train a global model without sharing local data, thereby effectively protecting data privacy. However, existing federated learning technologies are mostly focused on horizontal federated learning, that is, the data types of each participating node are the same and the data sets have the same feature dimensions. This approach cannot meet the needs of vertical scenarios that require the integration of different data types (such as images, text, and sensor data). Therefore, vertical federated learning has become a research hotspot. It is suitable for scenarios with different data features but requiring joint modeling. In this context, the network exposure function in the 5G network provides a standardized interface for cross-network function services, which can aggregate and schedule multiple network resources. Therefore, introducing NEF into the client selection process of VFL can make full use of the resource advantages of the 5G network and optimize the federated learning process while ensuring privacy.

[0003] However, the existing vertical federated learning technology has the following shortcomings in terms of client selection. First, the client selection is not accurate enough. The FL server lacks a comprehensive understanding of the real-time resource status, network connection status, and computing power of each client, resulting in the selected client being unable to participate in training efficiently, thereby slowing down the training progress of the overall model. Secondly, the support for privacy protection strategies is insufficient. The existing technology does not fully consider the different requirements of each client for privacy protection strategies (such as differential privacy, data encryption, etc.), resulting in the inability to effectively protect the data privacy of some clients. In addition, resource management is not flexible. Due to the lack of dynamic monitoring of client resources, some clients may encounter resource bottlenecks during the training process, affecting the model convergence speed and training effect.

[0004] Therefore, there is an urgent need for a vertical federated learning client intelligent selection training method and device that can solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a vertical federated learning client intelligent selection training method and device, which uses weighted scoring to comprehensively sort and prioritize FL clients in multiple dimensions to ensure efficient and accurate screening of the best clients to participate in federated learning.

[0006] In order to achieve the above-mentioned object, the present invention provides a vertical federated learning client intelligent selection training method, comprising: initiating a federated learning task, determining the task goal and task requirement of the federated learning task, generating a unique VFL process ID, setting initial model parameters and training parameters to obtain a task data packet including the task goal, task requirement, VFL process ID, initial model parameters and training parameters; obtaining status information of the FL client, wherein the status information includes the network status, data status and computer resources of the client according to the type, determining whether the FL client meets the task requirement according to the status information, and screening out the FL client that meets the task requirement; assigning corresponding weights to the status information according to the type, and screening out the FL client that meets the task requirement. The method further comprises the steps of: scoring each type of status information of the requested FL client, performing weighted calculation on each type of status information to obtain a total evaluation score of each FL client, sorting the FL clients according to the total evaluation score, and selecting the FL client with a high total evaluation score to participate in intelligent training; notifying the selected FL client to start a training task so that the FL client performs model training and returns the training result to the FL server; receiving the training result returned by the FL client, using the training result to update the model parameters of the global model through an aggregation algorithm, and returning the result to the FL client, evaluating and judging whether the model meets a preset end condition, and if so, terminating the training to generate a final model, and otherwise continuing the iterative training.

[0007] Preferably, the task objectives include the model type that determines the model structure and purpose, the performance indicators that determine the model training success criteria, and the data types used for local model training. The task requirements include data resource requirements, and the training parameters include training hyperparameters and training end conditions. Training hyperparameters are parameters that affect the speed and effect of model training, such as learning rate, batch size, and optimization algorithm.

[0008] Preferably, the state information also includes historical performance; the task objectives include privacy protection requirements, the task requirements also include support for privacy protection policies to ensure data training security, and the state information also includes privacy protection information on whether the corresponding privacy protection policies are supported according to the type. This solution can comprehensively examine whether the FL client can protect data security during model training.

[0009] Preferably, before sending the participation notification to the selected FL client, the training target and completion time requirement of the FL client task training are set, and the participation notification also includes the training target and completion time requirement, so that the FL client completes the model training according to the training target and completion time requirement and returns the training result; after calculating the model parameters of the global model.

[0010] Specifically, the model convergence speed is calculated, and the training target and completion time requirement are adjusted according to the model convergence speed. This solution can effectively improve the efficiency of each model training.

[0011] Preferably, the preset end condition is reaching a preset number of iterations or reaching target performance; after evaluating and determining whether the model has reached the preset end condition, the model convergence speed is also calculated, and the preset end condition is adjusted according to the model convergence speed. This solution enables the present invention to uniformly manage the transmission of multiple rounds of training parameters, adjust the overall number of iterations and termination conditions as needed, such as terminating the task in advance when the target performance is reached or the model converges slowly, ensuring that the training process is efficient and resources are used reasonably, and finally generating a global model that meets the task requirements, marking the completion of the task.

[0012] Preferably, when the FL client is performing model training, the status information of the FL client is continuously monitored, and the status information is transmitted to a preset prediction model. The computer resource usage rate of the FL client in the past first preset time period is predicted based on the computer resource usage rate of the FL client in the future second preset time period, and it is determined whether there is a possible resource bottleneck. If so, a FL client that has not performed the training task is selected according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training. This solution introduces a dynamic resource monitoring and FL client replacement mechanism to ensure efficient resource management during the training process and avoid task interruption or performance degradation.

[0013] Preferably, the present invention initiates a federated learning task through a FL server and uses the training results to update the model parameters of the global model through an aggregation algorithm. The FL server sends a task request containing a task data packet through a NEF network element to screen and select FL clients through the NEF network element.

[0014] The present invention also provides a vertical federated learning client intelligent selection training device, including one or more processors, a memory, and one or more programs. The one or more programs are stored in the memory and can be loaded into the one or more processors to execute the vertical federated learning client intelligent selection training method as described above.

[0015] The present invention also provides a vertical federated learning client intelligent selection training device, including a FL server and an NEF network element; the FL server initiates a federated learning task, determines the task goal and task requirement of the federated learning task, generates a unique VFL process ID, sets initial model parameters and training parameters to obtain a task data packet including the task goal, task requirement, VFL process ID, initial model parameters and training parameters, and initiates a task request including the task data packet to the NEF network element; the NEF network element obtains the status information of the FL client, the status information including the network status, data status, computer resources and historical performance of the client according to the type, the NEF network element determines whether the FL client meets the task requirement in the task request according to the status information, and screens out the FL client that meets the task requirement; the NEF network element assigns corresponding weights to the status information according to the type, and selects the FL client that meets the task requirement. The NEF network element scores each type of status information in the FL client, performs weighted calculation on the status information to obtain a total evaluation score of each FL client, sorts the FL clients according to the total evaluation score, and sends the screened FL client table and the scoring and sorting result to the FL server, so that the FL server selects the FL client with a high total evaluation score to participate in the intelligent training; the NEF network element receives the FL client information selected by the FL server, notifies the selected FL client to start the training task, so that the selected FL client performs model training, and returns the training result to the FL server; the FL server receives the training result returned by the FL client, uses the training result to update the model parameters of the global model through an aggregation algorithm, and returns it to the FL client, evaluates and determines whether the model meets the preset end condition, if so, terminates the training to generate a final model, otherwise, continues to perform iterative training.

[0016] Preferably, when the FL client performs model training, the NEF network element continuously monitors the status information of the FL client, transmits the status information to the preset prediction model, predicts the computer resource usage rate of the FL client in the future for the second preset time period based on the computer resource usage rate of the FL client in the past first preset time period, and determines whether there is a possible resource bottleneck. If so, select a FL client that has not performed the training task according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training. This solution introduces a dynamic resource monitoring and FL client replacement mechanism to ensure efficient resource management during the training process and avoid task interruption or performance degradation.

[0017] The present invention also provides a vertical federated learning client intelligent selection training device, including: a task initialization module, which determines the task goal and task requirement of the federated learning task, generates a unique VFL process ID, sets the initial model parameters and training parameters to obtain a task data packet including the task goal, task requirement, VFL process ID, initial model parameters and training parameters, and initiates a task request including the task data packet to the client screening module; a client screening module, which obtains the status information of the FL client after receiving the task request, the status information includes the network status, data status and computer resources of the client according to the type, the NEF network element determines whether the FL client meets the task requirement in the task request according to the status information, and screens out the FL clients that meet the task requirement; a priority sorting module, which sorts the status information according to the type Assign corresponding weights, score each type of status information in the screened FL clients that meet the task requirements, perform weighted calculation on each type of status information to obtain a total evaluation score of each FL client, sort the FL clients according to the total evaluation score, and select the FL clients with high total evaluation scores to participate in intelligent training; a task notification module sends a participation notification to the selected FL client to enable the FL client to perform model training, and returns the training result to the FL server; a model training and iteration module receives the training result returned by the FL client, uses the training result to update the model parameters of the global model through an aggregation algorithm, and returns it to the FL client, evaluates and determines whether the model meets the preset end condition, if so, ends the training to generate a final model, otherwise, continues to perform iterative training.

[0018] Preferably, the vertical federated learning client intelligent selection training device also includes a dynamic monitoring and replacement module. During the training process, the status information of the FL client is continuously monitored, and the status information is transmitted to the preset prediction model. The computer resource utilization rate of the FL client in the past first preset time period is predicted to determine whether there is a possible resource bottleneck. If so, a FL client that has not performed the training task is selected according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training. This solution introduces a dynamic resource monitoring and FL client replacement mechanism to ensure efficient resource management during the training process and avoid task interruption or performance degradation.

[0019] Compared with the prior art, on the one hand, the present invention uses a weighted scoring model to comprehensively sort and prioritize FL clients in multiple dimensions, ensuring efficient and accurate screening of the best clients to participate in federated learning. On the other hand, the present invention transmits all parameters and strategies at one time when sending a task request when the task is started, avoiding repeated notifications, simplifying the task startup process and improving efficiency. On the other hand, the present invention also sorts according to the historical performance of the FL client, which can comprehensively select suitable, excellent and stable FL clients, and timely eliminate the impact of unconsidered information on the model training of the FL client. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural diagram of a vertical federated learning client intelligent selection training device and a flow chart of a vertical federated learning client intelligent selection training method in one embodiment of the present invention.

[0021] Figure 2 It is a structural block diagram of a vertical federated learning client intelligent selection training device in another embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to explain the technical content, structural features, achieved objectives and effects of the present invention in detail, the following is a detailed description in conjunction with the implementation methods and the accompanying drawings.

[0023] refer to Figure 1 The present invention discloses a vertical federated learning client intelligent selection training device, including an FL server and an NEF network element. The vertical federated learning client intelligent selection training device can perform a vertical federated learning client intelligent selection training method, including steps S1 to S9.

[0024] S1, the FL server initiates a federated learning task, determines the task objectives and task requirements of the federated learning task, generates a unique VFL process ID (used to identify the task), sets the initial model parameters and training parameters, and obtains a task data packet including the task objectives, task requirements, VFL process ID, initial model parameters, and training parameters.

[0025] The task objectives include the model type that determines the model structure and purpose, the performance indicators that determine the model training success criteria, and the data types used for local model training. The task requirements include data resource requirements (whether the data held by the client meets the data type and data volume). The training parameters include training hyperparameters and training end conditions. The training end conditions include training rounds and termination conditions.

[0026] The task requirements also include privacy protection policy support to ensure data training security. The initial model data and training parameters are set through a standardized configuration module.

[0027] Refer to Table 1 below, which is a list of specific information of the task data packet in this embodiment:

[0028]

[0029] Table 1

[0030] S2, initiates a task request including a task data packet to the NEF network element.

[0031] In this embodiment, the FL server sends a task request including a task data packet to the NEF network element. After receiving the task request, the NEF network element selects and sorts the corresponding FL client. In this embodiment, the FL server sends the task request to the NEF network element through a standard interface.

[0032] S3, the NEF network element obtains the status information of the FL client, and the status information includes the client's network status, data status, computer resources (CPU usage, memory usage, storage space) and historical performance according to the type. Among them, the NEF network element directly collects the status information of the FL client, and the obtained status information is more comprehensive. Of course, in another embodiment, the FL client can also actively report the status information of the NEF network element itself.

[0033] The status information also includes privacy protection information and historical performance of whether the corresponding privacy protection policy is supported according to the type.

[0034] S4, the NEF network element determines whether the FL client meets the task requirement in the task request according to the status information, and selects the FL client that meets the task requirement.

[0035] In this embodiment, when the NEF network element collects the network status, computer resources and historical representation in the status information of the FL client, it automatically screens the FL clients that do not meet the requirements to ensure that the remaining FL clients have the required data resources, stable network connections and comply with the privacy protection policy (differential privacy or encrypted transmission). In addition, the NEF network element verifies the data availability of the FL client to ensure that they can provide data suitable for federated learning.

[0036] S5, the NEF network element assigns corresponding weights to the status information according to the type, and scores the status information of each type in the FL client that meets the task requirements.

[0037] Refer to Table 2 below, which shows the specific content, weight and scoring criteria of the status information in one embodiment.

[0038]

[0039] Table 2

[0040] This embodiment divides each state information into multiple levels, and each level is given a score. Specifically:

[0041] 1. Resource availability score (score of computer resources): 10 points for CPU usage below 50%, 7 points for 50% to 70%, 5 points for 70% to 90%, and 3 points for more than 90%. 10 points for memory usage below 50%, 7 points for 50% to 70%, 5 points for 70% to 90%, and 3 points for more than 90%. 10 points for free space of storage space greater than 80%, 7 points for 50% to 80%, 5 points for 20% to 50%, and 3 points for less than 20%. The average of the three is the resource availability score.

[0042] 2. Data quality score (data status score): If the client's data fully meets the task requirements and the data volume is sufficient (reaching the expected threshold), 10 points are awarded. If the data partially meets the task requirements, or the data volume is close to the threshold (80% to 100%), 7 points are awarded. If the data volume is small or the data does not fully match the task requirements, 3 points are awarded.

[0043] 3. Network status score: 10 points for network delay less than 50ms, 7 points for 50ms-100ms, 5 points for 100ms-150ms, and 3 points for more than 150ms. 10 points for bandwidth utilization rate less than 50%, 7 points for 50%-70%, 5 points for 70%-90%, and 3 points for more than 90%. Connection stability: 10 points for the number of recent connection interruptions of 0, 7 points for 1-2, and 5 points for 3 or more. The average of the three is taken.

[0044] 4. Privacy protection score (score of privacy protection information): If the client supports differential privacy and strong encrypted transmission, it will get 10 points. If it supports partial privacy protection mechanisms, such as only encrypted transmission but not supporting differential privacy, it will get 7 points. If the privacy protection mechanism is weak, such as only basic transmission encryption, it will get 5 points.

[0045] 5. Historical performance score: If the client's model converges faster than 80% of the task participants in the historical task, it will get 10 points; 7 points in the 60%-80% range; 5 points or lower if it is below 60%. If the accuracy of the model trained by the client is better than most other clients in the historical task, it will get 10 points; if the model accuracy is at a medium level, it will get 7 points; if the accuracy is low, it will get 5 points. In the historical task, the response time is always kept within the specified time, 10 points, occasional delays are 7 points, and frequent delays are 5 points. The average of the three is taken.

[0046] S6, the NEF network element performs weighted calculation on the state information to obtain a total evaluation score of each FL client, and the NEF network element sorts the FL clients according to the total evaluation score. In this embodiment, the weighted scoring model is as follows:

[0047] Total score =

[0048] ω1×resource availability score+ω2×data quality score+ω3×network status score+ω4×

[0049] Privacy protection score + ω5 × historical performance score.

[0050] Among them, ω1, ω2, ω3, ω4, and ω5 are weights of specific types of network status.

[0051] S7, sending the scoring and sorting results and the screened FL client list to the FL server.

[0052] S8, the FL server selects the FL clients with high total evaluation scores, and sends the selected FL client list to the NEF network element, and notifies the related FL clients to participate in the intelligent training through the NEF network element.

[0053] S9, the NEF network element notifies the selected FL client to start the training task, so that the FL client performs model training and returns the training result to the FL server.

[0054] Specifically, the NEF network element sends a participation notification including task requirements to the selected FL client, so that after receiving the participation notification, the FL client can perform model training according to the requirements and local data, and return the training results to the FL server.

[0055] Among them, before the NEF network element sends the participation notification containing the task requirements to the selected FL client, it also sets the training target and completion time requirement of the FL client task training, and the task requirements also include the training target and completion time requirement, so that the FL client completes the model training according to the training target and completion time requirement and returns the training result; after calculating the model parameters of the global model. Preferably, the NEF network element also calculates the model convergence speed, and adjusts the training target and completion time requirement according to the model convergence speed. This solution can effectively improve the efficiency of each model training.

[0056] Among them, the task requirements in the participation notification include: VFL process ID (identifying the current task), initial model parameters (such as weights, biases), data requirements (specific data types required for training), privacy protection requirements (such as specific requirements for encrypted transmission or differential privacy), training goals and completion time requirements (such as maximum training rounds, accuracy requirements).

[0057] When the FL client performs local model training, it uses local data to train the model according to the received task requirements and generates intermediate model parameters. The FL client sends the intermediate model parameters (such as local weight updates and loss values) back to the FL server as training results.

[0058] S10, the FL server receives the training result returned by the FL client, uses the training result to update the model parameters of the global model through an aggregation algorithm (such as weighted average), and returns it to the FL client.

[0059] The FL server aggregates the intermediate model parameters of each FL client and updates the global model based on these parameters. The FL server AF sends the updated global model parameters back to each FL client so that they can be used in the next round of model training.

[0060] S11, the FL server evaluates and determines whether the model has reached the preset end condition. If so, the training is terminated to generate the final model. If not, it continues the iterative training.

[0061] Preferably, the preset end condition is reaching a preset number of iterations or achieving target performance; after evaluating and determining whether the model has reached the preset end condition, the NEF network element also calculates the model convergence speed and adjusts the preset end condition based on the model convergence speed.

[0062] After each round of training, the FL server evaluates the convergence of the model and the task completion conditions. If the termination conditions (such as model accuracy or loss value) are not met, the iterative training continues; if the conditions are met, the training is terminated and the final model is generated. When the training is completed, the FL server AF sends a task termination notification to all participating clients and provides the final global model parameters.

[0063] In the above, the FL server gradually optimizes the global model through multiple rounds of iterations. Each client uses local data to train the model and returns the updated model parameters to the FL server. The FL server generates a new global model through an aggregation algorithm (such as weighted average) and distributes it back to the client for the next round of training. According to the model convergence speed, the FL server dynamically adjusts the number of iterations and termination conditions, such as terminating the task in advance when the target performance is reached or the model converges slowly, to ensure that the training process is efficient and resources are used reasonably. Finally, a global model that meets the task requirements is generated, marking the completion of the task.

[0064] Preferably, when the FL client is performing model training, the NEF network element continuously monitors the status information of the FL client, transmits the status information to the preset prediction model, predicts the computer resource utilization rate of the FL client in the future for the second preset time period based on the computer resource utilization rate of the FL client in the past first preset time period, and determines whether there is a possible resource bottleneck. If so, a FL client that has not performed the training task is selected according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training. This solution introduces a dynamic resource monitoring and FL client replacement mechanism to ensure efficient resource management during the training process and avoid task interruption or performance degradation. Among them, if a certain FL client cannot continue to complete the task, the NEF network element will replace it according to the list of alternative FL clients. The NEF network element then sends a task participation notification to the new FL client and sends the original task requirements. Among them, the NEF network element notifies the FL server of the information to replace the FL client.

[0065] The present invention also discloses a vertical federated learning client intelligent selection training device, including one or more processors, a memory, and one or more programs. The one or more programs are stored in the memory and can be loaded into the one or more processors to execute the vertical federated learning client intelligent selection training method as described above.

[0066] refer to Figure 2 The present invention also discloses a vertical federated learning client intelligent selection training device, including a task initialization module 11, a client screening module 12, a priority sorting module 13, a task notification module 14, and a model training and iteration module 15.

[0067] The task initialization module 11 determines the task objectives and task requirements of the federated learning task, generates a unique VFL process ID, sets the initial model parameters and training parameters, and initiates a task request containing a task data packet to the client screening module 12.

[0068] The client screening module 12 obtains the status information of the FL client, and the status information includes the network status, data status, computer resources and historical performance of the client according to the type. The NEF network element determines whether the FL client meets the task requirements in the task request based on the status information, and screens out the FL clients that meet the task requirements.

[0069] The priority sorting module 13 assigns corresponding weights to the status information according to the type, scores the status information of each type in the FL clients that meet the task requirements, performs weighted calculation on the status information to obtain the total evaluation score of each FL client, sorts the FL clients according to the total evaluation score, and selects the FL clients with high total evaluation scores to participate in intelligent training.

[0070] The task notification module 14 sends a participation notification to enable the FL client to perform model training and return the training result to the FL server.

[0071] Among them, the task requirements in the participation notification include: VFL process ID (identifying the current task), initial model parameters (such as weights, biases), data requirements (specific data types required for training), privacy protection requirements (such as specific requirements for encrypted transmission or differential privacy), training goals and completion time requirements (such as maximum training rounds, accuracy requirements).

[0072] When the FL client receives a participation notification containing task requirements, it performs local model training: according to the received task requirements, it uses local data to train the model and generates intermediate model parameters. The FL client sends the intermediate model parameters (such as local weight updates and loss values) back to the FL server as training results.

[0073] The model training and iteration module 15 receives the training results returned by the FL client, uses the training results to update the model parameters of the global model through an aggregation algorithm (such as weighted average), and returns it to the FL client to evaluate and determine whether the model meets the preset end condition. If so, the training is terminated to generate the final model, otherwise, iterative training continues.

[0074] Preferably, the vertical federated learning client intelligent selection training device also includes a dynamic monitoring and replacement module 16. During the training process, the dynamic monitoring and replacement module 16 continuously monitors the status information of the FL client, transmits the status information to the preset prediction model, predicts the computer resource utilization rate of the FL client in the future for the second preset time period based on the computer resource utilization rate of the FL client in the past first preset time period, and determines whether there is a possible resource bottleneck. If so, select a FL client that has not performed the training task according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training. This solution introduces a dynamic resource monitoring and FL client replacement mechanism to ensure efficient resource management during the training process and avoid task interruption or performance degradation.

[0075] FL server (Federated Learning Server): In vertical federated learning, as the coordination center of the federated learning task, it is responsible for managing the entire process of model training, including client selection, parameter aggregation, and task monitoring. NEF (Network Exposure Function): A network function in the 5G network that provides an access interface for network resources and services. In this method, NEF is responsible for selecting the appropriate FL client according to the requirements of the FL server AF, and performing task management and resource scheduling. VFL (Vertical Federated Learning): A federated learning mode that allows multiple participants to jointly train models without sharing local data, which is particularly suitable for scenarios with different data types. FL client (Federated Learning Client): A node participating in the federated learning task, responsible for local data processing and model training, and returning the training results to the FL server AF. Differential Privacy: A privacy protection strategy that ensures that individual data information is not leaked in the process of sharing model parameters. Data Encryption: Use encryption algorithms to protect data in transit to prevent unauthorized access or tampering. Initial Model Parameters: Global model parameters shared by all FL clients at the beginning of the federated learning task, used for model training locally on the client. Training Hyperparameters: Parameters that affect the speed and effect of model training, such as learning rate, batch size, and optimization algorithm. Weighted Scoring Model: A multi-dimensional scoring system used by NEF when selecting FL clients, which prioritizes the most suitable clients by scoring and ranking dimensions such as resource availability, data quality, and network status. LSTM (Long Short-Term Memory): A recursive neural network (RNN) used to predict future resource usage. Privacy Protection Strategy: Protecting data privacy during model training, usually including technical measures such as differential privacy and data encryption.

[0076] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the scope of the patent application of the present invention are still within the scope covered by the present invention.

Claims

1. A vertical federated learning client intelligent selection training method, characterized by: include: Initiate a federated learning task, determine the task objectives and task requirements of the federated learning task, generate a unique VFL process ID, set the initial model parameters and training parameters, and obtain a task data packet including the task objectives, task requirements, VFL process ID, initial model parameters, and training parameters; Obtaining status information of the FL client, wherein the status information includes the network status, data status, and computer resources of the client according to the type, determining whether the FL client meets the task requirements according to the status information, and selecting the FL client that meets the task requirements; Assigning corresponding weights to the state information according to the type, scoring each type of state information in the FL clients that meet the task requirements, performing weighted calculation on each type of the state information to obtain a total evaluation score of each FL client, sorting the FL clients according to the total evaluation score, and selecting the FL clients with high total evaluation scores to participate in intelligent training; Notify the selected FL client to start the training task, so that the FL client performs model training and returns the training result to the FL server; Receive the training results returned by the FL client, use the training results to update the model parameters of the global model through the aggregation algorithm, and return it to the FL client to evaluate and determine whether the model meets the preset end condition. If so, end the training to generate the final model, otherwise continue iterative training.

2. The vertical federated learning client intelligent selection training method according to claim 1, characterized in that: The task objectives include the model type that determines the model structure and purpose, the performance indicators that determine the success criteria for model training, and the data types used for local model training. The task requirements include data resource requirements. The training parameters include training hyperparameters and training end conditions.

3. The vertical federated learning client intelligent selection training method according to claim 1, characterized in that: The status information also includes historical performance; The task objectives include privacy protection requirements, the task requirements also include support for privacy protection policies to ensure the security of data training, and the status information also includes privacy protection information on whether the corresponding privacy protection policies are supported according to the type.

4. The vertical federated learning client intelligent selection training method according to claim 1, characterized in that: Before sending the participation notification to the selected FL client, the training target and completion time requirement of the FL client task training are also set. The participation notification also includes the training target and completion time requirement, so that the FL client completes the model training according to the training target and completion time requirement and returns the training result.

5. The vertical federated learning client intelligent selection training method according to claim 1, characterized in that: The preset end condition is reaching a preset number of iterations or reaching a target performance; after evaluating whether the model has reached the preset end condition, the model convergence speed is also calculated, and the preset end condition is adjusted according to the model convergence speed.

6. The vertical federated learning client intelligent selection training method according to claim 1, characterized in that: When the FL client is performing model training, the status information of the FL client is continuously monitored, and the status information is transmitted to a preset prediction model. The computer resource utilization rate of the FL client in the past first preset period of time is predicted to determine whether there is a possible resource bottleneck. If so, an FL client that has not performed the training task is selected according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training.

7. A vertical federated learning client intelligent selection training device, characterized by: It includes one or more processors, memories, and one or more programs. The one or more programs are stored in the memories and can be loaded into the one or more processors to execute the vertical federated learning client intelligent selection training method as described in any one of claims 1 to 5.

8. A vertical federated learning client intelligent selection training device, characterized by: Including FL server and NEF network element; The FL server initiates a federated learning task, determines the task objectives and task requirements of the federated learning task, generates a unique VFL process ID, sets initial model parameters and training parameters, obtains a task data packet including the task objectives, task requirements, VFL process ID, initial model parameters and training parameters, and initiates a task request including the task data packet to the NEF network element; The NEF network element obtains the status information of the FL client, wherein the status information includes the network status, data status, computer resources and historical performance of the client according to the type, and the NEF network element determines whether the FL client meets the task requirements in the task request according to the status information, and selects the FL client that meets the task requirements; The NEF network element assigns corresponding weights to the status information according to the type, scores the status information of each type in the FL client that meets the task requirements, performs weighted calculation on the status information to obtain a total evaluation score of each FL client, sorts the FL clients according to the total evaluation score, and sends the screened FL client table and the scoring and sorting results to the FL server, so that the FL server selects the FL client with a high total evaluation score to participate in intelligent training; The NEF network element receives the FL client information selected by the FL server, notifies the selected FL client to start a training task, so that the selected FL client performs model training, and returns the training result to the FL server; The FL server receives the training results returned by the FL client, uses the training results to update the model parameters of the global model through an aggregation algorithm, and returns it to the FL client to evaluate whether the model meets the preset end condition. If so, the training is terminated to generate the final model, otherwise iterative training continues.

9. A vertical federated learning client intelligent selection training device, characterized by: include: The task initialization module determines the task objectives and task requirements of the federated learning task, generates a unique VFL process ID, sets the initial model parameters and training parameters, obtains a task data packet including the task objectives, task requirements, VFL process ID, initial model parameters, and training parameters, and initiates a task request containing the task data packet to the client screening module; The client screening module obtains the status information of the FL client after receiving the task request, wherein the status information includes the network status, data status, and computer resources of the client according to the type, and the NEF network element determines whether the FL client meets the task requirements in the task request according to the status information, and screens out the FL client that meets the task requirements; a priority sorting module, which assigns corresponding weights to the status information according to the type, scores each type of status information among the screened FL clients that meet the task requirements, and performs weighted calculation on each type of status information to obtain a total evaluation score of each FL client, sorts the FL clients according to the total evaluation score, and selects the FL clients with high total evaluation scores to participate in intelligent training; A task notification module sends a participation notification to the selected FL client to enable the FL client to perform model training and return the training result to the FL server; The model training and iteration module receives the training results returned by the FL client, uses the training results to update the model parameters of the global model through an aggregation algorithm, and returns it to the FL client to evaluate and determine whether the model meets the preset end condition. If so, the training is terminated to generate the final model, otherwise iterative training continues.

10. The vertical federated learning client intelligent selection training device according to claim 9, characterized in that: The method further includes a dynamic monitoring and replacement module, which continuously monitors the status information of the FL client during the training process, transmits the status information to a preset prediction model, predicts the computer resource utilization rate of the FL client for a second preset period of time in the future based on the computer resource utilization rate of the FL client for a first preset period of time in the past, and determines whether there is a possible resource bottleneck. If so, an FL client that has not executed the training task is selected according to the ranking of the FL clients to replace the FL client that may have a resource bottleneck for data training.

Images