Asynchronous federated learning method based on queue scheduling
By using a queue scheduling method in asynchronous federated learning, clients and models are grouped by delay and staleness, fast clients help slow clients reduce model staleness, solving the problem of model staleness caused by device heterogeneity, and improving the accuracy and convergence speed of the global model.
Patent Information
- Application Number
- CN202510112096.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-05-27
AI Technical Summary
In asynchronous federated learning, the problem of model staleness caused by device heterogeneity reduces the accuracy of the global model.
The asynchronous federated learning method based on queue scheduling is adopted to divide the client into fast client and slow client, and group management is carried out according to the staleness of the model. The fast client is used to help the slow client reduce the staleness of the model, and model aggregation is performed through the queue scheduling strategy.
It effectively reduces the impact of outdated models on the global model, slows down the training deviation problem, speeds up the convergence time of the global model, and improves the accuracy of the global model.
Smart Images
Figure CN120046755A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of federated learning, and in particular, to an asynchronous federated learning method based on queue scheduling. Background Art
[0002] Federated learning is a distributed machine learning framework. Its core idea is to conduct local model training through each terminal device or node while ensuring local data storage and privacy, and upload model parameters or gradients to a central server for aggregation, thereby completing the training of the global model. However, traditional synchronous federated learning faces an efficiency bottleneck in practical applications. Due to the existence of stragglers, the server has to wait for all clients to complete training. To solve this problem, asynchronous federated learning emerged. It allows each node to upload model updates to the central server at any time after completing local training, without waiting for other nodes to synchronously complete. However, the device heterogeneity of asynchronous federated learning (i.e., the computing capabilities and communication conditions of client devices vary) may lead to varying degrees of staleness in the models uploaded by clients. The aggregation of stale models and the global model will significantly reduce the accuracy of the global model. For example, Figure 1 shows the impact of model staleness on the accuracy of the global model. It can be seen that as the model staleness increases, the accuracy of the global model significantly decreases. To solve the above problems, researchers have proposed various improvement schemes in recent years. One method is to reduce the weight of stale models during the aggregation process and give priority to the contributions of the latest models. This method calculates the staleness of the client model and uses it as a weight parameter for aggregation. However, this scheme may reduce the participation frequency of stale clients, further weakening the server's ability to learn from the data of these clients. In addition, another scheme introduces a semi-asynchronous framework, which alleviates the impact of stale models by adding partial synchronous operations in asynchronous federated learning. When the server detects that the client uploads a stale model, it will actively send the latest global model to replace the client's model. However, this method may lead to waste of resources because the time and computing power previously used by the client for training the model cannot be fully utilized. Summary of the Invention
[0003] The main object of the present invention is to design an asynchronous federated learning method based on queue scheduling, which conducts asynchronous federated learning based on model grouping and queue scheduling. By having fast clients assist slow clients in training models, the staleness of the models is reduced, thereby improving the training efficiency and accuracy of the global model.
[0004] The technical solution of the present invention is as follows: An asynchronous federated learning method based on queue scheduling, during the asynchronous federated training process, divides clients into fast clients and slow clients; divides the models uploaded by clients into high-staleness models and low-staleness models; according to the uploaded local model w iBased on the client type and the model type uploaded by the client, reclassify the local model w i for further classification; the server similarly generates queues corresponding to the local model w i categories to store the corresponding local models. The server selects different aggregation strategies or exchange strategies based on the type of the local model w i so that fast clients can participate in aggregation after reducing the staleness of high-staleness models to the determination requirements of low-staleness models.
[0005] Furthermore, calculate the total delay based on the model training delay and communication delay of the client, and divide the clients into fast clients and slow clients according to the total delay.
[0006] Furthermore, before the server initializes the model for distribution or before the model is distributed after aggregation is completed, the server records the current distribution time for the i-th client according to the current timestamp When the client uploads the model after local training on the model distributed by the server, the server records the current reception time for the i-th client according to the current timestamp According to calculate the delay of client i; record the delay of each client on the server, and calculate the average delay of client i during the asynchronous federated training process of submitting m local models to the server by recording the client delay of client i submitting m local models to the server:
[0007]
[0008] Sort the average delays of k clients according to their magnitudes to obtain the median of the average delays of k clients;
[0009]
[0010] If the average delay of client i is greater than the median T of the average delays threshold , it is classified as a slow client c s , and if the average delay of client i is less than the median T of the average delays threshold , it is classified as a fast client c f .
[0011] Furthermore, based on the difference between the local model uploaded by the client after training and the global model after aggregation on the server, divide the staleness of the local model uploaded by the client:
[0012] When client i uploads the local model w i to the server after completing local training, the server calculates the staleness λ of client i's model i, measure the difference between the local model of client i and the global model w g ; the server calculates the obsolescence of each client's uploaded model and calculates the average obsolescence λ according to avg ; when the obsolescence λ i of client i is greater than the average obsolescence λ avg , it is classified as a highly obsolete model w h ; when the obsolescence λ i of client i is less than the average obsolescence λ avg , it is classified as a lowly obsolete model w l .
[0013] Further, the reclassification principle of the local model w i is as follows:
[0014]
[0015] The server generates four queues corresponding to the categories of the local model w i for storing local models of the category for storing local models of the category ; for storing local models of the category ; for storing local models of the category ; for storing local models of the category .
[0016] Further, the aggregation strategy or exchange strategy is specifically as follows:
[0017] (1) When the local model uploaded by client i is
[0018] (a) When the queue is not an empty queue, take out a model w j from this queue in the order of putting, and send the model w i to client j that uploaded the model w j for training, and send w j to client i that uploaded the local model w i for training;
[0019] (b) When the queue is an empty queue, put the model w i into the queue ;
[0020] (2) When the local model uploaded by client i is
[0021] (a) Queue When the queue is not empty, take out a model w from the queue j and send the model w i to the client j that uploaded the model w for training, and send w j to the client i that uploaded the local model w for training; j (b) Queue i (b) Queue
[0022] (b) Queue When the queue is not empty, take out a model w from the queue j and send the local model w i to the client j that uploaded the model w for training, and send w j to the client i that uploaded the model w for training; j (c) Queue i (c) Queue
[0023] (c) Queue When both and i are empty queues, put the local model w into the queue
[0024] (3) The local model uploaded by client i is
[0025] (a) Queue When the queue is not empty, take out a model w from the queue j and use the Federated Averaging algorithm to aggregate the local model w i , the model w j and the global model w g to obtain a new global model Send to client i and client h for training;
[0026] (b) Queue When the queue is not empty, take out a model w from the queue j and use the Federated Averaging algorithm to aggregate the local model w i , the model w j and the global model w g to obtain a new global model Send to client i and client j for training;
[0027] (c) Queue When both and i are empty queues, put the local model w into the queue
[0028] (4) The model uploaded by client i is
[0029] (a) When the queue is not an empty queue, take out a model w from the queue j , and distribute the local model w i to client j that uploaded model w for training, and distribute w j to client i that uploaded model w for training; j to client i that uploaded model w for training; i for training;
[0030] (b) When the queue is not an empty queue, take out a model w from the queue j , and distribute the local model w i to client j that uploaded model w for training, and distribute w j to client i that uploaded model w for training; j to client i that uploaded model w for training; i for training;
[0031] (c) When the queue is not an empty queue, take out a model w from the queue j , and aggregate the local model w i , model w j and the global model w g using the federated averaging algorithm to obtain a new global model Distribute to client i and client j for training;
[0032] (d) When the queue is not an empty queue, take out a model w from the queue j , and aggregate the local model w i , model w j and the global model w g using the federated averaging algorithm to obtain a new global model Distribute to client i and client h for training;
[0033] (e) If the queue is all empty queues, then put the local model w i into the queue among them.
[0034] Advantages of the present invention: In the environment of asynchronous federated learning, due to the different local training delays and communication delays of clients with different device heterogeneities, the problem of the obsolescence of the models trained on slow clients will occur. The present invention proposes an asynchronous federated learning method based on queue scheduling. By using the client fast-slow division algorithm, clients are divided into fast and slow clients, and by using the model obsolescence division algorithm, models are divided into high and low obsolescence models. Combining the queue scheduling scheme, fast clients can be used to reduce the obsolescence of high-obsolescence models, and low-obsolescence models directly participate in aggregation. This scheme can effectively reduce the impact of obsolete models on the global model, while alleviating the training deviation problem, accelerating the convergence time of the global model, and improving the accuracy of the global model. Description of the Drawings
[0035] Figure 1 It is a comparative experimental diagram of the impact of obsolescence for the present invention;
[0036] Figure 2 It is a diagram of the experimental results of the method of the present invention;
[0037] Figure 3 It is the overall flowchart of the present invention;
[0038] Figure 4 It is the system framework diagram of the present invention. Detailed Embodiments
[0039] The asynchronous federated learning method based on queue scheduling of the present invention can be applied to the field of intelligent driving. In the field of intelligent driving, moving vehicles act as clients. The vehicles can send the local models on the vehicles to the central server deployed by the vehicle enterprise through the signal base station. A large amount of driving data accumulated by the vehicles, including road condition information, driving behavior data, etc., is used as the local data for training the federated learning model. These data contain rich user privacy and enterprise business secrets and cannot be directly shared. Therefore, vehicle enterprises can adopt federated learning technology to use the data of both parties for joint training of intelligent driving models without disclosing the original data. During the federated training process, the central server calculates the time for sending the model to the vehicle through the computing server, and the time for the vehicle to send the model back to the server. By calculating the time difference between the sending time and the sending-back time and according to Formula 3, the vehicle is divided into a fast vehicle or a slow vehicle. The central server calculates the obsolescence of the model uploaded by the vehicle according to Formula 4 and divides the model into a high-obsolescence model or a low-obsolescence model according to Formula 5. According to the fast-slow division of the vehicle and the high-low division of the obsolescence of the model, the queue scheduling asynchronous federated learning method is used for training, and the newly trained model is sent to the vehicle.
[0040] In this way, the model can learn a wider range of road conditions and driving scenarios, improving the safety and accuracy of autonomous driving. For example, its ability to recognize complex road conditions and respond to sudden traffic situations will be significantly enhanced. At the same time, the data privacy and commercial interests of each party are protected.
[0041] (I) Division of fast and slow clients
[0042] Based on a comprehensive consideration of the characteristics of federated training and the characteristics of the client training environment in asynchronous federated learning, an efficient algorithm for dividing fast and slow clients in asynchronous federated learning is designed. According to the sum of the training delay of the client in local model training and the communication delay with the central server, denoted as t i , the clients are specifically divided into fast clients or slow clients. First, before the server initializes the model for distribution or before distributing the model after aggregation, the server records the distribution time of the i-th client based on the current timestamp When the client uploads the model after local training on the model distributed by the server, the server records the reception time of the i-th client based on the current timestamp According to
[0043]
[0044] , the delay of client i is calculated. The delay of each client is recorded on the server. By recording the client delay of client i submitting m local models to the server, the average delay of client i in the asynchronous federated training process of these m model submissions is calculated:
[0045]
[0046] The average delays of k clients are sorted by size to obtain the median of the average delays of k clients
[0047]
[0048] If the average delay of client i is greater than the median T of the average delays threshold then it is classified as a slow client c s , and conversely, if the average delay of client i is less than the median T of the average delays threshold then it is classified as a fast client c f .
[0049] (II) Division based on the degree of model obsolescence
[0050] Based on comprehensively considering the characteristics of federated training and the characteristics of the client training environment in asynchronous federated learning, an efficient algorithm for classifying the degree of obsolescence of the asynchronous federated learning model is designed. According to the difference between the model after the client's federated training and the global model, the model uploaded by the client is specifically classified as a highly obsolete model or a lowly obsolete model. First, when client i finishes local training and uploads the local model w i to the server, the server will calculate the obsolescence degree λ
[0051]
[0052] of client i's model according to i , which measures the difference degree between client i's model and the global model w g . The server will calculate the obsolescence degree of each client's uploaded model, and calculate the average obsolescence degree λ
[0053]
[0054] according to avg . If the obsolescence degree λ i of client i is greater than the average obsolescence degree λ avg , it will be classified as a highly obsolete model w h . On the contrary, if the obsolescence degree λ i of client i is less than the average obsolescence degree λ avg , it will be classified as a lowly obsolete model w l .
[0055] (III) Queue Scheduling Asynchronous Federated Learning
[0056] At the beginning of asynchronous federated training, the server will initialize the global model w g . After initialization, it will communicate with k clients. The server will send the global model w g to k clients and record the sending timestamp for each client . Each client trains the sent model locally. When the training is completed, it will communicate with the server and upload its own local model w i . At this time, the server will record the timestamp when it receives the current model . Calculate the delay t i of client i according to formula (1), and then calculate the average delay of client i using formula (2) . Calculate the median T threshold of the client delay through formula (3). By comparing with T threshold , calculate the currently uploaded model w iClassify the client as a fast client or a slow client based on this. For model w i Calculate the obsolescence λ of the model according to formula (4) i , and use formula (5) to calculate the average obsolescence λ avg , and compare λ i with λ avg to calculate whether the model w uploaded by client i i is a high-obsolescence model or a low-obsolescence model. According to the type of the client that uploads model w i , re-classify model w i as follows
[0057]
[0058] The server will also generate four corresponding queues to store the corresponding models. The server will select different aggregation or exchange strategies according to the type of model w i . The main idea is to use model exchange to reduce the obsolescence of high-obsolescence models to low-obsolescence models by fast clients and then participate in aggregation. The specific rules are as follows:
[0059] (1) If the model uploaded by client i is
[0060] (a) If queue is not an empty queue, a model w j can be taken out of the queue, and the model w i is sent to client j that uploads model w j for training, and w j is sent to client i that uploads model w i for training.
[0061] (b) If queue is an empty queue, then the model w i is put into queue among
[0062] (2) If the model uploaded by client i is
[0063] (a) If queue is not an empty queue, a model w j can be taken out of the queue, and the model w i is sent to client j that uploads model w j for training, and w j is sent to client i that uploads model w i for training.
[0064] (b) If the queue is not an empty queue, a model w j can be taken out from the queue, and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client i that uploaded the model w i for training.
[0065] (c) If the queue and are both empty queues, then the model w i is put into the queue among them
[0066] (3) If the model uploaded by client i is
[0067] (a) If the queue is not an empty queue, a model w j can be taken out from the queue, and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model and is sent to client i and client j for training.
[0068] (b) If the queue is not an empty queue, a model w j can be taken out from the queue, and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model and is sent to client i and client j for training.
[0069] (c) If the queue and are both empty queues, then the model w i is put into the queue among them
[0070] (4) If the model uploaded by client i is
[0071] (a) If the queue is not an empty queue, a model w j can be taken out from the queue, and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client that uploaded the model wi The client i is trained.
[0072] (b) If the queue is not an empty queue, a model w can be taken out of the queue j , and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client i that uploaded the model w i for training.
[0073] (c) If the queue is not an empty queue, a model w can be taken out of the queue j , and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model The is sent to the client i and the client j for training.
[0074] (d) If the queue is not an empty queue, a model w can be taken out of the queue j , and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model The is sent to the client i and the client j for training.
[0075] (e) If the queues are all empty queues, then the model w i is put into the queue .
[0076] The following Algorithm 1 is the pseudocode of the model grouping exchange aggregation algorithm.
[0077]
[0078]
[0079]
[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments
[0081] As Figure 4 shown, Figure 2This is the system framework diagram of the present invention. The system includes a central server and multiple clients. The central server is responsible for distributing the global model, scheduling and aggregating the client models. The clients are responsible for locally training the client models sent by the central server and then asynchronously uploading the models to the central server.
[0082] As Figure 3 shown, an asynchronous federated learning method based on queue scheduling includes the following steps:
[0083] (1) Initialization: The central server initializes the global model;
[0084] (2) Global broadcast: The central server broadcasts the global model to all clients;
[0085] (3) Local model training and uploading: The clients use local data for training and upload the updated models to the server. For example
[0086] (4) Server model scheduling: The server schedules the models according to the model scheduling algorithm;
[0087] The specific processes of initialization, global broadcast, local model training and uploading, and server model scheduling are as follows:
[0088] Central server:
[0089] (1) In the first round, the central server initializes the global model w g , and distributes w g to all clients c i .
[0090] (2) In the t-th round, the server receives the local model w i uploaded by client c i .
[0091] (3) In the t-th round, the server calculates the average delay of client i through the time interval of one model reception and transmission, and divides client i into a fast client or a slow client by comparing it with T threshold on the server.
[0092] (4) In the t-th round, the server calculates λ i by comparing the model w g uploaded by the client with w i , and classifies the model as a high-staleness model or a low-staleness model by comparing λ i with λ avg .
[0093] (5) In the t-th round, the server classifies through .
[0094] (6) In the t-th round, if the queue is not an empty queue, a model w can be taken out from the queue j , and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client i that uploaded the model w i for training. If the queue is an empty queue, then the model w i is put into the queue ;
[0095] (7) In the t-th round, if the queue is not an empty queue, a model w can be taken out from the queue j , and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client i that uploaded the model w i for training. If the queue is not an empty queue, a model w can be taken out from the queue j , and the model w i is sent to the client j that uploaded the model w j for training, and w j is sent to the client i that uploaded the model w i for training. If the queues and are both empty queues, then the model w i is put into the queue .
[0096] (8) In the t-th round, if the queue is not an empty queue, a model w can be taken out from the queue j , and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model The is sent to the client i and the client j for training. If the queue is not an empty queue, a model w can be taken out from the queue j , and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model The Send it to client i and client j for training. If the queue and are both empty queues, then put the model w i into the queue .
[0097] (9) In the t-th round, if If the queue is not an empty queue, a model w j can be taken out of the queue, and the model w i is sent to client j that uploaded the model w j for training, and w j is sent to client i that uploaded the model w i for training. If the queue is not an empty queue, a model w j can be taken out of the queue, and the model w i is sent to client j that uploaded the model w j for training, and w j is sent to client i that uploaded the model w i for training. If the queue is not an empty queue, a model w j can be taken out of the queue, and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model Send to client i and client j for training. If the queue is not an empty queue, a model w j can be taken out of the queue, and the model w i , the model w j and the global model w g are aggregated using the federated averaging algorithm to obtain a new global model Send to client i and client j for training. If the queues are both empty queues, then put the model w i into the queue .
[0098] Client:
[0099] (1) In the t-th round, client i receives the model w g sent by the central server, and the client uses local data to train the model. After training, the model is uploaded to the server.
[0100] Experimental process:
[0101] This solution implements the algorithm framework on EasyFL. EasyFL is an open-source federated learning system implemented in Python, which can simulate the training of a large number of distributed edge clients. It uses a wide range of deep learning tools such as TensorFlow, PyTorch, etc. The experiments of this solution were completed on a personal PC consisting of a Nvidia GeForce GTX 1060 with 6GB GPU memory, a 6-core 2.8GHz Intel i5-8400 CPU, and 16GB DDR4 memory. It runs the 64-bit Windows 10 system, and the CUDA toolkit version is 11.0 and cuDNN v7.4.2.
[0102] The algorithm framework proposed by the present invention is implemented on the EasyFL platform and experimentally verified. EasyFL is an open-source federated learning framework that supports simulating large-scale distributed edge client environments and is compatible with mainstream deep learning tools (such as TensorFlow and PyTorch). The experiments were conducted on a personal PC equipped with a Nvidia GeForce GTX 1060 (6GB video memory), and the computing devices include a 6-core 2.8GHz Intel i5-8400 CPU and 16GB DDR4 memory. It runs the 64-bit Windows 10 operating system, the CUDA toolchain version is 11.0, and the cuDNN version is v7.4.2. The experiments used the CIFAR-10 dataset, selected CNN as the global model, and compared and tested the performance of the model in 500 communication rounds.
[0103] The experiments constructed a distributed federated learning scenario by simulating 100 heterogeneous clients, where the computing speed and communication bandwidth of some clients were restricted to simulate the differences in device capabilities in actual applications. In each round of training, after receiving the global model sent by the central server, the clients used the local data for training and asynchronously uploaded the updated model to the server. After receiving the model uploaded by the clients, the server recorded the model upload time and calculated the latency, average latency, and model staleness of the clients, and then calibrated the types of clients according to the fast and slow client classification algorithm, dividing them into fast clients and slow clients. At the same time, according to the level of model staleness, the uploaded local models were classified as high-staleness models or low-staleness models.
[0104] In the experiment, the server manages and processes the models through a queue scheduling mechanism. In response to the differences between fast clients and slow clients, the server adopts a dynamic scheduling strategy. For example, highly stale models are further trained with the assistance of fast clients to reduce staleness and then aggregated with the global model; low-staleness models directly participate in the update of the global model. Through such grouping and scheduling strategies, the experiment tested the performance of the proposed solution of the present invention in terms of the accuracy of the global model, the training convergence speed, and the model stability, and made a comparative analysis with the traditional asynchronous federated learning algorithm FedAsync. The experimental results show that the proposed queue scheduling asynchronous federated learning algorithm of the present invention has a smaller accuracy fluctuation range in the early stage of communication rounds and can more quickly improve the test accuracy of the global model after about 200 rounds. Compared with FedAsync, the accuracy of the proposed solution of the present invention approaches 0.9 at the end of 500 communication rounds and demonstrates higher efficiency and stronger stability during the convergence process. The results further verify the significant advantages of the present invention in addressing the issues of device heterogeneity and model staleness in asynchronous federated learning.
Claims
1. An asynchronous federated learning method based on queue scheduling, characterized in that: In the asynchronous federated training process, the clients are divided into fast clients and slow clients; the models uploaded by the clients are divided into high staleness models and low staleness models; according to the uploaded local model w i The client type and the model type uploaded by the client will be used to convert the local model into i Reclassify; the server generates the same local model w i The queue corresponding to the category is used to store the corresponding local model. The server uses the local model w i The type of the selected aggregation strategy or exchange strategy, so that the fast client will reduce the staleness of the high staleness model to the judgment requirement of the low staleness model and then participate in the aggregation.
2. The asynchronous federated learning method based on queue scheduling according to claim 1 is characterized in that: The total delay is calculated based on the client's model training delay and communication delay, and the clients are divided into fast clients and slow clients based on the total delay.
3. The asynchronous federated learning method based on queue scheduling according to claim 2 is characterized in that: Before the server initializes the model and sends it down, or before sending the model after completing aggregation, the server records the current sending time to the i-th client according to the current timestamp. When the client performs local training on the model sent by the server and then transmits the model back, the server records the current receipt time of the i-th client according to the current timestamp. according to Calculate the latency of client i; record the latency of each client on the server, and calculate the average latency of client i in the asynchronous federated training process of submitting m local models to the server by recording the client latency of client i: Sort the average latency of k clients by size and get the median average latency of k clients; The average delay of client i is greater than the median of the average delay T threshold When it is classified as a slow client c s , the average delay of client i is less than the median of the average delay T threshold When it is divided into fast client c f .
4. The asynchronous federated learning method based on queue scheduling according to claim 3 is characterized in that: According to the difference between the local model uploaded by the client after training and the global model after aggregation on the server, the staleness of the local model uploaded by the client is divided into the following categories: When client i completes local training, it will use the local model w i After uploading to the server, the server Calculate the model staleness λ of client i i , which measures the local model of client i and the global model w g The server calculates the staleness of each model uploaded by the client according to the difference between Calculate the average staleness λ avg ; The staleness λ of client i i Greater than the average staleness λ avg When , it is classified as a high staleness model w h , the staleness λ of client i i Less than the average staleness λ avg When , it is classified as a low staleness model w l .
5. The asynchronous federated learning method based on queue scheduling according to claim 4 is characterized in that: The local model w i The reclassification principles are as follows: The server generates four local models w i Queues corresponding to categories For storage category Local model of For storage category Local model of For storage category Local model of For storage category The local model.
6. The asynchronous federated learning method based on queue scheduling according to claim 5 is characterized in that: The aggregation or exchange strategy is as follows: (1) The local model uploaded by client i is (a) Queue When the queue is not empty, take out a model w from the queue according to the order in which it is put. j , the model w i Send to upload model w j Client j training, w j Send to upload local model w i Client i is trained; (b) Queue When the queue is empty, the model w i Put in queue among; (2) The local model uploaded by client i is (a) Queue When the queue is not empty, take a model w from the queue j , the model w i Send to upload model w j Client j training, w j Send to upload local model w i Client i is trained; (b) Queue When the queue is not empty, take a model w from the queue j , the local model w i Send to upload model w j Client j training, w j Send to upload model w i Client i is trained; (c) Queue and When all queues are empty, the local model w i Put in queue among; (3) The local model uploaded by client i is (a) Queue When the queue is not empty, take a model w from the queue j , the local model w i , model w j With the global model w g Use the federated average algorithm to aggregate and obtain a new global model Will Send it to client i and client j for training; (b) Queue When the queue is not empty, take a model w from the queue j , the local model w i , model w j With the global model w g Use the federated average algorithm to aggregate and obtain a new global model Will Send it to client i and client j for training; (c) Queue and Are all empty queues, then the local model w i Put in queue among; (4) The model uploaded by client i is (a) Queue When the queue is not empty, take a model w from the queue j , the local model w i Send to upload model w j Client j training, w j Send to upload model w i Client i is trained; (b) Queue When the queue is not empty, take a model w from the queue j , the local model w i Send to upload model w j Client j training, w j Send to upload model w i Client i is trained; (c) Queue When the queue is not empty, take a model w from the queue j , the local model w i , model w j With the global model w g Use the federated average algorithm to aggregate and obtain a new global model Will Send it to client i and client j for training; (d) Queue When the queue is not empty, take a model w from the queue j , the local model w i , model w j With the global model w g Use the federated average algorithm to aggregate and obtain a new global model Will Send it to client i and client j for training; (e) If the queue Are all empty queues, then the local model w i Put in queue among.