Asynchronous federated learning method based on sliding window in WiFi network

By adopting a sliding window mechanism in WiFi networks, controlling the upload order and aggregation process of client models, the channel contention and delay problems in the federated learning process in WiFi networks are solved, and the efficiency and performance of federated learning are improved.

CN119026657AActive Publication Date: 2024-11-26XIDIAN UNIV

Patent Information

Application Number
CN202410930271.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-26
Estimated Expiration
2044-07-11

AI Technical Summary

Technical Problem

During the federated learning process in WiFi networks, communication efficiency and model performance decline due to channel contention and network instability when the client uploads the model simultaneously, especially the problem of severe delay in global model updates.

Method used

Using an asynchronous federated learning method based on sliding windows, the server sets the sliding window size according to network conditions in each round of communication, slides the sliding window in multiple clients, only clients entering the window upload the model, and this round Aggregate the local model from the previous round into a global model and send it to clients that are about to leave the window, ensuring that only one client occupies the channel at a time.

Benefits of technology

By reducing channel competition and latency, the performance and efficiency of federated learning in WiFi networks are improved, transmission delay and convergence rate are balanced, and model compatibility and robustness are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026657A_ABST
    Figure CN119026657A_ABST
Patent Text Reader

Abstract

The invention provides an asynchronous federated learning method based on a sliding window in a WiFi (Wireless Fidelity) network, which is applied to a federated learning system in the WiFi network, and comprises the following steps: each client trains a local model of the client so as to promote translation and aggregation processes; the server sets the size of a sliding window according to network conditions in this round of communication, slides the sliding window in a plurality of clients and informs the clients entering the sliding window, the clients upload the latest local models of the clients, the server aggregates all the local models into a global model and sends the global model to the clients sliding out of the sliding window, and the global model is sent to the clients sliding out of the sliding window. And only one client is ensured to occupy the channel in any given time slot, so that the delay caused by backoff due to failure of channel competition is greatly reduced. According to the method, the total transmission delay and the convergence rate are balanced, and the size of the sliding window is dynamically adjusted by using a heuristic algorithm, so that the performance of WiFi network federated learning is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless local area networks, and in particular relates to an asynchronous federated learning method based on a sliding window in a WiFi network. Background Art

[0002] In recent years, machine learning technology has been widely used in computer vision, natural language processing, and speech recognition. The traditional model training method is to upload raw data to a cloud server, which then processes and returns the trained model to the client. However, this model not only consumes a lot of communication resources, but also seriously damages customer privacy. In order to solve this problem and keep the training data within the client device, Google's McMahan et al. proposed the Federated Learning (FL) framework in 2016. In this framework, the client uses local data to train machine learning models on its device. These updated local models are then sent to the parameter server, where they are aggregated into a global model. Subsequently, the server redistributes the global model back to the client as the basis for the next round of local model training. This process is repeated many times until the model converges. With the continuous development of wireless communication technology, WiFi networks have become an essential part of contemporary life and work. According to Cisco's forecast, by 2021, 50% of the world's Internet traffic will come from WiFi. The wide coverage and high-density deployment of WiFi access points ensure the continuous connection of devices, thus facilitating FL, even in a mobile environment. WiFi networks typically have high data transfer rates, which is very beneficial for the transmission of model parameters and the refreshing of the global model. In addition, WiFi networks can often circumvent the data costs associated with mobile networks, incentivizing clients to contribute their computing resources to FL. Finally, WiFi networks can be expanded by adding more access points, allowing FL to be implemented over a larger area and encouraging more users to participate.

[0003] Although FL has great potential in WiFi networks, many challenges still exist. Issues such as WiFi network instability and heterogeneity of federated learning devices affect the communication efficiency and model performance of FL in WiFi. In WiFi networks, a large number of clients uploading models simultaneously during the FL process will frequently trigger the backoff and retransmission mechanism in the carrier sense multiple access / collision avoidance protocol (CSMA / CA). In addition, under the synchronous update mechanism of the FL model, clients with poor WiFi conditions may become laggards, resulting in severe delays in global model updates. Therefore, it is increasingly important to solve the problem of FL model efficiency and performance degradation caused by repeated triggering of channel contention backoff mechanisms when clients upload models simultaneously in WiFi networks.

[0004] At present, there are three main types of update mechanisms for federated learning: synchronous mechanism, asynchronous mechanism, and semi-asynchronous mechanism. In the synchronous update mechanism, after receiving the local models from all or a selected part of the federated learning users, the parameter server aggregates these local models into a new global model, and then sends the updated global model to each user. In this process, the models of all users participating in the federated learning need to be received before the global model can be updated.

[0005] Problems such as WiFi network instability and the heterogeneity of federated devices affect the communication efficiency and model performance of federated learning in WiFi. In a WiFi network, when a large number of users try to upload their local models to update the global model at the same time during the federated learning process, it will cause network congestion and trigger the frequent triggering of the backoff retransmission mechanism. In addition, users with poor WiFi network conditions will become stragglers in the synchronous update mechanism of the federated learning model, resulting in serious delays in the global model update. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides an asynchronous federated learning method based on sliding windows in a WiFi network. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] In a first aspect, the present invention provides an asynchronous federated learning method based on a sliding window in a WiFi network, which is applied to a federated learning system in a WiFi network, wherein the federated learning system includes a server for model aggregation and a client set consisting of N clients and used for FL training, and the asynchronous federated learning method includes:

[0008] Each client trains its own local model using its own local dataset;

[0009] The server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the N clients, and informs the clients that enter the sliding window;

[0010] The client entering the sliding window uploads its latest local model to the server;

[0011] The server aggregates the local models of the user end in this round of communication and before this round of communication into a global model, and sends it to the client that slides out of the sliding window at the end of the previous round of communication.

[0012] In a second aspect, the present invention provides a federated learning system in a WiFi network, the federated learning system comprising a server for model aggregation and a client set consisting of N clients for FL training;

[0013] Among them, each client uses its own local data set to train its own local model; the server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the multiple clients, and informs the clients entering the sliding window; the clients entering the sliding window upload their latest local models to the server; the server aggregates the local models of this round of communication and before this round of communication into a global model, and sends it to the clients that enter this round of communication and slide out of the sliding window after the previous round of communication ends.

[0014] Beneficial effects:

[0015] The present invention provides an asynchronous federated learning method based on sliding windows in a WiFi network, which is applied to a federated learning system in a WiFi network. The asynchronous federated learning method includes: each client trains its own local model using its own local data set to promote the translation and aggregation process; the server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the multiple clients, and informs the clients entering the sliding window; the clients entering the sliding window upload their latest local models to the server; the server aggregates the local models of this round of communication and before this round of communication into a global model, and sends it to the client entering this round of communication and sliding out of the sliding window after the previous round of communication, ensuring that only one client occupies the channel in any given time slot, thereby greatly reducing the delay caused by backoff due to failure in the competition channel. The present invention achieves a balance between the total transmission delay and the convergence rate, and dynamically adjusts the size of the sliding window using a heuristic algorithm, thereby improving the performance of federated learning in the WiFi network. Finally, the effectiveness of the method of the present invention is evaluated by simulation, and it is improved in terms of convergence, compatibility, robustness and delay affecting the WiFi network.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic diagram of the FedSW architecture provided by the present invention;

[0018] Figure 2 It is a flow chart of an asynchronous federated learning method based on sliding windows in a WiFi network provided by the present invention;

[0019] Figure 3 It is a schematic diagram of the convergence of ResNet18 to CIFAR10 (IID) under different window sizes provided by the present invention;

[0020] Figure 4It is a schematic diagram of the convergence of ResNet18 on CIFAR10 (Non-IID) under different window sizes provided by the present invention;

[0021] Figure 5 It is a schematic diagram of the convergence of ResNet18 to Fedprox algorithm CIFAR10 (IID) under different window sizes provided by the present invention;

[0022] Figure 6 It is a schematic diagram of the convergence of the Fedprox algorithm of ResNet18 for CIFAR10 (Non-IID) under different window sizes provided by the present invention;

[0023] Figure 7 It is a schematic diagram of the convergence of the FedSW algorithm under different windows when the number of clients in the WiFi network is 20, provided by the present invention;

[0024] Figure 8 It is a schematic diagram of the transmission delay of FedSW and CSMA / CA under different numbers of clients in the WiFi network provided by the present invention. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0026] Existing technologies cannot effectively solve the problem of serious delays in global model updates. Therefore, it is necessary to comprehensively consider the instability of WiFi networks, user dynamics, and the asynchronous update mechanism of federated learning, and design a method to improve the efficiency, robustness, and compatibility of federated learning under WiFi networks, promote the development of federated learning in edge networks, and provide strong theoretical and technical support for its large-scale commercial use.

[0027] The present invention provides an asynchronous federated learning method based on sliding windows in a WiFi network, which is applied to a federated learning system in a WiFi network. Figure 1 , the federated learning system consists of a server for model aggregation and a client set U consisting of N clients for FL training. Each participating client u i Storing a local dataset Its size is D i Denotes. Define the sliding window size as W, that is, select the client entering the sliding window for local training. The trained local model is transmitted to the server. The server aggregates the models of other clients in the window into a global model, and then sends the global model to the client that is about to leave the window.

[0028] Among them, all clients have been sorted according to numbers, and the server uses a sliding window to slide among multiple clients, so that the clients enter the sliding window in sequence and cyclically.

[0029] First, combining Figure 1 and Figure 2 The present invention provides an asynchronous federated learning method based on sliding windows in a WiFi network, comprising:

[0030] Each client trains its own local model using its own local dataset;

[0031] The server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the N clients, and informs the clients that enter the sliding window;

[0032] In this round of communication, the server sets the size of the sliding window according to the training duration of the client and the number of local training times.

[0033] The client entering the sliding window uploads its latest local model to the server;

[0034] The server aggregates the local models of this round of communication and those before this round of communication into a global model, and sends the global model to the client that enters this round of communication after the previous round of communication ends and slides out of the sliding window.

[0035] Continue to refer Figure 1 , in each round of communication, the server determines the size W of the sliding window W according to its network conditions. Assume that there are five clients, represented by u0, u1, u2, u3 and u4. Client u0 that enters the sliding window performs local training and transmits the locally trained local model to the server. The local models of u2 and u4 are also uploaded to the server, aggregated into a global model, and then sent to client u1 that has just left the sliding window. In the subsequent round, the entering client u3 performs local training, and its local model, together with the previously uploaded local models of u4 and u0, is aggregated into a global model by the server and then sent to the output client u2. This process continues in a similar manner for other clients, with each client sending and receiving once per round, keeping the communication volume consistent with traditional FL. The process is repeated until the model converges.

[0036] In a specific implementation of the present invention, setting the size of the sliding window according to the training duration of the client and the number of local training times includes:

[0037] a. Using the training duration tr and the number of local training times E, derive the upper limit of the convergence rate ∈ of the client training local model;

[0038] In this step, according to the setting conditions of unbiased stochastic gradient, bounded local gradient variance, bounded global gradient variance, bounded gradient of loss function and Lipschitz continuity of gradient, combined with the training duration tr and the number of local training times E, the upper limit of the convergence rate ∈ of the client training local model is derived, which is expressed as

[0039]

[0040] The setting condition of unbiased stochastic gradient is that for each client i, the stochastic gradient g i (w) is unbiased, then The setting condition of bounded local gradient variance is that for each client i, the upper bound of the local gradient variance is The setting condition of bounded global gradient variance is that the upper bound of global gradient variance is W t \{j} represents the set of all clients in the sliding window except the client j that is about to leave; the setting condition of the bounded gradient of the loss function is that there is a constant G such that for all w and i, there is The Lipschitz continuity of the gradient requires that there exists a constant L such that for all x and y, L is a constant, η represents the learning rate, w (0) represents the initialized local model parameters, represents the upper bound of the local gradient variance, represents the upper bound of the global gradient variance, T represents the Tth round of training, W is the size of the sliding window, and F i (w) represents the loss function of the i-th user end, w represents the parameters of the local model, Indicates F i (·) Gradient with respect to x, Represents the function F i (·) Gradient with respect to y.

[0041] b. using the training delay of the client, determining the total time delay required for the sliding window to complete a full cycle for all clients;

[0042] The present invention considers the relationship between the size of the sliding window and the related delay. Considering the training delay tr of the client, the total time delay is expressed as:

[0043]

[0044] c. Use the upper limit of convergence rate ∈ and total time delay to construct the overall objective function;

[0045] In order to conceptualize the problem as a balancing act, the present invention must identify the relevant elements. Usually in optimization challenges, a balance needs to be struck between the upper limit of convergence ∈ and the time delay. Assuming that the goal of the present invention is to minimize the time delay while keeping the convergence rate ∈ within a certain acceptable range, a single-objective optimization may be sought to combine the two objectives by introducing a trade-off parameter β. The preference for the two objectives can be controlled by adjusting the value of β. Therefore, the overall objective function is expressed as:

[0046]

[0047] Here, β represents the trade-off parameter.

[0048] d. Solve the overall objective function in a discrete interval to obtain each possible value of the size of the sliding window, and use a heuristic algorithm to systematically evaluate each possible value to obtain the optimal size of the sliding window.

[0049] Considering the size of the sliding window, i.e., the variable W is restricted to a discrete interval [1, N-1] and must be an integer, a discrete optimization challenge is faced. Traditional analytical methods may not be sufficient because the problem cannot be solved by directly applying the derivative to zero due to the discrete nature of W. Therefore, the present invention must adopt a heuristic algorithm to systematically evaluate the objective function for each possible value of W within the specified range to determine the optimal solution, i.e., the optimal sliding window size.

[0050] In an optional implementation manner of the present invention, each client uses its own local data set to train its own local model, including:

[0051] Each client u i The parameter update formula is used to perform E rounds of local training in the tth round of communication to update the parameters of its own local model. The parameter update formula is expressed as:

[0052]

[0053] in, Indicates client u i The parameters of the local model trained in the t+1th round of communication, Indicates client u i The parameters of the local model trained in the tth round of communication, Indicates client u i Parameters of the local model after round e of training in round t of communication.

[0054] In the present invention, the server aggregates the local models of this round of communication and those before this round of communication into a global model using a parameter aggregation formula, and the parameter aggregation formula is expressed as:

[0055]

[0056] After the server of the present invention sends the global model to the client that enters the current round of communication after the previous round of communication ends and exits the sliding window, the asynchronous federated learning method further includes:

[0057] The client u that slides out of the sliding window j The local model is updated using the local model update formula, which is expressed as:

[0058]

[0059] in, Indicates client u j Parameters of the local model trained in round t+1 of communication.

[0060] In a second aspect, the present invention provides a federated learning system in a WiFi network, the federated learning system comprising a server for model aggregation and a client set consisting of N clients for FL training;

[0061] Among them, each client uses its own local data set to train its own local model; the server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the multiple clients, and informs the clients entering the sliding window; the clients entering the sliding window upload their latest local models to the server; the server aggregates the local models of this round of communication and before this round of communication into a global model, and sends it to the clients that enter this round of communication and slide out of the sliding window after the previous round of communication ends.

[0062] In order to comprehensively evaluate the efficiency of the proposed sliding window-based asynchronous federated learning method in WiFi networks, the present invention deeply analyzes its performance improvement.

[0063] The simulation software environment is based on Python 3.8.17, and PyTorch 1.12.1 is used as the backbone of the model development and training process. The widely used MNIST and CIFAR10 datasets are used in the experiments. In addition, IID and non-IID data are also used in the experiments. The performance of the proposed scheme is carefully examined in a series of neural network architectures (including convolutional neural networks, ResNet-18, LeNet-5 and VGG16) to ensure a robust evaluation of its applicability and efficiency: The present invention sets up a scenario in which five clients compete to upload ResNet-18 model parameters through a WiFi network. The present invention configures the communication bandwidth between the client and the access point to 100Mbps. Table 1 lists other simulation parameters. In addition, the present invention also introduces a global timeline to accurately record the start and end of the data packet transmission, thereby quantifying the total time required for all clients to complete a complete model parameter transmission cycle.

[0064] Table 1 Simulation parameters

[0065]

[0066] The simulation experiment of this invention evaluates the convergence performance of the sliding window-based asynchronous federated learning method in WiFi networks in a simulation environment, as well as the impact of window size on model accuracy. This experiment uses vanilla FL as a comparison benchmark to verify the convergence ability of FedSW. Figure 3 As shown, FedSW achieves the highest model accuracy on the CIFAR10 (IID) dataset with a sliding window size of W = 2 compared to the considered baselines at the same number of communication rounds. The convergence accuracy drops slightly for sliding window sizes W = 3 and W = 4, but still exceeds vanilla FL. In addition, Figure 4 The effect of different data distributions on the convergence performance of FedSW when training it on the CIFAR10 (non-IID) dataset is shown. In the non-IID setting, the performance with window size W=4 is comparable to vanilla FL, but the accuracy is lower with W=2 and W=3, which can be attributed to the less global information contained in the smaller window size in the non-IID dataset.

[0067] In order to further verify the compatibility of the sliding window-based asynchronous federated learning method in the WiFi network of the present invention, the present invention combines it with the FedProx algorithm and adds a sliding window mechanism. The present invention uses the ResNet-18 model to simulate the FedProx algorithm on the CIFAR-10 dataset. Figure 5 As shown in Figure 3, adding sliding windows of different sizes results in a higher final accuracy compared to the baseline FedProx algorithm without this modification. Figure 6It is shown that window sizes W = 3 and W = 4 achieve recommendable performance on the non-IID dataset, but the accuracy decreases when W = 2. However, even with smaller window sizes, the performance of the FedProx algorithm is better than the Vanilla FL method. These findings confirm that the sliding window scheme is not only compatible with the FedProx algorithm, but also enhances its adaptability.

[0068] In order to evaluate the impact of the number of clients on the efficiency of the sliding window-based asynchronous federated learning method in WiFi networks, the present invention conducts experiments with different numbers of clients during the FedSW training process. Figure 7 The performance of the FedSW algorithm in a WiFi network scenario with 20 clients is shown. The accuracy of the model exceeds that of the ordinary FL method in all tested window sizes, and the performance reaches a peak when the window size is W = 2. The reason is that when W = 2, the iid dataset contains more global information, and it can be clearly seen from the data that the convergence timeline is extended.

[0069] The present invention simulates a packet transmission sequence integrating the competitive back-off mechanism inherent in the CSMA / CA protocol to measure the total delay for all clients to successfully transmit their model parameters. Figure 8 It is shown that the time required to complete transmission through FedSW and CSMA / CA mechanisms increases with the increase in the number of clients. It is worth noting that the transmission delay of the present invention has been decreasing, which highlights its efficiency. The reason is that the present invention reduces the empty channel delay caused by frequent contention backoff.

[0070] In view of the problem of federated learning model efficiency and performance degradation caused by federated learning users uploading models at the same time in WiFi networks and frequently triggering the channel contention backoff mechanism, the present invention proposes an innovative asynchronous federated learning method based on sliding windows, which is built on the FedSW framework. Specifically, the present invention sets a sliding window at the wireless access point of the WiFi network, and the user's local model is aggregated and uploaded according to the sliding window translation. A model for the balance problem among the sliding window, user delay and convergence rate is established. By adjusting the sliding window length, the delay caused by channel preemption is reduced. The FedSW of the present invention can coordinate local training, model uploading, aggregation and distribution to keep consistent with the sliding window progress. In addition, the seamless integration of FedSW with the most advanced algorithm (SOTA) also proves its versatility. Simulation results show that FedSW always outperforms traditional benchmarks in terms of convergence, while significantly reducing latency.

[0071] It is worth noting that the terms "first" and "second" in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0072] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0073] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. An asynchronous federated learning method based on sliding windows in a WiFi network, characterized in that: A federated learning system applied in a WiFi network, the federated learning system comprising a server for model aggregation and a client set consisting of N clients for FL training, the asynchronous federated learning method comprising: Each client trains its own local model using its own local dataset; The server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the N clients, and informs the clients that enter the sliding window; The client entering the sliding window uploads its latest local model to the server; The server aggregates the local models of the user end in this round of communication and before this round of communication into a global model, and sends it to the client that slides out of the sliding window at the end of the previous round of communication.

2. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 1, characterized in that: The server setting the size of the sliding window according to the network conditions in the current round of communication includes: the server setting the size of the sliding window according to the training duration and the number of local training times of the client in the current round of communication.

3. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 2, characterized in that: The step of setting the size of the sliding window according to the training duration of the client and the number of local training times includes: a. Using the training duration tr and the number of local training times E, derive the upper limit of the convergence rate ∈ of the client training local model; b. using the training delay of the client, determining the total time delay required for the sliding window to complete a full cycle for all clients; c. Use the upper limit of convergence rate ∈ and total time delay to construct the overall objective function; d. Solve the overall objective function in a discrete interval to obtain each possible value of the size of the sliding window, and use a heuristic algorithm to systematically evaluate each possible value to obtain the optimal size of the sliding window.

4. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 3, characterized in that: include: According to the setting conditions of unbiased stochastic gradient, bounded local gradient variance, bounded global gradient variance, bounded gradient of loss function and Lipschitz continuity of gradient, combined with the training duration tr and the number of local training times E, the upper limit of the convergence rate ∈ of the client training local model is derived, which is expressed as The setting condition of unbiased stochastic gradient is that for each client i, the stochastic gradient g i (w) is unbiased, then The setting condition of bounded local gradient variance is that for each client i, the upper bound of the local gradient variance is The setting condition of bounded global gradient variance is that the upper bound of global gradient variance is W t \{j} represents the set of all clients in the sliding window except the client j that is about to leave; The bounded gradient of the loss function is set by the condition that there exists a constant G such that for all w and i, The Lipschitz continuity of the gradient requires that there exists a constant L such that for all x and y, L is a constant, η represents the learning rate, w (0) represents the initialized local model parameters, represents the upper bound of the local gradient variance, represents the upper bound of the global gradient variance, T represents the Tth round of training, W is the size of the sliding window, and F i (w) represents the loss function of the i-th user end, w represents the parameters of the local model, Indicates F i (·) Gradient with respect to x, Represents the function F i (·) Gradient with respect to y.

5. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 4, characterized in that: The total time delay is expressed as:

6. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 5, characterized in that: The overall objective function is expressed as: Here, β represents the trade-off parameter.

7. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 1, characterized in that: Each client uses its own local data set to train its own local model, including: Each client u i The parameter update formula is used to perform E rounds of local training in the tth round of communication to update the parameters of its own local model. The parameter update formula is expressed as: in, Indicates client u i The parameters of the local model trained in the t+1th round of communication, Indicates client u i The parameters of the local model trained in the tth round of communication, Indicates client u i Parameters of the local model after round e of training in round t of communication.

8. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 7, characterized in that: Aggregating the local models of the client in the current round of communication and before the current round of communication into a global model includes: The parameter aggregation formula is used to aggregate the local models of this round of communication and those before this round of communication into a global model. The parameter aggregation formula is expressed as:

9. The sliding window-based asynchronous federated learning method in a WiFi network according to claim 8, characterized in that: After sending to the client that slides out of the sliding window after the last round of communication, the asynchronous federated learning method further includes: The client u that slides out of the sliding window j The local model is updated using the local model update formula, which is expressed as: in, Indicates client u j Parameters of the local model trained in round t+1 of communication.

10. A federated learning system in a WiFi network, characterized in that: The federated learning system includes a server for model aggregation and a client set consisting of N clients for FL training; Among them, each client uses its own local data set to train its own local model; the server sets the size of the sliding window according to the network conditions in this round of communication, slides the sliding window among the multiple clients, and informs the clients entering the sliding window; the clients entering the sliding window upload their latest local models to the server; the server aggregates the local models of this round of communication and before this round of communication into a global model, and sends it to the clients that enter this round of communication and slide out of the sliding window after the previous round of communication ends.

Citation Information

Patent Citations

  • Federal continuous learning training method based on memory playback and differential privacy

    CN115081532A

  • Centralized client selection method for federal learning based on fuzzy logic

    CN116582539A

  • Client asynchronous federal learning method with adaptive partial training

    CN117875453A

  • System and method for decentralized federated learning

    US20210406782A1

  • A time-triggered federated learning algorithm

    WO2023175335A1

Cited By

  • Edge cloud data co-processing method based on dynamic sliding window and federal learning

    CN121284033A

  • An edge-cloud data collaborative processing method based on a dynamic sliding window and federated learning

    CN121284033B