Method for participating in fairness perception asynchronous federated learning based on predictive scheduling

Abstract

The invention belongs to the technical field of asynchronous federated learning, and discloses a fairness perception participation asynchronous federated learning method based on predictive scheduling. The method comprises the following steps: predicting training time delay required by a client to complete local training and participate in model aggregation, introducing a fairness control queue describing a long-term participation state of the client, constructing a scheduling criterion by jointly predicting the training time delay and the fairness state at a server side, dynamically selecting a client set participating in model aggregation, and performing model aggregation. And when the fairness constraint is about to take effect, predictive training restart is triggered so as to reduce the influence caused by the old degree of the model. According to the method, the long-term participation fairness of the client can be effectively improved in a heterogeneous environment, meanwhile, the better training efficiency and model convergence performance are kept, and the method has good applicability and practical value.

Description

A Participation Fairness-Aware Asynchronous Federated Learning Method Based on Predictive Scheduling Technical Field

[0001] This invention relates to the field of asynchronous federated learning technology, and more specifically to a participation fairness-aware asynchronous federated learning method based on predictive scheduling, used to control client participation scheduling during the asynchronous federated learning process. Background Technology

[0002] Asynchronous federated learning allows the server to aggregate updates from a subset of trained client models in real-time without waiting for all client models to complete their uploads, significantly improving training efficiency. However, in real-world applications such as connected vehicles, clients typically exist as vehicles, exhibiting high heterogeneity in computation, communication, and operational status. This leads to significant differences in the time required for different clients to complete model training and upload, categorizing clients as fast or slow based on their update speed. Fast clients, due to their rapid update speed, participate more frequently in global model updates, causing the global model to favor data distributions from fast clients. This problem is particularly pronounced in scenarios where client-side local data distribution is non-identical to identically distributed (Non-IID). Simultaneously, slower clients, under existing scheduling mechanisms, have a significantly lower probability of being selected for model aggregation than fast clients, resulting in uneven distribution of training opportunities. Furthermore, when some slow clients happen to hold scarce and critical data categories, their model updates may not be effectively aggregated for extended periods, hindering the global model's ability to fully learn the corresponding data features. This leads to decreased model performance and generalization ability, exacerbating biases during model training.

[0003] To address the above issues, existing research mainly improves the asynchronous federated learning process from different control dimensions, including: (1) a time-based control method, which adjusts the timing of client model updates, such as limiting the update frequency of clients with faster training speeds or introducing an update buffer mechanism to reduce the differences in update time among different clients; (2) a contribution weight adjustment method, which dynamically adjusts the weight of each client model update in the global model based on factors such as the frequency of client model updates, staleness, or data importance during the model aggregation stage; and (3) a participation decision control method, which selects and schedules clients in training rounds, prioritizing clients with higher computing power, communication speed, or data quality to participate in the training or aggregation process, thereby improving the overall convergence efficiency of the system.

[0004] While the aforementioned methods optimize the asynchronous federated learning process to some extent, they still fall short in terms of fairness in allocating opportunities for client participation. On one hand, some existing strategies, designed primarily to improve training efficiency, tend to reduce the participation frequency of slower-training clients during actual operation. For example, aggregation weighting mechanisms based on model staleness may weaken the impact of model updates from slower clients, while selection strategies based on client quality may consistently reduce the participation opportunities of resource-constrained clients, making it difficult to fully learn from high-value but infrequently updated data. On the other hand, existing research on fairness in federated learning largely focuses on fairness at the model performance or prediction result level. However, the problem of uneven allocation of opportunities for client participation may persist during asynchronous training, and relying solely on fairness constraints at the result level is insufficient for effective adjustment at the training process level. Therefore, it remains necessary to introduce a scheduling mechanism for the client participation process while ensuring the training efficiency of asynchronous federated learning. Summary of the Invention

[0005] To achieve a balance between client scheduling efficiency and long-term participation fairness in an asynchronous federated learning environment, this invention provides a participation fairness-aware asynchronous federated learning method based on predictive scheduling. By predicting the client's training completion time, the long-term participation fairness constraint is transformed into a single-round scheduling decision, ensuring the overall training efficiency of the system while achieving participation fairness for the client during the long-term training process.

[0006] Especially in application scenarios such as connected vehicles, edge computing, and large-scale distributed terminal collaborative learning, clients generally exhibit significant heterogeneity in terms of computing power, communication bandwidth, network connection stability, and local data scale. Affected by factors such as terminal mobility, network fluctuations, and differences in device performance, the time required for different clients to complete local training and model upload varies considerably. This forces the server to balance overall training time overhead with the fairness of long-term client participation during model aggregation. Therefore, this invention introduces a long-term training time metric for evaluating overall training efficiency. Its definition is: in, Indicates the number of training rounds. Indicates the first The actual training time for the wheel.

[0007] In asynchronous federated learning, to uniformly represent the impact of client scheduling choices on training latency, the first... The actual training time of the wheel The decision is made by the client with the longest predicted completion time among those selected to participate in the aggregation, defined as: in, Indicates the first The set of candidate clients for the wheel, For the client In the An indicator variable for whether a wheel is selected to participate in model aggregation. Indicates client No. The wheel is selected, otherwise ; Indicates client In the The remaining prediction time required for each round to complete local training and participate in model aggregation.

[0008] To prevent some clients from being systematically ignored during long-term training, this invention introduces a long-term participation fairness control mechanism. For each client... Constructing a virtual fairness deficit state This represents the fairness deficit accumulated by the client due to insufficient historical participation. This state is updated before each round of scheduling decisions begins, based on the results of the previous round of scheduling, as shown in formula (3): in, Indicates client The fairness objective parameter is used to quantify the expected participation "budget" per unit time during long-term training. It is the first The scheduling decision results of each round. Through this update mechanism, the fairness requirements for long-term participation are dynamically transformed into the state variables available for each round's decision. .

[0009] During each round of training, the server jointly considers the predicted training completion time of candidate clients. With fairness state parameters Constructing a system to guide the first The comprehensive cost function for round-robin client scheduling is defined as shown in formula (4): Formula (4) decomposes the system cost of a single round of scheduling decision into a training latency cost term and a long-term participation fairness risk term. The first term... The latency cost term represents the immediate impact of the selected client set in the current round on the system training time, where... This is a trade-off coefficient between training latency costs and long-term participation fairness risks, used to adjust the system's emphasis on training efficiency in scheduling decisions. Because the... The actual training time of the wheel The selection is based on the client with the longest predicted completion time among the chosen clients. Selecting a client with a longer training completion time may significantly increase the system training latency in the current round; due to the... The actual training time of the wheel The selection is based on the client with the longest predicted completion time. Selecting a client with a longer training completion time may significantly increase the system training latency for the current round. (Second item) This is a fairness risk term, representing the potential long-term participation imbalance risk that may arise during subsequent training due to the failure to schedule clients whose fairness deficit has exceeded a threshold. This term only applies to the client side. Not selected and its fairness deficit Exceeding the preset threshold A positive value is generated at certain times, thereby penalizing scheduling behavior that continues to ignore such clients.

[0010] Based on formula (4), the server does not select clients based on a single metric in each round of scheduling, but rather makes decisions by analyzing the marginal change in the overall cost caused by adding a single candidate client to the current round's aggregate set. Specifically, for any candidate client... Incorporating it into the current round of model aggregation will have two effects: firstly, it may increase the prediction completion time. This reduces the training latency costs; on the other hand, it avoids the fairness risks associated with the client not being selected in this round.

[0011] Therefore, scheduling decisions essentially transform into the following marginal trade-off process: when the client... When the latency increment cost of including the client in the current round of model aggregation is less than the fairness risk cost of continuing to ignore the client, selecting the client can effectively reduce the overall cost function. The value of will influence the server's selection of the client for this round of model aggregation; conversely, if selecting the client would significantly increase the system's training latency and fail to offset its fairness benefits, then the client will not be selected in the current round.

[0012] By explicitly introducing the aforementioned marginal cost trade-off mechanism in single-round scheduling decisions, the server can naturally internalize the long-term participation fairness requirement into a real-time computable scheduling decision basis without introducing cross-round constraint solving or global optimization calculations. This ensures the overall training efficiency of the system while suppressing the continuous accumulation of fairness deficits, achieving synergistic optimization of training efficiency and long-term participation fairness in asynchronous federated learning. Based on this, the present invention provides a participation fairness-aware asynchronous federated learning method based on predictive scheduling. The client-side local training process and the server-side scheduling decision process are executed intermittently within the asynchronous federated learning framework, including the following steps: First, the server initializes the global model and sets relevant parameters; the server distributes the global model and relevant parameters to all clients; Second, client-side local training and model upload; Step 2-1, the client performs local training on its local dataset based on the global model to obtain a local model update; Step 2-2, after completing local training, the client uploads the local model update to the server; Third, server-side scheduling subset decision; Step 3-1, the server predicts the client's training completion time; Step 3-2, the server updates the fairness status parameters of each client; Step 3-3, the server combines the training completion time prediction results and the fairness status parameters to determine the set of clients participating in model aggregation in the current round; Fourth, asynchronous model aggregation and iterative update; Step 4-1, the server aggregates the local models of the clients determined to participate in model aggregation to obtain the updated global model; Step 4-2, the server repeats steps 2 to 4-1 until a preset number of training rounds is reached or the global model meets the convergence condition.

[0013] The specific steps of the first step are as follows: The server initializes the global model and sets the trade-off parameters used to balance training latency and long-term participation fairness. The server is for all clients in the system. Set the corresponding fairness target parameters And initialize the fairness deficit state of each client. and fairness threshold ;in, This indicates the total number of clients in the system.

[0014] An additional step is added between step 2-1 and step 2-2: The client terminates the current training and restarts training based on the latest global model according to the server's instructions; the specific steps are as follows; on the client... If a restart command is received from the server during local training, then the client... Terminate the current local training task and use the latest global model parameters issued by the server. Restart local training.

[0015] The specific steps of step 3-1 are as follows: The server records the historical training time and communication time of each client, and predicts the client's time based on historical statistical information. The training completion time of each round; the server continuously records the local training time of each client in historical rounds. Communication time with model upload And calculate the client accordingly. Estimate the total runtime of completing a local training and model upload under the current training configuration. Full runtime estimate The exponential average method is used for updating, and its calculation method is as follows: in, The smoothing coefficient is used; the server simultaneously records the actual duration of each round of the system. ;No. The actual training time of the wheel It is determined by the client with the longest training completion time among the clients participating in model aggregation. in, Indicates the first The set of candidate clients for the wheel, For the client In the An indicator variable for whether a wheel is selected to participate in model aggregation. Indicates client No. The wheel is selected, otherwise ; Indicates client In the The remaining prediction time required for each round to complete local training and participate in model aggregation; the server records the client's prediction time. Current model obsolescence , indicating from the client The number of global model update rounds performed by the server since the last time global model parameters were received; server calculation client In the Remaining prediction time required for each round to complete local training and participate in model aggregation The calculation method is as follows: in, Indicated on the client The cumulative training time consumed by the system during the current model version's validity period; when When the client has completed local training or is in an idle state, it can immediately participate in model aggregation without increasing the system training latency of the current round.

[0016] The specific steps of step 3-2 are as follows; the server is each client in the system. Maintain the corresponding fairness state parameters The fairness state parameter is used to reflect the deviation between the actual frequency of client participation in model aggregation during historical training and its fairness objective; before each round of scheduling decisions begins, the server determines the scheduling result based on the previous round's results. Update the fairness state parameters for each client. The calculation method for this update mechanism is as follows: in, Indicates client Fairness objective parameters; The value represents the client's cumulative fairness deficit up to the current round; It has a dynamic accumulation characteristic: when a client is not selected to participate in model aggregation in multiple consecutive training rounds, its fairness state parameter... The fairness deficit gradually accumulates and increases; when a client is selected to participate in model aggregation, its fairness deficit is partially offset or cleared, thereby inhibiting further accumulation of fairness states; through this update mechanism, the server dynamically maps long-term participation fairness requirements into state variables directly used in single-round scheduling decisions without introducing cross-round constraint solutions; the updated fairness state parameters As one of the core inputs, it is directly fed into the current round. In the scheduling decision algorithm, the scheduling decision algorithm, based on jointly considering the training completion time prediction results, considers the fairness state parameters of each client. By dynamically adjusting their priority in scheduling decisions, clients with insufficient historical participation can receive a higher scheduling preference in subsequent rounds, thereby achieving a balance between the overall training efficiency of the system and the long-term fairness of client participation.

[0017] The step between steps 3-2 and 3-3 is to have the server perform predictive scheduling control and trigger the client to restart training. The specific steps are as follows: For clients not selected to participate in the current round of model aggregation... When its fairness state parameter As the data accumulates over multiple training rounds, the server adjusts the client's settings based on the changing trend of the fairness state parameter. The priority of participation in subsequent scheduling processes is predicted and evaluated; when the client The fairness state parameters satisfy At that time, among them The preset fairness threshold represents a critical state where the risk of insufficient client participation over a long period significantly impacts scheduling decisions. The server determines that the client has a significantly increased priority in subsequent scheduling processes. When the above condition is met, the server proactively informs the client... Distribute the latest global model parameters It also sends a restart command, instructing the client to terminate the current local training task and restart local training based on the latest global model parameters, in order to reduce the training latency risk and model obsolescence risk introduced when the client is subsequently selected to participate in model aggregation.

[0018] The specific steps of step 3-3 are as follows: The server uses the current candidate client set... As input, considering the various candidate clients in the first... Predicted training completion time of rounds With fairness state parameters To minimize the overall cost of a single round of scheduling, the [number]th round of scheduling is determined. The set of clients aggregated in the round-robin model has a scheduling metric function defined as: The scheduling optimization problem is solved using an online selection method based on a greedy principle; the server first selects the client based on the predicted training completion time of each candidate client. For the candidate client set The clients are sorted in ascending order to prioritize those with minimal impact on system training latency; subsequently, the server iterates through the candidate clients in the sorted order; for any candidate client... The server evaluated the settings respectively. and Comprehensive cost function in both cases The value changes; when set When the client is selected to participate in the current round of model aggregation, its corresponding fairness penalty is eliminated, but the training latency of the current round increases; when set At that time, the training latency of the current round remains unchanged, but when the fairness state parameter of the client... Exceeding the fairness threshold When appropriate fairness penalties are imposed, they should be introduced; when setting... The server selects a client to participate in the current round of model aggregation if it can reduce the overall cost function; otherwise, the server does not select the client. After all candidate clients have been traversed, the server obtains the result. The set of clients participating in the model aggregation round In scenarios where the number of participating clients is limited, a cap is set on the number of clients participating in each round of model aggregation. When the traversal is complete and no client is selected, the server will select at least one client with the shortest prediction training completion time to participate in model aggregation, in order to ensure the normal execution of the model aggregation process.

[0019] In this way, the server compares the changes in the overall cost before and after a client joins during each round of scheduling, thereby enabling online selection of clients. This effectively suppresses the problem of long-term imbalance in client participation while ensuring the overall training efficiency of the system.

[0020] This invention offers the following advantages: By predictively scheduling client training completion times, it effectively reduces overall training time in asynchronous federated learning environments, preventing system efficiency degradation due to long wait times for slow clients; by introducing fairness state parameters for long-term client participation and incorporating them into the scheduling decision process, it dynamically adjusts long-term client participation behavior, thereby improving the fairness of client participation during long-term training and preventing uneven distribution of training opportunities; by introducing predictive scheduling and restart control mechanisms, it updates the global model parameters used in local training before slow clients participate in model aggregation, helping to reduce the risk of model obsolescence that may occur when slow clients participate in model aggregation, enabling them to participate in model aggregation with lower time costs and less accuracy loss; furthermore, this invention does not require additional assumptions about client hardware conditions or network environments, possessing good versatility and scalability, and is suitable for asynchronous federated learning tasks in highly dynamic and heterogeneous computing environments such as vehicle networking. Attached Figure Description

[0021] Figure 1 is a flowchart of the participation fairness-aware asynchronous federated learning method based on predictive scheduling; Figure 2 is a comparison of the convergence speed of Faps and the benchmark algorithm in MNIST; (a) is the curve of model accuracy changing with training time when the data heterogeneity α=0.1; (b) is the curve of model accuracy changing with training time when the data heterogeneity α=0.5. Detailed Implementation

[0022] In this scenario, moving vehicles act as clients, communicating with a central server deployed by the automaker via signal base stations. The massive amounts of driving data accumulated by the vehicles, including road condition information and driving behavior data, serve as local data for training the federated learning model. This data contains rich user privacy and corporate trade secrets and cannot be directly shared; therefore, this method is used to achieve joint model training.

[0023] To achieve collaborative model training while protecting local data privacy, each vehicle trains its model locally using the data and uploads the model parameters or update results to the server, which then aggregates them based on an asynchronous federated learning mechanism. By employing the participation fairness-aware asynchronous federated learning method based on predictive scheduling proposed in this invention, the server can coordinate the timing and frequency of vehicle participation in model aggregation even when there are differences in vehicle computing power and communication conditions. This ensures overall training efficiency while improving the participation fairness of vehicles during long-term training. The local vehicle training process and the server-side scheduling decision-making process are executed concurrently within the asynchronous federated learning framework. The overall process includes the following steps: Step 1, server-side initialization.

[0024] The server initializes the global model and sets parameters to balance training latency with long-term participation fairness. Meanwhile, the server serves all vehicles in the system. Set the corresponding fairness target parameters And initialize the fairness deficit state of each vehicle. and fairness threshold .in, This represents the total number of vehicles in the system.

[0025] In each round of communication, the server sends messages to all vehicles in the system. Distribute global model parameters for the current round Current communication round and learning rate Training parameters, etc.

[0026] The second step involves local training on the client side and model uploading; vehicle After receiving the global model from the server, it is based on its local dataset. Perform a preset number of local iterative training rounds to obtain the updated local model. .

[0027] In the vehicle During local training, if a restart command is received from the server, then the vehicle... Terminate the current local training task and use the latest global model parameters issued by the server. Restart local training.

[0028] After completing local training, the vehicle uploads the local model update to the server.

[0029] The third step is server-side subset scheduling decision-making; the server records the historical training time and communication time of each vehicle, and predicts the vehicle's timing based on historical statistics. The training completion time for each round.

[0030] Specifically, the server continuously records the local training time for each vehicle in historical rounds. Communication time with model upload And calculate the vehicle accordingly. Estimate the total runtime of completing a local training and model upload under the current training configuration. .

[0031] In one implementation, the runtime estimate The exponential average method is used for updating, and its calculation method is as follows: in, This is the smoothing coefficient.

[0032] The server simultaneously records the actual duration of each round of the system. , No. The actual training time of the wheel It is determined by the vehicle with the longest training completion time among the vehicles participating in model aggregation, and its calculation method is as follows: Server records vehicles Current model obsolescence , indicating self-vehicle The number of global model update rounds performed by the server since the last time global model parameters were received.

[0033] Based on this, the server calculates the vehicle In the Remaining prediction time required for each round to complete local training and participate in model aggregation The calculation method is as follows: in, Indicates in the vehicle The cumulative training time consumed by the system during the current model version's validity period; when When the vehicle has completed local training or is idle, it can immediately participate in model aggregation without increasing the system training latency of the current round.

[0034] The server is for each vehicle in the system. Maintain the corresponding fairness state parameters The fairness state parameter is used to reflect the deviation between the actual frequency of a vehicle's participation in model aggregation during historical training and its fairness objective.

[0035] Before each round of scheduling decisions begins, the server determines the scheduling results based on the previous round's results. Calculate the fairness status of each vehicle. The calculation method is as follows: in, Indicates client Fairness objective parameters; The value of represents the cumulative fairness deficit of a vehicle up to the current round. The larger the value, the less frequently the vehicle participates in model aggregation during historical training relative to the fairness objective. It has a dynamic accumulation characteristic: when a vehicle is not selected to participate in model aggregation in multiple consecutive training rounds, its fairness state parameter... The fairness deficit will gradually accumulate and increase; when a vehicle is selected to participate in model aggregation, its fairness deficit will be partially offset or cleared, thereby inhibiting the further accumulation of fairness states. Through this update mechanism, the server can dynamically map long-term participation fairness requirements into state variables that can be directly used in single-round scheduling decisions without introducing cross-round constraint solutions.

[0036] Updated fairness state parameters It will be directly fed into the current round as one of the core inputs. In the scheduling decision algorithm, the scheduling decision algorithm, based on jointly considering the training completion time prediction results, considers the fairness state parameters of each vehicle. By dynamically adjusting their priority in scheduling decisions, vehicles that have historically participated insufficiently can receive a higher scheduling preference in subsequent rounds, thereby achieving a balance between the overall training efficiency of the system and the long-term fairness of vehicle participation.

[0037] For vehicles not selected to participate in the current round of model aggregation When its fairness state parameter As the data accumulates over multiple training rounds, the server adjusts the vehicle's settings based on the changing trend of the fairness state parameters. The participation priority in subsequent scheduling processes is predicted and evaluated.

[0038] When the vehicle The fairness state parameters satisfy At that time, among them The preset fairness threshold represents a critical state where the risk of long-term insufficient vehicle participation is about to have a significant impact on scheduling decisions. The server determines that the vehicle has a trend of significantly increased selection priority in subsequent scheduling processes.

[0039] When the above conditions are met, the server will send a message to the vehicle in advance. Distribute the latest global model parameters It sends a restart command, instructing the vehicle to terminate the current local training task and restart local training based on the latest global model parameters, so as to reduce the training latency and model obsolescence risks that may be introduced when the vehicle is subsequently selected to participate in model aggregation.

[0040] The server uses the currently available set of candidate vehicles. As input, comprehensively considering each candidate vehicle in the first... Predicted training completion time of rounds With fairness state parameters To minimize the overall cost of a single round of scheduling, the [number]th round of scheduling is determined. The scheduling metric function of the vehicle set aggregated by the round-robin participation model is defined as: This invention employs an online selection method based on a greedy principle to solve the scheduling optimization problem. The server first determines the scheduling optimization problem based on the predicted training completion time of each candidate vehicle. For the candidate vehicle set The vehicles are sorted in ascending order to prioritize vehicles that have a smaller impact on system training latency.

[0041] The server then iterates through the candidate vehicles in sorted order. For any given candidate vehicle... The server evaluated the settings respectively. and Comprehensive cost function in both cases The value changes. When set When the vehicle is selected to participate in the current round model aggregation, its corresponding fairness penalty is eliminated, but the training latency of the current round may increase; when setting At that time, the training delay of the current round remains unchanged, but when the fairness state parameter of the vehicle... Exceeding the fairness threshold In such cases, corresponding fairness penalties will be introduced.

[0042] When setting When the overall cost function can be reduced to a smaller value, the server selects the vehicle to participate in the current round of model aggregation; otherwise, the server does not select the vehicle. Through this method, the present invention achieves online vehicle selection in each round of scheduling based on a marginal trade-off between increased training latency and elimination of fairness risks.

[0043] After all candidate vehicles have been traversed, the server obtains the [number]th [vehicle]. The set of vehicles participating in the model aggregation In scenarios where the number of participating vehicles is limited, this invention can set an upper limit on the number of vehicles participating in each round of model aggregation. When the traversal is complete and no vehicle is selected, the server will select at least one vehicle with the shortest prediction training completion time to participate in model aggregation, in order to ensure the normal execution of the model aggregation process.

[0044] The fourth step is asynchronous model aggregation and iterative updates; the server receives the vehicle set. The model is updated, and weighted aggregation is performed to obtain the updated global model parameters. Aggregation can be represented as: in, For vehicles In the The aggregate weight of the round satisfies and Aggregate weights Can be matched with the age of the vehicle Correlation is used to reduce the impact of highly outdated updates on model accuracy.

[0045] The server repeats the above process until the global model reaches the preset number of training rounds or meets the convergence condition, and then outputs the final trained global model.

[0046] The effectiveness of the proposed participation fairness-aware asynchronous federated learning method based on predictive scheduling is verified below with specific embodiments. An image classification task is used to experimentally validate the method. This embodiment uses the MNIST dataset as the training dataset and conducts comparative experiments on different client scheduling and aggregation methods in an asynchronous federated learning environment.

[0047] In the experiment, to simulate the inconsistent distribution of client data in a real-world federated learning scenario, a Dirichlet distribution was used to partition the client's local data. The degree of data heterogeneity was controlled by adjusting the Dirichlet distribution parameter α. Specifically, α=0.5 and α=0.1 were set to correspond to medium and high data heterogeneity scenarios, respectively. Each client was allocated a fixed number of training and testing samples, and each client only used its local data for model training, without directly sharing the original data during the training process.

[0048] In terms of model settings, the MNIST dataset uses the LeNet-5 network architecture as the base model, with a global training epoch of 100 epochs and a client-side local training epoch of 5 epochs, and a learning rate of 0.01. To simulate the heterogeneity of client computing power and communication conditions, a random waiting time is introduced after each epoch of local training to reflect the differences in training and communication between different clients in the actual system.

[0049] First, the server initializes the global model and sets relevant parameters, including a tradeoff parameter for balancing training latency and long-term participation fairness, a fairness target parameter, and a fairness threshold. It also initializes the fairness state parameters for each client and then distributes the initial global model and related training parameters to all clients.

[0050] Secondly, the client performs local training.

[0051] First, the client performs local training on its local dataset based on the global model distributed by the server, and obtains the local model update. Second, if a restart command is received from the server during local training, the client terminates the current training task and restarts local training based on the latest global model. Third, after completing local training, the client asynchronously uploads the updated local model parameters to the server.

[0052] Then, the server performs participation scheduling and model aggregation decisions.

[0053] The first step involves the server predicting the time required for each client to complete local training and participate in model aggregation in the current round, based on the client's historical training and communication time information. The second step involves the server updating the fairness state parameters of each client based on their participation in previous rounds to characterize their long-term under-participation. The third step involves the server jointly considering the client's predicted training completion time and fairness state parameters to construct a scheduling metric function, and using this function to determine the set of clients participating in model aggregation in the current round. The fourth step involves the server prematurely distributing the latest global model to some clients when their fairness state parameters exceed a preset threshold, and triggering the clients to restart local training to reduce training latency and model obsolescence during subsequent aggregation.

[0054] Subsequently, the server performs asynchronous model aggregation on the model updates uploaded by the selected clients to obtain the updated global model.

[0055] Finally, the server repeats the above process of local training, scheduling decision-making, and model aggregation on the client side until the preset number of training rounds is reached or the global model meets the convergence condition, thus completing the asynchronous federated learning training.

[0056] Table 1. JFI at different levels of heterogeneity

[0057] To evaluate the impact of different methods on long-term client participation fairness during training, the Jain Fairness Index (JFI) was used for quantitative evaluation. The JFI value ranges from (0,1], with a value closer to 1 indicating a more balanced frequency of client participation in model training and aggregation, and better long-term participation fairness. Table 1 presents the comparison results of participation fairness for each method under two heterogeneous data conditions: α=0.1 and α=0.5.

[0058] Experimental results show that, under different data distribution conditions with varying degrees of heterogeneity, the JFI of the method proposed in this invention is close to, and significantly better than, the asynchronous federated learning method FedAsync and the buffer-based FedBuff method. This indicates that the method proposed in this invention can effectively alleviate the long-term uneven participation problem caused by client heterogeneity.

[0059] Figure 2 shows the comparison of model convergence performance of different methods on the MNIST dataset under two heterogeneous data conditions: α=0.1 and α=0.5. It can be seen that under both data distribution conditions, the model using the method of this invention can achieve rapid improvement in model accuracy within a short actual training time and maintain a stable convergence trend. Compared with the comparison methods, the method of this invention exhibits better training efficiency and convergence stability in different heterogeneous data scenarios, indicating that by predicting the client training completion time and introducing a participation fairness control mechanism, the timing and frequency of heterogeneous clients participating in model aggregation can be effectively coordinated.

[0060] The experimental results above demonstrate that the participation fairness-aware asynchronous federated learning method based on predictive scheduling proposed in this invention can balance overall training efficiency and model training performance in scenarios with varying degrees of data heterogeneity, exhibiting good versatility and robustness.

Claims

1. A participation fairness-aware asynchronous federated learning method based on predictive scheduling, characterized in that, The process includes the following steps: The client-side local training process and the server-side scheduling decision process are executed concurrently within the asynchronous federated learning framework, including the following steps: Step 1: The server initializes the global model and sets relevant parameters; the server distributes the global model and relevant parameters to all clients; Step 2: Client-side local training and model upload; Step 2-1: The client performs local training on its local dataset based on the global model, obtaining a local model update; Step 2-2: After completing local training, the client uploads the local model update to the server; Step 3: Server-side scheduling subset decision; Step 3-1: The server predicts the client's training completion time; Step 3-2: The server updates the fairness state parameters of each client; Step 3-3: The server combines the training completion time prediction results and the fairness state parameters to determine the set of clients participating in model aggregation in the current round; Step 4: Asynchronous model aggregation and iterative update; Step 4-1: The server aggregates the local models of the selected clients participating in model aggregation, obtaining the updated global model; Step 4-2: The server repeats steps 2 to 4-1 until the preset number of training rounds is reached or the global model meets the convergence condition.

2. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, The specific steps of the first step are as follows: The server initializes the global model and sets trade-off parameters to balance training latency and long-term participation fairness. The server is for all clients in the system. Set the corresponding fairness target parameters And initialize the fairness deficit state of each client. and fairness threshold ;in, This indicates the total number of clients in the system.

3. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, An additional step is added between step 2-1 and step 2-2: The client terminates the current training and restarts training based on the latest global model according to the server's instructions; the specific steps are as follows; on the client... If a restart command is received from the server during local training, then the client... Terminate the current local training task and use the latest global model parameters issued by the server. Restart local training.

4. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, The specific steps of step 3-1 are as follows: The server records the historical training time and communication time of each client, and predicts the client's time based on historical statistical information. The completion time of the training round; The server continuously records the local training time of each client in historical rounds. Communication time with model upload And calculate the client accordingly. Estimate the total runtime of completing a local training and model upload under the current training configuration. ； Full runtime estimate The exponential average method is used for updating, and its calculation method is as follows: ;in, The smoothing coefficient is used; the server simultaneously records the actual duration of each round of the system. ; the The actual training time of the wheel It is determined by the client with the longest training completion time among the clients participating in model aggregation. ;in, Indicates the first The set of candidate clients for the wheel, For the client In the An indicator variable for whether a wheel is selected to participate in model aggregation. Indicates client The The wheel is selected, otherwise ； Indicates client In the The remaining prediction time required for each round to complete local training and participate in model aggregation; the server records the client's prediction time. Current model obsolescence , indicating from the client The number of global model update rounds performed by the server since the last time global model parameters were received; server calculation client In the Remaining prediction time required for each round to complete local training and participate in model aggregation The calculation method is as follows: ;in, Indicated on the client The cumulative training time consumed by the system during the current model version's validity period; when When the client has completed local training or is in an idle state, it can immediately participate in model aggregation without increasing the system training latency of the current round.

5. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, The specific steps of step 3-2 are as follows; the server is each client in the system. Maintain the corresponding fairness state parameters The fairness state parameter is used to reflect the deviation between the actual frequency of a client's participation in model aggregation during historical training and its fairness objective. Before each round of scheduling decisions begins, the server determines the scheduling results based on the previous round's results. Calculate the fairness state parameters for each client. The calculation method is as follows: before the start of each round of scheduling decisions, the server calculates the results of the previous round of scheduling. Update the fairness state parameters for each client. The calculation method for this update mechanism is as follows: ;in, Indicates client Fairness objective parameters; The value represents the client's cumulative fairness deficit up to the current round; It has a dynamic accumulation characteristic: when a client is not selected to participate in model aggregation in multiple consecutive training rounds, its fairness state parameter... The fairness deficit gradually accumulates and increases; when a client is selected to participate in model aggregation, its fairness deficit is partially offset or cleared, thereby inhibiting further accumulation of fairness states; through this update mechanism, the server dynamically maps long-term participation fairness requirements into state variables directly used in single-round scheduling decisions without introducing cross-round constraint solutions; the updated fairness state parameters As one of the core inputs, it is directly fed into the current round. In the scheduling decision algorithm, the scheduling decision algorithm, based on jointly considering the training completion time prediction results, considers the fairness state parameters of each client. By dynamically adjusting their priority in scheduling decisions, clients with insufficient historical participation can receive a higher scheduling preference in subsequent rounds, thereby achieving a balance between the overall training efficiency of the system and the long-term fairness of client participation.

6. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, The step between steps 3-2 and 3-3 is to have the server perform predictive scheduling control and trigger the client to restart training. The specific steps are as follows: For clients not selected to participate in the current round of model aggregation... When its fairness state parameter As the data accumulates over multiple training rounds, the server adjusts the client's settings based on the changing trend of the fairness state parameter. The priority of participation in subsequent scheduling processes is predicted and evaluated; when the client The fairness state parameters satisfy At that time, among them The preset fairness threshold represents a critical state where the risk of insufficient client participation over a long period significantly impacts scheduling decisions. The server determines that the client has a significantly increased priority in subsequent scheduling processes. When the above condition is met, the server proactively informs the client... Distribute the latest global model parameters It also sends a restart command, instructing the client to terminate the current local training task and restart local training based on the latest global model parameters, in order to reduce the training latency risk and model obsolescence risk introduced when the client is subsequently selected to participate in model aggregation.

7. The participation fairness-aware asynchronous federated learning method based on predictive scheduling according to claim 1, characterized in that, The specific steps of step 3-3 are as follows: The server uses the current candidate client set... As input, considering the various candidate clients in the first... Predicted training completion time of rounds With fairness state parameters To minimize the overall cost of a single round of scheduling, the [number]th round of scheduling is determined. The set of clients aggregated in the round-robin model has a scheduling metric function defined as: The scheduling optimization problem is solved using an online selection method based on the greedy principle. The server first bases its training on the predicted completion time of each candidate client. For the candidate client set The clients in the system are sorted in ascending order to prioritize those that have a smaller impact on system training latency. Subsequently, the server iterates through the candidate clients in sorted order; for any candidate client... The server evaluated the settings respectively. and Comprehensive cost function in both cases The value changes; when set When the client is selected to participate in the current round of model aggregation, its corresponding fairness penalty is eliminated, but the training latency of the current round increases; when set At that time, the training latency of the current round remains unchanged, but when the fairness state parameter of the client... Exceeding the fairness threshold When appropriate fairness penalties are imposed, they should be introduced; when setting... The server selects a client to participate in the current round of model aggregation if it can reduce the overall cost function; otherwise, the server does not select the client. After all candidate clients have been traversed, the server obtains the result. The set of clients participating in the model aggregation round In scenarios where the number of participating clients is limited, a cap is set on the number of clients participating in each round of model aggregation. When the traversal is complete and no client is selected, the server will select at least one client with the shortest prediction training completion time to participate in model aggregation, in order to ensure the normal execution of the model aggregation process.

Images