Asynchronous federated learning method and system based on T-Step aggregation algorithm

Abstract

The present invention provides an asynchronous federated learning method and system based on the T-Step aggregation algorithm. All client devices willing to participate in federated learning can join the training process at any time. The central server updates the global model when it receives the client training results that meet the preset number of asynchronous steps, so that the global model can be updated as quickly as possible without waiting for slow clients. Furthermore, the method of this embodiment adopts the T-Step asynchronous step aggregation algorithm, so that "laggard" clients can also participate in the global model aggregation, and at the same time, the negative impact of the outdated model parameters of the "laggard" on the convergence of the global model is reduced by dynamically reducing the weight, so that the trained model has higher accuracy. In summary, the federated learning method of this embodiment encourages devices with heterogeneous resources to collaboratively train the global model in an asynchronous manner, and can speed up the training speed and model convergence speed.

Description

Technical Field

[0001] The present invention belongs to the field of federated learning technology, and specifically relates to an asynchronous federated learning method and system based on the T-Step asynchronous step aggregation algorithm. Background Art

[0002] Federated learning is a new distributed machine learning framework in which many clients (e.g., mobile devices or entire organizations) jointly train models under the coordination of a central server (e.g., a service provider), while maintaining the decentralization and distribution of training data. It embodies the principles of centralized collection and data minimization and can mitigate many systemic privacy risks and costs associated with traditional centralized machine learning. The typical federated learning process involves a central server initializing global model parameters. All participating clients download these parameters locally. After training on their local data, they upload the updated model parameters to the central server for aggregation, completing a round of model training. This process is repeated until the model converges or reaches a set upper limit for training rounds. In federated learning, each round of training can involve all or only a subset of clients. However, if a large number of client devices participate, uploading parameters to the central server may result in network performance bottlenecks. Furthermore, the central server must wait for updates from all participating clients before aggregating them, increasing training time for model convergence. However, if the number of devices participating in each round of training is small, the data samples participating in each training will only be a small part of the sample space. In order to cover the entire sample space as much as possible, this will prolong the model training cycle, making it difficult for the model to converge quickly. In order to find a balance between the two, there are usually two federated learning computing and communication architectures. One is the synchronous federated learning architecture represented by FevAvg. Under this architecture, each federated learning will randomly select a subset of clients from the entire client set to participate in this round of training. The central server needs to receive parameter updates from all clients participating in this round of training before starting global model weighted aggregation, such as Figure 10 shown.

[0003] Due to differences in hardware (such as CPU capacity and memory size), software (such as operating systems), and available data on devices, the computing power of client devices in a federated learning system is highly diverse. Furthermore, the widespread availability of communication technologies (such as 4G, 5G, and WiFi) and fluctuating network traffic mean that the network connection quality of devices participating in federated training is often unstable. Most user devices will only participate in federated learning training when they are charging and connected to WiFi; otherwise, they will pause or exit the entire training process. Consequently, the heterogeneity of device and network connection status will affect client devices training their local models at different speeds. The time it takes to upload parameter updates to the central server and the feedback delay will also vary significantly. This leads to the "straggler" problem, where the time spent on model training in each round of federated learning is often determined by the device that computes and uploads parameter updates the slowest among the participating client devices. Stragglers in federated learning significantly increase the time it takes for global model training to converge, reducing overall federated learning performance.

[0004] To address the "straggler" problem, existing synchronous federated learning architectures mostly improve and optimize device selection and global model weighted aggregation. Li et al. proposed selecting a smaller proportion of clients for training in each global iteration to mitigate the impact of "stragglers." However, this requires more training rounds for model convergence. Experiments have also shown that selecting a smaller number of clients per training round results in lower performance. FedProx addresses system heterogeneity by applying different local epochs to clients. However, selecting a perfect local epoch for each client is challenging in practice.

[0005] Although synchronous communication can accelerate federated learning training to a certain extent, the impact of the "laggard" problem is still difficult to eliminate. Another solution is the asynchronous federated learning architecture represented by FedAsyn. In this architecture, all clients train models in a fully distributed manner and upload parameter updates to the server. The server aggregates the global model after receiving the parameter updates from each client, such as Figure 11 As shown in the figure, the asynchronous approach avoids the need to wait for feedback from the slowest client during each training round. However, the central server aggregates the global model every time it receives an update from a client, which can lengthen the model convergence period. This is because each client device has a limited number of samples. To cover global data samples during training, more training rounds are required. Furthermore, devices with fast computing speeds and good network conditions participate more frequently in global model training, which can lead to overfitting of the global model on samples from these devices, reducing the model's predictive performance on global data samples. Summary of the Invention

[0006] The present invention is designed to address the above-mentioned issues and aims to provide a novel asynchronous federated learning method and corresponding system. This federated learning method enables servers to accelerate global model training without being affected by "laggard" clients. It also allows "laggards" to participate in model aggregation and mitigate the impact of their model staleness on the global model. The present invention adopts the following technical solutions:

[0007] The present invention provides an asynchronous federated learning method based on the T-Step aggregation algorithm, which is characterized by comprising the following steps:

[0008] Step S1: The central server initializes global model parameters and sends the global model parameters to all client devices participating in federated training. The central server maintains an asynchronous step length T.

[0009] In step S2, each client device starts local model training using its local data based on the received global model parameters, and sends the generated model parameter updates and the current number of local training rounds to the central server;

[0010] In step S3, the central server continuously receives the model parameter updates and the number of local training rounds sent by each client device. Once T model parameter updates are received, the central server performs global model aggregation using a T-Step asynchronous step-size aggregation algorithm to update the global model parameters.

[0011] Step S4: After the global model aggregation is completed, the number of global training rounds is increased by one, and the central server sends the latest global model parameters and the number of global training rounds to the client devices participating in this round of aggregation, and the client devices start the next round of local model training;

[0012] Step S5: repeat steps S2 to S4. When the global model converges to the preset target accuracy or the number of global training rounds reaches the preset upper threshold, the training ends.

[0013] The T-Step asynchronous step size aggregation algorithm compares the global number of training rounds and the local number of training rounds, and uses the inverse of the difference between the global number of training rounds and the local number of training rounds as a weight coefficient to eliminate the impact of outdated model parameter updates on global model aggregation.

[0014] The asynchronous federated learning method based on the T-Step aggregation algorithm provided by the present invention may also have such a technical feature, wherein, in step S3, global model aggregation is performed according to the following formula:

[0015]

[0016] in:

[0017]

[0018] Where, is the global model parameter after t rounds of update, t is the number of global training rounds, t k is the number of local training rounds for the kth client device, n k is the total number of samples of the kth client device, K is the total number of client devices, is the model parameter update uploaded by the k-th client device in the t-th round, C t is the set of client devices participating in the t-th round of global model aggregation.

[0019] The asynchronous federated learning method based on the T-Step aggregation algorithm provided by the present invention may also have such a technical feature, wherein the update of the outdated model parameters means that the corresponding number of local training rounds is much smaller than the number of global training rounds used as a benchmark, and the difference between the two is greater than a preset threshold.

[0020] The present invention provides an asynchronous federated learning system based on the T-Step aggregation algorithm, characterized in that it includes: a central server; and multiple client devices, which are communicatively connected to the central server, wherein the client devices include: a local database storing local data; and a local model training module for training local models, and the central server includes: a model parameter generation module for initializing global model parameters and sending the global model parameters to all the client devices participating in the federated training, the client devices start local model training based on the received global model parameters using their local data, and send the generated model parameter updates and the current number of local training rounds to the central server; a parameter storage module for storing and maintaining the asynchronous step size T, the number of global training rounds, the target accuracy, and the upper threshold limit of the number of global training rounds; a global model aggregation module for continuously receiving data sent by each of the client devices The model parameter updates and the local training round numbers are received. Once T model parameter updates are received, global model aggregation is performed through the T-Step asynchronous step-size aggregation algorithm to update the global model parameters; a global model sending module is used to send the latest global model parameters and the global training round numbers to the client devices participating in this round of aggregation after the global model aggregation is completed, and the client devices start the next round of local model training; and a training completion judgment module is used to judge whether the global model converges to the target accuracy or whether the global training round number reaches the upper threshold value. When the judgment is yes, the training ends, wherein the T-Step asynchronous step-size aggregation algorithm compares the global training round number and the local training round number, and uses the inverse of the difference between the global training round number and the local training round number as the weight coefficient to eliminate the influence of outdated model parameter updates on the global model aggregation.

[0021] Functions and effects of the invention

[0022] According to the asynchronous federated learning method and system based on the T-Step aggregation algorithm of the present invention, all client devices willing to participate in federated learning can join the training process at any time. When the central server receives the client training results that meet the preset number of asynchronous steps, it updates the global model, so that the global model can be updated as quickly as possible without waiting for slow clients. Furthermore, the method of this embodiment adopts the T-Step asynchronous step aggregation algorithm, so that the "laggard" clients can also participate in the global model aggregation, and at the same time, the negative impact of the outdated model parameters of the "laggard" on the convergence of the global model is reduced by dynamically reducing the weight, so that the trained model has higher accuracy. In summary, the federated learning method of the present invention encourages devices with heterogeneous resources to collaboratively train the global model in an asynchronous manner, and can speed up the training speed and model convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of an asynchronous federated learning method based on the T-Step aggregation algorithm in an embodiment of the present invention;

[0024] Figure 2 is a flowchart of an asynchronous federated learning method based on the T-Step aggregation algorithm in an embodiment of the present invention;

[0025] Figure 3 2 is a comparison chart of experimental results of two federated learning methods when ρ=0 in an embodiment of the present invention;

[0026] Figure 4 2 is a comparison chart of experimental results of two federated learning methods when ρ = 0.4 in an embodiment of the present invention;

[0027] Figure 5 2 is a comparison chart of experimental results of two federated learning methods when ρ = 0.5 in an embodiment of the present invention;

[0028] Figure 6 2 is a comparison chart of the experimental results of two federated learning methods when ρ = 0.7 in an embodiment of the present invention;

[0029] Figure 7 2 is a comparison chart of experimental results of two federated learning methods when σ=0.8 in an embodiment of the present invention;

[0030] Figure 8 This is a comparison chart of the model convergence time of two federated learning methods in an embodiment of the present invention;

[0031] Figure 9 This is a structural block diagram of an asynchronous federated learning system based on the T-Step aggregation algorithm in an embodiment of the present invention;

[0032] Figure 10 This is a schematic diagram of the synchronous federated learning architecture in the existing technology;

[0033] Figure 11 This is a schematic diagram of the asynchronous federated learning architecture in the existing technology. DETAILED DESCRIPTION

[0034] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the following is a detailed description of the asynchronous federated learning method and system based on the T-Step aggregation algorithm of the present invention in combination with embodiments and drawings.

[0035] <Example>

[0036] Figure 1 This is a schematic diagram of the asynchronous federated learning method based on the T-Step aggregation algorithm in this embodiment.

[0037] Figure 2This is a flowchart of the asynchronous federated learning method based on the T-Step aggregation algorithm in this embodiment.

[0038] like Figure 1-2 As shown in Figure 1, the asynchronous federated learning method based on the T-Step aggregation algorithm (also known as the AsyFed method) specifically includes the following steps:

[0039] In step S1, the central server initializes the global model parameters and sends them to all client devices participating in the federated training. The central server maintains an asynchronous step size T.

[0040] Right now Figure 1 The operation shown in ①.

[0041] In this embodiment, parameters such as the asynchronous step size T, the number of global training rounds, the target accuracy of the global model, and the upper threshold for the number of global training rounds are all stored and maintained by the central server. The number of local training rounds for each client is stored and maintained by each client.

[0042] In addition, if Figure 1 As shown, in this embodiment, the number of client devices is three, and the asynchronous step length T=2. In fact, the number of clients can be more, and other clients willing to participate in federated learning can also join the training process at any time, that is, the number of clients can change dynamically.

[0043] In step S2, each client device starts local model training using its local data based on the received global model parameters, and sends the generated model parameter updates and the current number of local training rounds to the central server.

[0044] Right now Figure 1 Follow the steps shown in ②.

[0045] In step S3, the central server continuously receives model parameter updates and local training round numbers from each client device. Once model parameter updates from T client devices are received, the central server performs global model aggregation using the T-Step asynchronous step aggregation algorithm to update the global model parameters.

[0046] Right now Figure 1 The operation shown in ③.

[0047] The T-Step asynchronous step aggregation algorithm is as follows:

[0048] Once the central server receives uploaded model parameter updates from T client devices, the central server updates the global model parameters.

[0049] use represents the local model parameters of the i-th client device, t represents the time index of the global model update, that is, the global update round, then the global model after t updates is recorded as Use C t Denotes the set of clients participating in the global model aggregation in round t. The collaborative training problem is defined below and expressed in the form of a mathematical formula.

[0050] Taking the M-classification problem as an example, it is defined in the feature space X and the label space Y = [M], where M = {1, 2, ..., M} means there are M classes in the sample space. The data samples {x, y} are distributed on clients with different distributions p. In the stage of training the classifier (i.e., the local model), the goal is to find a function f that maps X to the probability distribution S, where Typically, the function f is a neural network classification function that can be parameterized by the weights w of the neural network. i represents the probability of the i-th category for a given feature. The goal of local model training is to iteratively update w using training samples (i.e., local data of the client device) to minimize the loss function. In this embodiment, the cross entropy loss is used to define the overall loss l(w), which is defined as the following formula (1):

[0051]

[0052] The goal of local model training is to minimize the overall loss function, which is defined as the following formula (2):

[0053]

[0054] Assume that the number of data samples of the kth client is n k , the total number of clients is K, then the total amount of sample space is defined as the following formula (3):

[0055] n=∑ k∈K n k (3)

[0056] In order to satisfy the optimization formula (2), the update of w should be completed by the central server. Then the client updates the model parameters (i.e., the gradient difference) and the number of rounds of current training t k Sent to the central server, represents the model parameter update uploaded by the k-th client device in the t-th round. The central server aggregates the first T model parameter updates that arrive. Assuming that the global model has been updated t times, then the global model for the next stage can be calculated based on these T updates. The stale model parameter update refers to the number of local training rounds reported by the client (the number of local update rounds on the client) t kThe case where the number of benchmark global training rounds t is much smaller than that of the benchmark global training rounds t, that is, the difference between the two is greater than a preset threshold.

[0057] The T-Step asynchronous step aggregation algorithm in this embodiment mainly compares the current global training round number t and the local training round number t received from the client. k , and use the inverse of the difference between the two as the weight coefficient to eliminate the influence of outdated local model parameters on global model aggregation. Specifically, the aggregation operation is performed using the following formula (4):

[0058]

[0059] As shown in formula (4), are the global model parameters in round t, are the global model parameters in round t+1. Only the first T model parameter updates are included in the global model aggregation of this round. It is worth emphasizing that the T-Step asynchronous step-size aggregation algorithm has two advantages. First, devices with abundant data but limited resources and unstable networks can participate in federated learning training without wasting local computing resources and fully utilize their local data resources. Second, the weight decay method ensures that slow client devices do not significantly affect global parameters.

[0060] In step S4, after the global model aggregation is completed, the number of global training rounds is increased by one, and the central server sends the latest global model parameters and the number of global training rounds to each client device participating in this round of aggregation, and these client devices start the next round of local model training.

[0061] Right now Figure 1 The operations shown in ④ and ⑤.

[0062] Step S5: repeat the above steps S2-S4. When the global model converges to the preset target accuracy or the number of global training rounds reaches the preset upper threshold, the training ends.

[0063] Right now Figure 1 Step ⑥ and subsequent operations.

[0064] Figure 3-7 These are comparison charts of experimental results of the two federated learning methods in this embodiment under different data distributions; Figure 8 This is a comparison chart of the model convergence time of the two federated learning methods in this embodiment.

[0065] like Figure 3-8As shown, in this embodiment, experiments were conducted using the open-source MNIST dataset. A deep neural network classifier was trained using both the prior art FedAvg and the present embodiment's AsyFed, and the two were compared to evaluate the performance of the AsyFed method. The experiment also evaluated robustness under different data distributions. Samples distributed across different clients were divided according to the proportion of label categories, with the values of σ set to 0, 0.4, 0.5, 0.7, and 0.8, respectively. σ = 0 indicates that the data is independent and identically distributed, meaning that the data labels on each client are uniformly distributed. σ = 0.5 indicates that 50% of the data labels on the clients are of the same category, and the remaining 50% are uniformly distributed.

[0066] Figure 3-7 The experimental results of the impact of data distribution on the robustness of AsyFed are shown respectively. It can be seen that as the unevenness of data distribution increases, AsyFed has a significant advantage over FedAvg.

[0067] Figure 8 The model convergence time under different distributions is shown. The target accuracy is 99%. It can be seen that under different data distributions, AysFed's model convergence speed is 70.40%, 78.78%, 24.42%, 6.31% and 44.33% faster than FedAvg respectively.

[0068] Through the above comparative experiments, the performance and robustness of the AsyFed method of this embodiment are well demonstrated.

[0069] Figure 9 This is a structural block diagram of the asynchronous federated learning system based on the T-Step aggregation algorithm in this embodiment.

[0070] like Figure 9 As shown, this embodiment also provides an asynchronous federated learning system 10 based on the T-Step aggregation algorithm, including a central server 11 and a plurality of client devices 12 communicatively connected to the central server 11.

[0071] The central server 11 includes a model parameter generation module 111, a parameter storage module 112, a global model aggregation module 113, a global model delivery module 114, a training completion determination module 115, a server-side communication module 116, and a server-side control module 117. The model parameter generation module 111 generates initialized model parameters according to the method of step S1 above and distributes them to each client device 12; the parameter storage module 112 stores parameters such as the asynchronous step size T, the number of global training rounds, the target accuracy, and the upper threshold limit of the global training round; the global model aggregation module 113 aggregates the global model according to the method of step S3 above; the global model delivery module 114 delivers the global model to each client device 12 according to the method of step S4 above; the training completion determination module 115 determines whether the conditions have been met and whether the training has been completed according to the method of step S5 above; the server-side communication module 116 is used to communicate with each client device 12; and the server-side control module 117 controls the operation of each of the above modules.

[0072] The client device 12 includes a local database 121, a local model training module 122, a client-side communication module 123, and a client-side control module 124. The local database 121 stores data used for local model training; the local model training module 122 trains the local model according to the method described in step S2 above; the client-side communication module 123 is used to communicate with the central server 11; and the client-side control module 124 controls the operation of each of the above modules.

[0073] In this embodiment, parts not described in detail are well-known technologies in the art.

[0074] Example Function and Effect

[0075] According to the asynchronous federated learning method and system based on the T-Step aggregation algorithm provided by this embodiment, all client devices willing to participate in federated learning can join the training process at any time. When the central server receives the client training results that meet the preset number of asynchronous steps, it updates the global model, so that the global model can be updated as quickly as possible without waiting for slow clients. Furthermore, the method of this embodiment adopts the T-Step asynchronous step aggregation algorithm, so that "laggard" clients can also participate in the global model aggregation, and at the same time, the negative impact of the outdated model parameters of the "laggard" on the convergence of the global model is reduced by dynamically reducing the weight, so that the trained model is more accurate. In summary, the federated learning method of this embodiment encourages devices with heterogeneous resources to collaboratively train the global model in an asynchronous manner, and can speed up the training speed and model convergence speed.

[0076] In the embodiment, the effect of asynchronous step size on convergence speed was discussed experimentally. Experiments were conducted on different data sets, and the level of non-independent and identically distributed data was changed to verify the algorithm effect. After comparative experiments, the asynchronous federated learning method based on the T-Step aggregation algorithm of this embodiment significantly outperformed the standard synchronous federated learning scheme FedAvg in the prior art in terms of training time. Under both independent and identically distributed data distributions and non-independent and identically distributed data, the method of this embodiment can make the model converge faster, with performance advantages and robustness. It can solve the needs of multiple participants in actual production and life to safely and quickly train large-scale models, ensure data security, reduce the cost of multi-party training, and improve model training efficiency and model accuracy.

[0077] The above embodiments are only used to illustrate specific implementations of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. An asynchronous federated learning method based on the T-Step aggregation algorithm, characterized in that: The following steps are involved: Step S1: The central server initializes global model parameters and sends the global model parameters to all client devices participating in federated training. The central server maintains an asynchronous step length T. In step S2, each client device starts local model training using its local data based on the received global model parameters, and sends the generated model parameter updates and the current number of local training rounds to the central server; In step S3, the central server continuously receives the model parameter updates and the number of local training rounds sent by each client device. Once T model parameter updates are received, the central server performs global model aggregation using a T-Step asynchronous step-size aggregation algorithm to update the global model parameters. Step S4: After the global model aggregation is completed, the number of global training rounds is increased by one, and the central server sends the latest global model parameters and the number of global training rounds to the client devices participating in this round of aggregation, and the client devices start the next round of local model training; Step S5: repeat steps S2 to S4. When the global model converges to the preset target accuracy or the number of global training rounds reaches the preset upper threshold, the training ends. The T-Step asynchronous step size aggregation algorithm compares the global number of training rounds and the local number of training rounds, and uses the inverse of the difference between the global number of training rounds and the local number of training rounds as a weight coefficient to eliminate the impact of outdated model parameter updates on global model aggregation.

2. The asynchronous federated learning method based on the T-Step aggregation algorithm according to claim 1 is characterized by: in, In step S3, global model aggregation is performed according to the following formula: in: Where, is the global model parameter after t rounds of update, t is the number of global training rounds, t k is the number of local training rounds for the kth client device, n k is the total number of samples of the kth client device, K is the total number of client devices, is the model parameter update uploaded by the k-th client device in the t-th round, C t is the set of client devices participating in the t-th round of global model aggregation.

3. The asynchronous federated learning method based on the T-Step aggregation algorithm according to claim 1 is characterized by: in, The outdated model parameter update means that the number of local training rounds corresponding to the update is much smaller than the number of global training rounds used as a benchmark, and the difference between the two is greater than a preset threshold.

4. An asynchronous federated learning system based on the T-Step aggregation algorithm, characterized in that: include: Central server; as well as A plurality of client devices are connected in communication with the central server, The client device includes: A local database storing local data; and Local model training module, used to train local models, The central server includes: a model parameter generation module, configured to initialize global model parameters and send the global model parameters to all client devices participating in federated training; the client devices, based on the received global model parameters, start local model training using their local data, and send the generated model parameter updates and the current number of local training rounds to the central server; Parameter storage module, used to store and maintain the asynchronous step size T, the number of global training rounds, the target accuracy, and the upper threshold of the number of global training rounds; a global model aggregation module, configured to continuously receive the model parameter updates and the number of local training rounds sent by each of the client devices, and upon receiving T model parameter updates, perform global model aggregation using a T-Step asynchronous step-size aggregation algorithm to update the global model parameters; A global model delivery module is configured to deliver the latest global model parameters and the number of global training rounds to the client devices participating in the current round of aggregation after the global model aggregation is completed, so that the client devices start the next round of local model training; and The training completion judgment module judges whether the global model converges to the target accuracy or whether the number of global training rounds reaches the upper threshold. If the judgment is yes, the training ends. The T-Step asynchronous step size aggregation algorithm compares the global number of training rounds and the local number of training rounds, and uses the inverse of the difference between the global number of training rounds and the local number of training rounds as a weight coefficient to eliminate the impact of outdated model parameter updates on global model aggregation.

Images