Dynamic clustering federated learning method based on adaptive distribution similarity calculation

Abstract

The invention discloses a dynamic clustering federal learning method based on adaptive distribution similarity calculation. The method comprises the steps that a server receives local data feature vectors and parameter updates uploaded by clients; on the basis of the feature vectors and parameter updating, the self-adaptive distribution similarity between the clients is calculated, the self-adaptive distribution similarity comprises static data similarity and dynamic parameter similarity, and the weights of the static data similarity and the dynamic parameter similarity are dynamically balanced through a weight adjustment strategy; according to the self-adaptive distribution similarity, initial clustering division of a cold start stage is carried out on the client; in the iteration updating stage, cluster division is dynamically adjusted based on intra-group similarity updating and inter-group similarity updating until the result of continuous multi-round cluster division is stable; before each round of training, according to the pre-estimated training time and the residual battery capacity of the clients, executing hierarchical client selection to determine the clients participating in the training; and carrying out independent model training on the divided clusters until the model converges. According to the invention, the accuracy of clustering federal learning is improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of artificial intelligence, distributed machine learning, personalized learning, and model optimization, and particularly relates to a dynamic clustering federated learning method based on adaptive distribution similarity calculation. Background Art

[0002] With the rapid development of Internet of Things technology, a large amount of data is generated and needs to be uploaded to the cloud for processing, which has raised people's concerns about data privacy issues. Federated Learning (FL) can perform model training without sharing local data. The client performs local model training on its data, and then collaboratively constructs a global model by exchanging parameters, thus ensuring the privacy of the client data.

[0003] In FL, due to differences in factors such as geographical location, collection time, and other user characteristics, client data usually has different distributions, resulting in the problem of data heterogeneity (Non-IID), which affects the model accuracy and convergence speed of federated learning. The proposed Clustered Federated Learning (CFL) can effectively solve the data heterogeneity problem. In CFL, clients with similar data distributions are divided into a cluster, and a personalized model is trained for each cluster.

[0004] Existing CFL methods still have certain limitations. The primary one lies in how to divide each client into a suitable cluster. First, a distribution similarity metric under privacy protection is needed to measure the data similarity between clients. However, existing methods have a risk of privacy leakage, may require additional data sets and tests, and are not accurate enough.

[0005] Based on these distribution similarity metrics, CFL methods can be divided into two categories: one-time clustering methods and iterative clustering methods. Among them, the one-time federated clustering method only performs clustering once based on the pre-training result, which may lead to some clients being divided into inappropriate clusters, thus having a negative impact on the final model performance. In addition, since the data on edge devices may change over time, the one-time clustering method is not suitable for this dynamic scenario. And existing iterative clustering methods have large computational and communication overheads.

[0006] In addition, none of the existing CFL methods consider the system heterogeneity caused by different clients. Federated learning clients not only differ in data distribution but also in computing power, network bandwidth, and battery power. These factors pose further challenges, which may lead to uneven training speeds, synchronization problems, and model update delays. Therefore, in practical deployment scenarios where system heterogeneity is inevitable, the overall performance of the CFL framework will be severely affected. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a dynamic clustering federated learning method based on adaptive distribution similarity calculation to improve the accuracy of clustering federated learning.

[0008] To achieve the above object, the present invention provides a dynamic clustering federated learning method based on adaptive distribution similarity calculation, including:

[0009] The server receives the local data feature vectors and parameter updates uploaded by each client;

[0010] Based on the feature vectors and parameter updates, calculate the adaptive distribution similarity between clients, where the adaptive distribution similarity includes static data similarity and dynamic parameter similarity, and dynamically balance the weights of the two through a weight adjustment strategy;

[0011] According to the adaptive distribution similarity, perform an initial clustering division on the clients during the cold start phase;

[0012] In the iterative update phase, dynamically adjust the clustering division based on the within-group similarity update and the between-group similarity update until the clustering division results are stable for multiple consecutive rounds;

[0013] Before each round of training, perform hierarchical client selection according to the estimated training time and remaining battery power of the clients to determine the clients participating in the training;

[0014] Independently train the divided clusters until the model converges.

[0015] Optionally, the process of the initial clustering division in the cold start phase includes:

[0016] The client performs singular value decomposition on the local data, extracts a preset number of left singular vectors as feature vectors, and uploads them to the server;

[0017] The server calculates the initial similarity matrix between clients based on the feature vectors and the first-round parameter updates;

[0018] Apply a hierarchical clustering algorithm to the initial similarity matrix to generate an initial clustering division.

[0019] Optionally, the in-group similarity update process includes:

[0020] Clients within the same cluster participate in multiple rounds of local model training;

[0021] Update the parameter similarity between clients according to the latest parameter update calculation, and update the adaptive distribution similarity metric.

[0022] Optionally, the between-group similarity update process includes:

[0023] In global training, clients from different clusters participate in parameter updates;

[0024] Update the similarity metric value between clients based on the global training results and trigger reclustering.

[0025] Optionally, the parameter similarity update process includes:

[0026] Use the moving average method to combine the historical parameter update similarity and the current round parameter update similarity to generate the updated parameter similarity.

[0027] Optionally, the hierarchical client selection process includes:

[0028] Divide the clients within the same cluster into multiple groups with similar training times according to the estimated training time of the clients;

[0029] Give priority to selecting clients from the same group to participate in training.

[0030] Optionally, the hierarchical client selection process further includes:

[0031] Determine the selection priority of the clients according to the difference between the remaining battery power of the clients and the estimated energy consumption;

[0032] Clients with battery power higher than the estimated energy consumption have the opportunity to be selected, and the higher the remaining battery power, the higher the probability of being selected. Optionally, the calculation process of the estimated training time includes:

[0033] The client estimates the time required to complete local training and parameter upload based on its local computing power and network bandwidth.

[0034] Optionally, the calculation process of the estimated energy consumption includes:

[0035] The client estimates the total energy consumption of this round of training based on its local computing task volume and communication data volume.

[0036] Optionally, the determination process of the selection priority includes:

[0037] Allocate the selection probability of the clients proportionally according to the difference between the remaining battery power of the clients and the estimated energy consumption.

[0038] Technical effects of the present invention: The present invention discloses a dynamic clustering federated learning method based on adaptive distribution similarity calculation. On datasets such as FMNIST, CIFAR-10, and CIFAR-100, compared with algorithms such as FedAvg and FlexCFL, the test accuracy of the present invention is significantly improved, mainly due to the dual consideration of data features and parameter updates by the adaptive distribution similarity metric. Through the dynamic clustering algorithm, the defect that the existing CFL method cannot adapt to the dynamic changes of the client data distribution is solved, the problem of inaccurate one-time clustering division is avoided, and it is ensured that the clustering division is adaptively adjusted during the training process. The hierarchical client selection strategy reduces the waiting time caused by "stragglers" by grouping and selecting clients with similar training times, and experiments show that the average training time is significantly reduced; at the same time, the client selection is optimized by combining the remaining battery power to improve the system stability. The client only uploads the feature vectors extracted by singular value decomposition to avoid privacy leakage; at the same time, unnecessary parameter update rounds are reduced through dynamic clustering, reducing the communication cost. Description of the Drawings

[0039] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0040] Figure 1 It is a schematic flowchart of a dynamic clustering federated learning method based on adaptive distribution similarity calculation according to an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of client division in the cold start stage according to an embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of the hierarchical client selection process according to an embodiment of the present invention;

[0043] Figure 4 It is a schematic diagram of test accuracy comparison according to an embodiment of the present invention, where (a) is #label = 2 and (b) is #label = 3;

[0044] Figure 5 It is a schematic diagram of average training time comparison according to an embodiment of the present invention;

[0045] Figure 6 It is a schematic diagram of test accuracy comparison under different client selection strategies according to an embodiment of the present invention, where (a) is FMNIST, (b) is CIFAR-10, and (c) is CIFAR-100;

[0046] Figure 7 It is a flowchart of TDCFL according to an embodiment of the present invention. Detailed Embodiments

[0047] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0049] As Figure 1 shown, in this embodiment, a dynamic clustering federated learning method based on adaptive distribution similarity calculation is provided, including:

[0050] The server receives the local data feature vectors and parameter updates uploaded by each client;

[0051] Based on the feature vectors and parameter updates, calculate the adaptive distribution similarity between clients, where the adaptive distribution similarity includes static data similarity and dynamic parameter similarity, and dynamically balance the weights of the two through a weight adjustment strategy;

[0052] According to the adaptive distribution similarity, perform an initial clustering division on the clients in the cold start phase;

[0053] In the iterative update phase, dynamically adjust the clustering division based on the within-group similarity update and the between-group similarity update until the clustering division results are stable for multiple consecutive rounds;

[0054] Before each round of training, perform hierarchical client selection according to the estimated training time and remaining battery power of the clients to determine the clients participating in the training;

[0055] Perform independent model training on the divided clusters until the model converges.

[0056] Furthermore, the process of the initial clustering division in the cold start phase includes:

[0057] The client performs singular value decomposition on the local data, extracts a preset number of left singular vectors as feature vectors and uploads them to the server;

[0058] The server calculates the initial similarity matrix between clients based on the feature vectors and the first-round parameter updates;

[0059] Apply a hierarchical clustering algorithm to the initial similarity matrix to generate an initial clustering division.

[0060] Furthermore, the process of the within-group similarity update includes:

[0061] Clients within the same cluster participate in multiple rounds of local model training;

[0062] Update the parameter similarity between clients according to the latest parameters, and update the adaptive distribution similarity metric.

[0063] Furthermore, the inter-group similarity update process includes:

[0064] In global training, clients from different clusters participate in parameter updates;

[0065] Update the similarity metric value between clients based on the global training results and trigger reclustering.

[0066] Furthermore, the parameter similarity update process includes:

[0067] Adopt the moving average method to combine the historical parameter update similarity and the current round parameter update similarity to generate the updated parameter similarity.

[0068] Furthermore, the hierarchical client selection process includes:

[0069] Divide the clients within the same cluster into multiple groups with similar training times according to the estimated training time of the clients;

[0070] Give priority to selecting clients from the same group to participate in training.

[0071] Furthermore, the hierarchical client selection process also includes:

[0072] Determine the selection priority of the clients according to the difference between the remaining battery power and the estimated energy consumption of the clients;

[0073] Clients with battery power higher than the estimated energy consumption have the opportunity to be selected, and the higher the remaining battery power, the higher the probability of being selected. Furthermore, the calculation process of the estimated training time includes:

[0074] The client estimates the time required to complete local training and parameter upload based on local computing power and network bandwidth.

[0075] Furthermore, the calculation process of the estimated energy consumption includes:

[0076] The client estimates the total energy consumption of this round of training based on the local computing task volume and communication data volume.

[0077] Furthermore, the determination process of the selection priority includes:

[0078] Allocate the selection probability of the clients proportionally according to the difference between the remaining battery power and the estimated energy consumption of the clients.

[0079] Specifically, to solve the above problems, a dynamic clustering federated learning framework (TDCFL) is proposed, which improves the model accuracy and training efficiency of clustering federated learning. The present invention proposes an adaptive distribution similarity measure (ADSM), which takes into account both the data characteristics of each client and the weight updates of the model. As the training progresses, this measure is adaptively updated to accurately measure the data distribution similarity between clients. On this basis, the present invention designs a dynamic iterative clustering algorithm to adapt to clients with changing data distributions. In addition, a hierarchical client selection strategy is proposed to further improve the training efficiency. The overall architecture of TDCFL is as Figure 1 shown.

[0080] The adaptive distribution similarity measure consists of two parts: data similarity and parameter similarity.

[0081] (1) Data similarity (static): The present invention uses singular value decomposition (SVD) to obtain the eigenvectors representing the data characteristics of each client. The client data D i is represented as a d*n matrix, where d represents the dimension of each data sample and n represents the amount of data of this client. Apply SVD to the local data of the client and upload the p most important left singular vectors to the server. Then use the Golub-Werman subspace similarity distance to calculate the data similarity between clients C i and C j where M i and M j represent the feature matrices of two clients respectively.

[0082]

[0083] (2) Parameter similarity (dynamic): When a client completes one round of local training, the parameter updates are sent to the server. Then, the cosine similarity is used to calculate the parameter similarity between clients, where and represent the parameter updates of two clients at the t-th round respectively.

[0084]

[0085] (3) Adaptive distribution similarity metric: Combining these two parts gives the adaptive distribution similarity metric. Here, 0 ≤ α, β ≤ 1 and α + β = 1. The present invention designs a heuristic strategy to determine the values of these two parameters. First, at the beginning, a higher weight is assigned to data similarity. As training progresses, the parameter similarity gradually becomes more accurate in measuring the client distribution similarity. Therefore, the value of α is set to gradually decrease and the value of β is set to gradually increase. In each round of training, the ADSM term is dynamically updated based on the current training results, and finally the data distribution similarity of all clients is obtained.

[0086]

[0087] During model training, the update pattern of parameters may change significantly. Therefore, only a few rounds of pre-training and a one-time clustering method cannot guarantee satisfactory clustering division and adaptation to the changing environment. Therefore, based on the adaptive distribution similarity metric, the present invention proposes a dynamic clustering algorithm, which mainly includes two stages: cold start and iterative update.

[0088] (1) Cold start stage: In the first round of training, each participant first needs to perform singular value decomposition on the local data, and the obtained eigenvectors are transmitted to the server. At the same time, the weight updates obtained by the participants through local model training are also uploaded to the server. Then, based on this, the server calculates the similarity metric values between clients to form an initial similarity matrix M 0 , and then applies the hierarchical clustering algorithm to obtain the initial clustering.

[0089] In the cold start stage, due to the instability of the parameter update pattern, some clients may be assigned to inappropriate clusters, as Figure 2 shown. Therefore, next, the present invention introduces the clustering iterative update stage.

[0090] (2) Iterative update stage: The dynamic clustering strategy of the present invention updates the clustering division based on the update of the adaptive distribution similarity between clients. Since only some clients participate in training in each round, the present invention proposes two update strategies, namely within-group similarity update and between-group similarity update.

[0091] The steps of the within-group similarity update are as follows. After reclustering, each cluster uses the FedAvg algorithm for T′ rounds of model training. If clients C i and C j are selected simultaneously in the t-th round, the term s(C i , C j ) needs to be updated according to the latest training results t。In this case, clients from different clusters have no chance to update their similarity metrics, and inter-group similarity update is required. After the clustering model training in round T′, global training and inter-group similarity update are performed. In this way, clients belonging to different clusters can update their similarity metrics. Then re-clustering is performed, and new clusters are divided according to the current results. Intra-group similarity update, inter-group similarity update, and re-clustering processes are performed alternately until the clustering divisions remain unchanged for consecutive s rounds.

[0092] In the iterative update phase, how to calculate the update of the adaptive distribution similarity between two parties is a key issue. To retain the training history of the previous rounds and consider the influence of the current training results, the present invention uses the moving average method to obtain the parameters of client C l and C k at the t-th round

[0093]

[0094] where is the number of rounds that client C l and C k participated in the training simultaneously in the previous t - 1 rounds. The above formula indicates that if C l and C k participate in this round of training simultaneously, then will be updated to the sum of the historical similarity calculations in the previous t - 1 rounds and the current parameter update similarity result.

[0095] In clustering federated learning, few studies consider the system heterogeneity problem. In the framework of the present invention, a hierarchical client selection algorithm is introduced to determine the clients participating in the training in each round, improving the training efficiency of the model.

[0096] In federated learning, different clients have different CPU processing capabilities and network bandwidths. Therefore, the computing and communication times of different clients may vary greatly. In each round of training, model aggregation is only performed when all parties have completed local model training and parameter upload. Therefore, the training time of each round is determined by the "stragglers". To improve the efficiency of each round of training, an effective method is to select clients with similar computing and communication capabilities to participate in each round of training. Before the start of each round of training, clients need to evaluate their current computing capabilities and channel conditions and report their estimated training times

[0097]

[0098] After obtaining the estimated training time of each client, the present invention further divides the clients in each cluster into different groups, where clients with similar training times are assigned to the same group. When performing model training, client selection is first required. According to the client selection algorithm designed by the present invention, the first step is to determine which group of clients to select, and then further determine which clients in that group to select.

[0099] When the group to be selected is determined, the algorithm decides which clients can participate in this round of training based on the remaining battery power of the devices. Specifically, the clients with more remaining power have a greater chance of being selected. To obtain the estimated remaining battery power of client C in the t-th round, it is first necessary to estimate the total computational and communication energy consumption overhead of the client in this round of training. i For the estimated remaining battery power of client C in the t-th round, it is first necessary to estimate the total computational and communication energy consumption overhead of the client in this round of training.

[0100]

[0101] Finally, the selection probability of the clients in the group is as follows.

[0102]

[0103] Where is the estimated remaining battery power of client C. When the battery power of the client cannot support the training of the current round, the client will not be selected; otherwise, the selection probability of the client is obtained according to the amount of the remaining battery power. The hierarchical client selection process is as shown in i Figure 3 Figure 3 shown.

[0104] To prove the effectiveness of TDCFL, the present invention uses the model test accuracy as the performance evaluation index of the algorithm, and conducts comparative experiments on three datasets: FMNIST, CIFAR-10, and CIFAR-100. There are five objects for experimental comparison. FedAvg and FedProx are algorithms that only train one global model. FlexCFL is a one-time federated clustering algorithm. CFL and IFCA are two representative iterative federated clustering algorithms.

[0105] Figure 4 Shown in the figure is the experimental result on the CIFAR-10 dataset. This dataset contains a total of 10 classes. Suppose each client is randomly assigned 2 or 3 classes, and the LeNet-5 model is used for training. As shown in Figure 4As shown, the accuracy of the TDCFL algorithm is higher than that of FlexCFL and CFL. Both FlexCFL and CFL use the cosine similarity of model parameter updates as the distribution similarity metric for clients, while the adaptive distribution similarity metric (ADSM) in TDCFL not only considers parameter updates but also the data characteristics of each client, providing additional information for the clustering division process. Compared with TDCFL, when #label = 2, the accuracy of FlexCFL decreases by 10.82%. In addition, FlexCFL is a one-time federated clustering algorithm with poor adaptability to dynamic scenarios, which also reflects the superiority of the dynamic clustering strategy of TDCFL. The performance of the CFL algorithm is even worse, with an accuracy of only 54.62% when #label = 2.

[0106] The complete experimental results are shown in Table 1. The results show that in all cases, the algorithm of the present invention can outperform the comparative algorithms.

[0107] Table 1

[0108]

[0109] Figure 5 and Figure 6 Shown are the experimental results of the hierarchical client selection strategy in improving the model training efficiency.

[0110] Figure 5 The average training time per round under the random client selection strategy and the hierarchical client selection strategy is given, denoted as RS and HCS respectively. The average results of the first 50 rounds, 100 rounds, and 200 rounds of training are reported on the CIFAR-10 dataset when #label = 2. The hierarchical client selection strategy has higher training efficiency because in each round of training, it tends to select clients from the same group to participate in training, and these clients have similar training times. This can avoid the "straggler" problem and reduce the unnecessary waiting time of some clients, thus improving the overall training efficiency. On the contrary, in the random client selection strategy, the faster clients always need to wait for the slower clients to complete local training and parameter transmission, so the efficiency is lower. In addition, it tends to select clients with more battery power to participate in this round of training, avoiding the situation where the selected clients drop out due to battery exhaustion, and giving more clients the opportunity to be selected in subsequent training.

[0111] Figure 6 Given is the impact of the hierarchical client selection strategy on the test accuracy. The results show that compared with the traditional random client selection strategy, the hierarchical client selection strategy hardly affects the accuracy of the clustering model.

[0112] The overall process of TDCFL is as Figure 7 shown.

[0113] First, in the first round of training, each client calculates the adaptive distribution similarity metric between clients by uploading the eigenvectors obtained from singular value decomposition and the parameters updated through local model training, and then obtains the initial clustering partition through the hierarchical clustering algorithm.

[0114] Then, as the training progresses, the adaptive distribution similarity metric between clients is updated according to formula (4), and the clients are reclustered according to the results until the clustering results remain unchanged for consecutive S rounds, obtaining the final clustering partition.

[0115] After obtaining the final clustering partition, independent model training is performed in each cluster. Before the start of each round of training, the computing power, network condition, and battery power of the clients are evaluated, and the clients participating in this round of training are determined according to the hierarchical client selection strategy. Then, the local model training and global aggregation processes are iterated until the model converges, obtaining the models for each cluster.

[0116] The present invention discloses a dynamic clustering federated learning method based on adaptive distribution similarity calculation. On datasets such as FMNIST, CIFAR-10, and CIFAR-100, compared with algorithms such as FedAvg and FlexCFL, the test accuracy of the present invention is significantly improved, mainly due to the dual consideration of data features and parameter updates by the adaptive distribution similarity metric. Through the dynamic clustering algorithm, the defect that the existing CFL method cannot adapt to the dynamic changes of the client data distribution is solved, the problem of inaccurate one-time clustering partition is avoided, and it is ensured that the clustering partition is adaptively adjusted during the training process. The hierarchical client selection strategy reduces the waiting time caused by "stragglers" by grouping and selecting clients with similar training times, and the experimental results show that the average training time is significantly reduced; at the same time, the remaining battery power is combined to optimize the client selection, improving the system stability. The client only uploads the eigenvectors extracted by singular value decomposition, avoiding privacy leakage; at the same time, unnecessary parameter update rounds are reduced through dynamic clustering, reducing the communication cost.

[0117] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A dynamic clustering federated learning method based on adaptive distribution similarity calculation, characterized in that, Including: The server receives the local data feature vectors and parameter updates uploaded by each client; Based on the feature vectors and parameter updates, calculate the adaptive distribution similarity between clients, where the adaptive distribution similarity includes static data similarity and dynamic parameter similarity, and dynamically balance the weights of the two through a weight adjustment strategy; According to the adaptive distribution similarity, perform an initial clustering division for clients in the cold start phase; In the iterative update phase, dynamically adjust the clustering division based on the within-group similarity update and between-group similarity update until the clustering division results are stable for multiple consecutive rounds; Before each round of training, based on the estimated training time and remaining battery power of the client, perform hierarchical client selection to determine the clients participating in the training; Perform independent model training on the divided clusters until the model converges.

2. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, characterized in that, The process of the initial clustering division in the cold start phase includes: The client performs singular value decomposition on the local data, extracts a preset number of left singular vectors as feature vectors and uploads them to the server; The server calculates the initial similarity matrix between clients based on the feature vectors and the first-round parameter updates; Apply the hierarchical clustering algorithm to the initial similarity matrix to generate the initial clustering division.

3. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, wherein The process of the within-group similarity update includes: Clients within the same cluster participate in multiple rounds of local model training; Calculate the parameter similarity between clients based on the latest parameter updates, and update the adaptive distribution similarity metric.

4. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, wherein The process of the between-group similarity update includes: In global training, clients in different clusters participate in parameter updates; Update the similarity metric value between clients based on the global training results and trigger reclustering.

5. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, wherein, The process of the parameter similarity update includes: Adopt the moving average method to combine the historical parameter update similarity and the current round parameter update similarity to generate the updated parameter similarity.

6. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, characterized in that, The process of the hierarchical client selection includes: According to the estimated training time of the client, divide the clients within the same cluster into multiple groups with similar training times; Give priority to selecting clients from the same group to participate in the training.

7. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, characterized in that, The process of the hierarchical client selection also includes: Determine the selection priority of the client according to the difference between the remaining battery power and the estimated energy consumption of the client; Clients with battery power higher than the estimated energy consumption have the opportunity to be selected, and the higher the remaining battery power, the higher the probability of being selected.

8. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, characterized in that, The calculation process of the estimated training time includes: The client estimates the time required to complete local training and parameter upload based on the local computing power and network bandwidth.

9. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, characterized in that, The calculation process of the estimated energy consumption includes: The client estimates the total energy consumption of this round of training based on the local computing task volume and communication data volume.

10. The dynamic clustering federated learning method based on adaptive distribution similarity calculation according to claim 1, wherein The process of determining the selection priority includes: According to the difference between the remaining battery power and the estimated energy consumption of the client, allocate the selection probability of the client proportionally.

Images