Model training method based on client clustering hybrid federated split learning

Abstract

The application discloses a model training method based on client clustering hybrid federated split learning according to an embodiment of the application, comprising: each learning client clusters with the surrounding auxiliary clients according to a clustering strategy to obtain a plurality of learning clusters; a special client is selected as a cluster head in each learning cluster according to a cluster head selection strategy, so that the local base station allocates bandwidth to the cluster head according to a base station bandwidth allocation strategy; the cluster head obtains the global module issued by the local base station according to the bandwidth allocated by the local base station; the cluster head cuts the global model into a plurality of model segments according to a model cutting and allocation strategy, and distributes the model segments to all cluster members for cooperative training; the cluster head collects the local models trained by all cluster members in the cluster and uploads the local models to the local base station, so that the local base station aggregates the local models uploaded by the cluster head, and the user computing, storage and communication load is small, and the user parallelism and sensitivity to user non-independent and identically distributed data are high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a model training method based on client clustering and hybrid federated split learning. BACKGROUND

[0002] In recent years, machine learning technology based on deep neural networks has been widely used in intelligent transportation, Internet of Things, and electronic health fields due to its excellent learning and prediction capabilities. In order to process the massive data generated by intelligent devices while ensuring the learning performance of machine learning, centralized machine learning based on data centers and cloud computing provides an effective solution. The centralized machine method requires extensive deployment of devices on the network side to collect data, and uploads the collected data to a central processor with powerful computing capabilities to speed up model training. However, the centralized machine learning method stores training data in a single server, which may cause a risk of privacy data leakage, and transmitting a large number of data samples may cause serious transmission delay.

[0003] In order to solve the above problems, a distributed machine learning method called federated learning is proposed. The federated learning system uses a group of distributed client devices, each of which trains an ML model based on its local training data in parallel, which can effectively reduce the transmission delay. The client device only exchanges local model parameters with the FL server to aggregate the global model, while keeping the local training data, which helps to protect data privacy. However, the federated learning system requires a large amount of computing, storage, and communication resources to meet the frequent downloading and uploading of model parameters between the client and the FL server, which increases the cost of client computing, storage, and communication resources, further reducing the FL training performance and convergence speed.

[0004] In order to cope with the limited resources of the client, a distributed machine learning method called split learning is adopted. SL divides the model between the client and the server through a split layer. The client samples a small batch of data for local training and performs forward propagation (FP), and uploads the intermediate output data and label data to the SL server until reaching the output layer. The SL server updates its model based on the output result and label of the model, and sends the BP gradient data to the client to complete the client model update. Since SL only requires the client to complete part of the model training, the amount of computation of the client can be significantly reduced. However, the SL client needs to interact with the SL server frequently, and the SL server still needs to bear a considerable amount of computing workload. SUMMARY

[0005] The present application aims to at least partially solve one of the above technical problems in the art. To this end, one object of the present application is to provide a model training method based on client clustering for hybrid federated split learning, which not only makes the user computation, storage and communication load of federated learning small, but also makes the user parallelism and sensitivity to user non-independent and identically distributed data of split learning high.

[0006] To achieve the above object, the present application provides a model training method based on client clustering for hybrid federated split learning, comprising the following steps: each learning client clusters with the surrounding auxiliary clients according to a clustering strategy to obtain a plurality of learning clusters, wherein each learning cluster has and only has one learning client, and each learning cluster after clustering can not include auxiliary clients or include an unequal number of auxiliary clients; a special client in each learning cluster is selected as a cluster head according to a cluster head selection strategy, so that the local base station allocates bandwidth to the cluster head of each learning cluster according to a base station bandwidth allocation strategy; the cluster head of each learning cluster obtains the global module issued by the local base station according to the bandwidth allocated by the local base station, the cluster head of each learning cluster cuts the global model into a plurality of model segments according to a model cutting and allocation strategy, and distributes them to all cluster members in the learning cluster for collaborative training, so as to complete intra-cluster training; the cluster head of each learning cluster collects the local models trained by all cluster members in the cluster and uploads them to the local base station, so that the local base station aggregates the local models uploaded by the cluster head of each learning cluster to complete global model updating.

[0007] According to the model training method based on client clustering for hybrid federated split learning, first, each learning client clusters with the surrounding auxiliary clients according to a clustering strategy to obtain a plurality of learning clusters, wherein each learning cluster has and only has one learning client, and each learning cluster after clustering can not include auxiliary clients or include an unequal number of auxiliary clients; then, a special client in each learning cluster is selected as a cluster head according to a cluster head selection strategy, so that the local base station allocates bandwidth to the cluster head of each learning cluster according to a base station bandwidth allocation strategy; then, the cluster head of each learning cluster obtains the global module issued by the local base station according to the bandwidth allocated by the local base station, the cluster head of each learning cluster cuts the global model into a plurality of model segments according to a model cutting and allocation strategy, and distributes them to all cluster members in the learning cluster for collaborative training, so as to complete intra-cluster training; finally, the cluster head of each learning cluster collects the local models trained by all cluster members in the cluster and uploads them to the local base station, so that the local base station aggregates the local models uploaded by the cluster head of each learning cluster to complete global model updating; not only makes the user computation, storage and communication load of federated learning small, but also makes the user parallelism and sensitivity to user non-independent and identically distributed data of split learning high.

[0008] In addition, the model training method based on client-clustered hybrid federated segmentation learning proposed in the above embodiments of the present invention may also have the following additional technical features:

[0009] Optionally, the clustering strategy is a breadth-first search.

[0010] Optionally, the clustering strategy includes L learning clients participating in federated learning training using their own local data, and H auxiliary clients with idle computing and communication resources, whose set can be represented as follows: The set of all clusters can be represented as If a learning client has no available auxiliary clients nearby, then that learning client forms a separate cluster, assuming it is in the th... There are n learning clusters in total. l There are auxiliary clients, and the set of all cluster members can be represented as Where 0 represents the learning client, represented by a binary variable c (0-1). l,h Indicates the association state between the auxiliary client and the learning cluster, c l,h =1 indicates that the h-th auxiliary client is associated with the l-th learning cluster; otherwise, c l,h =0, therefore, the clustering strategy of the l-th learned cluster can be represented as a vector c. l ={c l,1 ,c l,2 ,…,c l,H Furthermore, any clustering strategy must satisfy the following conditions:

[0011] Optionally, a specific client is selected as the cluster head in each learning cluster according to the cluster head selection strategy, including: obtaining the direct D2D communication rate between cluster members in the learning cluster and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster; calculating the TCV value corresponding to each cluster member in the learning cluster based on the direct D2D communication rate between cluster members in the learning cluster and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster, and selecting the cluster member corresponding to the maximum value of the TCV value as the cluster head in that learning cluster.

[0012] Optionally, the TCV value for each cluster member in the learning cluster is calculated according to the following formula:

[0013]

[0014] Among them, TCV l，i Let n represent the TCV value of cluster member i in the learning cluster l, and N represent the total number of cluster members in the learning cluster l. l This represents the total number of auxiliary clients in learning cluster l. denotes the downlink unit communication rate between the local base station and the cluster head of learning cluster l; denotes the uplink unit communication rate between the cluster head of learning cluster l and the local base station.

[0015] Optionally, the local base station allocates bandwidth for the cluster head of each learning cluster according to the following formula:

[0016]

[0017] where the base station bandwidth allocation strategy can be denoted as vector b = {b1, b2,..., b L}, b l denotes the bandwidth allocated by the base station to learning cluster l, B denotes the total bandwidth available to the local base station, λ l denotes the unit delay cost of learning cluster l.

[0018] Optionally, the global module can split the model into model segments by splitting layers, and the set of all splitting layers can be denoted as A 0-1 binary variable s l,q is used to represent whether a splitting layer q ∈ Q is selected, s l,q = 1 indicates that the splitting layer q is selected in learning cluster l, otherwise s l,q = 0, and the model splitting and allocation strategy of learning cluster l can be denoted by vector s l = {s l,1 , s l,2 ,..., s l,Q} ; in each learning cluster, any cluster member has and only has one model segment, and can only train the model segment it owns and forward the output of the model segment it owns, where if a member in learning cluster l has no computing resources but has sufficient communication resources, it can be considered to be allocated a virtual model segment with a computing load of 0 and only participate in cooperative training as a communication relay, therefore, under a given clustering strategy c l , the model splitting and allocation strategy needs to satisfy the following conditions:

[0019] Optionally, according to the given clustering strategy c l , the computing capabilities of cluster members {f1,..., f nl+1}, the direct D2D communication rates between cluster members the computing load of the Q+1-layer model network and the corresponding communication load The computing load factor of each cluster member i ∈ c l in each learning cluster is calculated as The global delay cost minimization problem is modeled by building the delay of each item produced in the model training, wherein the model training includes global model download, intra-cluster training, local model and global model aggregation. In order to realize the die cutting and distribution strategy.

[0020] Optionally, further comprising: building the delay of each item produced in the model training, so as to complete the global delay cost minimization problem modeling, wherein the model training includes global model download, intra-cluster training, local model and global model aggregation; introducing optimal response dynamic theory, and gradually selecting the optimal clustering strategy for each learning cluster by calculating the utility value corresponding to each clustering strategy through strategy iteration. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of a model training method based on client clustering of hybrid federated split learning according to an embodiment of the application;

[0022] Figure 2 A model training framework diagram based on client clustering of hybrid federated split learning according to an embodiment of the application;

[0023] Figure 3 A learning cluster intra-cluster D2D communication topology diagram according to an embodiment of the application;

[0024] Figure 4 A flowchart of a heuristic-based joint TU selection and bandwidth allocation algorithm according to an embodiment of the application;

[0025] Figure 5 A flowchart of a distributed clustering algorithm based on optimal response dynamics according to an embodiment of the application. DETAILED DESCRIPTION

[0026] Embodiments of the application are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the application, and should not be understood as a limitation of the application.

[0027] In order to better understand the above technical solutions, exemplary embodiments of the application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the application are shown in the drawings, it should be understood that the application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the application and to fully convey the scope of the application to those skilled in the art.

[0028] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0029] Figure 1 This is a flowchart illustrating a model training method based on client-side clustering and hybrid federated segmentation learning according to an embodiment of the present invention. Figure 1 As shown, the model training method based on client-clustered hybrid federated segmentation learning includes the following steps:

[0030] S101, each learning client is clustered with its surrounding auxiliary clients according to the clustering strategy to obtain multiple learning clusters. Each learning cluster contains only one learning client, and each learning cluster may not contain auxiliary clients or may contain a different number of auxiliary clients.

[0031] In other words, such as Figure 2 As shown, each Helping Client (HC) forms a learning cluster with its surrounding Learning Clients (LC) through D2D communication. Through a clustering strategy, all LCs can form multiple learning clusters with all or some of the HCs.

[0032] As an example, a clustering strategy involves assuming that there are L LCs in the network that can participate in FL training using their local data, and that there are also H HCs with idle computing and communication resources, and their set can be represented as... Suppose that the set of all clusters can be represented as Assume that a learning cluster has exactly one LC, and each HC can be associated with at most one cluster; if an LC has no available HC nearby, then that LC forms a separate cluster; assume that in the... There are n learning clusters in total. l There are HCs, and the set of all cluster members can be represented as Where 0 represents the LC in the learning cluster l; further, a 0-1 binary variable c is used. l,h Indicates the association state between the HC and the learning cluster; c l,h =1 indicates that the h-th HC in H is associated with the l-th learning cluster; otherwise, c l,h =0; therefore, the clustering strategy of LC in cluster l can be represented as a vector c l ={c l,1 ,c l,2 ,...,c l,H Furthermore, any clustering strategy must satisfy the following conditions:

[0033] It should be noted that the distributed clustering problem is constructed into a clustering game problem by using game theory; for each cluster, the optimization goal is to find the optimal clustering strategy to minimize the delay cost of each cluster; define the LC strategy space of each cluster as C l , which is the set of all feasible clustering strategies c l ; in the traditional clustering strategy, a cluster is an unordered set of nodes, and unlike the traditional unordered clustering algorithm, any feasible clustering strategy is a Hamiltonian path, that is, an ordered set of nodes.

[0034] As a specific embodiment, a clustering strategy generation algorithm based on Breadth-First Search (BFS) is adopted, which intends to use the basic process of BFS search to generate the clustering strategy space C l of each LC, that is, to find all feasible clustering strategies with different lengths; in order to better illustrate the algorithm process, take the learning cluster D2D topology graph in Figure 3 as an example; there is a topology graph containing LC and 4 HC; first, node l is taken as the first element in the queue and the head node, and the first path containing only node l can be obtained, whose length is 0; then the three adjacent nodes are: HC-1, HC-3 and HC-4; based on this, three paths with length 1 can be obtained: LC-HC-1, LC-HC-3, LC-HC4; in this way, four Hamiltonian paths with length 2 can be obtained: LC-HC-4-HC-3, LC-HC-1-HC-2, LC-HC-3-HC-2, LC-HC-3-HC-4; finally, all feasible clustering strategies can be found by using the BFS algorithm.

[0035] S102, select a special client as the cluster head in each learning cluster according to the cluster head selection strategy, so that the local base station allocates bandwidth to the cluster head of each learning cluster according to the base station bandwidth allocation strategy.

[0036] That is, a special client in each learning cluster is selected as a transmission client (TC) to become a cluster head by using the cluster head selection strategy, and the cluster head completes local model uploading and global model downloading according to the bandwidth allocated by the base station, realizing inter-cluster parallel FL training.

[0037] As an example, a specific client is selected as the cluster head in each learning cluster according to the cluster head selection strategy, including: obtaining the direct D2D communication rate between cluster members in the learning cluster and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster; calculating the TCV value corresponding to each cluster member in the learning cluster based on the direct D2D communication rate between cluster members in the learning cluster and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster, and selecting the cluster member corresponding to the maximum value of the TCV value as the cluster head in the learning cluster.

[0038] It should be noted that in any learning cluster l, a 0-1 binary variable t is used. l,i Indicates cluster member Whether to be selected as the cluster head; when t l,i =1 indicates that i is selected as the cluster head; otherwise, t l,i =0; therefore, the cluster head selection strategy can be represented by a vector as:

[0039] t l ={t l,0 ,t l,1 ,...,t l,nl}

[0040] Since a learned cluster has one and only one cluster head, the cluster head selection strategy must satisfy the following conditions:

[0041]

[0042] Once a cluster head is selected for each learning cluster, the local base station needs to orthogonally allocate wireless communication bandwidth to the cluster heads of each learning cluster for the learning clusters to complete local model uploading and global model downloading. Assuming the total available bandwidth of the base station is B, its bandwidth allocation strategy can be represented as a vector b = {b1, b2, ..., b...} L}, where b l ∈b represents the bandwidth allocated by the local base station to learning cluster l; the bandwidth allocation strategy of the local base station must meet the following constraints:

[0043]

[0044] If member i in the learning cluster l is selected as TC, then from i to the remaining n l Minimum distribution latency for each client This can be achieved by calling Dijkstra's algorithm multiple times; since all learned clusters can be independently and in parallel determined by searching using Dijkstra's algorithm, the time complexity of this process is O(log n). Where e is the number of edges in the corresponding graph; however, due to the different unit communication rates and variable allocated bandwidth between different TCs and local base stations, the global model download latency... and local model uploading latency The optimal bandwidth allocation strategy will depend on the selected TUs and the allocated bandwidth; first, the optimal bandwidth allocation optimization strategy under the given cluster and MSA scheme and TU selection scheme The optimal bandwidth allocation strategy can be obtained by optimizing the following problem

[0045]

[0046] The optimal bandwidth allocated to the learning cluster l is:

[0047]

[0048] The above formula can guarantee the optimal bandwidth allocation strategy, but to obtain the optimal joint TC selection and bandwidth allocation strategy, it is necessary to traverse When L and n l are large, it will bring extremely high time complexity; therefore, a more efficient algorithm is needed to determine the TC of all clusters; based on this, first, an evaluation parameter called TCV (TC Value, TCV) is proposed to measure the gain as a TC.

[0049] Specifically, the calculation method of TCV of member i in learning cluster l is as follows:

[0050]

[0051] Where, R i,j represents the direct D2D communication rate between cluster members i and j; represents the downlink unit communication rate from the base station to the TC of the learning cluster l; represents the uplink unit communication rate between the TC of the learning cluster l and the base station.

[0052] As can be seen from the above formula, the basic idea of TCV is to measure the degree of the cluster member in the corresponding topology graph and the unit communication rate with the base station to indicate the gain as a TC; if a cluster member in the learning cluster has more direct links with other nodes and has a higher uplink and downlink unit communication rate with the base station, the cluster member has greater gain as a TC; by calculating the TCV of all cluster members in each learning cluster, then selecting the cluster member with the maximum TCV as the TC, the set of all cluster members TC can be obtained; then by calculating the optimal bandwidth allocation strategy at this time, the joint TC selection and bandwidth allocation scheme can be obtained, and the value of is calculated, and the above-mentioned heuristic-based joint TU selection and bandwidth allocation algorithm process is as shown in Figure 4 .

[0053] S103, the cluster head of each learning cluster obtains the global module issued by the local base station according to the bandwidth allocated by the local base station. The cluster head of each learning cluster divides the global model into multiple model segments according to the model segmentation and allocation strategy, and distributes them to all cluster members in the learning cluster for collaborative training in order to complete the intra-cluster training.

[0054] In other words, the cluster head of each learning cluster divides the model into multiple model segments according to the Model Splitting and Allocation (MSA) strategy, and distributes them to all cluster members for collaborative training, thus completing intra-cluster serial SL training.

[0055] It should be noted that, assuming the global model has a maximum of [number] possible [transmissions / accesses] The model is divided into individual layers. Given a model segment, the set of all sliced ​​layers can be represented as: Use a binary variable s (0-1) l,q Indicates whether the cutting layer q∈Q is selected, s l,q =1 indicates that the layer q is selected in the learning cluster l; otherwise, s l,q =0; therefore, the MSA strategy of learning cluster l can be obtained using vector s. l ={s l,1 ,s l,2 ,…,s l,Q} indicates that in each learning cluster, any cluster member has one and only one model segment, and can only train its own model segment and forward the output of its own model segment; where, if a cluster member in the learning cluster has no computing resources but has sufficient communication resources, it can be considered to be allocated a virtual model segment with zero computing load, and participate in collaborative training only as a communication relay; therefore, given the clustering policy c l A feasible MSA strategy must meet the following conditions:

[0056] Suppose that the learning cluster l randomly selects a clustering strategy c. l In this case, the optimal MSA strategy is to select n in Q. l The layering divides the model into n l Add 1 more model segment to optimize the following issues:

[0057]

[0058] The optimal solution of the MSA strategy can be obtained by, for example, an exhaustive algorithm, but when the scale of the model is large, the time complexity of the exhaustive algorithm is too high, resulting in low solving efficiency; in order to solve the problem efficiently, the embodiment proposes a heuristic algorithm based on a greedy strategy, the basic idea of the algorithm is to allocate the computing load to the cluster members according to the proportional relationship between the computing resources of the cluster members according to the proportional fairness, so as to select a suitable model split layer, the main process is as follows: define the clustering strategy c l , the computing capacity of the cluster members , the direct connection D2D communication rate between the cluster members , the computing load of the Q+1 layer model network , and the corresponding communication load , and initialize the currently selected split layer q=0, calculate the computing load factor of each cluster member i∈c l in the learning cluster , traverse each member in the learning cluster, calculate from the current split layer q to find the next new split layer q', so as to meet , so as to realize the model splitting and distribution strategy.

[0059] In S104, the cluster head of each learning cluster collects the trained local model of all cluster members in the cluster and uploads it to the local base station, so that the local base station aggregates the local model uploaded by the cluster head of each learning cluster to complete the global model update.

[0060] It should be noted that, as shown in Figure 2 , the learning cluster 1 includes one learning client and four auxiliary clients, the learning cluster 3 includes one learning client and three auxiliary clients, the cluster heads of the learning cluster 1 and the learning cluster 3 receive the global model issued by the local base station (federal learning server), the cluster head splits the global model, and distributes the corresponding model segment to the corresponding cluster member; the cluster member performs the forward propagation process, and sends the forward process calculation result to the next cluster member; after the forward process is completed, the last cluster member performs the backward propagation to update the model parameters, the cluster head collects the trained model of all cluster members and uploads the trained local model to the local base station; the local base station aggregates the local model uploaded by the cluster head to complete the global model update; and for the learning cluster 2, since it is a separate cluster, the learning client needs to complete the work of all clients.

[0061] As an example, the model training method based on client-side clustering hybrid federated segmentation learning also includes constructing various latency factors generated during model training in order to complete the modeling of the global latency cost minimization problem. The model training includes global model download, intra-cluster training, local model and global model aggregation. The optimal response dynamics theory is introduced, and by calculating the utility value corresponding to each clustering strategy, the strategy is iterated to gradually allow each learning cluster to select the optimal clustering strategy.

[0062] The global download model latency construction method is as follows:

[0063] At the beginning of each training round, given the clustering strategy, MSA strategy, TU selection strategy, and bandwidth allocation strategy, the TC of each learning cluster will first download the latest global model from the local base station; assuming the downlink communication rate from the local base station to the TCi of learning cluster l can be expressed as...

[0064]

[0065] in The unit communication rate is represented by P, where P is the base station's transmit power, N0 is the noise power, and h is the noise power. l,i This is the channel gain between TUi and the base station.

[0066] Therefore, the corresponding global model download latency can be expressed as:

[0067]

[0068] Here, Z represents the size of the data in a complete model.

[0069] The intra-cluster training delay is constructed as follows:

[0070] Within each learning cluster, the training process can be divided into four sub-stages: intra-cluster model distribution, FP training, BP training, and model collection. Once the TC obtains the global model, it first segments the model according to the MSA strategy and distributes the corresponding model segments to the relevant cluster members. Therefore, any cluster member... The size of the data in the model segment held is related to the MSA strategy. l The function can be represented as z i (s l Furthermore, when a model segment is sent from cluster member i to cluster member j via D2D communication, a binary indicator function `neighbor(i,j)` is used to indicate whether there is a direct link between them. `neighbor(i,j) = 1` indicates that they can be directly connected, otherwise `neighbor(i,j) = 0`. For example, in... Figure 3In the given communication topology, for LC and HC-3, neighbor = 1, while for LU and HU-2, neighbor = 0; based on the foregoing, the communication topology corresponding to a given feasible MSA strategy... Figure 1 It must be a connected graph; therefore, for any two cluster members that do not have a direct link, they can definitely be connected via multiple hops; therefore, when cluster member i forwards a model segment from cluster member j, the transmission delay between them is...

[0071]

[0072] Among them, R i,j This represents the D2D communication rate between cluster members i and j, while MinDelay(i,j) is a function that can calculate the minimum delay between non-directly connected cluster members i and j, for example, using Dijkstra's algorithm.

[0073] Therefore, given the clustering strategy, MSA strategy, and TU selection strategy, the total delay of model distribution within a cluster can be expressed as:

[0074]

[0075] After distributing each model segment to its respective cluster member, the FP process will begin. Before sending the forward training results to the next member, FP computation is required, and the computational load of FP primarily depends on the structure of the model segment. Assuming the FP computational load for processing a data sample on cluster member i can be expressed as a function of the MSA policy s... l The function, denoted as The computational load is measured in floating-point operations (FLOPs); therefore, the FP computation latency on cluster member i is:

[0076]

[0077] Where ζ represents the number of data samples in a mini-batch, f i This represents the computational power of cluster member i, measured in CPU revolutions per second (CPU revolutions per second). υ represents the number of FLOPs that can be processed in one CPU revolution. After FP computation is complete, the processed data samples (such as extracted image features) will be sent to the next cluster member j to continue the FP process. The size of the intermediate result data transmitted also depends on the MSA strategy s. l ,remember For the amount of data transmitted for a single data sample, the corresponding transmission delay is:

[0078]

[0079] Therefore, the total FP delay in a learning cluster l with a given MSA strategy and TC selection strategy is:

[0080]

[0081] After FP, the last cluster member will start BP, i.e., gradient descent calculation by computing loss and updating model parameters. The main difference between FP and BP is that the computation and transmission are in opposite order; therefore, the computation load on cluster member j in BP is and the corresponding transmission gradient data size is The corresponding BP computation and transmission delay is:

[0082]

[0083] The total BP delay in a training round is

[0084]

[0085] Assume that the local data set size of LC in learning cluster l is D l The total intra-cluster training delay when performing local training for ξ times is

[0086]

[0087] After FP and BP training are completed, the TC will collect all the trained model segments of the cluster members; therefore, the model collection and model distribution will have the same delay except for the opposite order. Therefore, the total intra-cluster training delay is

[0088]

[0089] The local model and global model aggregation delay construction method is as follows:

[0090] After intra-cluster training is completed, the TC of each learning cluster will upload the trained model to the local base station, i.e., the FL server; therefore, the upload rate between the TC of learning cluster l and the local base station is

[0091]

[0092] where P l,i represents the transmission power of the TC; therefore, the delay required for local model upload is:

[0093]

[0094] When the base station receives all the uploaded models from the clusters, it will aggregate all the local models into a global model through an aggregation algorithm; since the computation required for model aggregation is very small, the time required for model aggregation can be ignored.

[0095] Finally, given the clustering strategy, MSA strategy, TU selection strategy, and bandwidth allocation strategy, the total training delay required for learning cluster l to complete a training round can be expressed as

[0096]

[0097] After obtaining the training delay model of the learning cluster, assuming that the unit delay cost of learning cluster l is λ l , it can be used to characterize the comprehensive training cost brought by factors such as energy consumption, computing resources, and communication resources; on the basis of meeting the clustering strategy MSA strategy TC selection strategy, and the base station's bandwidth allocation strategy , the global training delay cost of all clusters is minimized; the delay cost of UAV cluster l is defined as Therefore, the global delay cost minimization problem can be modeled as follows:

[0098]

[0099] In addition, define any one clustering strategy c l ∈C l that each cluster can choose to obtain, therefore, the joint clustering strategy of all L clusters can be expressed as c = {c l ,c -l}, where c -l represents the joint clustering strategy of all clusters except learning cluster l; the utility function of learning cluster l is defined as:

[0100]

[0101] where, represents the utility of cluster j when learning cluster l does not participate in the game. Therefore, represents the impact of the strategy change of learning cluster l on the utility of other clusters.

[0102] As shown in the specific implementation Figure 5 , using the optimal response dynamics, given the clustering strategies c -l of the remaining clusters unchanged, the strategy c l ′ obtained by cluster l through the optimal response dynamics should satisfy the following:

[0103]

[0104] That is, the specific steps include: S1, inputting related parameters of global network and FL training; S2, calling a clustering strategy space of all clusters, each cluster randomly selecting a clustering strategy c from the respective clustering strategy space l ; S3, calling an utility value under the current clustering strategy, fixing other cluster strategies, and randomly selecting a learning cluster to dynamically select a new strategy according to the optimal response; S4, judging whether the utility value under the current clustering strategy is equal to the utility value under the random strategy, if yes, updating the clustering strategy after the current round as the optimal clustering strategy of the local model, and if no, updating the clustering strategy and setting , and returning to execute S3.

[0105] In summary, according to the model training method of the mixed federated split learning based on client clustering, first, each learning client clusters with the surrounding auxiliary clients according to the clustering strategy to obtain a plurality of learning clusters, wherein each learning cluster has and only has one learning client, and each learning cluster after clustering can not include auxiliary clients or include auxiliary clients in different quantities; then, a special client is selected as a cluster head in each learning cluster according to the cluster head selection strategy, so that the local base station allocates bandwidth to the cluster head of each learning cluster according to the base station bandwidth allocation strategy; then, the cluster head of each learning cluster obtains the global module issued by the local base station according to the bandwidth allocated by the local base station, the cluster head of each learning cluster splits the global model into a plurality of model segments according to the model splitting and allocation strategy, and distributes them to all cluster members in the learning cluster for collaborative training, so as to complete the intra-cluster training; finally, the cluster head of each learning cluster collects the local model trained by all cluster members in the cluster and uploads it to the local base station, so that the local base station aggregates the local model uploaded by the cluster head of each learning cluster to complete global model updating; not only the user computing, storage and communication load of federated learning is small, but also the user parallelism and sensitivity to non-independent and identically distributed data of split learning are high.

[0106] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] The present application is described in reference to the accompanying drawings, which use the legend: 100 - method, 200 - device (system), 300 - computer program product. Figure 1 one or more processes and / or blocks Figure 1 means for performing the functions specified in the flow or flows and / or blocks.

[0108] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more processes and / or blocks Figure 1 means for performing the functions specified in the flow or flows and / or blocks.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 means for performing the functions specified in the flow or flows and / or blocks.

[0110] It should be noted that the use of any of the terms "first", "second" or the like does not connote any order, quantity, or importance, but rather are used to distinguish one element from another. It should also be noted that any reference to claim, apparatus, system, method, device or the like can refer to a composite of the various elements in claim, apparatus, system, method, device or the like. Further, terms such as attache, attach, connecting, connecting, connection and so on are intended to encompass an attachment which is fixed, detachable or otherwise connectable / attachable.

[0111] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such variations and modifications as fall within the scope of the application.

[0112] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the application. Accordingly, it is intended that all such modifications and changes be included in the present application only as far as they come to be within the scope of the claims and their equivalents.

[0113] In the description of the present application, it is to be understood that the terms "first", "second", "third" and the like, merely mean different instances of the same thing, and do not imply relative importance or a number of the indicated technical features. Thus, a feature defined with "first", "second", "third" can include one or more of the features. In the description of the present application, the meaning of "plurality" is two or more, unless specifically defined otherwise.

[0114] In the present application, unless specifically defined and limited otherwise, the terms "mounting", "connection", "connecting", "fixed", and the like, should be understood broadly, for example, can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through intermediate medium; can be internal communication of two elements or interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0115] In the present application, unless specifically defined and limited otherwise, the first feature "on" or "under" the second feature can be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature "above", "over" and "on" the second feature can be that the first feature is directly above or obliquely above the second feature, or only means that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "under" and "under" the second feature can be that the first feature is directly below or obliquely below the second feature, or only means that the horizontal height of the first feature is less than that of the second feature.

[0116] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms should not be understood as necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0117] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that variations, modifications, substitutions and changes can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. A model training method based on client clustering hybrid federated split learning, characterized in that, The method comprises the following steps: Each learning client clusters with the surrounding auxiliary clients according to a clustering strategy to obtain a plurality of learning clusters, wherein each learning cluster has and only has one learning client, and each learning cluster after clustering can not include auxiliary clients or include auxiliary clients in different quantities; According to a cluster head selection strategy, a special client in each learning cluster is selected as a cluster head, so that the local base station allocates bandwidth to the cluster head of each learning cluster according to a base station bandwidth allocation strategy; The cluster head of each learning cluster obtains the global model issued by the local base station according to the bandwidth allocated by the local base station, the cluster head of each learning cluster cuts the global model into a plurality of model segments according to a model cutting and allocation strategy, and distributes them to all cluster members in the learning cluster for cooperative training, so as to complete intra-cluster training; The cluster head of each learning cluster collects the local models trained by all cluster members in the cluster and uploads them to the local base station, so that the local base station aggregates the local models uploaded by the cluster head of each learning cluster to complete global model updating; wherein the global module can split the model into model segments by a set of split layers, and the set of all split layers can be denoted as a 0-1 binary variable s l,q to indicate whether a split layer q ∈ Q is selected, s l,q = 1 indicates that the split layer q is selected in the learning cluster l, otherwise s l,q = 0, the model splitting and assignment strategy of the learning cluster l can be denoted by a vector s l = {s l,1 , s l,2 ,..., s l,Q}; in each learning cluster, any cluster member has and only has one model segment, and can only train the model segment owned by itself and forward the output of the model segment, wherein if a member in the learning cluster l has no computing resource but has sufficient communication resource, it can be considered to be assigned a virtual model segment with a computing load of 0 and only participate in cooperative training as a communication relay, thus, under a given clustering strategy c l , the model splitting and assignment strategy needs to satisfy the following conditions: wherein the clustering strategy c is given l computational power of the cluster members direct D2D communication rate between cluster members computational load of the Q+1 layer model network and the corresponding communication load computing the computational load factor of each cluster member i∈c l of each learning cluster traversing each cluster member in the entire learning cluster, computing from the current split layer q to find the next new split layer q' to satisfy in order to implement the modulo split and assignment strategy.

2. The method of claim 1, wherein the method is performed by a client device. The clustering strategy is breadth-first search.

3. The method of claim 2, wherein the method comprises: The clustering strategy includes L learning clients participating in federated learning training with their local data, and H helper clients with idle computing and communication resources, whose set can be denoted as The set of all clusters can be denoted as If there is no available helper client near a learning client, the learning client forms a separate cluster, assuming that there are n l helper clients in the l-th learning cluster, and the set of all cluster members can be denoted as where 0 represents a learning client, and a 0-1 binary variable c l,h represents the association state between a helper client and a learning cluster, c l,h = 1 indicates that the h-th helper client is associated with the l-th learning cluster, otherwise c l,h = 0, therefore, the clustering strategy of the l-th learning cluster can be represented as a vector c l = {c l,1 ,c l,2 ,...,c l,H}, and any clustering strategy needs to satisfy the following conditions: ​ 4. The model training method based on client clustering hybrid federated split learning according to claim 3, wherein, According to the cluster head selection strategy, a special client in each learning cluster is selected as a cluster head, comprising: Obtaining the direct D2D communication rate between the cluster members in the learning cluster, and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster; According to the direct D2D communication rate between the cluster members in the learning cluster and the uplink and downlink communication rate between the local base station and each cluster member in the learning cluster, the TCV value corresponding to each cluster member in the learning cluster is calculated, and the cluster member corresponding to the maximum value in the TCV value is selected as the cluster head in the learning cluster.

5. The method of claim 4, wherein the method is based on a client clustering-based hybrid federated split learning. The TCV value corresponding to each cluster member in the learning cluster is calculated according to the following formula: where TCV l,i denotes the TCV value corresponding to cluster member i in cluster l, denotes the total number of cluster members in cluster l, R i,j denotes the direct D2D communication rate between cluster members i and j, and a binary indicator function neighbor(i,j) is used to indicate whether there is a direct link between them when a model segment is transmitted from cluster member i to cluster member j through D2D communication, n l denotes the total number of helper clients in cluster l, denotes the downlink unit communication rate between the local base station and the cluster head of learning cluster l; denotes the uplink unit communication rate between the cluster head of learning cluster l and the local base station.

6. The method of claim 5, wherein the method is based on a client clustering-based hybrid federated split learning. The local base station allocates bandwidth to the cluster head of each learning cluster according to the following formula: where the base station bandwidth allocation strategy can be represented as a vector b = {b1, b2, …, bl}, bl L}, bl l ∈ b represents the bandwidth allocated by the base station to the learning cluster l, B represents the total bandwidth available to the local base station, and λ l represents the unit latency cost of the learning cluster l.

7. The method of claim 6, wherein the method is based on a client clustering-based hybrid federated split learning. Further comprising: Constructing each item of delay generated in model training to complete global delay cost minimization problem modeling, wherein the model training includes global model download, intra-cluster training, local model and global model aggregation; Introducing optimal response dynamic theory, calculating the utility value corresponding to each clustering strategy, and gradually selecting the optimal clustering strategy for each learning cluster through strategy iteration.

Images