A federated learning approach for edge heterogeneous environments

Abstract

The present invention proposes a federated learning method for edge heterogeneous environments, which belongs to the field of federated learning. The method includes: S1: the server detects the registration information of the edge client and performs performance detection on the edge client, removes the edge client that does not meet the system standard, selects the edge client that meets the requirements, and obtains the edge client set; S2: the edge client performs local model training, the server receives the training model, uses the wavelet clustering algorithm to partition the edge client, and obtains the edge client partition result that meets the data distribution; S3: the server randomly selects some edge clients from each edge client partition to participate in the training, the edge client adaptively performs local training, the server calculates the weight according to the local training situation of the edge client, performs model aggregation, and saves the trained model. The present invention can improve the accuracy and efficiency of model training, ensure the model training effect, and promote the development of edge intelligence technology.

Description

Technical Field

[0001] The present invention belongs to the field of federated learning, and in particular relates to a federated learning method for edge heterogeneous environments. Background Art

[0002] On the one hand, with the development of edge device performance and communication technology, the amount of data on edge networks has exploded, and edge computing as a new computing paradigm has attracted more and more attention; on the other hand, thanks to the progress of algorithms and hardware equipment, artificial intelligence technology has ushered in a new wave of development, and many researchers are seeking how to better develop and apply artificial intelligence technology. As a result, the organic combination of the two gave birth to edge intelligence technology.

[0003] There are two research directions for edge intelligence technology: one is how to use artificial intelligence technology to solve problems in edge computing; the other is how to train artificial intelligence models in an edge computing environment. For the latter, federated learning is a computing paradigm that is often mentioned. However, in actual edge environments, federated learning algorithms often face problems such as data heterogeneity and device heterogeneity, which will affect the accuracy and efficiency of model training. In addition, edge environments also have insufficient resources such as limited device performance.

[0004] To address the problem of data heterogeneity, current federated learning algorithms often use operations such as loss function optimization and knowledge distillation, which will increase the computing and communication burden on the client. To address the problem of device heterogeneity, current federated learning often uses asynchronous updates, or controls the amount of computation at the client level, and there are relatively few solutions discussed in combination with edge scenarios. Summary of the invention

[0005] The purpose of this invention is to propose a federated learning method for edge heterogeneous environments, so as to realize the reliable execution of federated learning tasks in edge heterogeneous environments under limited resources, so as to make up for the technical problems in the prior art that the data heterogeneity, device heterogeneity and limited resources in edge scenarios are not comprehensively considered, and the model training effect and efficiency of federated learning in edge heterogeneous environments are poor.

[0006] In order to solve the above technical problems, the present invention proposes a federated learning method for edge heterogeneous environments, comprising the following steps:

[0007] Step S1: The server detects the registration information of the edge client and performs performance testing on the edge client, removes the edge clients that do not meet the system standards, selects the edge clients that meet the requirements, and obtains an edge client set;

[0008] Step S2: The edge client performs local model training, the server receives the training model, and uses the wavelet clustering algorithm to partition the edge client to obtain an edge client partition result that conforms to the data distribution;

[0009] Step S3: The server randomly selects some edge clients from each edge client partition to participate in the training. The edge clients adaptively perform local training. The server calculates the weights according to the local training status of the edge clients, performs model aggregation, and finally saves the trained model.

[0010] Furthermore, step S1 includes the following steps:

[0011] Step S11: The server sets the highest CPU usage standard, the lowest GPU performance standard, and the highest acceptable network delay and the most unstable network fluctuation;

[0012] Step S12: The server performs performance testing on the edge client, including CPU testing, GPU testing, and network status testing, and records the test results;

[0013] Step S13: The server compares the performance test result with the set index, removes edge clients that do not meet the minimum system standard, selects edge clients that meet the requirements, and obtains an edge client set;

[0014] Step S14: Input the grid step size and density threshold parameters required for wavelet clustering into the server.

[0015] Further, step S21: edge client preheating; server uniformly generates and initializes the global model , the server will global model Sent to every edge client in the edge client set; except for the global model In addition, the server also sends the local iteration count To the edge client. The edge client uses the global model and the number of local iterations Perform local training and send the trained local model back to the server.

[0016] Step S22: extracting model parameters; the server obtains the local models returned by each edge client, and extracts parameter data of a specified model layer in each model;

[0017] Step S23: construct feature space; the server constructs and divides the mesh space according to the preset grid step size and the parameter data of the specified model layer; then puts each local model into each cell in the space, calculates the density of the cell, and constructs the feature space ;

[0018] Step S24: wavelet transform; the server sequentially transforms the feature space Perform wavelet transform on each dimension to achieve data compression and obtain feature space ;

[0019] Step S25: Mark the cell; if the density of the cell is greater than the preset density threshold, it is a dense cell; search the feature space The dense cells in the cells are marked as a cluster. If there are adjacent and dense cells, the adjacent and dense cells are marked as the same cluster;

[0020] Step S26: construct a mapping relationship table; according to the wavelet clustering algorithm, construct a cell mapping table before and after the wavelet change;

[0021] Step S27: Look up the table; look up the mapping relationship table, obtain each edge client cluster, partition the edge client, and obtain the edge client partition result that meets the data distribution situation.

[0022] Further, step S31: server initialization; the server randomly generates a new global model , global communication round Initialized to 0, setting the global communication round threshold.

[0023] Step S32: The server randomly selects edge clients to participate in the training from each edge client partition at a ratio of 40%-60%, and the server sends the global model and related parameters to the selected edge client; the related parameters include the specified theoretical local iteration rounds and aggregation time window;

[0024] Step S33: The edge client receives the global model , initialize the local iteration round e to 0;

[0025] Step S34: the edge client performs a round of local training, records the training time, and the local iteration round e is increased by 1;

[0026] Step S35: The edge client determines whether the local iteration round is less than the specified theoretical local iteration round. If so, execute step S34; otherwise, execute step S36;

[0027] Step S36: The edge client tests the network status with the server and estimates the transmission time, which is equal to the network delay * the size of the local training model. The end time of the next round of training is further estimated, which is equal to the transmission time + the training time + the current time. If the end time of the next round of training does not exceed the aggregation time window, then execute step S34; otherwise, execute step S37;

[0028] Step S37: The edge client sends the local training model and local iteration rounds to the server;

[0029] Step S38: The server waits within the time window. If an edge client does not respond after exceeding the preset time in the time window, it is determined that this round of global communication has failed and step S33 is executed. Otherwise, step S39 is executed.

[0030] Step S39: The server calculates the aggregation weight according to the local iteration number through the weight calculation formula. The server performs weighted aggregation on the local model returned by the edge client to obtain a new global model. , global communication round Self-increment;

[0031] Step S310: Training end judgment; server judgment Is it greater than the global communication round threshold? If less than, execute step S32; if greater, end the training and save the training results.

[0032] Furthermore, the weight calculation formula in step S39 is as follows:

[0033]

[0034] in, is the learning rate, For edge clients The local iteration round of is the number of local iterations that each edge client should theoretically perform, The number of clients participating in this round of training.

[0035] Compared with the prior art, the present invention has the following beneficial technical effects:

[0036] 1) Combining the wavelet clustering algorithm with the participant selection strategy solves the problem of poor model training effect caused by data heterogeneity and improves the accuracy of the training model in a low-cost way.

[0037] 2) Combining the adaptive training strategy with the weight design scheme can solve the problems of slow clients and stragglers caused by device heterogeneity while ensuring the training effect of the model.

[0038] 3) The present invention further integrates the above two strategies into a federated learning algorithm, taking into account the limited performance of edge environment clients. It solves the data and system heterogeneity problems that may be faced by federated learning in edge environments without adding additional computing burden to the clients, and has certain practical value.

[0039] 4) This invention can improve the accuracy and efficiency of model training, ensure the model training effect, and further promote the development of edge intelligence technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0041] Figure 1 It is the overall flow chart of the present invention.

[0042] Figure 2 This is a flow chart of the edge client partitioning function of the present invention.

[0043] Figure 3 This is a functional flow chart of the adaptive model training of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Federated learning in edge scenarios not only has problems of data heterogeneity and device heterogeneity, but also has restrictive conditions of limited client computing resources. The present invention proposes a federated learning method for edge heterogeneous environments, which consists of two parts: participant selection and dynamic weight calculation. It solves the problems of data heterogeneity and device heterogeneity in edge scenarios at the cost of lower computing burden on the client.

[0046] In terms of client selection, the wavelet clustering algorithm is used to cluster the clients based on the similarity between local models, so that the local models of clients in the same cluster can have high similarity; then clients are randomly selected from these partitions to participate in training, so as to avoid the impact of data heterogeneity on model training.

[0047] In terms of weight calculation, the server first sets an aggregation time window and the theoretical optimal number of local iterations, and sends them to the selected client together with the global model. After receiving the model and parameters, the client performs local model training with adaptive number of iterations based on its own computing device performance and network device status, and sends the results and local number of iterations back to the server. The server performs weighted aggregation based on the local number of iterations and weights to obtain the global model to be used in the next round of global communication.

[0048] Example:

[0049] The following is an embodiment for illustrating the scheme of the present invention in detail.

[0050] The present invention is based on the federated learning model training scenario of the handwritten character dataset. The overall system consists of 12 edge clients with local model training capabilities and an aggregation server.

[0051] In terms of edge clients, in addition to local training and communication capabilities, edge clients also collect and create handwritten character datasets on their own: 4 of them are mainly for the collection of Chinese characters; another 4 are for the collection of English characters; and the rest are for the collection of Russian characters.

[0052] In terms of the aggregation server, for the sake of fairness, the aggregation server is provided by a third-party platform. The server is responsible for edge client monitoring during training, publishing the global model, weighted aggregation of local models, and saving the final training results. In addition, the edge client needs to register with the server before participating in the training. The information and explanations that need to be submitted during registration are as follows:

[0053]

[0054] The overall flow chart of the present invention is as follows: Figure 1 As shown in the figure, it mainly includes three steps: edge client performance detection, edge client partitioning, and adaptive model training, and each step contains multiple small steps. Among them, the edge client performance detection link is that the server verifies the registration information of the edge client before the training task starts, and detects the edge client availability, occupancy rate and network status, and eliminates edge clients that do not meet the minimum system standards. Then comes the edge client partitioning link, which occurs after the edge client is preheated. The server uses the wavelet clustering algorithm to partition the edge client according to the data distribution. The adaptive model training link is that the edge client adaptively performs local model training to solve the differences in training and communication capabilities caused by device heterogeneity; then the server performs weighted aggregation to eliminate the impact of the heterogeneity of local iteration times.

[0055] The present invention proposes a federated learning method for edge heterogeneous environments, comprising the following steps:

[0056] Step S1: The server detects the registration information of the edge client and performs performance testing on the edge client, removes the edge clients that do not meet the system standards, selects the edge clients that meet the requirements, and obtains an edge client set.

[0057] Specifically, step S1 includes the following steps:

[0058] Step S11: The server sets the highest CPU usage standard, the lowest GPU performance standard, and the highest acceptable network latency and the most unstable network fluctuation.

[0059] The minimum system standards of the server can be set as needed. For example, the maximum CPU usage is set to 50%, the minimum GPU model is 3070, the maximum network delay is 100ms, and the minimum network fluctuation is 30% of the maximum average delay of the network delay standard deviation.

[0060] Step S12: The server performs performance testing on the edge client, including CPU testing, GPU testing, and network status testing, and records the test results.

[0061] Specifically, taking the Windows system as an example, during CPU detection, the basic information and occupancy rate of the client's CPU are obtained with the help of PowerShell scripts.

[0062]

[0063] When detecting the GPU, use the PowerShell script to obtain basic information about the client's GPU.

[0064]

[0065] The GPU occupancy rate varies depending on the graphics card type. Taking NVIDIA graphics card as an example, the following script can be used to obtain the overall occupancy rate.

[0066]

[0067] Network status detection is mainly to detect the network delay in real time. Network status detection can use a simple ping command or a more professional third-party tool to detect network status.

[0068] Step S13: The server compares the performance test results with the set indicators, removes edge clients that do not meet the minimum system standards, selects edge clients that meet the requirements, and obtains an edge client set.

[0069] Based on the device information obtained from CPU detection, GPU detection, and network status detection, determine whether the edge client meets the minimum performance indicators for participating in subsequent training tasks, eliminate edge clients that do not actually meet the system standards, and select edge clients that meet the requirements.

[0070] The judgment criteria are mainly based on three aspects: the edge client processor model, the average processor occupancy rate over a period of time, and the network delay status. In terms of processor model, assuming that the task requires the use of a GPU and the minimum model is 3070, processors with performance lower than this standard cannot participate in subsequent tasks. In terms of the average processor occupancy rate over a period of time, if the current occupancy rate of the edge client processor is greater than 50%, it is judged that it has other training tasks, and the edge client needs to terminate the task and release processor resources before it can participate in subsequent training tasks. In terms of network delay, if the network delay is greater than 100ms or the volatility is large, that is, the standard deviation of the network delay exceeds 30% of the average delay, it is also judged that it is not suitable for participating in subsequent training tasks.

[0071] Step S14: Input the grid step size and density threshold parameters required for wavelet clustering into the server.

[0072]

[0073] Step S2: The edge client performs local model training, the server receives the training model, and uses the wavelet clustering algorithm to partition the edge client to obtain an edge client partition result that conforms to the data distribution.

[0074] The edge client partitioning process of the present invention is as follows Figure 2 As shown. First, the edge client is preheated, then the parameter data of the specified number of layers of each local model is extracted, and then the wavelet clustering algorithm is applied according to the preset grid step size and density threshold to complete the client partition. Specifically, it includes the following steps:

[0075] Step S21: Preheat the edge client. The server uniformly generates and initializes the global model , the server will global model Sent to each edge client in the edge client set. Except for the global model In addition, the server also sends the local iteration count To the edge client. The edge client uses the global model and the number of local iterations Perform local training and send the trained local model back to the server.

[0076] Local Iterations It can be set reasonably according to needs. If it is a small model training, the number of local iterations Can be set to a smaller value, such as 2, 3. Unify the number of local iterations in the first round , which helps to control the local model training level of each client to be similar.

[0077] Step S22: Extract model parameters. The server obtains the local models returned by each edge client and extracts parameter data of the specified model layer in each model as input for subsequent wavelet clustering.

[0078] The selected specified model layer can be reasonably set according to needs. In order to reduce the calculation burden of subsequent servers, a model layer with a smaller parameter dimension can be selected.

[0079] Step S23: Construct feature space. The server constructs and divides the mesh space according to the preset mesh step size and the parameter data of the specified model layer; then puts each local model into each cell in the space, calculates the number of local models placed in each cell, that is, the density of the cell, and constructs the feature space. .

[0080] Step S24: Wavelet transform. The server transforms the feature space in sequence Perform wavelet transform on each dimension to achieve data compression and obtain feature space .

[0081] Specifically, the dimension is the dimension of the parameter vector of the specified model layer; the operation starts from the first dimension. Perform wavelet transform to obtain , and then from the second dimension, Perform wavelet transform, and so on.

[0082] Step S25: Mark the cell. If the density of the cell is greater than the preset density threshold, it is a dense cell. Search feature space The dense cells in the cells are marked as a cluster. If there are adjacent and dense cells, the adjacent and dense cells are marked as the same cluster.

[0083] Step S26: Constructing a mapping relationship table: According to the wavelet clustering algorithm, construct a cell mapping table before and after the wavelet change.

[0084] Step S27: Look up the table. Look up the mapping relationship table to obtain each edge client cluster, partition the edge client, and obtain the edge client partition result that meets the data distribution situation.

[0085] Step S3: The server randomly selects some edge clients from each edge client partition to participate in the training. The edge clients adaptively perform local training. The server calculates the weights according to the local training status of the edge clients, performs model aggregation, and finally saves the trained model.

[0086] The model training process of the present invention is as follows Figure 3 As shown in the figure, first the server agrees on a time window, publishes the global model and related parameters, and the client performs iterative rounds of adaptive local model training. The server then calculates the weights and performs weighted aggregation to obtain the next round of global model.

[0087] Step S31: Server initialization. The server randomly generates a new global model , global communication round Initialized to 0, setting the global communication round threshold.

[0088] Optionally, the global communication round threshold may be set to 200.

[0089] Step S32: The server randomly selects edge clients to participate in the training from each edge client partition at a ratio of 40%-60%, and the server sends the global model and related parameters to the selected edge client; the related parameters include the specified theoretical local iteration rounds and aggregation time window.

[0090] Step S33: The edge client receives the global model , initialize the local iteration round e to 0.

[0091] Step S34: the edge client performs a round of local training, records the training time, and increases the local iteration round e by 1.

[0092] Step S35: The edge client determines whether the local iteration round is less than the specified theoretical local iteration round. If so, execute step S34; otherwise, execute step S36.

[0093] Step S36: The edge client tests the network status between the edge client and the server and estimates the transmission time, which is equal to network delay * local training model size. The network delay can be updated by real-time ping, or a third-party tool can be used to detect the network status to obtain the network delay. The local training model size is the file size of the local training model. Further estimate the end time of the next round of training, which is equal to the transmission time + training time + current time. If the end time of the next round of training does not exceed the aggregation time window, execute step S34; otherwise, execute step S37.

[0094] Step S37: The edge client sends the local training model and local iteration rounds to the server.

[0095] Step S38: The server waits within the time window. If an edge client does not respond within the preset time of the time window, it is determined that the global communication has failed and step S33 is executed. Otherwise, step S39 is executed. The probability of edge clients that did not respond in this round being selected to participate in training is reduced. The preset time can be set to 5 minutes.

[0096] Step S39: The server calculates the aggregation weight according to the number of local iterations. The weight calculation formula is as follows:

[0097]

[0098] in, is the learning rate, For edge clients The local iteration round of is the number of local iterations that each edge client should theoretically perform, The number of clients participating in this round of training;

[0099] The server performs weighted aggregation on the local models returned by the edge clients to obtain a new global model , global communication round Self-increment.

[0100] Step S310: Training completion judgment. Server judgment Is it greater than the global communication round threshold? If less than, execute step S32; if greater, end the training and save the training results.

[0101] Edge client selection and training impose less additional computing burden on edge clients, which has certain adaptability to edge environments with limited device performance and resources. Based on this, the present invention combines the wavelet clustering algorithm to design a new weight calculation method, and then organically combines these strategies to solve the problem of poor model training effect and low efficiency of federated learning in edge heterogeneous environments.

[0102] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A federated learning method for edge heterogeneous environments, characterized in that: The steps include: Step S1: The server detects the registration information of the edge client and performs performance testing on the edge client, removes the edge clients that do not meet the system standards, selects the edge clients that meet the requirements, and obtains an edge client set; Step S2: The edge client performs local model training, the server receives the training model, and uses the wavelet clustering algorithm to partition the edge client to obtain an edge client partition result that conforms to the data distribution; Step S3: The server randomly selects some edge clients from each edge client partition to participate in the training. The edge clients adaptively perform local training. The server calculates the weights according to the local training conditions of the edge clients, performs model aggregation, and finally saves the trained model. Step S2 includes the following steps: Step S21: Preheat the edge client; the server uniformly generates and initializes the global model , the server will global model Sent to every edge client in the edge client set; except for the global model In addition, the server also sends the local iteration count To edge clients; Edge clients follow the global model and the number of local iterations Perform local training and send the trained local model back to the server; Step S22: extracting model parameters; The server obtains the local models returned by each edge client and extracts parameter data of the specified model layer in each model; Step S23: constructing feature space; The server constructs and divides the mesh space according to the preset grid step size and the parameter data of the specified model layer; then it puts each local model into each cell in the space, calculates the density of the cell, and constructs the feature space. ; Step S24: wavelet transform; the server sequentially transforms the feature space Perform wavelet transform on each dimension to achieve data compression and obtain feature space ; Step S25: Mark the cell; if the density of the cell is greater than the preset density threshold, it is a dense cell; search the feature space The dense cells in the cells are marked as a cluster. If there are adjacent and dense cells, the adjacent and dense cells are marked as the same cluster; Step S26: construct a mapping relationship table; According to the wavelet clustering algorithm, a cell mapping table before and after the wavelet change is constructed; Step S27: Look up the table; Look up the mapping relationship table, obtain each edge client cluster, partition the edge client, and obtain the edge client partition result that meets the data distribution.

2. The method for federated learning in edge heterogeneous environments according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: The server sets the highest CPU usage standard, the lowest GPU performance standard, and the highest acceptable network delay and the most unstable network fluctuation; Step S12: The server performs performance testing on the edge client, including CPU testing, GPU testing, and network status testing, and records the test results; Step S13: The server compares the performance test result with the set index, removes edge clients that do not meet the minimum system standard, selects edge clients that meet the requirements, and obtains an edge client set; Step S14: Input the grid step size and density threshold parameters required for wavelet clustering into the server.

3. The method for federated learning in edge heterogeneous environments according to claim 2, characterized in that: Step S3 includes the following steps: Step S31: Server initialization; the server randomly generates a new global model , global communication round Initialize to 0 and set the global communication round threshold; Step S32: The server randomly selects edge clients to participate in the training from each edge client partition at a ratio of 40%-60%, and the server sends the global model and related parameters to the selected edge client; the related parameters include the specified theoretical local iteration rounds and aggregation time window; Step S33: The edge client receives the global model , initialize the local iteration round e to 0; Step S34: the edge client performs a round of local training, records the training time, and the local iteration round e is increased by 1; Step S35: The edge client determines whether the local iteration round is less than the specified theoretical local iteration round. If so, execute step S34; Otherwise, execute step S36; Step S36: the edge client tests the network status with the server, estimates the transmission time, which is equal to network delay * local training model size, and further estimates the end time of the next round of training, which is equal to transmission time + training time + current time. If the end time of the next round of training does not exceed the aggregation time window, execute step S34; Otherwise, execute step S37; Step S37: The edge client sends the local training model and local iteration rounds to the server; Step S38: The server waits within the time window. If an edge client does not respond after exceeding the preset time in the time window, it is determined that this round of global communication has failed and step S33 is executed. Otherwise, step S39 is executed. Step S39: The server calculates the aggregation weight according to the local iteration number through the weight calculation formula. The server performs weighted aggregation on the local model returned by the edge client to obtain a new global model. , global communication round Self-increment; Step S310: judging whether training is finished; Server judgment Is it greater than the global communication round threshold? If less than, execute step S32; if greater, end the training and save the training results.

4. The method for federated learning in edge heterogeneous environments according to claim 3, characterized in that: The weight calculation formula in step S39 is as follows: in, is the learning rate, For edge clients The local iteration round of is the number of local iterations that each edge client should theoretically perform, The number of clients participating in this round of training.

Images