Federal learning efficient training optimization method for 6G network

Abstract

The invention discloses a federal learning efficient training optimization method for a 6G network, and the method comprises the steps: transmitting a global model parameter to a client through a parameter server in a downlink communication stage, and transmitting the global model parameter to the client through a local training stage; according to the customized local training strategy, the convergence speed is improved by training the specified number of iterations at a local client and adding constraint terms in a target function, so that the total number of rounds of communication is reduced; in the uplink communication stage, part of clients participate in clients with relatively high rule selection calculation efficiency and upload training results of the clients at the same time, so that the communication cost of each round is reduced; in the model aggregation stage, a parameter server distributes reasonable weights for different local parameters based on a weighted aggregation strategy of data quality so as to improve model precision, and finally outputs optimized global model parameters so as to realize efficient model training and parameter updating, and an efficient and reliable solution is provided for distributed learning in a 6G network.

Description

Technical Field

[0001] The invention relates to the field of communication network technology, and specifically to an efficient training optimization method for federated learning for 6G networks. Background Art

[0002] Globally, as the exploration of the 6th generation mobile communication technology (6G) continues to deepen, artificial intelligence (AI) technology has been widely used in daily life and industrial scenarios. With its ultra-high speed, extremely low latency and wide connectivity, 6G networks are expected to make full use of a large amount of sensitive data to achieve a fully connected world. This vision not only brings unprecedented convenience to people, but also promotes the integration of 6G and AI endogenous intelligence. At the same time, the high efficiency and high reliability of 6G networks provide strong support for distributed computing. In 6G networks, data is usually stored in a large number of edge devices, and the data is dispersed and the amount of data is huge. The traditional centralized data processing method not only exposes user privacy, but also may lead to security risks during data transmission. Therefore, how to achieve efficient processing and utilization of data while protecting user privacy has become a challenge that must be faced in the development of 6G networks.

[0003] Federated learning technology clients train local deep neural network models using their private data. Local model updates are then sent to the central server, while private data remains on the client. After collecting all local updates, the central server is responsible for aggregating a new global model and providing it to the client for the next round of model training. This distributed training iteration is repeated until the global model converges to a satisfactory test accuracy or iterates to a specified round. By using federated learning, the privacy of the client can be effectively protected because no private information is shared between the client and the central server.

[0004] Although federated learning has great advantages in communication efficiency and privacy protection, traditional federated learning still faces major problems such as excessive communication overhead and slow convergence speed. These problems stem from the following three aspects:

[0005] First, in terms of federated learning model training, the data generated or collected by each client device in the network is different, and the statistical characteristics of these data obey non-independent identically distributed (Non-IID, non independentidentically distributed), which means that the local data distribution P between different clients i, j i ≠P jThis non-IID reduces the training efficiency of typical stochastic gradient descent (SGD)-based algorithms and seriously affects the convergence rate of the global model.

[0006] Secondly, the traditional federated learning system does not have a strict mechanism for screening participants, and generally adopts a random strategy to select clients to participate in the aggregation. However, due to the heterogeneity, uncontrollability, and distribution of clients, using a random strategy to receive client aggregation global models may lead to an influx of low-quality clients, thereby destroying the overall performance of the federated learning system. Low-quality clients have poor data quality or the data involved in training is not independent and identically distributed. The model parameters provided by such unreliable clients may also lead to a decline in model performance.

[0007] Finally, the existing aggregation process in federated learning also significantly affects the overall communication overhead. Traditional aggregation methods, such as simple averaging strategies, although relatively straightforward to implement, often fail to fully consider the quality and diversity of individual client data. This simple aggregation strategy may result in limited performance of the global model and fail to effectively utilize the data provided by each client, especially when the data is unevenly distributed or the number of clients is large. In addition, simple averaging may also ignore the important information contributed by some clients during the training process, which in turn affects the generalization ability of the model.

[0008] Therefore, it is necessary to study an efficient federated learning training optimization method for 6G networks to reduce communication overhead and ensure convergence speed. In order to achieve a better trade-off between local computing and communication overhead, we allow different clients to have different local training rounds so that their update directions are as close to the global model as possible. In the model training phase, we adopt a customized local training strategy to reduce the number of communication rounds. By adding constraints to the objective function, we can optimize the number of local training rounds and improve the convergence speed. Since the computational efficiency of each client is different, we should select clients with high computational efficiency to participate in the aggregation, which is summarized as the "partial client participation rule" to reduce the communication overhead per round. In addition, we build a data quality assessment model based on the client's data volume, gradient similarity, and data diversity, and aggregate local models with different weights calculated by the data quality assessment model into a global model, assign appropriate weights to local models, accelerate model convergence, and improve the accuracy of the global model. Summary of the invention

[0009] The present invention provides an efficient training optimization method for federated learning for 6G networks, aiming to solve the limitations of traditional federated learning in model training in terms of communication efficiency and model convergence speed due to the heterogeneity of data distribution of different clients. The present invention reduces the number of communication rounds through customized local training strategies, introduces some client participation rules to reduce the cost of each round of communication, and adopts a weighted aggregation strategy based on data quality to improve model accuracy. While maintaining the accuracy of the model, it can significantly reduce the number of communication rounds and participating clients in each round of communication, thereby reducing the overall communication cost and accelerating the convergence speed of the model.

[0010] A federated learning efficient training optimization method for 6G networks includes the following steps:

[0011] Step S1: Construct a federated learning training framework for 6G networks, which is used in scenarios where the heterogeneity of client data leads to huge communication costs and slow model convergence;

[0012] Step S2: Introduce a customized local training strategy to increase the convergence speed by adding constraints to the objective function of the local client to reduce the total number of communication rounds;

[0013] Step S3: Setting the rules for some clients to participate in the training, and selecting clients with higher computational efficiency to upload their training results at the same time, thereby reducing the communication cost of each round;

[0014] Step S4: Design a weight calculation method. In a non-IID setting, by analyzing the three key factors of client data volume, gradient similarity, and data diversity, a data quality assessment model is constructed and the aggregation weights are adjusted to improve model accuracy.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The present invention proposes an efficient training optimization method for federated learning for 6G networks. Because the data generated or collected by client devices is heterogeneous, measures need to be taken during the training process to reduce this heterogeneity. Because the adoption of a random strategy to accept local models for aggregation will lead to an influx of low-quality clients, it is necessary to select high-quality clients to participate in the aggregation. Because the traditional average aggregation method fails to fully consider the quality and diversity of client data, a more reasonable weight calculation method needs to be designed. The federated learning training optimization method of the present invention considers the influencing factors more comprehensively, and provides an efficient and reliable solution for distributed learning in a 6G network environment, especially in processing large amounts of sensitive data and realizing a fully interconnected world. It has significant advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a training scenario diagram of the federated learning model provided by the present invention;

[0018] Figure 2 It is a flow chart of a method for efficient training optimization of federated learning for 6G networks of the present invention;

[0019] Figure 3 It is a schematic diagram comparing the effects of different Dirichlet distribution parameters;

[0020] Figure 4 This is a schematic diagram comparing the effects of the parameter ε of different local training strategies on accuracy;

[0021] Figure 5 It is a schematic diagram of the performance comparison of different algorithms under the actual Non-IID setting;

[0022] Figure 6 It is a schematic diagram of the performance comparison of different algorithms under the pathological Non-IID setting;

[0023] Figure 7 This is a schematic diagram comparing the communication overhead of different algorithms after 100 rounds of training. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, those skilled in the art will fall within the scope defined by the appended claims of the present application to the modification of various equivalent forms of the present invention.

[0025] The present invention Figure 1 Take the federated learning model training system shown in the figure as an example. Figure 1 In the model architecture of the federated learning system shown, a central server is connected to multiple clients, where clients 1 to K each hold their own data sets and perform local training. During the training process, the server selects a portion of clients with high computational efficiency to participate in the weighted aggregation of the model based on the data quality assessment model to build a global model. This data quality assessment model comprehensively considers the data volume, gradient similarity, and data diversity of each client, evaluates the quality of the data by calculating the KL divergence, and assigns reasonable aggregation weights to the local model of each client accordingly. Finally, the server sends the aggregated global model parameters to all clients for the next round of local training and model updates. This process is repeated until the global model converges.

[0026] Based on the scene graph, the present invention proposes an efficient training optimization method for federated learning in 6G networks. In order to achieve efficient and accurate federated learning model training in 6G networks, this method specifically targets different indicators of different clients in 6G networks, and combines customized local training strategies, partial client participation rules, and weighted aggregation strategies based on data quality to effectively reduce communication overhead and accelerate the convergence of the model. This method not only reduces the number of communication rounds, but also optimizes the aggregation process of the global model through an accurate data quality assessment model, improves the accuracy and robustness of the model when processing non-independent and identically distributed data, and provides an efficient and reliable solution for distributed learning in future wireless communication networks. The method comprises the following steps:

[0027] Step S1: Downlink communication phase: The federated learning process starts with the downlink communication phase, which converts the global model parameters ω t Downloaded from the parameter server to the K clients in round t.

[0028] Step S2: Local training phase: Each client calculates an updated model based on the collected data. Considering the data heterogeneity of different clients, each client will calculate the gradient of the loss function based on its own data set and model parameters, and solve the optimal number of local iterations to achieve higher accuracy with fewer iterations.

[0029] Step S3: Uplink communication phase: First, evaluate the computational efficiency of each client based on the model training time and data set size of the previous round, and select the model parameters of the n(t) clients with the highest computational efficiency to upload to the parameter server. This phase directly determines the size of each round of communication overhead. After multiple communication rounds, the number of participants has a crucial impact on the total communication overhead of this phase. The downlink and uplink communication phases constitute a complete communication round, which ends with the uplink communication round.

[0030] Step S4: Aggregation phase: Each device transmits the latest model parameters to the parameter server. After the transmitted message arrives at the parameter server, the aggregation phase begins. The server needs to wait for the transmission data sent by all participants to the parameter server, calculate three different indicators and build a data quality assessment model, use data quality as a weight to weight and aggregate the model parameters of different clients, and finally distribute the aggregated global model back to the local client for the next round of local training.

[0031] Specific steps:

[0032] Step S1: In the initial stage of federated learning, the training data set is first divided. Different Dirichlet distribution parameters represent different data heterogeneity. Then the parameter server sends the initial global model parameters to all clients participating in the tth round of training. This step is the initial communication link in the federated learning process, and its purpose is to ensure that all participating clients can start local training from the same model parameters.

[0033] Step S2: The client uses its private dataset to train the local model. At this stage, each client calculates the gradient according to the characteristics of its dataset and adjusts the number of local iterations according to data heterogeneity to achieve higher model accuracy with fewer iterations. This step is a key step in federated learning because it involves the client using local data to perform personalized optimization of the global model.

[0034] Step S3: After completing local training, the client will participate in uplink communication based on its computational efficiency evaluation results. Only the n(t) clients with the highest computational efficiency are selected to upload their model parameters to the parameter server. This step determines the cost of each round of communication, because only some clients participate in data upload, thereby reducing the overall communication burden.

[0035] Step S4: After receiving the model parameters uploaded by the selected client, the parameter server enters the aggregation phase. The server will calculate the quality of each client's data and use these weights to weight and aggregate all uploaded model parameters. The aggregated global model parameters will be updated and distributed to all clients in preparation for the next round of federated learning. This step is the aggregation link in federated learning, and its purpose is to integrate the model updates of each client to form a global model update.

[0036] 1. In view of the divergence of weights between local models and global models caused by the heterogeneity of data distribution of different clients, a suitable local training strategy is specified for each client, different numbers of training cycles are provided, and a constraint term is introduced to effectively reduce the gradient diversity between local loss functions, accelerate the convergence of the model, and save communication costs. The specific steps are:

[0037] (1) Adaptive local iteration strategy: Different devices in a federated network usually have different resource constraints in terms of computing hardware, network connectivity, and computing power levels. Therefore, it is unrealistic to force each device to perform the same number of local iterations as in the federated averaging algorithm. We allow a variable number of iteration cycles to be performed locally based on the system resources available to the device, and then aggregate partial training results sent from stragglers, implicitly accommodating different numbers of iterations for different devices. , we assume It is through Sub-local training The local model parameters of client k are in This customized local training strategy achieves higher accuracy with fewer local iterations.

[0038] (2) Loss-constrained optimization: While tolerating different amounts of work performed across devices can help mitigate the negative impact of system heterogeneity, excessive local updates may still potentially cause the method to diverge due to the underlying heterogeneous data. We propose to add a proximal term to the local subproblem to effectively constrain the impact of varying local updates. Specifically, instead of simply minimizing the local loss function, device k uses its selected local solver to approximately minimize the following objective equation:

[0039]

[0040] Among them, g k (ω k ;ω t ) represents the objective equation constructed after adding the regularization term to the local loss function. The problem is transformed from minimizing the local loss function to minimizing the objective equation g k (ω k ;ω t ), F k (ω k ) represents the loss function of client k, ε is the penalty factor of the constraint term, ω k represents the local model parameters of client k in round t training, ω t Represents the global model parameters after the aggregation of the t-th round of training.

[0041] 2. In a 6G network environment, due to the limitation of uplink communication bandwidth, only a few clients can upload training results to the parameter server at the same time, which reduces the training speed and may make the global model biased towards clients with large data volumes. To solve this problem, we propose a client selection strategy based on computational efficiency. By evaluating the computing power and data volume of the client, we select clients with high computational efficiency to participate in model training and parameter upload, so as to reduce communication overhead and ensure that the global model evenly reflects the data characteristics of all clients, thereby improving the generalization ability and accuracy of the model. The specific steps are as follows:

[0042] (1) Client computing efficiency evaluation phase: using the local training time ζ of the previous round k and local dataset size |D k |, calculate the local computational efficiency e of client k k for:

[0043] Assume local data volume |D k | is constant, the local computation efficiency of client k in the tth communication round is expressed as where t latest is the most recent communication round, t latest <T. Since local training is not performed on all clients in the initialization phase, the local data volume of client k is used |D k | to initialize the local computing efficiency e k for:

[0044] In general, the client's initial local computing efficiency is much higher than its actual local computing efficiency, that is, Therefore, at the beginning of an iteration, the distribution of clients with larger local data volumes is more dispersed. In other words, clients that have failed to complete training are more likely to be selected, thereby quickly obtaining true local computing efficiency.

[0045] (2) Client selection based on computational efficiency: Based on the information uploaded by the local client, the probability value p k It is used to represent the probability of a client being selected. Clients with higher computational efficiency tend to have larger p k , the parameter server calculates the probability value p for each client k As shown below:

[0046] In each round, the subset of selected clients is different, which makes the global model consider as much client information as possible. Therefore, the global loss function is transformed into the following formula:

[0047]

[0048] 3. In the federated learning scenario, the non-independent and identically distributed characteristics of client data lead to challenges in global model accuracy and convergence. If data heterogeneity is not handled properly, it will reduce the accuracy of the global model and slow down the convergence. Existing methods often ignore this and mistakenly give too high weights to low-quality models, thereby affecting the performance of the global model. Therefore, we propose a new weighted aggregation method that evaluates data quality based on indicators such as data volume, gradient similarity, and data diversity, and uses KL divergence to quantify the differences between data sets to reasonably assign aggregation weights to local models. This method strengthens the impact of high-quality data, improves the accuracy and convergence of the global model, and maintains training efficiency. The specific steps are as follows:

[0049] (1) Calculation stages of different indicators:

[0050] a) We will client kThe data volume is defined as the normalized value of the ratio of client data to total data, as shown in the following formula:

[0051]

[0052] b) In the scenario of federated learning, the contribution of client k’s data to the global model training can also be described by gradient similarity. Gradient similarity measures the data quality of client k by comparing the gradient of the local model with the gradient of the global model. From the perspective of optimization, the main goal of federated learning is to obtain the optimal global model parameters by minimizing one or more loss functions. The gradient used in the fitting process reflects the direction of optimization to a certain extent. Therefore, if the training gradient generated by client k’s data is closer to the assumed optimal gradient direction, it is considered to be more similar to the optimal model and the data quality is higher. Therefore, in the non-independent and identically distributed setting, we consider using gradient similarity to alleviate the client drift phenomenon. The gradient similarity γ of client k is constructed by aggregating and normalizing the cosine similarity of multiple rounds. k :

[0053]

[0054] in, represents the local model gradient vector of the client in round t, u t Represents the tth global model gradient vector, and iter represents the total number of gradient iterations.

[0055] c) When the data provided in federated learning is sufficiently reliable, we find that the diversity of data distribution can reflect the quality of the data to a certain extent. Higher data quality corresponds to richer data diversity. In the scenario of federated learning, each data sample contains multiple feature dimensions. The Gram matrix G can calculate the feature representation vector x of the data sample. i ∈X k The volume of the parallel polyhedron constructed in the Euclidean space. From previous studies, we know that when the volume of the constructed parallel polyhedron is larger, the characteristic distribution of the data is more divergent, corresponding to higher data quality. Therefore, we use the Gram determinant to measure the diversity of the client's local data and normalize the data diversity, as shown in the following formula:

[0056]

[0057] (2) Index weight evaluation stage: A data quality evaluation model including data volume, gradient similarity and data diversity is constructed through KL divergence to evaluate the impact of the above indicators.

[0058] set up γ and σ are the weights of data volume, gradient similarity and data diversity respectively. ρ γ and ρ σ To measure the importance of these three indicators. If the proportion of models with large data distribution skewness is reduced during the model aggregation process, the adverse impact of low-quality local models on the overall model performance can be reduced, and the accuracy of the global model can be improved. At the same time, if the amount of data of multiple clients is the same, the weight of the data amount The impact of is not significant, so we can give a data quality evaluation model, assuming that the probability distribution W is the target distribution, indicating that when the three indicators A set of weights when γ and σ are indistinguishable:

[0059]

[0060] (3) Data quality assessment model construction phase: Data quality assessment model C for client k k Defined as follows:

[0061]

[0062] Model C is evaluated by client data quality k , the server collects all locally trained models from the clients and aggregates them using the normalization of the data quality assessment model as weights to update the global model; the t-round global model is obtained by:

[0063]

[0064] Among them, C k represents the data quality assessment model of client k, C i represents the data quality assessment model for different clients i, where i is the client index, represents the local model parameters of client k in the tth round of training.

[0065] 4. We designed an efficient training optimization method for federated learning in 6G networks. In order to achieve efficient and accurate federated learning model training in 6G networks, this method specifically targets different indicators of different clients in 6G networks. By combining customized local training strategies, some client participation rules, and weighted aggregation strategies based on data quality, it effectively reduces communication overhead and accelerates the convergence of the model. This method not only reduces the number of communication rounds, but also optimizes the aggregation process of the global model through an accurate data quality assessment model, improves the accuracy and robustness of the model when processing non-independent and identically distributed data, and provides an efficient and reliable solution for distributed learning in future wireless communication networks.

[0066] The general setting of federated learning is that different clients have the same neural network model in the initial stage. These distributed devices iteratively train the local model with their private data for a fixed number of times. A complete round of communication includes the following stages:

[0067] (1) Downlink communication phase: The federated learning process starts with the downlink communication phase, which initializes the global model parameters ω t Downloaded from the parameter server to the K clients in the tth round as the starting point for model training.

[0068] (2) Local training phase: Each client calculates an updated model based on the collected data. Considering the data heterogeneity of different clients, each client calculates the gradient of the loss function based on its own data set and model parameters, and solves the optimal number of local iterations to achieve higher accuracy with fewer iterations.

[0069] (3) Uplink communication phase: First, the computational efficiency of each client is evaluated based on the model training time and data set size of the previous round, and the model parameters of the n(t) clients with the highest computational efficiency are selected and uploaded to the parameter server. This phase directly determines the size of the communication overhead in each round. After multiple communication rounds, the number of participants has a crucial impact on the total communication overhead of this phase. The downlink and uplink communication phases constitute a complete communication round, which ends with the uplink communication round.

[0070] (4) Aggregation phase: Each device transmits the latest model parameters to the parameter server. After the transmission message arrives at the parameter server, the aggregation phase begins. The server needs to wait for the transmission data sent by all participants to the parameter server, calculate three different indicators and build a data quality assessment model, use the data quality assessment model as a weight to aggregate the model parameters of different clients, and finally distribute the aggregated global model back to the local client for the next round of local training.

[0071] First, we demonstrate how statistical heterogeneity affects convergence. We use two different dataset samples (FashionMNIST and CIFAR10) to generate samples using different Dirichlet distribution concentration parameters β for training to compare the performance of the algorithm under different heterogeneity conditions. The smaller the value of the Dirichlet parameter β, the higher the statistical heterogeneity of the generated sample labels.

[0072] pass Figure 3From the simulation results, it can be observed that as the concentration parameter β of the Dirichlet distribution decreases, that is, the data heterogeneity increases, the accuracy of the method of the present invention is significantly improved, and the convergence speed is accelerated. This phenomenon shows that the present method has significant advantages in processing non-independent and identically distributed data, especially when facing highly heterogeneous data distribution. Therefore, the present method has better performance for heterogeneous data and can effectively solve the weight divergence phenomenon caused by data heterogeneity in personalized federated learning.

[0073] Next, we examine the impact of different local training parameters on the performance, and fix the local training period to E = 20. For different data sets, the local training parameter ε is set to 0, 0.3, 0.8, and 1 to represent different degrees of penalty.

[0074] Figure 4 The experimental results show that when the penalty coefficient weight is larger, the ε value is larger, and it has better performance in different situations. Therefore, according to the above results, ε is set to 1 in subsequent experiments. According to the experimental results, the customized local training strategy introduced in the method is suitable for real settings with different statistical heterogeneity. Therefore, the method of the present invention reduces the total number of communication rounds T and reduces the total communication cost through the customized local training strategy.

[0075] The proposed CEFED method is compared with several related FL algorithms, and compared with FedAvg, FedProx, SCAFFOLD, and PerAVG algorithms to highlight the advantages of this method. We compare the performance of CEFED and the baseline method in 100 rounds of communication under the actual non-IID setting. In order to intuitively demonstrate the performance of the CEFED algorithm, Figure 5 As shown in the figure, the test accuracy of the CEFED algorithm in 100 iterations is compared with that of the baseline method. The results show that the CEFED algorithm not only significantly outperforms all baseline methods on the Fashion-MNIST and CIFAR-10 datasets, but also exhibits smoother performance fluctuations. This stability is attributed to the set consistency regularization term, which effectively alleviates the weight divergence problem caused by non-IID data, further demonstrating the superiority and practical value of CEFED in processing complex data environments.

[0076] In contrast, although the actual non-IID scenario based on Dirichlet partitioning has a more diverse and richer category distribution, which helps the model capture more diversified feature expressions and thus improves the generalization ability of the model, at the same time, the fixed number of categories may limit the model's ability to accurately distinguish the distribution of different client data. Figure 6As shown, it vividly demonstrates the comparison of the test accuracy between the proposed method and the benchmark method during 100 rounds of iterations, further highlighting the performance advantage of CEFED in the pathological non-IID setting.

[0077] like Figure 7 As shown, a comparison of the communication overhead of different algorithms for 100 rounds of training on two data sets is also carried out. The CEFED algorithm proposed in the present invention shows significant advantages in reducing communication overhead. By introducing a customized local training strategy, the algorithm can significantly reduce the required communication rounds while maintaining model accuracy. In addition, by implementing partial client participation rules, the CEFED algorithm allows only some clients to upload their local models in each round of training. This strategy effectively reduces the number of clients participating in each round of communication, thereby reducing communication costs. More importantly, based on the weighted aggregation strategy of data quality, the algorithm assigns reasonable weights to different local parameters by evaluating the amount of client data, gradient similarity, and data diversity, which not only improves the model accuracy, but also reduces the number of times communication is required to correct the model.

[0078] The present invention discloses a federated learning efficient training optimization method for 6G networks, which includes three key modules: customized local training strategy, partial client participation rules, and weighted aggregation strategy based on data quality. The working process of the method is as follows: first, in the downlink communication stage, the global model parameters are sent to the client by the parameter server, and secondly, in the local training stage, the customized local training strategy improves the convergence speed by training the specified number of iterations in the local client and adding constraints in the objective function to reduce the total number of communication rounds; then, in the uplink communication stage, the partial client participation rules select clients with higher computational efficiency and upload their training results at the same time, thereby reducing the communication cost of each round; finally, in the model aggregation stage, the parameter server uses a weighted aggregation strategy based on data quality to improve the model accuracy by assigning reasonable weights to different local parameters, and finally outputs the optimized global model parameters to achieve efficient model training and parameter update, providing an efficient and reliable solution for distributed learning in 6G networks.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A federated learning efficient training optimization method for 6G networks, comprising the following steps: Step S1: Construct a federated learning training framework for 6G networks, which is used in scenarios where the heterogeneity of client data leads to huge communication costs and slow model convergence; Step S2: Introduce a customized local training strategy to increase the convergence speed by adding constraints to the objective function of the local client to reduce the total number of communication rounds; Step S3: Setting the rules for some clients to participate in the training, and selecting clients with higher computational efficiency to upload their training results at the same time, thereby reducing the communication cost of each round; Step S4: Design a weight calculation method. In a non-IID setting, by analyzing the three key factors of client data volume, gradient similarity, and data diversity, a data quality assessment model is constructed and the aggregation weights are adjusted to improve model accuracy.

2. The method for efficient training optimization of federated learning for 6G networks according to claim 1 is characterized in that: In the above step S1, federated learning means that different clients have the same neural network model in the initial stage, and the distributed devices composed of different clients use their private data to iteratively train the local model for a fixed number of times. A complete round of communication includes the following stages: Step S11: Downlink communication phase: The federated learning process starts with the downlink communication phase, which converts the global model parameter ω t Download from the parameter server to the K clients in round t; Step S12: Local training phase: Each client calculates an updated model based on the collected data; considering the data heterogeneity of different clients, each client calculates the gradient of the loss function based on its own data set and model parameters, and solves the optimal number of local iterations to achieve higher accuracy with fewer iterations; Step S13: Uplink communication phase: first, the computational efficiency of each client is evaluated according to the model training time and data set size of the previous round, and the model parameters of the n(t) clients with the highest computational efficiency are selected and uploaded to the parameter server; This phase directly determines the size of each round of communication overhead; after multiple communication rounds, the number of participants has a crucial impact on the total communication overhead of this phase; the downlink and uplink communication phases constitute a complete communication round, which ends with the uplink communication round; Step S14: Aggregation phase: Each device transmits the latest model parameters to the parameter server; After the transmitted message arrives at the parameter server, the aggregation phase begins; The server needs to wait for the transmission data sent by all participants to the parameter server, calculate three different indicators and build a data quality assessment model, use the data quality as a weight to aggregate the model parameters of different clients, and finally distribute the updated output back to the device.

3. The method for efficient training optimization of federated learning for 6G networks according to claim 2 is characterized in that: In the above step S2, a customized local training strategy is implemented by allowing different iteration cycles to be executed locally across clients; a constraint term is introduced to effectively reduce the gradient diversity between local functions; client k does not only minimize the local loss function F k (·), but uses its local loss function to approximately minimize the following objective equation: Among them, g k (ω k ;ω t ) represents the objective equation constructed after adding the regularization term to the local loss function. The problem is transformed from minimizing the local loss function to minimizing the objective equation g k (ω k ;ω t ), F k (ω k ) represents the loss function of client k, ε is the penalty factor of the constraint term, ω k represents the local model parameters of client k in round t training, ω t Represents the global model parameters after the aggregation of the t-th round of training.

4. The method for efficient training optimization of federated learning for 6G networks according to claim 3 is characterized in that: In the above step S2, for the target equation, the variable θ represents different local iteration times. Indicates the iteration on client k in the tth federated training times; assuming It is through The local model parameters of client k after local training, in This customized local training strategy achieves higher accuracy with fewer local iterations.

5. The method for efficient training optimization of federated learning for 6G networks according to claim 4 is characterized in that: In the above step S3, the specific process of the partial client participation rule is as follows: in each round of communication, a subset of n(t) clients is selected, denoted as S t ; The local model on the client is used to average the global model loss function after the aggregation stage on each selected client Optimize the local objective function; where D k represents the amount of data of client k, and D represents the total amount of data. In the uplink communication phase, the client uploads the local model parameters and transmits the size of the local data set and the training time of the local model in the previous round of communication to the parameter server.

6. The method for efficient training optimization of federated learning for 6G networks according to claim 5 is characterized in that: In the above step S3, the local training time ζ of the previous round is used k and local dataset size |D k |, calculate the local computational efficiency e of client k k for: Based on the information uploaded by the local client, the probability value p k It is used to represent the probability of a client being selected. Clients with higher computational efficiency tend to have larger p k , the parameter server calculates the probability value p for each client k As shown below: In each round, the subset of selected clients is different, which makes the global model consider as much client information as possible, and the expression of the global loss function is transformed into the following:

7. The method for efficient training optimization of federated learning for 6G networks according to claim 6 is characterized in that: In the above step S4, the quality of the data is first evaluated, and then the KL divergence-based model is used to determine the data relevance, and the aggregation weight of the model in federated learning is adjusted accordingly; the specific indicator is the normalized value of the data volume Gradient similarity normalized value γ k , the normalized value of data diversity σ k ; The data volume of client k is defined as the normalized value of the ratio of client data to total data In the scenario of federated learning, the contribution of client k's data to the global model training can also be described by gradient similarity. Gradient similarity measures the data quality of client k by comparing the gradient of the local model with the global model gradient. From the perspective of optimization, the main goal of federated learning is to obtain the optimal global model parameters by minimizing one or more loss functions. The gradient used in the fitting process reflects the direction of optimization to a certain extent. Therefore, if the training gradient generated by client k's data is closer to the assumed optimal gradient direction, it is considered to be more similar to the optimal model and the data quality is higher. Therefore, in the non-independent and identically distributed setting, we consider using gradient similarity to alleviate the client drift phenomenon. The gradient similarity γ of client k is constructed by aggregating and normalizing the cosine similarity of multiple rounds. k : in, represents the local model gradient vector of the client in round t, u t Represents the tth global model gradient vector, and iter represents the total number of gradient iterations. When the data provided in federated learning is sufficiently reliable, we find that the diversity of data distribution can reflect the quality of data to a certain extent. Higher data quality corresponds to richer data diversity. In the scenario of federated learning, each data sample contains multiple feature dimensions. The Gram matrix G can calculate the feature representation vector x of the data sample. i ∈X k The volume of the parallel polyhedron constructed in the Euclidean space. From previous studies, we know that when the volume of the constructed parallel polyhedron is larger, the feature distribution of the data is more divergent, corresponding to higher data quality. Therefore, we use the Gram determinant to measure the diversity of the client local data σ k , and normalize the data diversity as shown below:

8. The method for efficient training optimization of federated learning for 6G networks according to claim 7, characterized in that: In the above step S4, a data quality assessment model including data volume, gradient similarity and data diversity is constructed through KL divergence; γ and σ are the weights of data volume, gradient similarity and data diversity, respectively, and are expressed as ρ γ and ρ σ To measure the importance of these three indicators; Provide an evaluation model for data quality Assume that the probability distribution W is the target distribution, which means that when the three indicators A set of weights when γ and σ are indistinguishable.

9. The method for efficient training optimization of federated learning for 6G networks according to claim 8, characterized in that: In the above step S4, the data quality model C of client k k As shown below: Model C is evaluated by client data quality k ,The server collects all locally trained models from clients and aggregates them with the normalization of the data quality assessment model as weights to update the global model; The t-round global model is obtained by the following formula: Among them, C k represents the data quality assessment model of client k, C i represents the data quality assessment model for different clients i, where i is the client index, represents the local model parameters of client k in the tth round of training.

Images