An adaptive communication federated learning method suitable for a dynamic network environment

Abstract

The application discloses a self-adaptive communication federated learning method suitable for a dynamic network environment, and is applied to the technical field of edge intelligent calculation, and aims at the problems of low network resource utilization and large communication cost caused by the inherent network resource limitation, network dynamics and heterogeneity of edge calculation in the prior art. The application is based on a decentralized federated learning training model, each client only needs to communicate with a neighbor client, so that the network resources can be more fully utilized, and then a parameter screening mechanism is provided, so that the clients only need to exchange screened parameters, and the communication cost is reduced. The application can reconstruct a communication topology in a dynamic network environment, and can efficiently train a model to reach an expected precision.

Description

Technical Field

[0001] This invention belongs to the field of edge intelligent computing technology, and specifically relates to an adaptive communication federated learning technology. Background Technology

[0002] With the rapid development of 5G and the Internet of Things (IoT), and the widespread adoption of smartphones, wearable devices, and mobile networks, edge devices are ubiquitous in modern society. These edge devices continuously generate vast amounts of private data, such as photos and voice recordings, which are invaluable for developing intelligent applications. Simultaneously, edge computing (EC) stores data locally, pushing more computing power to the network edge for data analysis. Within EC, federated learning (FL) has emerged as an new paradigm. Federated learning frameworks can fully utilize the rich private data on edge devices without compromising privacy, training machine learning (ML) models while maintaining data decentralization. FL is a distributed machine learning framework incorporating secure encryption technologies. Such frameworks are envisioned as promising solutions for achieving ubiquitous intelligence and are well-suited for applications in data security, healthcare, and medical analytics, thus attracting significant attention.

[0003] FL architecture is mainly divided into two types: centralized federated learning (CFL) using a parameter server and decentralized federated learning (DFL) using gossip communication. CFL centralizes the model parameters trained by all clients using local data to a central server, aggregates the parameters to obtain a global model, and then distributes it to each client. This model inevitably suffers from single point of failure risk, trust dependencies, and communication bottlenecks. To address these issues, the DFL model abandons the parameter server, exchanging parameters only between adjacent clients. In this way, DFL improves fault tolerance because clients continuously update their knowledge of available nodes, enhancing network robustness and reducing the risk of single point of failure. Furthermore, DFL improves trust issues by distributing trust among clients, which determines the overall trust of the federation, thus significantly reducing single point attacks. DFL also allows for a more even distribution of communication load among clients, thereby mitigating communication bottlenecks to some extent.

[0004] However, DFL faces three main challenges in achieving efficient model training: 1) Limited and heterogeneous resources. In large-scale real-world scenarios, training clients are typically distributed across different geographical locations, with significant differences in bandwidth, local data volume, and computing power. Optimizing resource allocation among clients during DFL training remains an open question. 2) Dynamic network conditions. In practical applications, clients usually connect to the network via wireless links. Due to client mobility and link instability, the state of these wireless links is often time-varying. Maintaining a good convergence rate while dealing with constantly changing network conditions presents a new challenge for DFL. 3) High communication overhead. Furthermore, modern models often involve millions of parameters, resulting in high communication costs during iterative communication, posing a significant obstacle to the feasibility of DFL. The literature A. Koloskova, T. Lin, SUStich, and M. Jaggi, “Decentralized deep learning with arbitrary communication compression,” in Proc. Int. Conf. Learn. Representations, 2020, pp. 1–22, designs a client selection scheme based on heterogeneous client resources. However, it still requires each client to communicate with multiple connected peers and uses a fixed communication topology to ensure model convergence, lacking a comprehensive consideration of network dynamics. The literature A. Abdelmoniem and M. Canini, “Dc2: Delay-aware compression control for distributed machine learning,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications, IEEE, 2021, pp. 1–10, employs quantization techniques to reduce communication load, but assigns a fixed or identical compression ratio to all parameters without considering the importance of specific parameters in the model, thus severely reducing training efficiency. Therefore, researching an adaptive communication federated learning method suitable for dynamic network environments has significant practical implications. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a federated learning method for heterogeneous clients in dynamic network environments, which adaptively selects peers for communication connections and filters parameters based on their importance.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is an adaptive communication federated learning method suitable for dynamic network environments, comprising the following steps:

[0007] The steps to establish a decentralized federated learning framework are as follows: adopt a network topology without a central server, each client in the network hosts a copy of the global model parameters and uses local data for training, while in each round of training, only one neighboring client is selected to exchange model parameters.

[0008] Coordinator configuration steps: The coordinator makes a comprehensive assessment of the network resources of each client. Before the client starts a round of model parameter iteration, it makes a comprehensive assessment of the network resources of each client, selects a neighboring client as the communication object for the next round of model parameter iteration, and broadcasts the completed communication topology to each client.

[0009] Model training steps: Each client iterates the model parameters in each round according to the consensus optimization algorithm;

[0010] The model parameter exchange steps based on dynamic thresholds are as follows: After each client completes a round of model parameter iteration, it determines whether the change in the model parameters in this round reaches the set parameter change threshold. If so, the parameters that meet the change threshold are sent to the selected neighboring clients. If not, the parameters that do not meet the change threshold are not sent, and the client starts the next round of model parameter iteration.

[0011] This invention integrates client selection and parameter filtering to accelerate DFL training, makes a comprehensive evaluation of network resources, and thereby completes the adaptive peer selection of clients. At the same time, it simplifies the communication topology construction during the iteration process to the greatest extent and takes into account the parameter filtering during model exchange, thereby minimizing the communication overhead of training.

[0012] The beneficial effects of this invention are: low communication overhead, fast convergence speed and improved network resource utilization, and it can be directly applied to Internet of Things (IoT) technology. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method process in the embodiment.

[0014] Figure 2 This section compares the convergence performance of the method of this invention with the baseline on the IID dataset, including comparisons with the method of this invention, FedAvg algorithm, DPSGD algorithm, Choco algorithm, and GossipFL algorithm.

[0015] in, Figure 2 (a) shows the training accuracy variation curve of the MNIST dataset on CNN. Figure 2(b) shows the training accuracy variation curve of the CIFAR-10 dataset on CNN. Figure 2 (c) shows the training accuracy variation curve of the CIFAR-10 dataset on ResNet-20.

[0016] Figure 3 This section compares the communication overhead of the method in this invention with the baseline on the IID dataset, including comparisons with the method of this invention, the FedAvg algorithm, the DPSGD algorithm, the Choco algorithm, and the GossipFL algorithm.

[0017] in, Figure 3 (a) shows the communication overhead curve of the MNIST dataset on a CNN. Figure 3 (b) shows the communication overhead curve of the CIFAR-10 dataset on a CNN. Figure 3 (c) shows the communication overhead curve of the CIFAR-10 dataset on ResNet-20.

[0018] Figure 4 This section compares the communication overhead of the method of this invention with the baseline on the IID dataset to achieve the target accuracy, including comparisons with the algorithm of this invention, FedAvg algorithm, DPSGD algorithm, Choco algorithm, and GossipFL algorithm.

[0019] in, Figure 4 (a) is a bar chart showing the communication overhead for achieving the target accuracy on CNN using the MNIST dataset. Figure 4 (b) is a bar chart showing the communication overhead for achieving the target accuracy on CNNs using the CIFAR-10 dataset. Figure 4 (c) is a bar chart showing the communication overhead for achieving the target accuracy on ResNet-20 using the CIFAR-10 dataset.

[0020] Figure 5 This section compares the convergence performance of the method of this invention with the baseline on a non-IID dataset, including comparisons with the algorithm of this invention, FedAvg algorithm, DPSGD algorithm, Choco algorithm, and GossipFL algorithm.

[0021] in, Figure 5 (a) shows the training accuracy variation curve of the MNIST dataset on CNN. Figure 5 (b) shows the training accuracy variation curve of the CIFAR-10 dataset on CNN.

[0022] Figure 6 This section compares the communication overhead of the method of this invention with the baseline on a non-IID dataset, including comparisons with the algorithm of this invention, FedAvg algorithm, DPSGD algorithm, Choco algorithm, and GossipFL algorithm.

[0023] in, Figure 6 (a) shows the communication overhead curve of the MNIST dataset on a CNN. Figure 6 (b) shows the communication overhead curve of the CIFAR-10 dataset on a CNN.

[0024] Figure 7 This section compares the communication overhead of the method of this invention with the baseline on a non-IID dataset to achieve the target accuracy, including comparisons with the algorithm of this invention, FedAvg algorithm, DPSGD algorithm, Choco algorithm, and GossipFL algorithm.

[0025] in, Figure 7 (a) is a bar chart showing the communication overhead for achieving the target accuracy on CNN using the MNIST dataset. Figure 7 (b) is a bar chart showing the communication overhead for achieving the target accuracy on the CIFAR-10 dataset using CNN. Detailed Implementation

[0026] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0027] The method illustration of the embodiment is shown in the figure below. Figure 1 As shown, the specific implementation steps include:

[0028] Step 1: Data Sampling

[0029] The MNIST and CIFAR10 datasets were selected. MNIST contains 70,000 grayscale handwritten digit images (28×28 pixels) divided into 10 classes, with 60,000 images used for training and 10,000 for testing. CIFAR10 contains 60,000 color images (32×32 pixels), also divided into 10 classes, with 50,000 images used for training and 10,000 for testing. Except for the case where model performance was evaluated on non-IID data, all datasets were uniformly distributed among clients by default.

[0030] Step 2: Establish a decentralized federated learning framework

[0031] Federated learning frameworks can train on large, distributed datasets collected on clients. Decentralization means that each client uses its local data to train a copy of its local model parameters. Decentralized federated learning consists of a set of distributed clients. Each client They all have their own local datasets D iIn the absence of a parameter server, clients collaboratively optimize ML models using local datasets. In this invention, each client only needs to exchange models with one neighboring client and does not need to share its original data. Formally, the loss function f for client i... i (x) can be represented as:

[0032]

[0033] Where x is a local copy of the model parameters, D i Let be the dataset of client i, and f(x; ξ) be the loss function computed for a specific sample ξ. In the decentralized federated learning framework, each client trains a local copy of its model x using its local dataset, and then exchanges model parameters with a selected neighbor client to ensure that the two clients jointly train the same model.

[0034] Step 3: Calculate and transmit information

[0035] The algorithm includes a coordinator, primarily used to calculate and transmit global information needed during training, such as the gossip matrix. Initially, the coordinator initializes the connection with the client. After the client begins training, the coordinator enters an iterative loop. Each iteration includes the following three steps:

[0036] 1) The coordinator aggregates lightweight information from clients, such as bandwidth, data volume, and computing power;

[0037] 2) The coordinator generates the octagonal matrix G. k This is used to determine the communication topology in the next iteration and to save the latest timestamp matrix T. k ;

[0038] 3) The coordinator broadcasts G to all clients. k The coordinator then notifies the client to proceed with the next iteration. Compared to model parameters that can number in the tens or hundreds of millions of bytes, the information aggregated by the coordinator is extremely concise, less than a thousand bytes. Therefore, communication between the client and the coordinator does not create any communication bottlenecks and can be easily scaled.

[0039] Step 4: Select Client

[0040] To reduce communication overhead, this invention ensures that each client communicates with only one neighboring client. When selecting a client, the coordinator comprehensively considers the network resources of different clients, including bandwidth, local data volume, and computing power. This time-varying information is processed by the coordinator in each iteration, transmitted to each client, and guides the client to dynamically select neighbors for communication.

[0041] Because various resource indicators differ significantly, using the entropy weighting method in the coordinator to process information allows for the objective determination of the weights of different indicators, ultimately yielding a comprehensive evaluation indicator γ. Specifically, this involves considering the bandwidth B of client i's neighbor, client j. ij For example, let L be the number of neighboring clients of client i. First, through... Calculate the sample weight, where This represents the proportion of bandwidth samples from neighboring client j of client i, reflecting the relative importance of the metric. Then, based on... Calculate the information entropy of the bandwidth metric for client i. A higher information entropy indicates greater variability in this metric and a greater contribution to the overall evaluation. The coefficient μ is typically taken as... The information entropy of the local data volume V and computing power P of client i are also calculated in the same way as the bandwidth B, i.e. *={B,V,P}. The next step is to determine the weights of each indicator. Indicators with greater differences have higher weights, among which Σ 3 (1-entropy * ) represents the (1-entropy) of three indicators. * The summation is performed on {B, V, P}. Finally, the comprehensive evaluation γ of client i is calculated as follows:

[0042] γ=γ B ω B +γ V ω V +γ P ω P

[0043] Where γ * ω represents the sample proportion. * This represents the weight of the indicators. Combining the entropy weight method allows for the most objective evaluation of clients based on three indicators: bandwidth, local data volume, and computing power. This maximizes network resource utilization and further improves convergence speed. Finally, the neighboring client with the largest overall evaluation γ is selected as the client for communication in this round.

[0044] Step 5: Train the model

[0045] Decentralized federated learning consists of a set of distributed clients. Each client They all share the same consensus algorithm, and the training objective of the model is to minimize the loss function. Formally, the loss function for client i can be expressed as: Where x is a local copy of the model parameters, D iLet f(x; ξ) be the dataset of client i, and f(x; ξ) be the loss function computed for a specific sample ξ. FL seeks to minimize the global loss function F(x), transforming it into an optimization problem. Where S is the number of clients participating in model training, and D i D and D represent the local dataset and global dataset of client i, respectively.

[0046] Specifically, in each iteration, client i completes one round of training locally to obtain a local model. Client i receives and utilizes the model parameters of its neighbor client j in the current round (kth iteration). Complete local model update Client i first uses the Bagua matrix G k Select neighbor client j in this round of communication (the k-th iteration) to receive the neighbor model after parameter filtering. And based on G k Perform a weighted average:

[0047]

[0048] Among them G k,ij Represents the octagonal matrix G k The element values ​​corresponding to client i and client j, when client j is the selected neighbor client in this round, G k,ij If it is 1, then G is not. k,ij The value is 0. Since only one neighboring client is selected for communication in each iteration, only client i is in the octagonal matrix G. k The corresponding i-th row has only one element with a value of 1, and the rest are 0.

[0049] Then, client i updates its local parameters and obtains the results in the (k+1)th training iteration.

[0050]

[0051] Where λ is the learning rate. For local models Corresponding loss function The gradient. In the next (k+1)th iteration, the model... The parameters are transmitted.

[0052] Step 6: Set dynamic threshold

[0053] In the model training process shown in step 5, many parameters do not change or only change slightly in a certain iteration. When the parameter change is too small, its impact on model convergence is negligible, and these parameters do not need to be sent to neighbors. As the number of algorithm iterations increases, the number of parameters that need to be sent to neighboring clients decreases, which can significantly save communication costs. The key to reasonably controlling the number of parameters sent between clients is to set a parameter change threshold. Only parameters whose change is higher than this threshold are sent; those below the threshold are not sent. If clients i and j have some updated parameters in a certain iteration, but the change does not reach the threshold, the values ​​of these parameters can be updated locally and used in the next iteration.

[0054] First, determine the cumulative parameter error APE, using R... k This means that the update parameters for all missing neighbors are limited to no more than APE. The iteration parameter error in each iteration is called IPE, denoted by r. k This indicates that, in the k-th iteration, client i sends... Right now Instead but In the k-th iteration, the APE of client i is:

[0055]

[0056] The model iteration process for client i can be represented as:

[0057]

[0058] Therefore, it is possible to find for:

[0059]

[0060] when When the quantity is sufficiently small, the above equation can be approximated using derivatives:

[0061]

[0062] Suppose that the second gradient of the loss function corresponding to the local model parameters of client i has a bound of U, i.e. Then we have:

[0063]

[0064] The maximum value of APE can then be obtained as:

[0065]

[0066] The threshold for APE after k iterations is R. kThe change during training is less than The parameters do not need to be sent; this expression is the dynamic threshold.

[0067] Step 7: Exchange model parameters

[0068] Based on the dynamic threshold limit set in step 6, each client filters the parameters that are allowed to be transmitted, and then passes the parameters that meet the threshold limit to the neighboring client selected in step 4, ensuring that the same model can be trained in the end.

[0069] The effectiveness of this invention is further illustrated by the following simulations and experiments:

[0070] Experimental results:

[0071] This invention uses convolutional neural networks (CNNs) with different configurations trained on different datasets: a CNN model applied to the MNIST dataset, a CNN model for the CIFAR10 dataset, and a ResNet-20 model for CIFAR10. The batch sizes for MNIST-CNN, CIFAR10-CNN, and ResNet-20 are set to 32, 50, and 16, respectively, and the number of epochs are set to 100, 300, and 150, respectively, to ensure model convergence. Furthermore, the initial learning rates for MNIST-CNN, CIFAR10-CNN, and ResNet-20 are 0.05, 0.04, and 0.1, respectively, and all decay at a rate of 0.992.

[0072] In the experiments of this invention, four classic algorithms were selected as performance comparison baselines: FedAvg, D-PSGD, ChocoPSGD, and GossipFL. Among them, FedAvg is a popular centralized algorithm. For decentralized algorithms, D-PSGD and ChocoPSGD (compression ratio c=100) were selected, as well as SNAP, which also uses a parameter selection scheme. GossipFL, currently the most advanced algorithm, was also selected.

[0073] In this invention, the following three metrics are used to evaluate algorithm performance: 1) test accuracy; 2) convergence speed; and 3) traffic consumption. Specifically, in each training round, the client calculates the accuracy of its local model on the test dataset and reports it to the coordinator for averaging. The final average test accuracy of all models is then used as the overall test accuracy. To more fully demonstrate the performance comparison of the algorithms, this invention compares the epochs and traffic consumption required by different algorithms to achieve the same target test accuracy.

[0074] First, the top-1 accuracy of the algorithm of this invention and the baseline were compared on IID and non-IID datasets. The three evaluation models trained on the IID dataset by the algorithm of this invention all achieved the same level of accuracy as the state-of-the-art algorithms, such as... Figure 2 As shown, the algorithm of this invention achieved top-1 accuracies of 99.33%, 82.57%, and 90.61% respectively during the training of the three evaluation models, verifying the convergence of the algorithm. Furthermore, the algorithm was also validated on a non-IID dataset. Real-world training data often exhibits non-IID data distributions, which typically make training more difficult. Figure 5 As shown, in the non-IID dataset, the MNIST dataset has a lower training difficulty, and most algorithms only suffer slight accuracy loss. Among them, the algorithm of this invention suffers the least loss, achieving almost the same accuracy. However, almost all algorithms exhibit significant accuracy loss on the CIFAR10 dataset, which has non-IID characteristics, but the algorithm of this invention still achieves higher accuracy in this scenario. The comprehensive experimental results demonstrate that the algorithm of this invention achieves the same level of convergence performance on IID data and achieves better convergence performance on non-IID data.

[0075] We selected the highest achievable accuracy among all algorithms as the target accuracy. We then compared the number of epochs required to reach the same accuracy; fewer epochs indicate faster convergence. In the IID dataset, we set the target accuracy for MNIST-CNN to 98%, and for CIFAR10-CNN and ResNet-20, to 80%. The algorithm performance comparison is as follows: Figure 2 As shown, the algorithm of this invention reaches the target accuracy faster than other algorithms. In the training of the MNIST-CNN evaluation model, all algorithms except FedAvg reach the target accuracy extremely quickly, with the algorithm of this invention reaching it almost simultaneously with the fastest, Choco-PSGD. In the more challenging CIFAR10-CNN and ResNet-20 datasets, the algorithm of this invention achieves the same accuracy as other algorithms while improving training speed by 11.7% and 29.6%, respectively. Performance comparisons of the algorithm on the non-IID dataset are also provided. Figure 6 As shown, the algorithm of this invention achieved the fastest convergence speed in the training of all evaluated models. The target accuracy of MNIST-CNN and CIFAR10-CNN was 95% and 40%, respectively. Choco-PSGD failed to converge, DPSGD also faced significant convergence difficulties, and FedAvg converged but performed poorly. Compared to GossipFL, the best-performing baseline, the algorithm of this invention improved convergence speed by up to 18.1%. Experimental results demonstrate the advantages of the algorithm of this invention in accelerating model training.

[0076] We compared the communication overhead of different algorithms to demonstrate the communication efficiency of the algorithm in this invention. Figure 3 and Figure 6 The accuracy of each algorithm relative to cumulative communication overhead is shown. It can be seen that the algorithm of this invention achieves the same accuracy on both IID and non-IID datasets while reducing traffic consumption. Figure 4 and Figure 7 This paper more intuitively demonstrates the comparison of communication overhead required to achieve the same accuracy, showing that the algorithm of this invention has an advantage in communication overhead across various datasets. The algorithm of this invention outperforms the state-of-the-art GossipFL algorithm, especially in ResNet-20 training, where communication overhead is reduced by 12.2%. Compared to Choco-PSGD, which performs well in compressed models, the algorithm of this invention saves up to 56.9% of communication traffic. Compared to FedAvg and DPSGD, algorithms susceptible to high communication traffic, the algorithm of this invention shows a significant improvement, with communication overhead differing by two orders of magnitude. The algorithm of this invention also performs excellently on non-IID datasets. With target accuracies of 95% for MNIST-CNN and 30% for CIFAR10-CNN, the algorithm of this invention saves 15.5% of traffic to achieve the same accuracy as GossipFL. FedAvg, on the other hand, requires approximately 10 times the communication overhead of the algorithm of this invention, and DPSGD and Choco-PSGD even struggle to achieve the target accuracy. In summary, the algorithm of this invention requires less bandwidth to achieve the same level of accuracy as other algorithms, and has a significant advantage in communication efficiency.

[0077] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. An adaptive communication federated learning method suitable for dynamic network environments, characterized in that, Including the following steps: The steps to establish a decentralized federated learning framework are as follows: adopt a network topology without a central server, each client in the network hosts a copy of the global model parameters and uses local data for training, while in each round of training, only one neighboring client is selected to exchange model parameters. Coordinator configuration steps: The coordinator makes a comprehensive assessment of the network resources of each client. Before the client starts a round of model parameter iteration, it makes a comprehensive assessment of the network resources of each client, selects a neighboring client as the communication object for the next round of model parameter iteration, and broadcasts the completed communication topology to each client. The coordinator collects the bandwidth, local data volume and computing power of each client to make a comprehensive evaluation of the network resources of each client, and selects the neighbor client with the largest comprehensive evaluation value as the neighbor client for the next round of exchange model parameters. Model training steps: Each client iterates the model parameters in each round according to the consensus optimization algorithm; The model parameter exchange steps based on dynamic thresholds are as follows: After each client completes a round of model parameter iteration, it determines whether the change in the model parameters in this round reaches the set parameter change threshold. If so, the parameters that meet the change threshold are sent to the selected neighboring clients. If not, the parameters that do not meet the change threshold are not sent, and the client starts the next round of model parameter iteration. The parameter change threshold is a dynamic threshold; the parameter change threshold is... ;in, This represents the current iteration number. For the first Learning rate for each iteration To limit The threshold of cumulative parameter error after iterations This is the bound for the existence of the second gradient of the loss function corresponding to the local model parameters on the client side.

2. The method as described in claim 1, characterized in that, A comprehensive assessment of the network resources of each client was conducted. The calculation is as follows: ； Among them, indicator code , Indicates bandwidth. Indicates the amount of local data. Indicates computing power. Indicators The proportion of the sample Indicators The weight of the indicators.

3. The method as described in claim 2, characterized in that, Client Indicators sample proportion Represented as Client Indicator weights Among them, the client Indicators Information entropy ; Indicates to the client The three indicators Accumulate; Indicates the client Neighbor client Metrics, client The number of neighbor clients is , is a coefficient.

4. The method as described in claim 3, characterized in that, coefficient Pick .

Images