Research On Communication Frequency Optimization Of Cloud-Edge-Device Collaborative Federated Learning

As an emerging distributed machine learning technique,federated learning enables the training of highly accurate machine learning models without sharing private data.It has experienced rapid development in the field of machine learning in recent years.The core idea of federated learning is to perform distributed model training among multiple end nodes that possess local datasets.By exchanging model gradients or parameters without sharing local training data,a balance is achieved between data sharing for model training and data privacy protection.However,in federated learning,the end nodes need to communicate frequently with the cloud server for model updates and iterations,resulting in high communication overhead.To reduce this overhead,cloud-edge-device collaborative federated learning based on a dual-layer parameter server architecture has emerged.It utilizes edge servers to perform local aggregation of model parameters,thereby reducing the communication overhead between the end and the cloud.Nevertheless,the local iteration frequency of end nodes and the model aggregation frequency of edge servers not only affect the communication overhead but also impact the convergence speed and accuracy of the models.These factors are crucial for efficient model training in cloud-edge-device collaborative federated learning.Therefore,this thesis focuses on optimizing the communication frequency of cloud-edge-device collaborative federated learning in two typical scenarios: resource homogeneity and multi-dimensional resource heterogeneity.Considering the importance of timely completion of high-quality model training,this thesis primarily explores how to reduce communication overhead and improve model training accuracy by optimizing the communication frequency within time constraints.The specific work is as follows.First,for the optimization of communication frequency in the resource homogeneity scenario of cloud-edge-device collaborative federated learning,this thesis investigates the relationship between the global average loss function and the communication frequency.It proposes a heuristic-based communication frequency optimization algorithm to achieve effective improvement in model accuracy within the given training time.The core idea is to greedily select the communication frequency for the end and the edge nodes in each global training round,aiming to minimize the global average loss function.To accomplish this,the algorithm quantifies the impact of any given set of communication frequency values on the global average loss function by analyzing the differences in average loss functions across the cloud,edge,and device.It then calculates a set of communication frequency values with the lowest global average loss function from a series of possible communication frequency settings.Experimental comparisons show that under the condition of non-independent and identically distributed datasets,the set of communication frequency values with the smallest upper bound of the average loss function achieves higher convergence accuracy.Moreover,these communication frequency values achieve an accuracy improvement of up to 8.42% in the logistic regression(LR)model and up to 6.96% in the convolutional neural network(CNN)model compared to other values.Next,for the optimization of communication frequency in the multi-dimensional resource heterogeneity scenario of cloud-edge-device collaborative federated learning,this thesis considers the reality that end nodes are typically energy-constrained.It proposes a joint optimization algorithm for energy allocation and communication frequency control to improve model training accuracy under the constraints of training time and energy consumption.The core idea is to implement differentiated energy allocation and communication frequency control strategies based on the differences in the contribution of different end nodes to model convergence.More energy is allocated to edge nodes with higher contribution for more computation and communication,thereby promoting faster model iteration and convergence.To achieve this,the algorithm quantifies the energy consumption and contribution of each edge node and determines the energy allocation for end nodes as well as the communication frequency between the edge-device and edge-cloud based on the comprehensive consideration of computational energy consumption and communication time.Experimental comparisons show that under the condition of non-independent and identically distributed datasets,the proposed algorithm achieves the highest accuracy improvement of up to 11.64% compared to existing algorithms.Furthermore,while maintaining the same model accuracy,the proposed algorithm reduces energy consumption by up to 14.33% compared to existing algorithms.