Abstract
The combination of artificial intelligence (AI) and mobile edge computing (MEC) is considered as a promising evolution path of the future wireless networks. As a modellevel coordination learning paradigm, federated learning can make full use of the distributed computation resource in the MEC systems, which allows the users to keep their private data locally. Though federated learning provides a promising paradigm to enable network edge intelligence in the future sixth generation systems, there are still some important issues that need to be addressed. One of the main challenges is that both the performance and training efficiency of federated learning cannot be guaranteed due to the unreliable wireless transmission circumstance and resource constraints in the MEC systems. To solve this problem, we study the optimization design of federated learning with unreliable transmissions in the MEC systems at first. In particular, we formulate an optimization problem to manage the tradeoff between model accuracy and cost. Moreover, we design a joint optimization algorithm to optimize the model compression, sample selection, and user selection strategies, which can approach the stationary optimal solution in a computationally efficient way. Besides theoretical analysis, we also provide numerical simulation and experiment results to evaluate the performance of our proposed ptimization scheme, which show that both the accuracy loss and model cost of federated learning in the MEC systems can be reduced significantly by employing our proposed algorithm. The second challenge is that due to the high dynamics of wireless circumstances and user behavior, the collected training data is non-independent and identically distributed (non-IID), which causes severe performance degradation of federated learning. To solve this problem, we further study federated learning with non-IID data in wireless networks. In particular, we firstly derive an upper bound of expected weight divergence, and propose a novel federated averaging scheme to reduce the data distribution divergence of non-IID data. Secondly, we introduce a data sharing scheme to further harmonize the distribution divergence, and design a joint optimization algorithm to keep a sophisticated balance between the model accuracy and the cost. Finally, we also provide the simulation results based on a common-used image data set to evaluate the performance of our proposed schemes, which can achieve significant performance gains with a small price of latency and energy consumption. Few works have considered the impact of user mobility on learning performance, which is a critical factor in wireless networks. To fill this research gap, firstly, we develop a theoretical model to characterize the hierarchical federated learning (HFL) algorithm with mobile users. In our model, the network consists of multiple mobile users that have their positions vary over time, and each user is connected with one edge server, while all the edge servers can further exchange information with one cloud server. All the mobile users will download the latest shared model from this assigned edge server and then perform the local training. During the local training process, they may stay in or move out the coverage area of the assigned edge server. Secondly, we provide the convergence analysis of HFL with user mobility. Based on this analysis, we propose a mobility-aware cluster federated learning (MACFL) model by redesigning the local update and model aggregation schemes. Finally, we provide experiments to v evaluate the learning performance of HFL with user mobility under different settings of network parameters, and compare with our proposed algorithm. The results show that our proposed MACFL can enhance the learning performance, especially for three different cases: the case of users with non-IID data, the case of users with high mobility, and the cases with a small number of users. Moreover, our results provide useful insight to select the values of key parameters in the hierarchical federated learning algorithm in MEC systems.