Federated learning for edge computing: A survey

Abstract

New technologies bring opportunities to deploy AI and machine learning to the edge of the network, allowing edge devices to train simple models that can then be deployed in practice. Federated learning (FL) is a distributed machine learning technique to create a global model by learning from multiple decentralized edge clients. Although FL methods offer several advantages, including scalability and data privacy, they also introduce some risks and drawbacks in terms of computational complexity in the case of heterogeneous devices. Internet of Things (IoT) devices may have limited computing resources, poorer connection quality, or may use different operating systems. This paper provides an overview of the methods used in FL with a focus on edge devices with limited computational resources. This paper also presents FL frameworks that are currently popular and that provide communication between clients and servers. In this context, various topics are described, which include contributions and trends in the literature. This includes basic models and designs of system architecture, possibilities of application in practice, privacy and security, and resource management. Challenges related to the computational requirements of edge devices such as hardware heterogeneity, communication overload or limited resources of devices are discussed.