A Survey on Federated Learning Algorithms in IoT networks

The capability of mobile devices to sense and compute at an advanced level has significantly improved in recent years, and combined with the advancements in Deep Learning (DL) has led to a vast range of opportunities for different practical applications such as healthcare and vehicular networks. However, traditional cloud-based Machine Learning (ML) approaches require the data to be stored in a cloud server, resulting in critical issues such as poor communication efficiency, high latency and privacy concerns. The idea to bring intelligence closer to the edge was proposed. However, the conventional technologies that enable Machine Learning (ML) at edge networks require sharing of personal data, which defies strict data privacy regulations. In response, Federated Learning (FL) has been introduced. In FL, devices use their local data to train a ML model in a decentralized way. Instead of sharing raw data, devices iteratively optimize the model locally. Then, the resulting model updates are shared to a centralized server for aggregation. However, in a complex mobile edge network that includes various devices with different constraints, some challenges may arise. Implementing Federated Learning (FL) on a large scale presents difficulties with regards to communication costs, resource allocation, as well as privacy and security concerns. To address these challenges, this survey paper provides an overview of the fundamentals and background of FL and its implementation challenges. We focus on reviewing client selection frameworks for Federated Learning (FL), then we shed light on state-of-the-art techniques that tackle the issue of data heterogeneity. Lastly, we examine the application of FL in the context of IoT networks for connected vehicles.