Abstract
In light of heightened data privacy concerns and the enhanced capabilities of edge devices, the concept of federated learning (FL) has emerged as a new distributed machine learning paradigm. This approach facilitates model training while circumventing the need to transfer data to a centralized server, thereby preserving privacy and improving communication efficiency. However, implementing FL at the network edge is challenging due to storage limitations and resource constraints. Additionally, the streaming nature of data introduces increased complexity to FL frameworks. The continual influx of streaming data underscores the necessity of considering FL within the context of a long-term process. Therefore it is essentially to design some online algorithm to deal with the problems faced in long-term FL process. This dissertation presents a comprehensive analysis of the unique challenges associated with online optimization within FL frameworks when dealing with streaming data. The study delineates three distinct issues encountered in FL within a streaming data setting, focusing on resource allocation, data distribution, and learning objectives.