Federated Learning: Challenges, SoTA, Performance Improvements and Application Domains

Abstract

Federated Learning has emerged as a revolutionary technology in Machine Learning (ML), enabling collaborative training of models in a distributed environment while ensuring privacy and security. This work discusses the topic of FL by providing insights into its various dimensions, perspectives, and components, leading to a comprehensive understanding of the technology. The survey begins by introducing the basic principles of FL and provides a high-level taxonomy of its methods. It continues by presenting application domains and associating challenges, categories and their applications. This mapping allows for an understanding of how particular challenges manifest in different contexts and applications. The main body delves into the various aspects of FL, including centralized and decentralized variants, methods for improving efficiency and effectiveness, and concerns regarding security, privacy, dynamic conditions, fairness, scalability and integration with other new technologies. Ultimately, the goal is to present recent advancements in these areas, along with new challenges and opportunities for future exploration. FL is poised to reshape the landscape of intelligent systems while promoting data privacy in decentralized and collaborative learning. Finally, this survey can serve as a reference point for methodological improvements as it highlights the strengths and weaknesses of existing approaches.