Exploring privacy mechanisms and metrics in federated learning

Abstract The federated learning (FL) principle ensures multiple clients jointly develop a machine learning model without exchanging their local data. Various government enacted prohibition policies on data exchange between organizations have led to the need for privacy-preserved federated learning. Many industries have cultivated this idea of model development through federated learning to enhance performance and accuracy. This paper offers a detailed overview of the background of FL, highlighting existing aggregation algorithms, frameworks, implementation aspects, and dataset repositories, establishing itself as an essential reference for researchers in the field. The paper thoroughly reviews existing centralized and decentralized FL approaches proposed in the literature and gives an overview about the methodology, privacy techniques implemented and limitations to guide other researchers to advance their research in the field of federated learning. The paper discusses the critical role of privacy-enhancing technologies like differential privacy (DP), homomorphic encryption (HE), and secure multiparty computation (SMPC) in federated learning highlighting their effectiveness in safeguarding sensitive data while optimizing the balance between privacy, communication efficiency, and computational cost. The paper explores the applications of federated learning in privacy-sensitive areas like natural language processing (NLP), healthcare, and Internet of Things (IoT) with edge computing. We believe our work provides a novel addition by identifying privacy evaluation metrics and spotlighting the measures in terms of data privacy and correctness, communication cost, computational cost and scalability. Furthermore, it identifies emerging challenges and suggests promising research directions in the federated learning domain.