Game Theory for Security, Privacy, and Trustworthiness in Federated Learning: A Systematization of Knowledge
Abstract
Rather than aggregating private datasets at a central server, Federated Learning (FL) facilitates localized, collaborative Machine Learning (ML) training, thereby reducing communication overhead and maintaining strict data privacy. Due to these structural advantages, FL has been broadly integrated into fields such as healthcare, edge computing, smart city infrastructure, vehicular networks, and the Internet of Things (IoT). However, its decentralized design raises security, privacy, and resource challenges, enabling attacks such as inference attacks, model poisoning, free-riding, and data poisoning. Game Theory (GT) addresses these by modeling competitive and cooperative agent behaviors, supporting client incentive design, resource allocation, profit maximization, threat detection, and robust data privacy management. In this paper, we present a Systematization of Knowledge (SoK) of over recent studies on FL and GT based models, application domains, and security objectives such as adversarial threats, privacy attacks, and malicious behaviors, while highlighting their key contributions, limitations, and future research directions over the last three years.