Federated Learning Approaches for Anomaly Detection in IoT Environments: A Systematic Literature Review

Abstract

The rapid proliferation of Internet of Things (IoT) devices is driving significant data generation—and, consequently, a crucial demand for reliable, privacy-centric anomaly detection (AD). Federated Learning (FL) enables collaborative model training without centralizing sensitive data, addressing fundamental constraints encountered by conventional, server-oriented approaches. Adhering to the PRISMA 2020 framework, we rigorously examined 60 well-documented studies (2017–2026) on FL-oriented AD in IoT. We scrutinize architectures, learning paradigms, datasets, applications, privacy mechanisms, communication efficacy, and evaluation metrics, and systematically classify the domain into a taxonomy encompassing architectural frameworks, privacy-preserving strategies, communication protocols, and detection algorithms. In scholarly publications, FL methodologies frequently align with centralized references while preserving data confidentiality, yielding detection accuracy metrics ranging from 80% to 99.89%, with the majority of approaches exceeding 90%. Simultaneously, unresolved challenges persist: non-IID and heterogeneous data, communication and energy expenditures, resilience under adversarial circumstances, and real-world scalability. Emerging hybrid methodologies that amalgamate FL with complementary techniques (e.g., model compression, secure aggregation, and multi-modal fusion) demonstrate promise. This review delves into the contemporary landscape, exposes inadequacies, and charts explicit research directions to enhance the security and reliability of IoT environments.