Federated Learning (FL) has emerged as a groundbreaking distributed machine learning paradigm that enables collaborative model training while preserving data privacy. This comprehensive review examines FL's evolution from its inception to current state-of-the-art approaches, addressing both theoretical foundations and practical applications. We analyze the core FL framework, highlighting its advantages over centralized learning in terms of privacy preservation, reduced communication overhead, and edge computing capabilities. The paper explores key algorithmic advancements including Federated Averaging (FedAvg) and its variants (FedProx, SCAFFOLD), which tackle challenges like data heterogeneity and client drift. We discuss FL's transformative applications across healthcare, finance, and IoT domains, where data privacy is paramount. Major challenges are critically examined, including communication bottlenecks, straggler effects, security vulnerabilities, and the complexities of non-IID data distributions. The review evaluates privacy-enhancing technologies such as differential privacy and homomorphic encryption, analyzing their trade-offs between privacy guarantees and model performance. Looking forward, we identify promising research directions: adaptive personalization techniques, integration with large language models, blockchain-assisted security frameworks, and standardization efforts for broader adoption. Ethical considerations and regulatory compliance aspects are also addressed, providing a holistic perspective on FL's role in shaping responsible AI development. This review serves as both a technical reference and a roadmap for future innovation in federated learning systems.
Wajahat et al. (Mon,) studied this question.