Federated Learning (FL) has emerged as a privacy-preserving distributed learning paradigm that enables collaborative model training without sharing raw data. Despite its advantages, ensuring trustworthy FL deployment remains challenging due to privacy leakage risks, adversarial attacks, communication overhead, system heterogeneity, and Non- Independent and Identically Distributed (non-IID) data distributions. This review provides a comprehensive analysis of recent advancements in FL security mechanisms, including Differential Privacy (DP), secure multiparty computation, Homomorphic Encryption (HE), Byzantine-robust aggregation, blockchain-based trust integration, model compression, and post-quantum cryptographic approaches within centralized client–server architectures. The study systematically categorizes defense strategies based on their functional objectives and introduces a structured threat taxonomy to clarify attacker models and vulnerabilities. While existing works present layered security frameworks and extensive attack–defense taxonomies, most rely heavily on gradient-based optimization (e.g., Federated Averaging (FedAvg)) and lack large-scale empirical validation under realistic deployment conditions. Moreover, standardized benchmarking and multi-objective evaluation across privacy, robustness, scalability, and computational cost remain limited. This review identifies critical research gaps and emphasizes the need for integrated, deployment-aware, and empirically validated frameworks to support secure, scalable, and practical FL systems in real-world environments.
Krishna et al. (Thu,) studied this question.