Systematic Review Of Privacy-Preserving Federated Learning: Methods, Architectures, Performance, And Challenges

Abstract

Federated learning enables collaborative model training without centralizing raw data, but model updates, gradients, communication metadata, and network topology can still expose sensitive information. This article presents a systematic literature review of 281 studies published between 2021 and 2026 on privacy-preserving federated learning. The review followed the PRISMA 2020 flow diagram for systematic reviews and applied title and abstract screening, full-text eligibility assessment, quality assessment, data extraction, and thematic synthesis. Extracted data covered privacy methods, federated learning architectures, performance outcomes, privacy analysis, limitations, open challenges, application contexts, and study design. The findings show that differential privacy is the most frequently reported approach, followed by homomorphic encryption, secure aggregation, secure multi-party computation, blockchain, trusted execution environments, and communication-efficient privacy techniques. The synthesis indicates that stronger privacy protection often introduces trade-offs in accuracy, utility, latency, computational overhead, communication cost, scalability, and fairness. Persistent challenges include privacy-utility trade-offs, non-IID data, topology-based leakage, attack resistance, heterogeneous clients, and inconsistent evaluation standards. The article contributes an integrated map of methods, architectures, empirical outcomes, and implementation challenges to guide future privacy-preserving federated learning research.