Recent advances in privacy-preserving federated learning for financial applications

Financial institutions increasingly rely on machine learning (ML) for fraud detection, credit scoring, anti-money laundering, and personalised services, yet regulatory constraints and competitive data silos limit centralised model training. This study presents a systematic review of recent advances (2024–2026) in privacy-preserving federated learning (FL) for financial applications, addressing the growing tension between collaborative intelligence and regulatory data constraints. This review synthesises architectural paradigms (cross-device, cross-silo, and hybrid FL), optimisation strategies (FedAvg, FedProx, personalisation, and knowledge transfer), and security foundations including differential privacy (DP), homomorphic encryption (HE), and secure multi-party computation (SMPC). We analyse empirical evidence on non-independent and identically distributed (non-IID) robustness, privacy-utility trade-offs, encrypted aggregation overhead, and performance deviation from centralised baselines. Results indicate that properly engineered federated systems can achieve near-centralised predictive performance, though heterogeneity, communication cost, and governance complexity remain critical constraints. The study further identifies benchmarking inconsistencies, limited real-world cross-institution deployments, and insufficient standardisation of evaluation metrics. By integrating architectural, algorithmic, and cryptographic perspectives, this review provides a structured foundation for advancing secure, scalable, and regulation-compliant federated intelligence in modern financial ecosystems.