TITLE:
Evaluating Data-Protection Mechanisms in Federated Learning: A Performance-Centric Analysis
AUTHORS:
Mitende Nicholus Nyapete, Richard Omolo, Newton Masinde
KEYWORDS:
Differential Privacy, Homomorphic Encryption, Secure Multi-Party Computation, Secure Aggregation, Federated Averaging, Model Distillation, Data Anonymization
JOURNAL NAME:
Journal of Information Security,
Vol.17 No.4,
August
14,
2026
ABSTRACT: Background: Federated learning enables multiple participants to train ML models collaboratively without exposing or sharing raw data, making FL attractive approach for privacy-sensible applications. To reduce residual leakage risks in shared model updates, studies have proposed several data-protection mechanisms, including differential privacy, homomorphic encryption, secure multi-party computation, secure aggregation protocols, federated averaging with secure aggregation, model distillation, regularization, and client-side anonymization. While these mechanisms are regularly evaluated for the strength of guarantees they provide, their performance implications: computational overhead, communication cost, model accuracy, and scalability are comparatively under-examined as a harmonized basis for practical selection. Method: The study conducted a systematic comparative analysis of eight data-protection mechanisms in federated learning, evaluating them using performance metrics. Drawing on a structured analysis of the most recent innovations, the study developed a comparison framework, applied it consistently across mechanisms, and presented the resulting trade-offs in tabular and narrative form. Results: The study found that lightweight mechanisms: secure aggregation, model distillation, and regularization techniques offer the most favourable performance profiles for resource-constrained, large-scale deployments, while the cryptographic approach: homomorphic encryption and secure multi-party computation impose substantial computational and communication burdens that limit technique scalability. Conclusion: The study concludes that, with practical guidance for mechanism selection based on deployment context, it outlines directions for future performance-oriented research in federated learning.