Paper 2026/416

An Ultra-Robust Privacy Preserving Scheme for Federated Learning using Distributed Homomorphic Encryption

Abstract

Federated Learning is an emerging machine learning paradigm that enables distributed model training directly at data sources and transmitting only model updates, thereby reducing communication bottlenecks and mitigating risks associated with raw data exposure. Despite these advantages, recent advances have demonstrated that privacy in federated learning remains limited and subject to inference attacks that exploit shared model updates to extract sensitive information. To address this limitation, we propose Robust and Resilient Federated Learning using Distributed Homomorphic Encryption (RRFL-DHE), a privacy-preserving federated learning framework that combines a distributed homomorphic encryption scheme with threshold linear secret sharing. The framework enables clients to encrypt their model updates to allow secure aggregation without exposing individual contributions. To maintain resilience against client dropouts, RRFL-DHE incorporates a dropout management protocol, maintaining training continuity and accurate global model reconstruction. To assess our framework, we provide a rigorous security proof against a semi-honest server model and evaluate RRFL-DHE on non-IID MNIST and Fashion MNIST datasets using SVM and CNN models. The results show that RRFL-DHE preserves model utility with less than 1% deviation compared to the FedAvg approach, while outperforming the xMK-CKKS approach by approximately 15% in accuracy. These findings highlight the importance of RRFL-DHE as a promising solution for distributed computing, while preserving privacy, maintaining utility, and ensuring resilience against dropouts.