Byzantine-Robust Federated Learning with Verifiable Secure Aggregation for Multi-Party Healthcare Collaborations

Federated learning (FL) has become a potential way for computer scientists from different healthcare institutions to work together on machine learning projects while keeping data private. But the fact that there are "Byzantine" participants—those who might do bad things or send tainted updates—makes it hard for shared healthcare models to be honest and work together. This paper presents a novel Byzantine-robust federated learning framework with verifiable secure aggregation specifically designed for multi-party healthcare collaborations. Our approach combines differential privacy mechanisms, cryptographic secure aggregation protocols, and robust aggregation algorithms to ensure both privacy preservation and resilience against Byzantine attacks. We introduce a verifiable computation scheme that allows participants to validate the integrity of aggregated updates without compromising individual contributions. Experimental results on three healthcare datasets demonstrate that our framework achieves superior robustness against various Byzantine attack scenarios while maintaining comparable accuracy to non-Byzantine federated learning approaches. The proposed system reduces the impact of Byzantine participants by up to 89% compared to baseline federated averaging, while preserving differential privacy with ε = 1.2. Our framework provides a practical solution for secure and robust collaborative learning in healthcare environments where trust assumptions cannot be guaranteed.