Federated Learning Governance using Eclipse Dataspace Components Connectors
Angela Mitrovska, Marcel Fernandez Rosas, Pooyan Safari, Behnam Shariati, Johannes Fischer, Ronald E. Freund · 2024
Federated Learning (FL) has emerged as a promising privacy-preserving approach for training Machine Learning (ML) models in industries handling sensitive data, such as healthcare. However, the practical application of FL in healthcare faces challenges due to complex privacy regulations and the lack of inherent robust governance mechanisms to ensure compliance and protect data ownership. To address these limitations, we propose a governed FL framework that leverages the Eclipse Dataspace Components (EDC) connector to enforce compliance and data sovereignty through the implementation of data access and usage policies. We demonstrate the effectiveness of our approach by training an in-hospital mortality prediction model and an Alzheimer’s Disease (AD) detection model using FL with Federated Averaging (FedAvg) and Secure Aggregation (SecAgg). Our results demonstrate that the governed FL framework enables the training of models with performance comparable to those trained on centrally aggregated data. At the same time, it effectively introduces governance through policy enforcement, making it a practical and compliant solution for healthcare applications.