A federated learning architecture for secure and private neuroimaging analysis

Area of Science:

• Neuroimaging
• Biomedical Data Analysis
• Machine Learning

Background:

• The rapid growth of biomedical data presents challenges for multi-site analysis due to security, privacy, and regulatory concerns.
• Federated learning offers a solution by enabling distributed model training without direct data sharing.

Purpose of the Study:

• To introduce MetisFL, a federated learning architecture designed for secure and private distributed training of neural network models.
• To evaluate the performance of MetisFL in neuroimaging tasks within heterogeneous federated environments.

Main Methods:

• Utilizing federated learning for distributed neural network training, where local models are trained on private data.
• Implementing robust security measures including encrypted parameter transmission and fully homomorphic encryption for global model aggregation.
• Employing information-theoretic methods to mitigate potential data leakage and prevent adversarial attacks like model inversion and membership inference.

Main Results:

• Demonstrated the effectiveness of MetisFL in performing secure and private federated learning for neuroimaging applications.
• Successfully applied the architecture to predict Alzheimer's disease and estimate Brain Age Gap (BrainAGE) from magnetic resonance imaging (MRI) data.
• Validated performance in challenging, heterogeneous federated settings with varying data distributions and quantities across sites.

Conclusions:

• MetisFL provides a secure and private framework for collaborative analysis of sensitive biomedical data, particularly in neuroimaging.
• The architecture effectively addresses challenges posed by data heterogeneity and privacy concerns in multi-site studies.
• Federated learning, as implemented by MetisFL, is a viable and powerful approach for advancing medical research through distributed data analysis.