Decentralized Inter-Hospital Motor Imagery Detection using Robust Multiparty Model Aggregation

Complex deep learning problems in domains like medicine require large volumes of high quality supervised training data obtained using a multi-institutional collaborative effort. However, regulations in healthcare enforce stringent data privacy requirements and legal restrictions that penalize the transmission of data (or metadata, models, identities) to servers that lie outside its jurisdictional boundary. To achieve global model training without mobilizing training data, or model weights/gradients, we propose a serverless collaborative model training approach called DECS. To illustrate our research, we take the inter-hospital Motor Imagery (MI) detection use-case. We start by introducing the MI detection problem and its need for collaborative model training. Then, we identify a set of ideal non-functional requirements and explore how direct applications of existing private learning methods fail to meet them. Finally, we develop a compliant solution that performs serverless collaborative model training that is also resilient to active adversaries. Here, multiple local MI models from different hospitals are aggregated into a global model, without transmitting the data, models, or even gradients. We compare our accuracies, in presence of malicious clients, and observe that they are on par, and conclude by discussing the trade-offs incurred in private learning.