Efficient Privacy-Preserving Medical Cross-Silo Federated Learning

Federated Learning (FL) is becoming increasingly important for advancing machine learning in the medical domain, as it enables collaborative model training across multiple institutions without requiring them to share raw patient data. Unlike traditional centralized machine learning, which gathers data and trains the model in a single location, FL shares model updates among multiple independent participants, thereby keeping their data private. However, this decentralized structure-while privacy-preserving in principle-significantly expands the attack surface. Model updates exchanged between sites introduce new vulnerabilities, increasing the risk of unauthorized access, model poisoning, server-side eavesdropping, and potential leakage of sensitive patient information. Although established homomorphic encryption (HE) schemes such as CKKS, BFV, and BGV offer strong theoretical guarantees for secure computation, their high computational overhead often renders them impractical for real-world medical applications. To address this, we introduce a novel Secure Multi-Party Computation (SMPC)-based protocol tailored specifically for FL in the healthcare context. Our approach holistically rethinks the entire FL-SMPC pipeline, incorporating concurrency at critical stages to improve runtime efficiency while preserving strong security guarantees for the aggregated global model on the server side-an often overlooked aspect in existing SMPC approaches.