Secure federated learning applied to medical imaging with fully homomorphic encryption

Abstract

This study explores the convergence of Federated Learning (FL) and Fully Homomorphic Encryption (FHE) through an innovative approach applied to a confidential dataset composed of mammograms from Belgian medical records. Our goal is to clarify the feasibility and challenges associated with integrating FHE into the context of Federated Learning, with a particular focus on evaluating the memory constraints inherent in FHE when using sensitive medical data. The results highlight notable limitations in terms of memory usage, underscoring the need for ongoing research to optimize FHE in real-world applications. Despite these challenges, our research demonstrates that FHE maintains comparable performance in terms of Receiver Operating Characteristic (ROC) curves, affirming the robustness of our approach in secure machine learning applications, especially in sectors where data confidentiality, such as medical data management, is imperative. The conclusions not only shed light on the technical limitations of FHE but also emphasize its potential for practical applications. By combining Federated Learning with FHE, our model preserves data confidentiality while ensuring the security of exchanges between participants and the central server