Federated and secure cloud services for building medical image classifiers on an intercontinental infrastructure

Abstract Medical data processing has found a new dimension with the extensive use of machine-learning techniques to classify and extract features. Machine learning strongly benefits from computing accelerators. However, such accelerators are not easily available at hospital premises, although they can be easily found on public cloud infrastructures or research centers. Nevertheless, the sensitivity of medical data poses several challenges on the access to such data, requiring security guarantees and isolation. In this paper we present an architecture that addresses this problem. It keeps critical data encrypted in memory and disk, which can only be accessed inside trusted execution environments protected by hardware extensions. Data is anonymized inside these environments and securely transferred to external sites that host accelerator devices, keeping the same network space and reducing security risks even in untrusted cloud backends. Results on the processing of data in different scenarios are presented and discussed. The results are demonstrated on a geographically-wide deployment provided by the ATMOSPHERE project.