Achieving Privacy-Preserving Diagnosis with Federated Learning in LEO Satellite Constellation

LEO satellite constellations can support the federated learning-based diagnosis model construction for epidemics occurring in remote less developed areas, like Malaria. However, the model trained directly from the patients’ medical data may leak their physical conditions, and LEO satellite constellations also suffer from the long propagation delay and limited energy supply. To tackle the challenges, we suggest a new privacy-preserving model update framework in federated learning, which adapts to the highly dynamic topology of the LEO satellite constellation. To protect each individual piece of model update, we propose this privacy-preserving scheme which combines a symmetric homomorphic cryptosystem and a verifiable secret sharing scheme, where the security goals of privacy preservation and authentication can be achieved. We demonstrate the feasibility and evaluate the effectiveness of our proposed privacy-preserving diagnosis model on the real dataset, namely the Malaria Cell Image Dataset, and simulation results demonstrate that our proposed privacy-preserving scheme mostly improves the computational complexity in contrast to a scheme exploiting the Paillier cryptosystem.