Testing Federated Learning on Health and Wellbeing Data

Nowadays artificial intelligence is used in healthcare applications, too. Based on current research personalized medicine could transform the healthcare domain. Thus, medical data from users must be collected and used for training models. In order to preserve the privacy of data, federated learning represents a good candidate. This paper proposes an extension of the federated learning model that is evaluated for learning over aa distributed dataset. The proposed architecture is a client-server, where the clients are clustered by the server, according to their data similarity (without exposing data to the server). The server stores the clusters models and manages the clients. Different tests were performed on three datasets: CIFAR-10, MNIST and a non-standard one - a sleep dataset. Results show that an increase of the convergence rate was obtained (in case of the MNIST dataset was 50 times faster). Also, the method has the ability to learn patterns from the data, by keeping data locally.