Personalised Federated Learning on Heterogeneous Feature Spaces

Abstract

Personalised federated learning (FL) approaches assume that raw data of all clients are defined in a common space \emph{i.e.} all clients store their data according to the same schema. For real-world applications, this assumption is restrictive as clients, having their own systems to collect and then store data, may use {\em heterogeneous} data representations. To bridge the gap between the assumption of a shared subspace and the more realistic situation of client-specific spaces, we propose a general framework coined FLIC that maps client's data onto a common feature space via local embedding functions, in a federated manner. Preservation of class information in the latent space is ensured by a distribution alignment with respect to a learned reference distribution. We provide the algorithmic details of FLIC as well as theoretical insights supporting the relevance of our methodology. We compare its performances against FL benchmarks involving heterogeneous input features spaces. Notably, we are the first to present a successful application of FL to Brain-Computer Interface signals acquired on a different number of sensors.