Abstract
Federated learning (FL), a recent collaborative learning paradigm, has raised plenty of opportunities to bring learning to the edge, but it also faces many obstacles. We focus on the challenges explicitly imposed by heterogeneity within FL in a pervasive computing environment. We summarize three main methods that the scientific community has proposed to tackle this challenge (improved aggregation techniques, regularization of clients learning, and clustering of similar clients). The research objectives of this thesis work are to develop upon these proposed methods to build robust FL schemes to benefit from user diversity while also mitigating its detrimental effects. To fulfill this objective, we follow a three-step approach: (a) analyze and evaluate different FL approaches for pervasive environments (b) experiments and the proposal of the three different fields of FL for heterogeneity (c) integration of the three proposed methods on real-devices.