Information Fusion 88 (2022) 263–280
265
M.F. Criado et al.
actual methods that face this challenge, their advantages and their
drawbacks. In this paper, we will focus on those matters.
Statistical heterogeneity of the data is a major issue that needs to
be faced in order to construct and deploy a FL model. A bunch of
devices training over different local datasets may produce updates in
a wide range, thus leading to an undesired result after aggregation, or
even worse, impeding the model to converge at all. To prevent these
obstacles a common assumption in decentralized learning is considering
that the data of the different participants is Independent and Identically
Distributed (IID) [
3
,
5
,
31
]. This means that data collected by different
Dostları ilə paylaş: