Federated Learning: A Hype or a Trend?
Anna Wilbik
a
Department of Advanced Computing Sciences (DACS),
Maastricht University, Maastricht, The Netherlands
1 EXTENDED ABSTRACT
Federated learning is a emerging technology that at-
tracts growing attention from academia and industry.
Almost 3000 papers on this topic are already indexed
in Web of Science, and almost half of them was pub-
lished last year. Also many companies, from various
industries, e.g. ICT, telecom, healthcare, manufactur-
ing, finance are assessing the new possibilities.
Federated learning enables a collaboration be-
tween multiple parties to jointly train a machine learn-
ing model without exchanging the local data (by: Pe-
ter Kairouz and McMahan, 2021). Because the data
are not exchanged between parties, it is considered
a privacy preserving approach. The collaboration in
learning is considered successful, if for at least one
party the performance of the federated model is bet-
ter than the performance of the local model (Li et al.,
2019). It aims to help organizations in situations,
when a single party does not have sufficient amount
of data.
Federated learning has came a long way since it
was first proposed by McMahan in 2016 (McMahan
et al., 2017). Generally, FL can be divided into differ-
ent scenarios based on how the data is partitioned or
distributed among the data owners, i.e., horizontally
or vertically. Horizontal federated learning is used
when different parties collect the same features but
from different subjects. A common example of hori-
zontal federated learning is a group of hospitals col-
laborating to build a model that can predict a health
risk for their patients, based on agreed data. Verti-
cal federated learning is used when multiple parties
share not the features, but the subjects, like e.g., a
telecom company collaborating with a home enter-
tainment company (cable tv provider), or an airline
collaborating with a car rental agency.
Federated learning is still facing many challenges.
For some issues, especially in the context of hori-
zontal federated learning, here were proposed vari-
ous approaches to deal with problems such as for in-
stance algorithm convergence, communication over-
a
https://orcid.org/0000-0002-1989-0301
head, data heterogeneity, or security and privacy risks,
especially in the context of adversary attacks. Yet,
still those solutions are fragmented, and do not cover
the whole spectrum of the problem, e.g. there are
different, complementary strategies to deal with the
non-iid data (Zhu et al., 2021). But there are also
some basic challenges that needs some attention, such
as supporting machine learning workflows includ-
ing hyperparameter searches. Also currently most
of the implemented federated learning methods em-
ploy the empirical risk minimization formulations.
The tree-based methods, online learning, Bayesian
learning are still not investigated. There may be
needs for developments in the areas of other learn-
ing types, e.g., reinforcement learning, unsupervised
and semi-supervised, active learning. Other chal-
lenges, like data alignment, are only partially rec-
ognized. Entity alignment is an important topic in
vertical federated learning (e.g.,(Scannapieco et al.,
2007)), while almost neglected in horizontal federated
learning (Pekala et al., 2022). Moreover recent de-
velopments in XAI and ethical computing open addi-
tional possibilities in terms of addressing model fair-
ness or assessing a party contribution to the model.
This vast research effort did not remained unno-
ticed, and the federated learning was added to the
Gartner’s Hype Cycle for Privacy in 2021 (Moore,
2021), see Figure 1. A hype cycle is a graphical repre-
sentation of a common pattern that a technology goes
through from conception to maturity and widespread
adoption. The five stages in the hype cycle are Tech-
nology/Innovation Trigger, Peak of Inflated Expecta-
tions, Trough of Disillusionment, Slope of Enlighten-
ment and Plateau of Productivity.
Federated learning is still at the first stage. Here
the technology is at its infancy, with early proof-
of-concept stories and significant publicity. How-
ever, often no usable products exist or the commer-
cial viability is unproven. So far, Horizontal Fed-
erated Learning has been successfully deployed in
google keyboard, where a mobile phone can bet-
ter predict the next word typed by the owner (Hard
et al., 2018). Moreover there were several suc-
Wilbik, A.
Federated Learning: A Hype or a Trend?.
DOI: 10.5220/0011598500003332
In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 9-11
ISBN: 978-989-758-611-8; ISSN: 2184-3236
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
9