![](https://csdnimg.cn/release/download_crawler_static/86331163/bg1.jpg)
Client Selection for Federated Learning with
Heterogeneous Resources in Mobile Edge
Takayuki Nishio
Graduate School of Informatics,
Kyoto University, Japan
Email: nishio@i.kyoto-u.ac.jp
Ryo Yonetani
Institute of Industrial Science,
The University of Tokyo, Japan
Email: yonetani@iis.u-tokyo.ac.jp
Abstract—We envision a mobile edge computing (MEC) frame-
work for machine learning (ML) technologies, which leverages
distributed client data and computation resources for training
high-performance ML models while preserving client privacy.
Toward this future goal, this work aims to extend Federated
Learning (FL), a decentralized learning framework that enables
privacy-preserving training of models, to work with heteroge-
neous clients in a practical cellular network. The FL protocol
iteratively asks random clients to download a trainable model
from a server, update it with own data, and upload the updated
model to the server, while asking the server to aggregate multiple
client updates to further improve the model. While clients in
this protocol are free from disclosing own private data, the
overall training process can become inefficient when some clients
are with limited computational resources (i.e., requiring longer
update time) or under poor wireless channel conditions (longer
upload time). Our new FL protocol, which we refer to as
FedCS, mitigates this problem and performs FL efficiently while
actively managing clients based on their resource conditions.
Specifically, FedCS solves a client selection problem with resource
constraints, which allows the server to aggregate as many client
updates as possible and to accelerate performance improvement
in ML models. We conducted an experimental evaluation using
publicly-available large-scale image datasets to train deep neural
networks on MEC environment simulations. The experimental
results show that FedCS is able to complete its training process in
a significantly shorter time compared to the original FL protocol.
I. INTRODUCTION
A variety of modern AI products are powered by cutting-
edge machine learning (ML) technologies, which range from
face detection and language translation installed on smart-
phones to voice recognition and speech synthesis used in
virtual assistants such as Amazon Alexa and Google Home.
Therefore, the development of such AI products typically
necessitates large-scale data, which are essential for training
high-performance ML models such as a deep neural network.
Arguably, a massive amount of IoT devices, smartphones,
and autonomous vehicles with high-resolution sensors, all of
which are connected to a high-speed network, can serve as
promising data collection infrastructure in the near future
(e.g., [1]). Researchers in the field of communication and
mobile computing have started to interact with data science
communities in the last decade and have proposed mobile edge
computing (MEC) frameworks that can be used for large-scale
data collection and processing [2].
Typically, MEC frameworks assume that all data resources
are transferred from data collection clients (IoT devices,
smartphones, and connected vehicles) to computational infras-
tructure (high-performance servers) through cellular networks
to perform their tasks [3], [4]. However, this assumption is
not always acceptable when private human activity data are
1. Downloading model parameters
3. Uploading the new parameters
2. Updating the model with own data
Server
MEC platform
4. Aggregating client updates
Base station
Clients
Cellular network
Fig. 1. Federated learning [5] enables one to train machine learning
models on private client data through the iterative communications of model
parameters between a server and clients. How can we implement this training
process in practical cellular networks with heterogeneous clients?
collected, such as life-logging videos, a history of e-mail
conversations, and recorded phone calls. On one hand, such
private activity data would be a key factor for improving
the quality of AI products that support our daily life, which
include not only AI-related apps on smartphones and virtual
assistants but also AI-powered smart cities. On the other hand,
uploading these data directly to computational infrastructure
is problematic as the data could be eavesdropped by malicious
users in a network to compromise client’s privacy.
To address this fundamental privacy concern, one work has
recently been presented by the ML community: Federated
Learning (FL) [5]. As illustrated in Fig. 1, FL iteratively asks
random clients to 1) download parameters of a trainable model
from a certain server, 2) update the model with their own
data, and 3) upload the new model parameters to the server,
while asking the server to 4) aggregate multiple client updates
to further improve the model. In exchange for requiring data
collection clients to install a certain level of computational
resources (e.g., a laptop equipped with reasonable GPUs,
autonomous vehicles with moderate computational capaci-
ties [1]), the FL protocol allows the clients to keep their data
secure in their local storage.
In this work, we focus on the implementation of the
abovementioned FL protocol in practical MEC frameworks.
We believe that our work will influence the future development
platform of various AI products that require a large amount
of private activity data to train ML models. In particular, we
consider the problem of running FL in a cellular network used
by heterogeneous mobile devices with different data resources,
computational capabilities, and wireless channel conditions.
Unfortunately, a direct application of existing FL protocols
without any consideration of such heterogeneous client prop-
arXiv:1804.08333v2 [cs.NI] 30 Oct 2018