Page 13 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 5 – Internet of Everything

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 5

FEDERATED LEARNING FOR IOE ENVIRONMENTS: A SERVICE PROVIDER REVENUE
MAXIMIZATION FRAMEWORK

1
1
1
Benedetta Picano , Romano Fantacci , Tommaso Pecorella , Adnan Rashid 1
1
Dpt. Information Engineering, Università di Firenze, Italy
NOTE: Corresponding author: Benedetta Picano, benedetta.picano@uni i.it

Abstract – In accordance with the Internet of Everything (IoE) paradigm, millions of people and billions of devices are ex‑
pected to be connected to each other, giving rise to an ever increasing demand for application services with a strict quality of
service requirements. Therefore, service providers are dealing with the functional integration of the classical cloud comput‑
ing architecture with edge computing networks. However, the intrinsic limited capacity of the edge computing nodes implies
the need for proper virtual functions’ allocations to improve user satisfaction and service ful illment. In this sense, demand
prediction is crucial in services management and exploitation. The main challenge here consists of the high variability of
application requests that result in inaccurate forecasts. Federated learning has recently emerged as a solution to train math‑
ematical learning models on the users’ site. This paper investigates the application of federated learning to virtual functions
demand prediction in IoE based edge‑cloud computing systems, to preserve the data security and maximise service provider
revenue. Additionally, the paper proposes a virtual function placement based on the services demand prediction provided
by the federated learning module. A matching‑based tasks allocation is proposed. Finally, numerical results validate the
proposed approach, compared with a chaos theory prediction scheme.

Keywords – Edge computing, federated learning, Internet of Everything, matching theory, revenue maximization, virtual
function placement

1. INTRODUCTION Indeed, a proper resource exploitation planning is essen‑
tial to guarantee elevated levels of network ef iciency,
The emergence of new network paradigms such as Edge
Computing (EC) [1, 2, 3, 4], for which the limitations typ‑ user satisfaction and consequent high SP revenues, as
highlighted by literature such as [13], [14]. In particu‑
ical of the cloud architecture have been bypassed mov‑
lar, having an a priori knowledge about the data low ser‑
ing computation nodes to the network edges close to the
vice demand can be properly exploited to perform suit‑
end users, has given rise to a wide range of challenges in
able resource infrastructure planning with maximum in‑
many research areas [5, 6]. Consequently, several new
come. In order to pursue this objective, Machine Learn‑
issues, such as user mobility, heterogeneity in Quality
ing (ML) [15, 16, 17, 18, 19] has emerged by providing
of Service (QoS) or service requirements, massive vol‑
many techniques to perform data behavior interpretation
ume of data, user privacy, diversity on data types and
and analysis. The ability of ML techniques in catching
so on, have led to numerous efforts from both academia
data trends, patterns and hidden features, has ensured its
and industry in providing highly effective and ef icient
applicability to many problems. However, although the
solutions [7, 8, 9, 10, 11]. In particular, there exists a
signi icant branch of literature regarding possible solu‑ knowledge and extrapolation of user data characteristics
tions to improve EC Network (ECN) performance in or‑ positively impacts many application areas, it may result
in being non‑compliant with some speci ic user privacy
der to guarantee a high level of user satisfaction and to
provide dynamic and lexible network resource alloca‑ constraints [20]. In this respect, if on the one hand the
users’ data analysis may lead to remarkable advantages
tion and decision‑making strategies. Within this context, in reference to the network resources planning and ex‑
the Internet of Everything (IoE) paradigm, in which peo‑
ploitation, on the other the user data gathering may trig‑
ple, process, data, and things are connected and exchange
ger userdissention, due toprivacyconcernsand violation.
data,has given rise to systems with increasing complexity
Within this context, a data‑manipulation framework able
and applications involving strict real‑time requirements
to collect users’ data without contravening users’ privacy
and sensitive data [12], heterogeneous traf ic. Generally
speaking, heterogeneity in data low types implies dif‑ is a priority. In this respect, Federated Learning (FL) [21,
20, 8, 22, 23, 24] has recently emerged as a promising
ferent QoS or service requirements. Furthermore, from
a Service Provider (SP) perspective, such diversity trig‑ tool to perform, locally on the users’ devices, statistical
and mathematical training models based on ML method‑
gers new data low management policies, service provi‑
ologies without losing users privacy constraints. The FL
sion costs and selling prices. In this respect, the SP rev‑ framework consists of the devices level, generally indi‑
enue maximization is strictly related to the adopted man‑
cated in literature as clients, and a central server unit
agement and administration policy.
© International Telecommunication Union, 2021 1