Page 14 - 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

which aggregates and merges the data preliminary pro‑ The r organized follow 2
cessed by the clients. Typically, FL has the following mat‑ an in‑depth review of the related literature is presented.
ters to face with [25] Section 3, discusses the problem statement, while in Sec‑
tion 4 the FL framework and the placement strategy are
• Non‑Independent Identically Distributed Data present experimental results
The clients have different training datasets, therefore are analyzed and the alternative CT predictive approach
a single dataset cannot be considered representative explained. Finally, the conclusions are presented in
of the other clients datasets; Section 6.
• Unbalanced Datasets Different clients have differ‑ 2. RELATED WORKS
ent datasets, and each dataset may have a diverse
number of elements in comparison to other clients Recently techniques hav f extensiv applica‑
datasets; tions in big data analysis in fog/edge networks research
area.
• Large‑Scale Distribution The number of clients
An overview of the ML techniques applied to fog is pre‑
involved in the FL training procedure is generally
sented in paper [26 Then, paper [26] investigates the
higher than the amount of data processed at the
ability of the ML strategies in detecting malicious attack‑
client level;
ers in fog networks, while paper [27] focuses on the ML
t evaluat advantages deri fr an
• Limited Communication Mobile devices may or
may not be available for data training and the com‑ edge caching solution, taking into account user satisfac‑
tion perspective and energy ef iciency. The improvement
putational capability or communication conditions
in sensing reliability and network latency is the aim of pa‑
could be poor.
per [28], in which the authors implement a multi‑hidden
In reference to the proposed contextualization, we have multi‑layer convolutional neural network solution to pro‑
assumed here that sensitive user data may be derived vide data authentication crowd‑sensing en‑
from historical users functions utilization. In this per‑ vironment The tr decisions strat combined with
spective, sharing data about daily users habits may ex‑ k‑near neighbors method [29 in
pose the users to undue risks. For this reason, the FL which authors deal with the position‑based con idential‑
framework may represent a useful tool to counteract such ity problem in high real‑time industrial application sce‑
a problem. However, a deep investigation of the privacy narios.
issues are out of the scope of this paper. The paper pro‑ In a different way, SP maximization is the objective of pa‑
poses the application of the FL framework, in order to per [30], in which a deep supervised learning approach is
forecast the service demands, without losing the user pri‑ applied to perform the minimization of the total network
vacy constraints, in an IoE scenario. Moreover, on the ba‑ cost A fog blockchain network is analyzed in paper [31],
sis of service demand forecasting, this paper proposes a which formulates a solution based on the auction theory,
suitable Virtual Functions (VFs) placement both on the where deep learning is applied to the maximization of the
ECN and cloud. Summarizing, the contributions of this pa‑ edge computing SP revenue.
per are Additionally, distributed ML is adopted in papers [32, 33,
34, 35 In paper [32], a distributed version of the well‑
• Application of the FL strategy to forecast the network
known support vector machine method is implemented
VFs demand, in order to take into account the users
to investigate its applicability The reinforcement learn‑
privacy;
mor Q‑learning ap‑
plied in paper [33], in order to minimize the users’ outage
• Formulation of the SP maximum revenue problem,
in heterogeneous cellular networks scenarios. The con‑
by considering Service Requests (SRs) with a differ‑
trol in crowd‑sensing problem is the main objective of pa‑
ent priority and hence, different cost and price. In
per [34], exploiting the human in the loop methodology to
particular, the SP can accept the data SRs with low
propose a hierarchical crowd sensing framework with the
priority if all the high priority lows have been satis‑
aim of reducing cloud congestion and promoting the bal‑
ied;
ancing of the data tr ic. Then, the distributed stochas‑
tic variance reduced gradient is applied in paper [35], in
• Proposal of a VFs placement strategy and a suitable
which a target accuracy is ixed, and the optimization of
matching‑based SRs allocation algorithm based on
the number of collection points to make data analysis pro‑
the considered FL and the previously provided VFs
vided. Furthermore, paper [35] proposes the minimiza‑
forecasting scheme;
tion of the amount of network traf ic sent towards the col‑
• Performance evaluation of the proposed approach lection In a different way, the maximization of SP
and the comparisons with a centralized Chaos The‑ pro it in a Mobile Edge Computing (MEC) blockchain net‑
ory (CT)‑based prediction scheme, by resorting to work has been studied in paper [31], in which an auction
extensive computer simulation runs. strat deep formulat to

2 © International Telecommunication Union, 2021