IEEE Consumer Electronics Magazine - November/December 2022 - 62

Security, Trust and Privacy Solutions for Intelligent Internet of Vehicular Things
Immediate Response
Typically, most vehicular applications necessitate
making real-time decisions, hence making
such decisions on a central server is impractical
because of the incurred high latency. In this
regard, the convergence of FL and VEC enables
vehicular participants to be aware of both local
and global learned knowledge, thereby can provide
decisions with lower latency to satisfy the
requirement of real-time IoVT applications.
Heterogeneity
Taking into consideration the utilization of
vehicular data from a wide range of heterogeneous
vehicular participants along with alleviating
the security and privacy threats approach, the FL
opens the way toward the development of a new
kind of intelligent vehicular applications implausible
to realize using traditional approaches.
PROPOSED FED-GCRN FRAMEWORK
This section presents the methodology of the
proposed Fed-GCRN framework for predicting
the traffic flows in IoVT environment (smart
transportation, smart cities).
Problem Formulation
For traffic forecasting, the TFF is formulated
as a multistep forecasting problem. Where innumerable
traffic series consists of Nof related univariate
time series signified as X:t ¼fX0
t g; where X:t 2 RN1 with N sources per
t ;X1
... ;XN
time stage t. Then the main goal is to forecast
the upcoming traffic elements (t steps) of the
interrelated traffic series according to the experiential
ancient values (T steps). Thus, the TFF
can be formulated as the problem of discovering
a method F to estimate the next t values according
to the previous T steps of traffic data. For
modeling the spatial relationships among various
traffic series, the traffic data is encoded into
graph representation G ¼ðV; E;AÞ, in which
the traffic series' sources are represented with
N of vertices denoted as V, while E denotes the
set of interconnecting edges and A 2 RNN represent
the adjacent matrix of the G indicating
the vicinity among the traffic series. Thus, the
TFF task can be formulated as finding a function
62
t ;
that satisfies the following mapping: F : ðfXt;
... ;XtT1g; GÞ!fXtþ1; ... ;Xtþtg.
Model Design
Recently, the graph convolutional network
(GCN) has been exploited to model the spatial
relationships among traffic series by transforming
the features of a given vertex shared among all
vertices. Yet, sharing parameters is demonstrated
to provide a sub-optimal solution for TFF problems
owing to the dynamic nature of traffic data
and various components of the node that might
impact traffic. In particular, the traffic series in
two neighboring vertices could also introduce different
representations at some time intervals
owing to their local characteristics. Besides, the
traffic sequence from two separate vertices could
still reveal contradictory representations. Therefore,
modeling only the common representations
between all vertices is insufficient for efficient traffic
regression in IoVT environments, and it is necessary
to preserve an exclusive parameter space
per vertex to model node-related information,
which is complex to optimize and might lead to
overfitting. Inspired by the work of Bai et al.,16
node adaptive parameter learning (NAPL) module
is included to solve the limitations of conventional
GCN by sketching intuitions from the matrix factorization
rather than from parameter matrices.
Especially, the NAPL learns two matrices namely
vertex-embedding matrix EG
i and a weight pool
matrix, where each vertex extracts parameters
from a big joint weight pool based on the vertex
embedding, which can be considered as learning
vertex-related representation from a group of
possible representations exposed from traffic
streams altogether. In addition, instead of using
similarity or distance function to compute adjacency
matrices, the data adaptive graph generation
is employed to automatically deduce the
hidden spatial interrelationships from traffic
series essential for TFF. We refer to the beforementioned
set GCN as enhance GCN (ECGN).
Moreover, the temporal representation is also
vital for efficient TFF, hence the GRU is redesigned
to model temporal representation sequential
dependencies in traffic series. In particular, the
GCRN architecture substitutes the multilayer perceptron
of the GRU with the ECGN to learn and
extract vertex-related representations.
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - November/December 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - November/December 2022
