IEEE Consumer Electronics Magazine - January/February 2023 - 112

Security, Trust and Privacy Solutions for Intelligent Internet of Vehicular Things
Table 2. Summary of contemporary typical AI-based IDSs for the IoV.
proposed an intelligent IDS based on the previous
study, which was widely suitable for detecting
CAN bus and external network attacks. The treebased
ML algorithms, including random forest
(RF), decision tree, extra tree, and extreme gradient
boost, were used in the design. The stacked
detection model was with an accuracy of more
than 99%, but with a longer execution time compared
to any single model. After using the synthetic
minority oversampling technique for
reducing the category imbalance and computational
cost, however, high detection performance
and low delay cannot be satisfied simultaneously.
Besides traditional ML algorithms, many
advanced ML extension algorithms and models,
for instance, deep learning (DL), have been applied
to the study of IDS for IoV.15 DL is a representation
learning algorithm based on data, and it has many
advantages that traditional ML methods do not
have. It is able to use nonsupervised or semisupervised
feature learning and hierarchical feature
selection algorithms instead of manual characteristic
acquisition. Many DL models have been
applied to IoV intrusion detection, such as ANN,
DNN,DBN, CNN, RNN, and GAN.
ANN/MLP: ANN has played a vital role in scientific
research fields such as target classification
and data fitting with its strong self-learning and
self-adaptive capabilities. Since themodel was proposed,
various neural networks have appeared.3,8
112
Researchers have applied various neural network
models to the research of IoV intrusion detection.9
MLP was the initial ANN design.15 Anzer et al.16
implemented a distributed intrusion detection system
(DIDS) based on MLP to monitor malicious
traffic in IoV. The researchers planned to install
the DIDS in OBUs. Their IDS successfully detected
different types of attacks, such as DoS, probe, R2L,
and U2R. From the experiments, it can be understood
that using a small number of selected features
also has high detection accuracy, while
reducing latency and complexity. However, there
is still the limitation that the network must be
deep enough to obtain better results.
DNN: DNN has more hidden layers than traditional
ANN, the role of these hidden layers
is very significant, and they give DNN superiority
that can come from extracting high-level
features from raw data. DNN has been extensively
researched in the field of AI and is
widely used in practical applications (e.g.,
image processing, speech recognition, and
computer vision). The following types of models
are extensions and improvements of the
DNN model.
DBN: DBN is a DNN model superimposed by
restricted Boltzmann machines; it is an effective
means to deal with the problems of low speed
and overfitting in the learning process of DNN.
Vitalkar et al.8 used DBN to realize intrusion
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - January/February 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2023
