IEEE Consumer Electronics Magazine - January/February 2023 - 113

detection on VANET. Their IDS output results
can be binary classification and multiclass classification.
Currently, there is not much research
on applying DBN models alone as an algorithm
for IoV intrusion detection, because DBN is more
suitable as a means to understand the ''philosophy''
and ''mindset'' of DL.
CNN/DCNN: CNN is a DNN with a convolutional
structure, and its application in the field of image
processing is remarkable due to its characteristics
of weight sharing and affine invariance.15 Many
researchers have applied CNNs to spatial mobile
data analysis. Peng et al.3 have designed a lightweight
CNN-based IDS for vehicle terminals. By
learning the data features in the network traffic
graph, an image model of normal network traffic
and abnormal network traffic was established, and
the model was used to detect network data in real
time. DCNN is structurally similar to CNN, but uses
more convolutional layers and a larger parameter
space to fit large-scale datasets. Nie et al.12
designed a data-driven IDS based on DCNN. The
author assumed that the coverage region of all
RSUs was nonoverlapping, and that each RSU independently
identified the invasion of the relevant
OBU. By analyzing the link load behavior of RSU in
IoV, various intrusion behaviors that caused irregular
fluctuations in traffic flow were detected. The
abovementioned method has better performance
in terms of accuracy, but ignores the consideration
of latency.
RNN: RNN is designed for modeling sequential
data. However, the gradient disappearance
and explosion problems make it difficult to
train traditional RNN models. LSTM alleviates
these problems by introducing a set of
''gates,'' so it is very effective in classifying
time series data.17 Ashraf et al.5 designed an
LSTM autoencoding algorithm to identify multiple
types of intrusion events from IoV networks.
Zeng et al.11 combined the LSTM model
and the CNN model to design their IDS, which was
used to automatically detect the traffic passing
through the OBUs. The method could achieve
higher detection performance with lower resource
requirements. Mobile wireless networks generate
large amounts of continuous data, and exploring
RNN family-based intrusion detection techniques
is promising for solving the problem of network
attacks in IoV.
January/February 2023
GAN: In addition to the previous learning
models, new DL techniques with better performance
are being introduced. For instance, GAN
has been actively applied to intrusion detection.7
The advantage of GAN is that it can produce
lifelike artifacts from target distributions,15
and it can train any missing data through the
reinforcement learning algorithm in the GAN
model.18 Considering that the emerging information
physics system technology has the problems
of unbalanced and incomplete sample data,
which brings difficulties to the model training,
Shahriar et al.18 used GAN to generate synthetic
samples, and IDS was trained based on the original
samples and these generated samples, which
also solved the problem of unbalanced or missing
data. The traditional GAN training process is
highly sensitive to model structure, learning
rate, and other hyperparameters. Therefore,
researchers need to use many ''tricks'' to ensure
convergence and data fidelity.
Although anomaly-based IDSs detect the
emerging attacks efficiently, it also inevitably has a
high false alarm rate. AI technologies can greatly
improve the detection accuracy of anomaly-based
IDSs. However, it requires massive computing
resources and storage resources.2 In the current
research, there are few intrusion detection
schemes for IoV that can achieve high detection
performance, low latency, lightweight, and strong
generalization ability at the same time. Besides,
how to obtain an effective dataset that can be used
for training is also a challenge for relevant
research. In addition, the security threats faced by
AI technologies themselves have to be considered
in the current research. For example, DL is vulnerable
to adversarial examples (attackers deliberately
design human inputs to trick the learning model
into making mistakes).15 Therefore, building DLbased
intrusion detection models that are robust
to adversarial cases is a must, but it is also very
challenging.
OPEN CHALLENGES AND FUTURE
RESEARCH DIRECTIONS
Although a lot of research achievements have
been achieved in the field of Intelligence IoV IDS,
there are still many open issues worth discussing
in the future.
113
IEEE Consumer Electronics Magazine - January/February 2023

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