IEEE Consumer Electronics Magazine - November/December 2022 - 17

runtime to detect cyber-attacks. This approach
demonstrated superior performance in accuracy,
F1 score, false-positive rate (FPR), and receiver
operating characteristics curve-area under the
curve (AUC) compared to various state-of-the-art
statistical, proximity-based, and machine learning
(ML) based IDS approaches under different attack
scenarios. Linghu and Li15 constructed a weighted
state graph (WSG) in the offline phase with the
CAN message IDs as the vertices of the graph and
the edges of the graph representing the time-varying
state transitions of CAN data frames. At runtime,
a subgraph was generated for a sliding
window of CAN IDs, and an anomaly was triggered
if the generated subgraph was not within the range
of the WSG created during training. This technique
focuses on detecting a wide range of attacks and
minimizing FPR. A temporal convolutional neural
attention (TCNA) network based lightweight IDS
was presented by Thiruloga et al.,17 to learn very
long-term dependencies between messages in an
in-vehicle network. This was used to effectively
learn the normal system behavior of the vehicle at
design time. A decision tree (DT) based classifier
was used to learn the model deviations that correspond
to the normal vehicle operation at design
time. At runtime, the trained TCNA and DT models
were used to observe for deviations using the deviation
score metric to detect cyber-attacks. This
technique outperformed all the comparison works
in theMatthews correlation coefficient metric.
IDSs for VANET Security
Several other efforts have focused on AIbased
IDS in VANET environments, particularly
during vehicle-to-vehicle (V2V) and vehicle-toinfrastructure
(V2I) communication.
An AI-based IDS framework was introduced by
Boddupalli et al.21 to detect and mitigate anomalous
vehicle responses of cooperative adaptive
cruise control (CACC) modules in a V2V network.
An ML-based predictor was trained to predict the
acceleration of the preceding vehicle, and an
anomaly alert was raised when the observed
acceleration exceeded the prediction by a certain
threshold. The framework then assessed the
threshold deviation level and would override the
CACC when anomalies were detected. On-road
driving data, such as the preceding vehicle acceleration,
intervehicle space, and vehicle velocity,
November/December 2022
were periodically collected to train the MLmodel.
A data-driven IDS to detect attacks occurring in
roadside units (RSUs) communicating with a vehicle
was presented by Nie et al.19 It used a convolutional
neural network (CNN) based architecture to
extract the link load features of the RSU to detect
attacks. This technique demonstrated high accuracy
compared to neural network, support vector
machine, and principal component analysis models.
However, this technique was evaluated only
with a limited set of attacks. A hybrid IDS framework
was proposed by Aloqaily et al.20 that
performed a three-phase data traffic analysis,
reduction, and classification on the received cloud
service requests to detect intrusions. A combination
of a deep belief neural network (DBN) and a
DT classifier was presented to perform data
dimensionality reduction and attack classification.
This technique achieved high detection accuracy
and lower FPR when evaluated under DoS,
probe, remote to user, and user to root attacks.
However, this approach incurs large detection
overhead. An LSTM-based autoencoder model
was proposed by Ashraf et al.22 to detect intrusion
in V2X networks. The LSTM model tries to learn
the temporal representation of the compressed
network traffic data, and any deviation from the
learned representation is classified as an attack.
The authors claim their technique can detect
intrusions in both in-vehicle and external networks
with high accuracy.
A summary of both the in-vehicle network and
the VANET-level IDSs discussed in this article is
presented in Table 1.Moreover, the techniques by
Kukkala et al.,16 Thiruloga et al.,17 and Kukkala
et al.18 have explored the feasibility of AI-based
IDS on a real-world automotive ECU. They were
able to detect various cyber-attacks with minimal
computational, memory, and power overhead.
CYBERSECURITY ROADMAP FOR
AUTONOMOUS VEHICLES
Despite the many promising state-of-the-art
AI-based IDS techniques that show compelling
results, several issues still need to be addressed
to make future autonomous vehicles truly
secure. Here, we discuss various key elements in
our roadmap (see Table 2) toward achieving
autonomous automotive cybersecurity goals in
17
IEEE Consumer Electronics Magazine - November/December 2022

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