Received 11 October 2023, accepted 23 December 2023, date of publication 5 January 2024, date of current version 22 January 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3350271

Enhancing Road Safety and Cybersecurity in

Traffic Management Systems: Leveraging the
Potential of Reinforcement Learning
ISHITA AGARWAL 1 , AANCHAL SINGH1 , ARAN AGARWAL 1 , SHRUTI MISHRA 2 , SANDEEP
KUMAR SATAPATHY 3 , SUNG-BAE CHO3 , (Senior Member, IEEE), MANAS RANJAN PRUSTY 4 ,
AND SACHI NANDAN MOHANTY 5 , (Senior Member, IEEE)
1 School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India
2 Centre for Advanced Data Science, Vellore Institute of Technology, Chennai, Tamil Nadu 600127, India
3 Department of Computer Science, Yonsei University, Seodaemun, Seoul 03722, South Korea
4 Centre for Cyber-Physical Systems, Vellore Institute of Technology, Chennai 600127, India
5 School of Computer Science and Engineering (SCOPE), VIT-AP University, Amaravati, Andhra Pradesh 522237, India

Corresponding author: Shruti Mishra (shrutim2129@[Link])

This work was supported by the Vellore Institute of Technology, India.

ABSTRACT With the increasing reliance on technology in traffic management systems, ensuring road
safety and protecting the integrity of these systems against cyber threats have become critical concerns.
This research paper investigates the potential of reinforcement learning techniques in enhancing both road
safety and cyber security of traffic management systems. The paper explores the theoretical foundations
of reinforcement learning, discusses its applications in traffic management, and presents case studies and
empirical evidence demonstrating its effectiveness in improving road safety and mitigating cyber security
risks. The findings indicate that reinforcement learning can contribute to the development of intelligent and
secure traffic management systems, thus minimizing accidents and protecting against cyber-attacks.

INDEX TERMS Cyber security, traffic management systems, reinforcement learning, road safety.

I. INTRODUCTION Simultaneously, the digitalization of traffic management

Road safety and cyber security are of paramount importance systems has introduced new vulnerabilities and risks associ-
in the context of traffic management systems. As traffic man- ated with cyber threats [2]. Traffic management systems are
agement systems become increasingly reliant on technology, now interconnected and exposed to potential cyber-attacks
there is a critical need to ensure the safety of road users that can compromise the safety and efficiency of these sys-
and protect the integrity of these systems from cyber threats. tems. Ensuring the cyber security of traffic management
This research paper intends to investigate the significance systems is crucial to maintaining their reliability and pre-
of reinforcement learning in enhancing both road safety and venting disruptions that could have severe consequences
cyber security in traffic management systems. Road safety is for road safety. Several challenges exist in ensuring road
a global concern due to the significant human and economic safety and cyber security in traffic management systems.
losses caused by traffic accidents. Traditional approaches to First, road safety faces challenges such as distracted driving,
road safety have relied on traffic regulations, infrastructure aggressive behaviour, non-compliance with traffic rules, and
improvements, and driver education [1]. While these mea- complex traffic scenarios. These factors contribute to acci-
sures have made valuable contributions, the complexity of dents and pose significant risks to road users. In terms of
modern traffic environments demands innovative solutions cyber security, traffic management systems face vulnerabil-
that can adapt to dynamic conditions and minimize accidents. ities due to their interconnected nature. Unauthorized access,
data breaches, malware attacks, and system manipulation are
The associate editor coordinating the review of this manuscript and all potential threats that can compromise the integrity and
approving it for publication was Yang Liu . availability of the systems. Ensuring the cyber security of
2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 12, 2024 For more information, see [Link] 9963
 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

traffic management systems requires proactive measures to II. RELATED WORK

detect, prevent, and respond to yber-attacks effectively [3]. There have been some works related to this field. Some of
Reinforcement learning offers a promising approach to them are as discussed below:
address the challenges in road safety and cyber security of Ouallane et al. [1] presented a global vision of road traffic
traffic management systems. There is immense potential of management solutions proposed in the literature, including
reinforcement learning in developing intelligent driver assis- routing mechanisms, traffic light-based approaches, and net-
tance systems that can analyse traffic patterns, identify risky work traffic management strategies. It provides a classifica-
situations, and make proactive decisions to prevent acci- tion and evaluation of these solutions, along with highlighting
dents [4]. Moreover, reinforcement learning can contribute to future research directions in urban road traffic management.
the cyber security of traffic management systems by devel- Reinforcement Learning has proven to be a effective AI
oping intrusion detection and prevention systems that learn mechanism as Botvinick et al. [2] stated that recent advance-
from network traffic data and identify anomalous behav- ments in AI research resulting in potent methodologies
ior indicative of cyber-attacks [3]. Additionally, resilient for deep reinforcement learning, these approaches amalga-
and adaptive systems can be designed using reinforcement mate representation learning with incentive-guided actions.
learning techniques to respond effectively to evolving cyber Although initial concerns centered on the substantial volume
threats [5]. Additionally, in the field of traffic signal con- of training data necessary, subsequent AI research has intro-
trol using connected vehicles, a study in 2018 highlighted duced techniques facilitating the more efficient learning of
the vulnerabilities of connected vehicle-based traffic signal deep reinforcement systems.
control when dealing with congestion attacks [6]. Moreover, Pattanaik et al. [3] discussed adversarial attacks specifi-
a recent paper in 2022 summarized IoT and AI-driven road cally designed for Reinforcement Learning (RL) and demon-
traffic management strategies, underscoring the potential of strates their effectiveness in degrading the performance
these technologies to improve traffic management [1]. of Deep Reinforcement Learning algorithms (DRL). The
By leveraging the capabilities of reinforcement learning, attacks, even when naively engineered, successfully degrade
traffic management systems can become more intelligent, the performance of DRL algorithms. By incorporating these
adaptive, and secure. The ability to learn from data and adapt attacks into the training process, the robustness of RL
to dynamic conditions can significantly improve road safety algorithms such as Deep Double Q learning and Deep
by reducing accidents and enhancing the cyber security of Deterministic Policy Gradients is significantly improved,
these systems. In the following sections of this research paper, as evidenced by experiments on RL benchmarks like Cart-
we will delve into the theoretical foundations of reinforce- pole, Mountain Car, Hopper, and Half Cheetah environments.
ment learning, explore its applications in traffic management Chen et al. [4] proposed a reinforcement learning stands as
systems, and present case studies and empirical evidence that a pivotal technology in contemporary artificial intelligence
demonstrate its effectiveness in enhancing road safety and domains, encompassing applications in both the gaming
cyber security. The objective of this study is to add to the industry and connected and automated vehicle systems.
expanding knowledge within this domain and offer perspec- Nonetheless, concerns are growing regarding the security
tives on the prospective advantages and obstacles linked to of reinforcement learning systems, particularly due to the
incorporating reinforcement learning into traffic management identification of effective adversarial attacks directed at neu-
systems. ral network policies within this framework. Chen et al. [6]
The key findings of the study indicate that reinforcement proposed a study on the imminent alteration of present-day
learning-based intelligent driver assistance systems offer sub- transportation systems is on the horizon, thanks to Con-
stantial potential for enhancing road safety. They achieve nected Vehicle (CV) technology, which establishes links
this by effectively identifying potential collision risks, aiding between vehicles and transportation infrastructure via wire-
in safe decision-making processes, and ultimately reduc- less communication. This heightened connectivity has shown
ing the occurrence of accidents. Additionally, the use of the potential to significantly enhance transportation mobility
reinforcement learning algorithms in adaptive traffic control efficiency, but it also unveils a pathway for potential cyber-
proves advantageous in optimizing traffic signal timings, attacks.
thereby reducing congestion and improving overall traffic Lin et al. [7] introduced a pair of strategies for targeting
flow efficiency when compared to traditional fixed-timing agents trained through deep reinforcement learning algo-
plans. Furthermore, the application of reinforcement learn- rithms with adversarial examples: the strategically timed
ing techniques enables real-time decision-making for safe attack and the enchanting attack. The strategically timed
and efficient lane-changing and merging manoeuvres, fur- attack is designed to lower the agent’s reward by precisely
ther contributing to improved traffic flow and a decreased targeting it during specific time steps, thus decreasing the
likelihood of accidents. These findings underscore the sig- likelihood of being detected. A novel method was proposed
nificance of integrating reinforcement learning into traffic to determine when to craft and apply adversarial exam-
management systems for safer and more efficient road ples. The enchanting attack lures the agent to specific target
networks. states by combining a generative model and a planning

9964 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

algorithm, generating a sequence of actions that entices the in flow monitoring tasks, achieving load balancing and
agent to follow. Experimental results on agents trained with per-flow monitoring. It utilizes memory-efficient two-stage
DQN and A3C algorithms in Atari games demonstrate the Bloom filters to represent monitoring rules, ensuring sys-
efficacy of the strategically timed attack, achieving simi- tem scalability. The centralized SDN control is employed
lar reductions in reward as the uniform attack with fewer for installing, updating, and reconstructing the filters in the
attacks. The enchanting attack successfully lures the agent switch data plane. Experimental evaluation demonstrates that
towards the designated target states with a success rate DCM achieves high measurement accuracy compared to
exceeding 70%. existing solutions with the same memory budget.
Zhang et al. [5] address the vulnerability of deep reinforce- Zulqarnain et al. [12] focused on active traffic management
ment learning (DRL) agents to natural measurement errors (ATM) systems and their vulnerability to cyberattacks, espe-
and adversarial noises in their observations. They highlight cially with the integration of Internet of Things (IoT) devices.
that these deviations from true states can lead to suboptimal A prototype ATM system and real-time cyberattack monitor-
actions by the agent. While conventional techniques aimed at ing system were developed and evaluated on a section of I-66
bolstering resilience in classification tasks prove ineffective in Northern Virginia. The evaluation demonstrated that the
for Deep Reinforcement Learning (DRL), the authors present ATM system improved vehicle speed, but when subjected to
the concept of a state-adversarial Markov decision process cyberattacks, its effectiveness was negated. The monitoring
(SA-MDP) to probe this issue. They introduce a theoreti- system helped mitigate the impact of cyberattacks, highlight-
cally grounded approach for policy regularization that can be ing the need for revisiting ATM system design for enhanced
applied to various DRL algorithms, such as deep determin- cybersecurity.
istic policy gradient (DDPG), proximal policy optimization
(PPO), and deep Q networks (DQN), suitable for both dis-
crete and continuous action control scenarios. This proposed
technique notably enhances the resilience of DDPG, PPO,
and DQN agents against potent white box adversarial attacks,
encompassing novel attacks introduced within the study.
Additionally, the authors note that the adoption of a robust
policy tangibly enhances the overall performance of DRL
agents across diverse environments.
Artificial Intelligence (AI) and Machine Learning (ML)
algorithms play a crucial role in enhancing road safety
management systems. The Smart Road Traffic Management
System (SRTMS) leverages AI to detect unsafe driving
patterns and promptly inform the authorities. Real-time mon-
itoring of human activities is facilitated through the Internet
of Things (IoT), utilizing sensor equipped IoT devices.
Blockchain (BC) technology automates secure and decen-
tralized information sharing between IoT nodes, while AI
enables intelligent decision-making capabilities, resembling
human cognition. Together, these technologies form a pow-
erful framework for efficient and intelligent road traffic
management [8]. Sheikh et al. [9] provides a comprehensive
overview of Vehicular Ad Hoc Networks (VANETs), cov-
ering their architecture, security, challenges, authentication
schemes, mobility simulation, and safety applications, incor-
porating the latest trends in the field.
Putra et al. [10] describe that the internet network plays a
crucial role in all aspects of modern society, and the concept
of a smart city internet system is vital for addressing urban
challenges. With proper precautionary methods and intelli-
gent monitoring through IoT technologies, such as motion
sensors, ultrasonic sensors, PIR sensors, and speed sensors,
cities can achieve orderly traffic systems, efficient transporta-
tion, and improved safety measures. Yu et al. [11] presents
DCM, a Distributed and Collaborative Monitoring system
for network traffic. DCM enables switches to collaborate

VOLUME 12, 2024 9965

 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

FIGURE 3. RP2 pipeline [14].

the incurrence of penalties over a duration. This process

involves the agent delving into the environment, obtaining
feedback through rewards or consequences, and leveraging
this insight to refine its ability to make informed decisions.

B. INTRODUCTION TO ADVERSARIAL ATTACKS

Instances of adversarial attacks in the realm of reinforcement
learning are illustrated here. The uppermost row demon-
strates instances of adversarial attack within the context of
Atari games. The initial image portrays the unaltered, pristine
game background, whereas the subsequent images show-
case manipulated game backgrounds, labelled as ‘‘adversarial
examples.’’ Remarkably, Huang et al. [13] and colleagues
unveiled that these adversarial examples, which remain
FIGURE 1. Original vs perturbed adversary [4]. imperceptible to humans, wield a considerable impact on
game outcomes. Correspondingly, the lower row offers
instances of adversarial attacks within the sphere of auto-
mated path planning. Much like the preceding row, the
primary image represents the unmodified pathfinding map,
while the ensuing two images exhibit adversarial examples
generated through the addition of noise. Chen et al. [4] found
that under such adversarial conditions, the trained agent
struggled to navigate accurately.

C. MARKOV DECISION PROCESSES (MDPs)

Markov Decision Processes (MDPs) provide a mathematical
framework for modelling sequential decision-making prob-
lems in reinforcement learning. MDPs consist of a set of
FIGURE 2. Stop sign on left is something which doesn’t seem suspicious states, actions, transition probabilities, rewards, and a dis-
to the human eye and the image on the right is perturbed [14]. count factor. The states represent the different configurations
of the environment, actions are the possible choices that
the agent can make, and transition probabilities describe the
III. GENERAL METHODOLOGY likelihood of transitioning from one state to another after
A. INTRODUCTION TO REINFORCEMENT LEARNING taking a specific action. Rewards represent the immediate
Reinforcement learning constitutes a subdivision of machine feedback given to the agent for each action, and the discount
learning that centers on instructing intelligent agents to factor determines the trade-off between immediate and future
execute sequential judgments within an environment. Its rewards [15].
foundation rests upon the notion of acquiring knowledge A Markov Decision Process is defined by a tuple (S, A, P,
through engagement and responses from the surroundings. R, γ ), where:
Within this framework, an agent acquires the skill of execut- S is the set of states.
ing actions to optimize the accumulation of rewards or curtail A is the set of actions.

9966 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

P(s′ |s, a) is the transition probability, representing the

probability of transitioning to state s’ from state s when taking
action a.
R(s, a, s′ ) is the reward function, representing the imme-
diate reward received after taking action a in state s and
transitioning to state s′ .
γ (gamma) is the discount factor, a value between 0 and
1 that discounts the importance of future rewards.
Policy (π):
A policy π is a mapping from states to probabilities of
selecting each action in that state.
Deterministic policy: π(s) = a (a single action is chosen
for each state).
Stochastic policy: π(a|s) (the probability of selecting
action a in state s). FIGURE 4. Policy improvement factors [4].
State-Value Function (V π(s)):
The state-value function represents the expected cumula-
tive reward the agent can obtain from a particular state under Optimal Action-Value Function (Q)
a given policy (as shown in eq. 1). The optimal action-value function represents the maximum
expected cumulative reward the agent can obtain from each
V π(s)
X state-action pair by following the optimal policy (as shown in
= Eπ[ _(∧ {}t = 0) ∧ (∞)γ ∧ t ∗ R(t + 1)|S(0) = s] eq. 6).
(1) Q ∗ (s, a) = 6(s′ ) P s′ | s, a


 
Action-Value Function (Q π(s, a)): ∗ R s, a, s + γ ∗ max Q ∗ s , a
′

′ ′


(6)
The action-value function represents the expected cumula- (a′ )
tive reward the agent can obtain from a state-action pair under
a given policy (as shown in eq. 2). D. Q-LEARNING AND POLICY ITERATION ALGORITHMS
X Q-learning stands as a widely employed algorithm within
Qπ(s, a) = Eπ[ _(∧ {}t = 0)∧ (∞)γ ∧ t ∗ R(t + 1)|S(0) the scope of reinforcement learning, aiming to acquire an
= s, A(0) = a] (2) optimal policy for effective decision-making within a Markov
Decision Process (MDP). Central to Q-learning is the concept
Bellman Expectation Equation for Value Function: of approximating the worth of each state-action combination,
This equation expresses the value of a state as the expected denoted as the Q-value. This Q-value signifies the projected
sum of the immediate reward and the value of the next state, total rewards that an agent can amass by executing a distinct
following the policy π (as shown in eq 3). action from a given state while adhering to a specific policy.
The process of Q-learning involves a step-by-step refine-
V π (s) = 6a π (a | s) ∗ 6(s′ ) P s′ | s, a


ment of Q-values, incorporating observed rewards and the
∗ R s, a, s′ + γ ∗ V π s′



(3) highest anticipated forthcoming rewards. As time progresses,
Bellman Expectation Equation for Action-Value Function: the agent gradually converges toward an optimal policy that
This equation relates the value of a state-action pair to the maximizes the projected cumulative rewards [15].
expected sum of the immediate reward and the value of the Policy iteration is another approach to solving MDPs.
next state-action pair, following the policy π (as shown in eq. It involves two main steps: policy evaluation and policy
4). improvement. In policy evaluation, the value function of a
given policy is estimated by iteratively updating the value
Qπ (s, a) = 6(s′ ) P s′ | s, a ∗ R s, a, s′



of each state based on the expected future rewards. Policy
+γ ∗ 6(a′ ) π a′ | s′ ∗ Qπ s′ , a′ improvement (as shown in fig 4) then uses the estimated value



(4)
function to generate a new policy that is greedily optimized
Optimal Value Function (V ) with respect to the current value function. This iterative pro-
The optimal value function represents the maximum cess continues until the policy converges to an optimal policy
expected cumulative reward the agent can obtain from that maximizes the expected rewards [15].
each state by following the optimal policy (as shown Therefore, the probability for each adversarial point ai can
in eq. 5). be concluded as shown in eq. 7:
V ∗ (s) = max 6(s′ ) P s′ | s, a ∗ R s, a, s′ + γ ∗ V ∗ s′



 X4
a pai = j = 1pjai = ω1 · aie + ω2 · d ′ 1i
(5) + ω3 · d ′ 2i + ω4 · θ ′ i (7)

VOLUME 12, 2024 9967

 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

E. DEEP REINFORCEMENT LEARNING continuous learning and adaptation, adaptive signal timing
Deep reinforcement learning combines reinforcement learn- systems can effectively respond to changing traffic patterns,
ing algorithms with deep neural networks. Deep neural reduce congestion, and enhance traffic efficiency.
networks, often referred to as deep Q-networks (DQNs), are
used to approximate the Q-values in high-dimensional state 3) TRAFFIC CONGESTION MANAGEMENT
and action spaces. Deep reinforcement learning enables the Reinforcement learning techniques can be utilized to address
agent to learn directly from raw sensory input, such as images the challenges of traffic congestion management. Congestion
or sensor data, without explicitly engineering features. It has arises from various factors, such as road incidents, bot-
shown significant success in domains with complex and high- tlenecks, and unpredictable traffic patterns. Reinforcement
dimensional environments, such as playing video games and learning algorithms can learn effective control policies to
controlling robotic systems. mitigate congestion by optimizing traffic flow and rerouting
Deep reinforcement learning algorithms, such as Deep Q- strategies. By considering factors such as traffic volume,
Networks (DQN), Proximal Policy Optimization (PPO), and historical congestion patterns, and incident detection, rein-
Advantage Actor-Critic (A2C), utilize neural networks as forcement learning agents can make decisions that help
function approximators to estimate Q-values or policy func- alleviate congestion hotspots and distribute traffic more
tions. These algorithms use techniques such as experience evenly across the road network. This can result in reduced
replay, target networks, and exploration strategies to stabilize travel times, enhanced mobility, and improved overall traffic
the learning process and improve sample efficiency. By uti- conditions.
lizing the theoretical foundations of reinforcement learning,
including Markov Decision Processes, Q-learning, policy
4) INCIDENT DETECTION AND RESPONSE
iteration, and deep reinforcement learning, researchers and
Reinforcement learning can contribute to incident detection
practitioners can develop intelligent agents that learn opti-
and response systems in traffic management. Timely detec-
mal policies in complex and dynamic environments. These
tion and efficient response to incidents, such as accidents,
foundations provide the basis for understanding and applying
breakdowns, or road hazards, are crucial for minimizing
reinforcement learning techniques to enhance road safety and
the impact on traffic flow and ensuring road safety. Rein-
cyber security in traffic management systems.
forcement learning algorithms can learn to analyze real-time
sensor data, including traffic cameras, vehicle trajectories,
F. APPLICATIONS OF REINFORCEMENT LEARNING IN and environmental sensors, to identify abnormal patterns or
TRAFFIC MANAGEMENT events indicative of incidents. Once an incident is detected,
1) INTELLIGENT TRAFFIC CONTROL SYSTEMS the system can use reinforcement learning to determine
Reinforcement learning can be applied to develop intelli- optimal response actions, such as rerouting traffic, dispatch-
gent traffic control systems that optimize traffic flow and ing emergency services, or implementing traffic diversions.
reduce congestion. Traditional traffic control systems often By integrating reinforcement learning into incident manage-
rely on fixed timing plans or pre-defined algorithms, which ment systems, the response time can be reduced, and the
may not adapt well to changing traffic patterns. Reinforce- overall impact on traffic flow can be mitigated.
ment learning enables traffic control systems to learn from By leveraging reinforcement learning techniques in these
real-time data and make adaptive decisions to improve traffic applications, traffic management systems can become more
efficiency. By modelling the traffic network as an MDP, the intelligent, adaptive, and efficient. These approaches have
reinforcement learning agent can learn optimal control poli- the potential to significantly improve road safety, reduce
cies that dynamically adjust signal timings at intersections congestion, and enhance the overall performance of trans-
based on current traffic conditions, such as traffic volume, portation networks. However, the successful deployment of
congestion levels, and pedestrian demand. This approach can these applications requires careful consideration of real-
significantly reduce delays, improve travel time, and enhance world constraints, system scalability, and coordination with
overall traffic flow. other components of the traffic management ecosystem.

2) ADAPTIVE SIGNAL TIMING IV. GENERAL USE CASES

Reinforcement learning can also be employed for adaptive A. ENHANCING ROAD SAFETY USING REINFORCEMENT
signal timing, where the timing of traffic signals is dynami- LEARNING
cally adjusted based on real-time traffic conditions. By using 1) DEVELOPING INTELLLIGENT DRIVER ASSISTANCE
reinforcement learning algorithms, the traffic signal con- SYSTEMS
troller can learn to optimize signal timings to minimize delays Reinforcement learning can play a crucial role in developing
and maximize traffic throughput. The agent observes the intelligent driver assistance systems (DAS) that enhance road
current traffic state, such as the number of vehicles in differ- safety. DAS leverage sensors, cameras, and other data sources
ent lanes, queue lengths, and approaching traffic, and takes to monitor the surrounding environment and assist drivers in
actions to adjust the signal timings accordingly. Through making safe decisions. By applying reinforcement learning

9968 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

techniques, DAS can learn optimal driving policies by ana- Algorithm 1 Define a function that takes input data as input.
lyzing real-time data and feedback from the environment. For Within the function:
example, reinforcement learning can be used to train DAS Store the pre-processed data in a variable.
to detect and respond to potential collision risks, maintain Return the pre-processed data.
safe distances from other vehicles, and navigate complex traf-
fic scenarios. Through continuous learning and adaptation, Define a function to extract features that takes pre-processed
intelligent DAS can assist drivers in avoiding accidents and data as input.
mitigating risks on the road.
Within the extract features function:
Extract pre-processed data and its features.
2) IMPROVING TRAFFIC FLOW AND REDUCING ACCIDENTS
Store the extracted features.
THROUGH OPTIMAL CONTROL POLICIES
Return features.
Reinforcement learning can be utilized to improve traffic Define a function to apply Detection Algorithm that takes
flow and reduce accidents by developing optimal control features as input.
policies for traffic management systems. By modelling traffic
as an MDP, reinforcement learning agents can learn control Within the function:
strategies that minimize congestion and improve overall traf- Apply the Reinforcement Learning detection algorithm to
fic conditions. These agents can make real-time decisions identify adversarial or malicious content based on features.
regarding traffic signal timings, lane management, and speed Store the detection result.
limits to optimize traffic flow and minimize the likelihood Return the detection result.
of accidents. For instance, reinforcement learning can be Define a function to detect Adversarial Content that takes
employed to determine the most effective signal phasing input Data as input.
and timing plans at intersections, considering factors such Within the function:
as traffic volume, pedestrian activity, and historical traffic Call the pre-process function with input Data as input and
patterns. By continuously learning and adapting to changing store the pre-processed data that will be generated.
traffic conditions, reinforcement learning-based control poli- Call the extract Features function with pre-processed Data
cies can lead to smoother traffic flow and reduced accident as input and store the result as features.
rates. Call the apply Detection Algorithm function with features as
input and store the result.
3) REAL-TIME DECISION-MAKING FOR SAFE AND EFFICIENT Return the stored result.
LANE-CHANGING AND MERGING
Reinforcement learning can enable real-time decision-
making for safe and efficient lane-changing and merging B. CYBER SECURITY OF TRAFFIC MANAGEMENT SYSTEMS
maneuvers. These maneuvers often pose challenges and risks, 1) VULNERABILITIES AND THREATS IN TRAFFIC
especially in congested traffic. By training reinforcement MANAGEMENT SYSTEMS
learning agents with rich sensory data, such as vehicle tra- Traffic management systems are vulnerable to various cyber
jectories, sensor readings, and contextual information, they threats that can compromise their security and integrity.
can learn to make informed decisions regarding when and These systems often rely on interconnected components,
how to change lanes or merge into traffic. The agents can including control systems, communication networks, and
consider factors such as vehicle speeds, distances, and safety data processing platforms. Vulnerabilities can arise from
gaps to make decisions that optimize traffic flow and mini- inadequate security measures, poor network segregation, out-
mize collision risks. Through reinforcement learning, these dated software, or weak authentication mechanisms. Threats
decision-making models can improve the efficiency and to traffic management systems can include unauthorized
safety of lane-changing and merging maneuvers, thereby access, denial-of-service attacks, data breaches, tampering
reducing the chances of accidents and enhancing overall road with traffic signals or sensors, and the injection of false
safety. information. Understanding the vulnerabilities and threats is
By leveraging reinforcement learning techniques in these crucial for developing effective cyber security measures.
areas, road safety can be significantly enhanced. The con-
tinuous learning and adaptation capabilities of reinforcement 2) REINFORCEMENT LEARNING FOR INTRUSION
learning allow for the development of intelligent systems that DETECTION AND PREVENTION
adapt to changing road conditions, learn from experience, Reinforcement learning can be utilized for intrusion detection
and make informed decisions to prevent accidents. However, and prevention in traffic management systems. By analysing
the deployment of reinforcement learning-based systems for network traffic data, reinforcement learning algorithms
road safety requires addressing challenges such as real-time can learn patterns and behaviours associated with normal
processing, ensuring system reliability, and integrating with system operation. Deviations from normal behaviour can
existing transportation infrastructure and regulations. be flagged as potential intrusion attempts. Reinforcement

VOLUME 12, 2024 9969

 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

Algorithm 2 3) ADVERSARIAL REINFORCEMENT LEARNING FOR ROBUST

Define the Environment: SYSTEM PROTECTION
Use the ‘[Link]‘ function to create the environment based on the environ-
ment specifications.
Adversarial reinforcement learning is an approach that
focuses on training reinforcement learning agents to be robust
Define the Defense Agent: against adversarial attacks. In the context of traffic manage-
Create a class called ‘DefenseAgent‘. ment systems, adversarial reinforcement learning can be used
In the ‘_init_‘ function: to develop intelligent agents that can anticipate and defend
Initialize the defense agent with the ‘state_size‘ and ‘action_size‘.
Build the model using the ‘_build_model‘ method.
against cyber-attacks. Adversarial agents can simulate poten-
Implement the ‘_build_model‘ method: tial attack scenarios and learn effective defence strategies to
Create a sequential model using the ‘Sequential‘ class from protect the integrity and availability of the traffic management
‘[Link]‘. system. These agents can detect adversarial behaviour, iden-
Add layers to the model using the ‘Dense‘ class from ‘[Link]‘. tify attack patterns, and take appropriate actions to mitigate
Compile the model using the appropriate loss function and optimizer.
Return the model.
the impact of the attacks. Adversarial reinforcement learning
Implement the ‘act‘ method: techniques help improve the resilience of traffic management
Take the current state as input and return an action based on a chosen systems against sophisticated and targeted cyber-attacks.
strategy (e.g., random action selection). To enhance the cyber security of traffic management
Implement the ‘train‘ method: systems, a multi-layered approach is necessary. It includes
Take ‘state‘, ‘action‘, ‘reward‘, ‘next_state‘, and ‘done‘ as input.
Add code for training the agent using a reinforcement learning
implementing secure network architectures, employing
algorithm (e.g., Q-learning, REINFORCE, etc.). strong encryption protocols, conducting regular security
Set up the Defence Agent: audits, and providing training and awareness programs for
Obtain the ‘state_size‘ and ‘action_size‘ from the environment’s observation personnel. Reinforcement learning techniques can com-
space and action space, respectively. plement these measures by enabling intelligent intrusion
Initialize the defense agent by creating an instance of the ‘DefenseAgent‘
class with the obtained sizes.
detection, real-time threat analysis, and robust defense mech-
Training Loop: anisms. It is important to continuously update and improve
Set the number of episodes (‘EPISODES‘) for the training loop. the reinforcement learning models to keep pace with emerg-
Iterate over the range of episodes. ing cyber threats and ensure the long-term security of traffic
Inside each episode loop: management systems.
Reset the environment and obtain the initial state.
Reshape the state to match the expected input shape of the defense agent’s
model.
V. PROPOSED METHODOLOGY
Set the ‘done‘ flag to False.
Execute the environment steps until the episode is done: Reinforcement learning can be employed for training a
Get the defence agent’s action by calling the ‘act‘ method with the defence mechanism to mitigate adversarial attacks after
current state as input. detection. Here’s an updated outline of the approach:
Execute the chosen action in the environment.
Obtain the next state, reward, done flag, and additional information
from the environment step. A. ADVERSARIAL ATTACK DETECTION
Reshape the next state to match the expected input shape of the defense
Use anomaly detection or pattern recognition techniques to
agent’s model.
Train the defence agent by calling the ‘train‘ method with the relevant identify unusual or adversarial patterns in the input data.
information. Once an attack is detected, trigger the defence mechanism.
Update the current state with the next state. In the main code execution, receive input data from the
Break the loop if the episode is done. vehicle detection system and store it as input Data.
Evaluation:
Call the detect Adversarial Content function with input
After the training loop, add code to evaluate the trained defence agent’s
performance. This involves measuring metrics such as accuracy, precision, Data as input and store the result.
recall, or any relevant evaluation criteria. If the result stored is True, trigger an alert or response
Save the Trained Model: mechanism and take appropriate actions to mitigate the
Save the trained defence agent’s model using the ‘save‘ method from impact of the content.
‘[Link]‘.

B. REINFORCEMENT LEARNING FOR DEFENCE

◦ Create an environment that simulates the interaction
learning agents can continuously update their models and between the defence agent and potential adversarial
adapt to new attack techniques, making them effective attacks.
in detecting previously unseen or evolving cyber threats. ◦ Define actions that the defence agent can take to coun-
Once an intrusion is detected, appropriate response actions teract or mitigate the effects of adversarial attacks.
can be taken, such as isolating the affected components, ◦ Design a reward system that encourages the defence
alerting security personnel, or initiating incident response agent to take actions that minimize the impact of attacks
procedures. and maximize correct detections.

9970 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

VII. PROPOSED USE CASES

A. CASE STUDY: INTELLIGENT DRIVER ASSISTANCE
SYSTEMS (IDAS)
Description: In this case study, a reinforcement learning-
based IDAS was developed to enhance road safety. The
system utilized sensor data from vehicles, such as cameras,
radar, and lidar, to detect potential collision risks and assist
drivers in making safe decisions. Reinforcement learning
algorithms were trained using real-world driving data to learn
optimal driving policies for collision avoidance and safe lane-
changing maneuver’s.
Experimental Results: The experimental results demon-
strated that the reinforcement learning-based IDAS outper-
formed traditional rule-based systems in terms of collision
avoidance and lane-changing safety. Evaluation metrics, such
as collision rates, near-miss incidents, and successful lane
changes, were used to assess the system’s performance.
Comparative Analysis: A comparative analysis was con-
ducted to compare the performance of the reinforcement
learning-based IDAS with traditional approaches, such as
FIGURE 5. Proposed model.
rule-based systems or expert-designed algorithms. The anal-
ysis showed that the reinforcement learning-based approach
◦ Train the defence agent using reinforcement learning achieved higher accuracy, adaptability to diverse driving sce-
algorithms to learn effective defense strategies. narios, and improved overall road safety.

VI. PROPOSED MODEL

This activity diagram as shown in fig 5 provides a visual B. CASE STUDY: ADAPTIVE TRAFFIC SIGNAL CONTROL
representation of the workflow, highlighting the activities and Description: This case study focused on using reinforcement
decision points involved in the vehicle detection system’s learning for adaptive traffic signal control to improve traf-
adversarial attack detection and defense using reinforcement fic flow and reduce congestion. The reinforcement learning
learning. agent was trained to learn optimal signal timing policies based
Steps: on real-time traffic data, including traffic volumes, queues,
1. The process starts with the ‘‘Start’’ node. and historical patterns. The goal was to minimize delays,
2. The ‘‘Collect input data from vehicle detection system’’ reduce travel time, and improve overall traffic conditions.
activity is performed. Experimental Results: The experimental results demon-
3. The ‘‘Vehicle Detection’’ activity is executed. strated that the reinforcement learning-based adaptive traffic
4. The ‘‘Anomaly Detection’’ activity is performed to signal control outperformed fixed-timing plans in terms of
identify any anomalies in the vehicle detection results. traffic flow efficiency. Evaluation metrics such as average
5. The ‘‘Adversarial Attack Detection’’ activity is exe- travel time, traffic throughput, and congestion levels were
cuted to detect adversarial attacks. used to measure the system’s performance.
6. If an attack is detected, the ‘‘Trigger Defence Mecha- Comparative Analysis: A comparative analysis was con-
nism’’ activity is performed. ducted to compare the performance of the reinforcement
7. The ‘‘Reinforcement Learning Environment’’ activ- learning-based adaptive traffic signal control with traditional
ity is created to simulate the interaction between the fixed-timing plans. The analysis revealed that the adaptive
defence agent and potential adversarial attacks. control approach resulted in reduced travel times, increased
8. The ‘‘Design Reward System’’ activity is carried out to traffic throughput, and decreased congestion, outperforming
define the rewards for the defence agent’s actions. traditional fixed-timing plans.
9. The defence agent is trained through the ‘‘Train
Defence Agent’’ activity using reinforcement learning
algorithms. C. CASE STUDY: CYBER SECURITY OF TRAFFIC
10. The ‘‘Evaluation and Testing’’ activity is performed to MANAGEMENT SYSTEMS
assess the performance of the trained defence agent. Description: This case study focused on leveraging rein-
11. The ‘‘Deployment and Monitoring’’ activity is carried forcement learning for cyber security in traffic management
out to integrate the defence agent into the vehicle detec- systems. Reinforcement learning agents were trained to
tion system and continuously monitor its performance. detect and prevent cyber-attacks on traffic management
12. The process ends with the ‘‘Stop’’ node. systems by analysing network traffic data and identifying

VOLUME 12, 2024 9971

 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

anomalous patterns associated with intrusion attempts or C. INTEGRATION WITH EXISTING TRANSPORTATION
malicious activities. INFRASTRUCTURE AND SYSTEMS
Experimental Results: The experimental results demon- Integrating intelligent traffic management systems with exist-
strated the effectiveness of the reinforcement learning-based ing transportation infrastructure and systems can pose chal-
intrusion detection system in accurately detecting cyber- lenges due to legacy systems, interoperability issues, and
attacks with low false positive rates. Evaluation metrics, such coordination between different stakeholders. It is important to
as detection accuracy, false positive rates, and attack identifi- establish open standards and protocols to facilitate seamless
cation time, were used to assess the system’s performance. integration and interoperability among various components.
Comparative Analysis: A comparative analysis was con- Collaboration and coordination between traffic management
ducted to compare the performance of the reinforcement authorities, transportation agencies, and technology providers
learning-based intrusion detection system with tradi- are crucial to ensure smooth integration and maximize the
tional signature-based detection systems or anomaly-based benefits of intelligent traffic management systems.
methods. The analysis showed that the reinforcement
learning-based approach achieved higher detection rates, D. REGULATORY AND LEGAL CONSIDERATIONS
faster response times, and improved overall cyber security The deployment of intelligent traffic management systems
compared to traditional approaches. raises regulatory and legal considerations. Regulations must
In these case studies, the experimental results and eval- address the operation and responsibility of autonomous
uation metrics were used to quantify the effectiveness and systems, data ownership, liability, and privacy protection.
performance of the reinforcement learning-based approaches Developing appropriate regulations and standards to ensure
for road safety and cyber security. Comparative analysis the safe and responsible use of these systems is neces-
with traditional approaches provided insights into the advan- sary. Collaboration between policymakers, regulatory bodies,
tages and improvements offered by reinforcement learning industry experts, and legal professionals is essential to estab-
techniques. The experimental results and comparative anal- lish a regulatory framework that promotes innovation while
ysis highlighted the potential of reinforcement learning in safeguarding public safety and privacy.
enhancing road safety and cyber security, demonstrating its
superiority over traditional methods in terms of accuracy, E. FUTURE DIRECTIONS
adaptability, and efficiency. 1) MACHINE LEARNING EXPLAINABILITY
VIII. PROPOSED USE CASES
Enhancing transparency and interpretability of reinforcement
learning algorithms is crucial for gaining public trust and
A. ADDRESSING ETHICAL AND PRIVACY CONCERNS IN
regulatory compliance. Research should focus on developing
INTELLIGENT TRAFFIC MANAGEMENT SYSTEMS
explainable reinforcement learning models that provide clear
As intelligent traffic management systems rely on data collec-
rationales for decision-making, making them more under-
tion and analysis, addressing ethical and privacy concerns is
standable and accountable.
paramount. It is essential to establish transparent data collec-
tion and usage policies, ensuring that data is anonymized and
2) MULTI-AGENT SYSTEMS
handled in compliance with privacy regulations. Additionally,
attention should be given to potential biases in data collection Traffic management systems often involve multiple agents,
and algorithmic decision-making to avoid discriminatory out- such as vehicles, pedestrians, and infrastructure components.
comes. Developing robust ethical frameworks and engaging Future research should explore reinforcement learning tech-
stakeholders can help address these concerns and build public niques for multi-agent systems to address complex interac-
trust in intelligent traffic management systems tions, coordination, and cooperation among different entities,
thereby improving overall traffic efficiency and safety.
B. ENSURING RESILIENCE AGAINST SOPHISTICATED
CYBER ATTACKS 3) ADVERSARIAL REINFORCEMENT LEARNING
As traffic management systems become more connected Advancing research on adversarial reinforcement learning
and rely on digital infrastructure, the risk of sophisticated can enhance the resilience of traffic management systems
cyber-attacks increases. It is crucial to implement robust against sophisticated cyber-attacks. Developing intelligent
cybersecurity measures to protect against potential vul- agents that can detect and defend against adversarial attacks
nerabilities. This includes implementing strong encryption, in real-time can significantly improve the security and relia-
intrusion detection systems, and continuous monitoring of bility of traffic management systems.
network traffic. Regular security audits and proactive vulner-
ability assessments can help identify and address potential 4) REAL-TIME DATA FUSION AND SENSOR INTEGRATION
weaknesses. Additionally, promoting a culture of cyberse- Integrating data from diverse sources, such as connected
curity awareness and training among system administrators vehicles, traffic sensors, and surveillance cameras, can pro-
and personnel is essential to enhance the resilience of traffic vide a comprehensive view of the traffic environment. Future
management systems against cyber threats research should focus on developing reinforcement learning

9972 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

approaches that effectively fuse and utilize real-time data through reinforcement learning techniques, improving traffic
from multiple sources to improve decision-making in traffic flow and reducing the likelihood of accidents
management systems
Addressing these challenges and advancing research in
these future directions can pave the way for more effective
B. POTENTIAL IMPACT OF REINFORCEMENT LEARNING
and secure intelligent traffic management systems that pri-
The integration of reinforcement learning in traffic manage-
oritize road safety, privacy, and efficiency while complying
ment systems has the potential to revolutionize road safety
with regulations and fostering public trust.
and cyber security. It enables the development of intelligent
systems that adapt to dynamic conditions, learn from real-
IX. CONCLUSION time data, and make informed decisions to prevent accidents
In this research paper, we have explored the potential of and mitigate cyber threats. Reinforcement learning tech-
reinforcement learning in enhancing road safety and cyber niques provide a flexible and adaptive approach, allowing
security in traffic management systems. We discussed the traffic management systems to continuously improve and
theoretical foundations of reinforcement learning, including optimize their performance in real-world scenarios.
concepts such as Markov Decision Processes, Q-learning,
policy iteration, and deep reinforcement learning. We also
examined its applications in traffic management, including
C. RECOMMENDATIONS FOR FUTURE RESEARCH AND
intelligent driver assistance systems, improving traffic flow,
IMPLEMENTATION
and real-time decision-making for safe lane-changing and
1. Ethical and Privacy Considerations: Future research
merging.
should focus on addressing ethical and privacy concerns
To enhance road safety, leveraging reinforcement learning
associated with intelligent traffic management systems. This
(RL) within traffic management systems proves instrumental.
includes developing transparent data collection and usage
Traditional approaches, such as regulations and infrastructure
policies, ensuring fairness and non-discrimination, and pro-
improvements, have limitations in addressing the complex-
tecting user privacy while maximizing the benefits of data-
ities of modern traffic environments. RL offers a dynamic
driven approaches.
solution by enabling the development of intelligent driver
2. Resilience against Cyber Attacks: Further research is
assistance systems. These systems analyze traffic patterns,
needed to enhance the resilience of traffic management sys-
identify risky situations, and make proactive decisions to pre-
tems against sophisticated cyber-attacks. This includes the
vent accidents. By learning from data, RL-based systems can
development of robust intrusion detection and prevention sys-
adapt to dynamic conditions, effectively reducing accident
tems, as well as the application of adversarial reinforcement
occurrences.
learning techniques to anticipate and defend against evolving
Furthermore, in adaptive traffic control, RL algorithms
cyber threats.
optimize signal timings, minimizing congestion and improv-
3. Integration with Existing Infrastructure: Future research
ing overall traffic flow efficiency compared to fixed-timing
should explore methods for seamless integration of intelli-
plans. Real-time decision-making facilitated by RL tech-
gent traffic management systems with existing transportation
niques enhances safe lane-changing and merging maneuvers,
infrastructure and systems. This involves considering interop-
contributing to improved traffic flow and decreased accident
erability, standardization, and coordination among different
likelihood. Integrating RL into traffic management systems
stakeholders to ensure the compatibility and effectiveness of
not only addresses challenges in road safety but also estab-
these systems.
lishes more intelligent, adaptive, and secure systems. These
4. Regulatory and Legal Frameworks: The implementation
findings underscore the transformative potential of reinforce-
of intelligent traffic management systems requires the estab-
ment learning in creating safer and more efficient road
lishment of appropriate regulatory and legal frameworks.
networks.
Research should focus on developing guidelines and stan-
dards to ensure the safe and responsible deployment of these
A. KEY FINDINGS systems, addressing liability, privacy, and security concerns.
1. Reinforcement learning-based intelligent driver assistance In conclusion, reinforcement learning offers immense
systems can significantly improve road safety by detecting potential for enhancing road safety and cyber security in
potential collision risks, assisting in safe decision-making, traffic management systems. By leveraging its capabilities,
and mitigating accidents. we can develop intelligent systems that adapt, learn, and
2. Adaptive traffic control using reinforcement learning optimize their performance to create safer and more efficient
algorithms can optimize traffic signal timings, reduce con- transportation networks. Future research and implementa-
gestion, and enhance overall traffic flow efficiency compared tion efforts should address ethical concerns, strengthen cyber
to traditional fixed-timing plans. security, integrate with existing infrastructure, and establish
3. Real-time decision-making for safe and efficient regulatory frameworks to realize the full potential of rein-
lane-changing and merging maneuvers can be achieved forcement learning in traffic management.

VOLUME 12, 2024 9973

 I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

FUNDING AND/OR CONFLICTS OF AANCHAL SINGH is currently pursuing the

INTERESTS/COMPETING INTERESTS [Link]. degree in computer science engineering
with the Vellore Institute of Technology, Chennai,
The authors declare that they have no competing interests or India. Her research interests include data analytics
personal relationships that could have appeared to influence and deep learning and its applications.
the work reported in this article.

REFERENCES
[1] A. A. Ouallane, A. Bahnasse, A. Bakali, and M. Talea, ‘‘Overview of road
traffic management solutions based on IoT and AI,’’ Proc. Comput. Sci.,
vol. 198, pp. 518–523, Jan. 2022.
[2] M. Botvinick, S. Ritter, J. X. Wang, Z. Kurth-Nelson, C. Blundell, and
D. Hassabis, ‘‘Reinforcement learning, fast and slow,’’ Trends Cognit. Sci.,
vol. 23, no. 5, pp. 408–422, May 2019.
ARAN AGARWAL is currently pursuing the
[3] A. Pattanaik, Z. Tang, S. Liu, G. Bommannan, and G. Chowdhary,
‘‘Robust deep reinforcement learning with adversarial attacks,’’ 2017, [Link]. degree in computer science engineering
arXiv:1712.03632. with the Vellore Institute of Technology, Chennai,
[4] T. Chen, J. Liu, Y. Xiang, W. Niu, E. Tong, and Z. Han, ‘‘Adversarial India. His research interests include data analytics
attack and defense in reinforcement learning-from AI security view,’’ and deep learning and its applications.
Cybersecurity, vol. 2, no. 1, pp. 1–22, Dec. 2019.
[5] H. Zhang, H. Chen, C. Xiao, B. Li, M. Liu, D. Boning, and C. J. Hsieh,
‘‘Robust deep reinforcement learning against adversarial perturbations on
state observations,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 33, 2020,
pp. 21024–21037.
[6] Q. A. Chen, Y. Yin, Y. Feng, Z. M. Mao, and H. X. Liu, ‘‘Exposing con-
gestion attack on emerging connected vehicle based traffic signal control,’’
in Proc. Netw. Distrib. Syst. Secur. Symp., 2018, pp. 1–15.
[7] Y.-C. Lin, Z.-W. Hong, Y.-H. Liao, M.-L. Shih, M.-Y. Liu, and M. Sun,
‘‘Tactics of adversarial attack on deep reinforcement learning agents,’’ SHRUTI MISHRA received the Ph.D. degree in
2017, arXiv:1703.06748.
computer science and engineering from Siksha ‘O’
[8] A. Sharma, Y. Awasthi, and S. Kumar, ‘‘The role of blockchain, AI and IoT
Anusandhan University, Bhubaneswar, Odisha,
for smart road traffic management system,’’ in Proc. IEEE India Council
Int. Subsections Conf. (INDISCON), Oct. 2020, pp. 289–296.
India. She has been an Associate Professor with
[9] M. S. Sheikh and J. Liang, ‘‘A comprehensive survey on VANET security the Department of Computer Science and Engi-
services in traffic management system,’’ Wireless Commun. Mobile Com- neering, Vignana Bharathi Institute of Technology,
put., vol. 2019, pp. 1–23, Sep. 2019. Hyderabad. She is currently an Assistant Profes-
[10] A. S. Putra and H. L. H. S. Warnars, ‘‘Intelligent traffic monitoring system sor (Senior) with the Centre of Advanced Data
(ITMS) for smart city based on IoT monitoring,’’ in Proc. Indonesian Science, Vellore Institute of Technology, Chennai.
Assoc. Pattern Recognit. Int. Conf. (INAPR), Sep. 2018, pp. 161–165. She has more than 30 articles in both national and
[11] Y. Yu, C. Qian, and X. Li, ‘‘Distributed and collaborative traffic monitoring international to her credit along with three books. She has guided more than
in software defined networks,’’ in Proc. 3rd workshop Hot Topics Softw. 40 postgraduate and undergraduate students. She has served as a reviewer for
Defined Netw., Aug. 2014, pp. 85–90. many reputed journals and conferences. She is a guest editor of many reputed
[12] Z. H. Khattak, H. Park, S. Hong, R. A. Boateng, and B. L. Smith, ‘‘Investi- publishers, such as Elsevier.
gating cybersecurity issues in active traffic management systems,’’ Transp.
Res. Rec., J. Transp. Res. Board, vol. 2672, no. 19, pp. 79–90, Dec. 2018.
[13] S. Huang, N. Papernot, I. Goodfellow, Y. Duan, and P. Abbeel, ‘‘Adversar-
ial attacks on neural network policies,’’ 2017, arXiv:1702.0228.
[14] K. Eykholt, I. Evtimov, E. Fernandes, B. Li, A. Rahmati, C. Xiao,
A. Prakash, T. Kohno, and D. Song, ‘‘Robust physical-world attacks on SANDEEP KUMAR SATAPATHY received the
deep learning visual classification,’’ in Proc. IEEE/CVF Conf. Comput. Vis. Ph.D. degree in data mining and machine learning.
Pattern Recognit., Jun. 2018, pp. 1625–1634. His Ph.D. thesis include a detailed classification
[15] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. of brain EEG signals using machine learning tech-
Cambridge, MA, USA: MIT Press, 2018.
niques. He was an Associate Professor with the
Department of Computer Science and Engineering
and the Head of the Department of Information
Technology, Vignana Bharathi Institute of Tech-
nology, Hyderabad. He worked as an Associate
Professor at Centre for Advanced Data Science,
VIT University, Chennai. He is currently a Post-Doctoral Fellow with Yonsei
University, Seoul, South Korea. He is highly engrossed into the areas of deep
learning, image processing, and machine learning. He has many research
ISHITA AGARWAL is currently pursuing the publications to his credit, such as more than 40 research articles, three books,
[Link]. degree in computer science engineering and many book chapters in various peer-reviewed journals. He has guided
with the Vellore Institute of Technology, Chennai, more than 15 master’s thesis. He has also authored two books, such as
India. Her research interests include data analytics, Frequent Pattern Discovery from Gene Expression Data: An Experimental
cloud computing, and machine learning and its Approach (Elsevier) and EEG Brain Signal Classification for Epileptic
applications. Seizure Disorder Detection (Elsevier). He has been a member of various
academic committees within the institution. He is a member of many profes-
sional organizations and society. He has been an active reviewer of various
peer-reviewed journals and prestigious conferences. He has also reviewed
many research articles and books in Elsevier for possible publication.

9974 VOLUME 12, 2024

I. Agarwal et al.: Enhancing Road Safety and Cybersecurity in Traffic Management Systems

SUNG-BAE CHO (Senior Member, IEEE) SACHI NANDAN MOHANTY (Senior Member,
received the B.S. degree in computer science IEEE) received the first Ph.D. degree from IIT
from Yonsei University, Seoul, South Korea, and Kharagpur, India, in 2015, and the second Ph.D.
the M.S. and Ph.D. degrees in computer science degree from IIT Kanpur, in 2019. He has guided
from the Korea Advanced Institute of Science six Ph.D. scholars. He received the MHRD Schol-
and Technology (KAIST), Daejeon, South Korea. arship from the Government of India, for first
He was an Invited Researcher with the Human Ph.D. study. He has published 60 international
Information Processing Research Laboratories, journals of international repute. He has edited
Advanced Telecommunications Research (ATR) 24 books in association with Springer and Wiley.
Institute, Kyoto, Japan, from 1993 to 1995; and His research interests include data mining, big data
a Visiting Scholar with The University of New South Wales, Canberra, analysis, cognitive science, fuzzy decision making, brain–computer inter-
Australia, in 1998. He was also a Visiting Professor with The University face, cognition, and computational intelligence. He was elected as a fellow
of British Columbia, Vancouver, Canada, from 2005 to 2006; and King of the Institute of Engineers and a Senior Member of the IEEE Computer
Mongkut’s University of Technology, Thonburi, Bangkok, Thailand, in 2013. Society Hyderabad Chapter. He has received three Best Paper Awards during
Since 1995, he has been a Professor with the Department of Computer the Ph.D. degree at IIT Kharagpur from the International Conference in
Science, Yonsei University, and a Underwood Distinguished Professor, since Beijing, China; and the other from the International Conference on Soft Com-
2021. He has published over 230 journal articles and over 680 conference puting Applications organized by IIT Rookee, in 2013. He was a recipient
papers. His research interests include neural networks, pattern recognition, of the Best Thesis Award (First Prize Award) from the Computer Society of
intelligent man–machine interfaces, evolutionary computation, and artificial India, in 2015. He is also a Reviewer of Robotics and Autonomous Systems
life. He was a recipient of the Richard E. Merwin Prize from the IEEE (Elsevier), Computational and Structural Biotechnology Journal (Elsevier),
Computer Society, in 1993. He received several distinguished investigators Artificial Intelligence Review (Springer), and Spatial Information Research
awards from the Korea Information Science Society, in 2005, and the (Springer).
GaheonSindoricoh, in 2017. He was also a recipient of the Service Merit
Medal from the Korean Government, in 2022.

MANAS RANJAN PRUSTY received the Ph.D.

degree in computer science and engineering from
the Homi Bhabha National Institute (HBNI),
Indira Gandhi Centre for Atomic Research
(IGCAR), in 2017. Currently, he is an Assistant
Professor (Senior Grade) with the Vellore Institute
of Technology, Chennai, Tamil Nadu, where he
is also deputed as a Research Faculty Member
with the Centre for Cyber Physical Systems. Pre-
viously, he was an Assistant Professor with the
SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu; and
Siksha ‘O’ Anusandhan Deemed to be University, Bhubaneswar, Odisha.
He has also with Tata Consultancy Services (TCS) as an Assistant System
Engineer, from 2010 to 2011. He is Certified NASSCOM Trainer on Analyst
Security Operations Centre (SSC/Q0909). He has published a patent and
many research articles in reputed SCI and Scopus-indexed peer review
journals. His research interests include machine learning and deep Learning.
His ongoing areas of research interests include smart agriculture, health care
diagnosis, disease detection in poultry, and plant disease prediction. He is
an Active Reviewer of standard international journals, such as Springer,
Elsevier, IEEE ACCESS, and Emerald publications.

VOLUME 12, 2024 9975