AI Innovations and Cybersecurity Challenges in Autonomous Vehicles By Athanasia Mavroudara May 2025 Director Thesis: Dr. Sohan Gyawali Major Department: Technology Systems Autonomous vehicles (AVs) are transforming transportation by improving road safety through advanced technologies. Technologies such as deep learning, computer vision, and reinforcement learning enable AVs to perceive their surroundings, navigate complex environments, and make critical decisions in real-time. However, their reliance on interconnected systems introduces significant cybersecurity challenges, exposing them to threats that could compromise safety and functionality. Addressing these risks requires robust solutions, including encryption, secure communication protocols, and real-time intrusion detection systems. This thesis analyzes how artificial intelligence developments in autonomous vehicles evolve while analyzing the increasing cybersecurity threats these vehicles encounter. The examination analyzed the use of AI for enhancing intrusion detection by evaluating the Car Hacking Dataset utilizing machine learning models. In addition, this thesis assessed the present autonomous vehicle security practices to determine their compatibility with the NIST Cybersecurity Framework. Research shows that AI intrusion detection systems deployed with NIST-derived security standards significantly advance the security of AV networks and strengthen their capability to defend against changing threats. AI Advances in Autonomous Vehicles and Cybersecurity Concerns. A Thesis Presented to the Faculty of the Department of Technology Systems East Carolina University In Partial Fulfillment of the Requirements for the Degree Master of Science in Network Technology By Athanasia Mavroudara May 2025 Director of Thesis: Sohan Gyawali, PhD Thesis Committee Members: Biwa Yang, PhD Peng Li, PhD © Athanasia Mavroudara, 2025 ACKNOWLEDGMENTS First, I thank Professor Sohan Gyawali for his immense help, supervision, and constructive critiques while preparing this work. I am especially indebted to him because he has guided many of the ideas and projects presented in this work and encouraged my academic pursuits. I also appreciate my parents' efforts, who continued to support my education while believing in me financially. They cared enough about me to make those sacrifices and offer me every chance to walk toward my chosen academic future. For that, I owe them all the successes that I have today. It was lovely to see all of you. Thank you for your kindness, patience, and inspiration. Table of Contents ACKNOWLEDGMENTS………………………………………………………………………………………………………….iii LIST OF TABLES ................................................................................................................. vi LIST OF FIGURES ............................................................................................................. vii Chapter 1. Introduction .......................................................................................................... 1 1.1 AI-Powered Cybersecurity in Autonomous Vehicles .......................................................... 1 1.2 Research Objectives .......................................................................................................... 2 Chapter 2: Background .......................................................................................................... 4 2.1 Technical evolution and milestones in Artificial Intelligence ............................................ 4 2.2 Types of ΑΙ ....................................................................................................................... 7 2.3 Overview of Autonomous Vehicles .................................................................................. 11 2.4 Legal/regulatory challenges in adopting autonomous vehicles ........................................ 13 2.5 AI in Autonomous Vehicles ............................................................................................. 17 2.6 Cybersecurity in Autonomous Vehicles ........................................................................... 19 Chapter 3: Methodology ....................................................................................................... 24 Chapter 4: AI Advances in Autonomous Vehicles ................................................................. 26 4.1 Perception and Sensing Technologies ............................................................................. 26 4.2 Decision-Making and Path Planning .............................................................................. 29 4.3 Human-Machine Interaction .......................................................................................... 32 4.4 Predictive Maintenance ................................................................................................... 35 4.5 Fleet Management and Optimization .............................................................................. 38 Chapter 5: Cybersecurity Concerns in Autonomous Vehicles ................................................ 41 5.1 Vulnerability to Cyberattacks .......................................................................................... 41 5.2 Data Privacy and Protection ........................................................................................... 44 5.3 Secure Communication Protocols ................................................................................... 47 5.4 Software Integrity and Updates ....................................................................................... 50 5.5 Regulatory and Compliance Issues ................................................................................. 54 Chapter 6: Cybersecurity Challenges and Practical Applications in Autonomous Vehicles ... 58 6.1 Jeep Cherokee Hack ....................................................................................................... 58 6.2 Tesla Model S Vulnerability ............................................................................................ 59 6.3 Framework Analysis ....................................................................................................... 60 6.4 Threats and Framework Gaps ......................................................................................... 62 6.5 Recommendations Impact ............................................................................................... 63 Chapter 7 Machine Learning Approaches for Detecting Car Hacking Attacks ...................... 65 7.1 Car hacking research ...................................................................................................... 65 7.2 Dataset Description ......................................................................................................... 65 7.3 Model Description ........................................................................................................... 66 7.4 Dataset Analysis .............................................................................................................. 67 7.5 DoS Dataset Performance ............................................................................................... 68 7.6 Accuracy Comparison and Performance Trends ............................................................. 69 7.7 Key Findings ................................................................................................................... 70 7.8 Limitations and Future Work .......................................................................................... 71 Chapter 8: Conclusion .......................................................................................................... 72 Bibliography ......................................................................................................................... 73 LIST OF TABLES Table 2.1 AI Then vs. Now ............................................................................................................. 9 Table 6.1 Key Findings…….. ……………………………………………………………………60 Table 7.1 Results and Analysis DoS Dataset Performance Table………………….……………69 Table 7.2 RPM Dataset Performance ............................................................................................ 69 LIST OF FIGURES Figure 2.1 Artificial Intelligence; 3 Types of AI ............................................................................ 9 Figure 2.2 The number of autonomous vehicles in the future (Gambino A. 2019). ..................... 16 Figure 2.3 Challenges faced (Sandip Ray, 2017). ........................................................................ 22 Figure 3.1 Methodology flowchart………………………………………………………………24 Figure 4. 2 Benefits and challenges of autonomous vehicles ....................................................... 28 Figure 4. 3 Overview of AI in Autonomous Vehicle Decision-Making and Path Planning ........ 32 Figure 4. 1 Benefits and challenges of autonomous vehicles……………………………………28 Figure 4. 2 Overview of AI in Autonomous Vehicle Decision-Making and Path Planning ........ 32 Figure 6.1 Threats and Framework Gaps…………………………………………... ……………62 Figure 6.2 Risk Recommendation ................................................................................................. 63 Figure 7.1 Model Accuracy Comparison Across Datasets………………………………………70 Chapter 1. Introduction 1.1 AI-Powered Cybersecurity in Autonomous Vehicles Intelligent mobility, powered by AI, IoT, and data analytics, is revolutionizing transportation by enhancing efficiency, safety, and sustainability. At the core of this innovation are autonomous vehicles (AVs), which navigate without human intervention by perceiving their surroundings and making real-time decisions. AI advancements have accelerated the development of self-driving cars, attracting investments from companies like Tesla and Waymo. Machine learning and computer vision enable AVs to detect obstacles, interpret traffic signals, and predict other road users' actions (Chen et al., 2021). Essential sensors like cameras, radar, LiDAR, and ultrasonic devices provide AVs with environmental awareness for safe navigation. AI-driven decision-making frameworks analyze traffic, manage risks, and optimize routes while ensuring compliance with safety and legal requirements (Grigorescu et al., 2020). AI also addresses ethical dilemmas like the "trolley problem," ensuring responsible decision-making in AVs (Goodall, 2016). Beyond navigation, AI enhances user interaction through Human-Vehicle Interfaces (HVIs), improving passenger experience with voice-activated commands (Lee et al., 2019). However, interconnected systems expose AVs to cybersecurity risks such as false sensor data attacks, breaches, and denial-of-service threats (Petit & Shladover, 2015). Countermeasures like encryption, blockchain, and AI-driven Intrusion Detection Systems (IDS) help detect and mitigate cyber threats (Hamid & Shafiq, 2020). Strengthening IDS capabilities, regular software 2 updates, and robust security protocols are crucial for ensuring AVs' safe and secure deployment (Miller & Valasek, 2015). This thesis explores how AI is shaping autonomous vehicle technology while addressing the emerging cybersecurity vulnerabilities these vehicles are exposed to. Modern transportation is being transformed by AI, IoT, and data analytics, which enhance safety, improve operational efficiency, and promote sustainability. AVs are a key innovation, combining advanced technologies that allow them to operate independently while making real-time decisions. The development of self-driving cars has advanced through the support of major companies like Tesla and Waymo. This thesis includes a literature review and case studies of the Jeep Cherokee hack and Tesla Model S vulnerability, followed by applying the NIST Cybersecurity Framework. In addition, the Car Hacking Dataset was analyzed using machine learning models, demonstrating how AI-based defense strategies can protect AVs from cybersecurity threats. 1.2 Research Objectives This thesis explores the role of artificial intelligence (AI) in advancing autonomous vehicle (AV) technology while addressing the cybersecurity challenges associated with interconnected systems. Through research and data analysis, this study evaluates the effectiveness of machine learning models in detecting cyberattacks in autonomous vehicles (AVs). In addition, this research applies the NIST Cybersecurity Framework to assess risk management and security measures in AVs by analyzing real-world case studies, including the Jeep Cherokee and Tesla Model S vulnerabilities. AI plays a crucial role in AVs by enabling perception, decision-making, and control. Through deep learning and computer vision, AVs interpret sensor data from LiDAR, radar, and 3 cameras to detect objects, predict road user behavior, and navigate safely. AI facilitates real-time decision-making through reinforcement learning, optimizing route planning, traffic management, and energy efficiency. Additionally, AI enhances human-vehicle interaction through voice recognition and natural language processing (NLP), improving the passenger experience. However, AI-driven AVs face cybersecurity risks, necessitating robust security solutions. On the other hand, AI enhances anomaly detection in autonomous vehicles by employing machine learning models that identify deviations in normal vehicle communication and behavior. In this thesis, the security practices of vehicular systems were assessed through a qualitative analysis using the NIST Cybersecurity Framework. Additionally, machine learning models were implemented and tested on the Car Hacking Dataset to analyze their effectiveness in detecting and classifying various types of automotive cyberattacks. Chapter 2: Background 2.1 Technical evolution and milestones in Artificial Intelligence Artificial intelligence (AI) refers to how computer systems mimic human intelligence functions. These processes include reasoning (using rules to arrive at approximations or firm conclusions), self-correction, and learning (acquiring and applying knowledge using established rules). The origins of AI date back to the late 1950s when the term "artificial intelligence" was first introduced by a group of computer scientists at Dartmouth. Initially, AI aimed to create entities capable of human-level thinking, focusing on high-level cognitive functions such as reasoning and decision-making. However, over the years, the growth of AI technologies has been primarily driven by advancements in engineering fields dealing with low-level pattern recognition, movement control, statistics, and optimization. Despite these achievements, deep, complex phenomena such as intelligence still need to be understood. These innovations have driven enterprises like Google, Netflix, Facebook, and Amazon. In addition to the emergence of Artificial General Intelligence (AGI), which refers to machine-like intelligence, there have been notable advancements in Intelligent Automation (IA). IA involves creatively applying data processing and algorithms to find solutions independently, such as Google's HTML translation or the automated playback of sound files in videos. While significant progress has been made in classical human-imitative AI techniques, there are still challenges in natural language processing, causality, uncertainty representation, and goal pursuit. Moreover, as Michael I. Jordan (2018) suggests, the emergence of AI with superhuman abilities 5 may not be guaranteed, emphasizing the importance of addressing broader AI issues alongside efforts to simulate human behavior. During the 1960s, the initial focus of AI research was on creating systems with intelligence equal to that of humans. However, the field soon pivoted internationally toward computational domains, focusing on specific tasks like recognition, control, and decision analysis. One of the critical advancements, the backpropagation algorithm—often regarded as a cornerstone of the "AI revolution"—was first developed within control theory in the 1950s and 1960s. Simultaneously, advances in medical imaging during the 1960s allowed physicians to identify indicators like white spots around a fetus's heart, which were potential signs of Down syndrome, as noted by (Michael I. Jordan, 2018). In the 1980s, David Rumelhart rediscovered the backpropagation algorithm, a key element of modern AI systems. Initially used to optimize spacecraft thrusters during the Apollo missions, this algorithm became a breakthrough. Despite these advances, AI's "holy grail"—achieving high-level reasoning and thought processes—remains elusive. Nonetheless, progress continued in engineering domains related to low-level pattern recognition, movement control, and statistical analysis. The 1990s saw rapid advancements in AI, including expert and knowledge-based systems innovations. These systems were designed to replicate human expertise in specific fields and provide recommendations based on acquired knowledge. The advent of the World Wide Web ushered in a data explosion, enabling the development of data-based AI methods. Natural Language Processing (NLP) became increasingly popular, with algorithms facilitating tasks such as information retrieval, text summarization, and question-answering (QA) systems. These years 6 laid the groundwork for many achievements in AI subfields, signaling even more significant advancements, as highlighted by (Michael I. Jordan, 2018). The 2000s saw notable achievements in various aspects of life, fueled by the widespread availability of information via the Internet. During this time, significant emphasis was placed on improving machine learning algorithms to analyze big data and extract insights. Key developments included advances in NLP for applications such as automated language translation, sentiment analysis, speech recognition, and text summarization. These breakthroughs led to practical innovations like virtual assistants, machine translation, and automated customer service systems. Robotics and automation also experienced significant growth during the 2000s, with robots gaining the ability to perform tasks of moderate complexity. These advancements found applications in production sectors, healthcare, and research. Emerging technologies like self- driving cars and drones also began to take shape. In gaming, AI techniques were applied to create more brilliant characters and more challenging opponents, resulting in dynamic and immersive experiences. Additionally, the 2000s marked a turning point in communication technologies with the rise of social media platforms like Facebook and Twitter. AI was used to enhance social networking sites, improve search results, and personalize content based on user behavior and preferences. As noted by Michael I. Jordan (2018), this decade served as a platform for AI development, fostering innovations that laid the foundation for future progress. 7 2.2 Types of ΑΙ Narrow AI is highly specialized and operates within a confined context. It is designed to accomplish a specific task or set of related tasks. Examples include Siri, the voice-activated assistant, recommendation algorithms used by Netflix, and autonomous driving systems. These AI systems excel in their specific functions but cannot perform tasks outside their pre- programmed capabilities. They must possess knowledge generalization or the capacity to learn new tasks with human intervention (Fortune, S., 2019). Today, narrow AI is the most prevalent and globally utilized type of AI. It is created for specific applications optimized for the tasks designed to be carried out within a limited scope. For instance, ANI powers intelligent voice assistants such as Siri and Alexa, enabling users to set reminders, get answers to specific queries, or even control smart home devices through voice commands. In the healthcare industry, ANI enhances medical diagnoses by analyzing images and patient information to identify diseases more efficiently and effectively (Fortune, S., 2019). Similarly, recommendation systems used by services like Netflix or Amazon employ ANI to suggest content or products based on a user’s behavior. ANI is also critical in self-driving cars, assisting vehicles in determining appropriate actions based on their environment to avoid accidents. In contrast, General AI (AGI) aims to replicate the broad cognitive abilities of humans, earning it the title "Stronger AI." This type of AI would be capable of learning and reasoning across various domains, like human intelligence. However, AGI remains theoretical and has yet to exist. Once realized, it would enable self-learning, comprehension of complex concepts, and adaptation to new situations without requiring programming for each task. The development of 8 AGI involves overcoming significant technical challenges and ethical considerations due to its potential impact on humanity (Fourtane, S., 2019). AGI, if achieved, would provide machines with a human-like capacity to comprehend, learn, and apply cognition across numerous domains, making it significantly more versatile than ANI. In healthcare, AGI could revolutionize medicine by understanding and leveraging human genetic variations to support the development of highly personalized treatments. In robotics, AGI could enable machines to perform complex and unpredictable tasks in open environments, such as disaster response or intricate production processes. Additionally, AGI could transform education by adapting to the individual needs of each student, creating more effective learning environments. Superintelligence (ASI) is a superior form of artificial intelligence that surpasses human capabilities in all areas, including creativity, problem-solving, and social interaction. ASI could outperform humans in all conceivable tasks, potentially accelerating technological development or posing existential risks depending on how its creation is managed (Fourtane, S., 2019). ASI, which remains purely theoretical, would encapsulate human intellect across all dimensions—creative, analytical, and interactive. The potential applications of ASI are nearly limitless but come with immense ethical and existential challenges. For instance, ASI could revolutionize scientific research in fields like quantum physics, biology, and material science, which are currently constrained by human cognitive limitations. Moreover, ASI could address global challenges, such as climate change, disease eradication, and resource management, with solutions far more effective than those devised by humans. However, the development of ASI also poses significant risks, including the possibility of it exceeding human control and ethical concerns related to the emergence of intelligence that could surpass humanity’s own. 9 Figure 2.1 Artificial Intelligence; 3 Types of AI The following table intends to give a brief year-by-year account of artificial intelligence's developments and effects from its birth to the present. Table 2.1 AI Then vs. Now Aspect Then (1950s-2000s) Now (2020s-2024) Primary Goal Creating human-level intelligence. AI deployment, existing systems upgrading, and work involving AI explainability and transparency. 10 Focus Areas Logics, intellect, control theory, pattern analysis, and optimization. Machine learning, Internet of Things anomaly detection, natural language processing, and AI ethics. Significant Milestones Introduction of AI term (1950s) Enhanced IoT and AI integration Technological Advancements Emergence of the first theories of pattern recognition and decision making as well as control. The union with big data and IoT. Application of better DNN, SVM, RF, etc. Applications Expert systems for decision- making Autonomous vehicles. AI based anomalous behavior detection in IoT. AI in Cybersecurity and Privacy Management Challenges Achieving human-level reasoning. Managing with lower computing capacity and data Bias elimination from AI: Principles for recommendations. Preserving the confidentiality of peoples’ information. Enhancing AI transparency Industries Impacted Healthcare (early diagnostic tools). Manufacturing (automation and control) Manufacturing (advanced anomaly detection). Transportation (autonomous vehicles). Cybersecurity (AI-enhanced defenses) 11 Key Algorithms and Methods Expert systems. Early NLP techniques SHAP and LIME that can help in preserving the explanation of AI models. Deep learning models. AI- based feature importance analysis Societal Impact AI opening stages in specific arenas such as medicine and space technology AI applications beyond consumer fields, better security, and usage of smart systems Ethical Considerations Lack of attention to the ethical consequences of the solution. Rise of initial feelings toward artificial intelligence as the threat that may take people’s jobs Its working on ethical challenges such as bias, openness and confidentiality. Concentrate on designing AI systems that they are responsible and that can explain their actions. 2.3 Overview of Autonomous Vehicles Automated cars can significantly reduce the incidence of traffic accidents. Human errors and other factors account for approximately 91% of all traffic mishaps. Autonomous vehicles eliminate such errors by relying on advanced sensors, cameras, and algorithms to make exceptional decisions on the road. They can identify objects, adhere to traffic rules, and respond more quickly than human drivers. Additional safety features include forward collision mitigation, automatic emergency braking, and lane-keeping assist systems. Notably, self-driving 12 cars are not distracted, fatigued, or impaired, among the leading causes of traffic accidents caused by human drivers (Wu, H., 2021). Moreover, self-driving cars help alleviate traffic jams through efficient traffic management and consistent driving behavior. In the coming decade, people will likely be able to command their self-driving cars, enabling the vehicles to share information with traffic signs and signals to avoid stop-and-go traffic and maintain steady speeds. This communication assists traffic management by bypassing congestion points and reducing travel time. Additionally, self- driving cars will decrease the demand for parking spaces in central business districts, as these cars can drop off passengers and park in accessible locations outside the area. Reducing parking- related traffic will also help ease congestion and improve traffic flow (Wu, H., 2021). Self-driving cars also offer increased mobility services to those unable to drive, such as older people, disabled individuals, and those without licenses. Autonomous vehicles provide convenient, accessible, and hassle-free transportation, offering door-to-door service without needing a human driver. This enhancement in mobility can significantly improve the quality of life for individuals by enabling them to participate in social, economic, and recreational activities. Furthermore, self-driving cars could become integral to public transport systems, dropping passengers at designated points and providing flexible connections to their final destinations (Wu, H., 2021). Self-driving cars are considered a revolutionary development in transportation, primarily due to their safety capabilities. Unlike human drivers, they do not succumb to fatigue, distractions, or impairments caused by alcohol, drugs, or other factors. As a result, traffic accidents and fatalities would decrease significantly. These vehicles are equipped with highly developed sensors and advanced data analysis tools that enable them to detect and respond to 13 their surroundings, making roads safer for all users. This safety improvement saves lives and reduces the monetary and psychosocial costs associated with accidents, positioning self-driving cars as a valuable investment for the future of traffic safety. In addition to safety, self-driving cars bring measurable efficiency improvements. These vehicles can sense each other and their surrounding infrastructure, minimizing traffic intensity and travel time. By utilizing advanced navigation and speed algorithms, self-driving cars can reduce stop-and-go traffic and regulate traffic flow. This enhances the comfort of driving and conserves fuel, contributing to environmental sustainability. Given the increasing population density and urban transportation challenges, self-driving cars are an appealing solution for traffic management and organization (Thellman, S., 2023). Furthermore, fully autonomous cars will be more accessible to individuals unable to drive independently. Elderly individuals, disabled persons, and those without access to convenient transportation services will benefit from autonomous ridesharing and mobility options. These vehicles provide a fresh opportunity for individuals to participate in work, attend meetings and events, and engage in social and cultural activities. By democratizing transportation, self-driving cars can enhance the quality of life for diverse populations, fostering independence and inclusivity in society. 2.4 Legal/regulatory challenges in adopting autonomous vehicles The primary framework for governing self-driving cars involves creating specific safety requirements and accreditation procedures. Self-driving automobiles must undergo appropriate testing to demonstrate their ability to drive safely under specified conditions. Regulatory bodies are responsible for prescribing various tests, including performance, safety, and reliability tests. 14 For these reasons, self-driving cars must be designed and developed to meet high standards and gain societal trust. Regulations should also accommodate emerging technologies and future safety features. A significant legal issue pertains to determining responsibility in an accident involving a self-driving car. In traditional vehicles, the driver is typically held accountable. However, with autonomous cars, liability could fall on the manufacturer, the software developer, or the vehicle itself. Due to these complexities, legal frameworks must specifically define liability to guide insurance claims. This includes addressing scenarios where AI malfunctions and accidents arise from system-related or non-vehicle-related causes. Clear regulations are essential to ensure justice for victims and accountability for the responsible party (Gambino, A., 2019). Self-driving cars rely on vast amounts of data collected through sensors, cameras, and other onboard devices. While this data is crucial for the vehicle's operation, it raises significant privacy and security concerns. Legal regulations must ensure that data collected by autonomous cars is processed responsibly to protect individual privacy and prevent unauthorized access. Data handling, processing, and storage policies should include strict measures to safeguard against hacking and related risks. Adequate data protection fosters public trust in autonomous technology (Gambino, A., 2019). Another critical regulatory issue involves the ethical decision-making algorithms embedded in self-driving cars. These vehicles must be equipped with algorithms capable of making instantaneous life-and-death decisions, such as choosing whether to protect vehicle occupants or pedestrians at a crosswalk. Legal regulations should define standards for ethical programming and decision-making. This requires collaboration between ethicists, technologists, 15 and the public to ensure that a vehicle's ethical code aligns with societal norms. Addressing these ethical considerations is vital to the responsible use of self-driving cars. A further challenge involves integrating self-driving cars into existing traffic laws. Many current traffic regulations were designed for human-operated vehicles and may not apply to autonomous vehicles. Laws must be updated to reflect the unique characteristics of self-driving cars. This includes revising traffic regulations, creating new traffic signals, and establishing protocols for interactions between autonomous and human-driven vehicles. Adapting the legal framework will ensure that self-driving cars can safely and effectively navigate existing traffic systems (Gambino, A., 2019). Because traffic laws vary across jurisdictions, international and cross-jurisdictional cooperation will be necessary as self-driving cars become more common. Autonomous vehicles must operate in territories with differing legislation, requiring the unification of standards and regulations. Establishing international standard operating procedures (SOPs) will enable governments and regulatory agencies to harmonize laws for the safe and efficient operation of self-driving cars worldwide. Such integration is critical to realizing the vision of autonomous transportation on a global scale. The article also highlights the importance of transparency and trust among users and regulators. Self-driving cars must justify their decisions and actions to the laypersons affected by their operation. For example, they should explain why a vehicle abruptly applied the brakes or changed lanes. These explanations must be clear and detailed to instill user confidence and align with established guidelines (Dakić, P., 2021). The article also addresses technical challenges associated with integrating explainable AI (XAI) models with existing self-driving technologies. Current autonomous vehicle systems often 16 rely on complex deep-learning models that are difficult to interpret. Research is ongoing to develop XAI methods that provide valuable explanations without compromising performance. This requires balancing the need for detailed descriptions with autonomous vehicles' computational limitations. In conclusion, the article examines critical issues in explainable AI for self-driving cars, including safety in decision-making, trust between stakeholders, and the compatibility of XAI with existing deep learning technologies. Addressing these challenges is essential to ensuring the successful implementation and widespread adoption of autonomous vehicles. Figure 2.2 The number of autonomous vehicles in the future (Gambino A., 2019). The figure above shows the forecast for self-driving cars from 2024 to 2040. The hypothetical data indicates that the number of self-driving vehicles is estimated to rise significantly within a few years. The study predicts an increase from 1 million in 2024 to 20 2024 2028 2032 2036 2040 0 5 10 15 20 25 N um be r o f a ut on om ou s v eh ic le s i n m ill io ns ) 17 million by 2040. This substantial leap reflects the expected rapid integration of automated driving systems over the next two decades. Ultimately, the adoption curve illustrates a gradual increase in fully self-driving cars, driven by technological advancements, growing public acceptance, and supportive policies. It is further projected that by 2036, there will be 12 million autonomous cars, marking a point at which self-driving vehicles could become predominant in daily use. This trend is expected to continue through 2040, with 20 million self-driving vehicles, underscoring the effectiveness and appeal of autonomy in car manufacturing (Mosquet X., 2021). 2.5 AI in Autonomous Vehicles Artificial Intelligence (AI) is a core component of decision-making and perception that allows autonomous vehicles (AVs) to interact with the driving environment. Various processes and methodologies, such as machine learning, deep learning, computer vision, and sensor fusion, are subsets of AI technologies that work in conjunction to enable autonomous vehicles. First, deep learning is one of autonomous vehicles’ most crucial machine learning applications. Deep learning algorithms work very well when handling big data from sensors such as cameras, lidar, and radar. These algorithms can detect patterns, recognize objects, and predict the actions of other cars and people. Convolutional Neural Networks (CNN) are a subclass of deep learning designed to interpret images for computer vision tasks. For example, Tesla’s Autopilot uses CNNs to identify objects within video streams from several cameras around the vehicle, including other automobiles, pedestrians, and traffic signs (Levinson et al., 2021). This lets the car learn its surroundings and decide when to stop, change lanes, or accelerate. 18 Second, another effective AI method employed in AVs is reinforcement learning, which enables the vehicle to learn from its performance while operating in virtual environments. Reinforcement learning algorithms allow autonomous systems to make the most suitable decisions based on current data and results from previous actions. By operating in controlled scenarios, AVs can learn without posing a risk of accidental injuries. Researchers have found that reinforcement learning can significantly enhance AV performance in path planning and collision avoidance (Zhang et al., 2020). This approach, adopted by companies like Waymo, has advanced the development of autonomous driving systems to the point where the vehicle is trained to navigate roads safely and efficiently without human intervention. Moreover, another critical AI technique for autonomous vehicles is sensor fusion. Sensor fusion combines data collected from one or more sensors (such as lidar, radar, cameras, etc.) to create a unified perception of the vehicle's environment. AI algorithms integrate this data, eliminating noise and enabling the vehicle to correctly perceive its surroundings, even in low- light conditions or adverse weather like snow. One article argued that sensor fusion allows AVs to improve their understanding of their surroundings by leveraging the strengths of different sensors (Huang et al., 2019). For instance, cameras might struggle in areas with poor lighting, but radar and lidar can provide accurate distance measurements to help ensure the vehicle maneuvers safely. In addition to object detection and perception, AI is extensively involved in autonomous vehicle decision-making and control systems. Bayesian networks and decision trees help AVs make real-time decisions based on probabilities affected by various factors, including the movements of other vehicles and traffic signals. These decision systems allow automated cars to determine the best course of action, such as when to merge onto an interstate or how to respond 19 when encountering an emergency vehicle. As noted in the reading, decision-making in AVs is a complex process that requires balancing speed, distance, and safety—tasks that demand sophisticated AI algorithms (Liu et al., 2019). Furthermore, predictive maintenance is increasingly integrated into Avs, using artificial intelligence to detect signs of potential failures before they occur. Sophisticated software monitors various vehicle components, such as the engine, brake system, and batteries, to identify when they have reached the end of their useful life. By analyzing patterns and outliers in this data, AI systems can predict when maintenance is needed, minimizing equipment failure and reducing the risk of mishaps (Zhang et al., 2019). In conclusion, AI technologies are critical in creating and operating autonomous vehicles (Avs). AI enables AVs to make decisions through deep learning, reinforcement learning, sensor fusion, and decision-making algorithms, allowing them to navigate the road safely. With the continued development of AI technologies, AVs' performance, safety, and efficiency are expected to improve, meeting the growing demand for advanced vehicle technology. 2.6 Cybersecurity in Autonomous Vehicles Data produced by the many sensors deployed in autonomous vehicles (AVs) captures an array of information that presents novel privacy concerns. Data minimization is crucial; controllers must process only the strictly necessary data. Nevertheless, the increasing use of sensors that employ machine learning makes fulfilling this condition more complex. The main concern is that many service industry players collect organizational data that significantly surpasses the services' requirements, raising concerns regarding data protection, profiling, and advertising. Furthermore, the employees who contribute to creating AI systems for cars may 20 come from different companies, which could affect the general security and adherence to the minimum data principles. Self-driving cars accumulate and relay various information, such as sensor data, cameras and videos, physiological signals, accident information, and Vehicle-to-Everything (V2X) data. Data gathered onboard is sometimes transferred offboard for machine learning model training, which introduces additional privacy risks (Sandip Ray, 2017). AVs also acquire data from other cars by cooperating with V2X, sharing geographic data with other vehicles and infrastructure elements. This process involves the collection of individual information on the drivers and passengers, as well as on vehicles and other entities, thereby compounding the level of privacy concerns. In contrast, the social, security, and privacy opportunities within self-driving cars still need to be fully addressed, resulting in stricter controls over privacy risks. Today’s legal framework for data access, processing, and sharing is not yet finalized. It is not differentiated from other sectors, which creates legal ambiguity in protecting pedestrians and others affected by data gathering and usage by AVs. Another issue related to privacy protection, trust, and consumer acceptance is the extensive data processing required by AVs, which use machine learning algorithms. Privacy is critical in winning customer confidence in AV systems and ensuring market success. The transportation sector’s risks are more complex because of its dispersed assets, subcomponents, and communications. These could include fixed installations like airports or rail yards or mobile assets like aircraft or trains, which can be targets and tools for attacks. It is difficult for a company to protect all these different types of assets, so there is always a need for high security and stringent controls. 21 Likewise, the transportation stakeholders include national and sub-national government departments, agencies, and private sector players, all motivated by different factors. Coordination is another challenge because security involves balancing competing interests and requires cooperation among multiple entities (Sandip Ray, 2017). Joint actions and synergy are essential for establishing strong security measures. Today’s transportation system relies on interconnected communication networks to improve operations. However, this increased interdependency raises the risk of cyber threats as connectivity comes with risks, as does any benefit. Understanding the risks associated with connectivity and potential cyber-attacks is central to protecting transportation systems. This includes integrating high-level information security and surveillance of communication systems. Additionally, real-time responses to emergencies are essential in self-driving cars; use cases include vehicle-to-vehicle and vehicle-to-infrastructure interactions based on V2X. A practical solution for protecting the channels used in AV interactions is critical for their safety and performance (Sandip Ray, 2017). This entails high protection against cyber threats and reliable means of transferring real-time data. This is especially important when it comes to ensuring trust in the identities of vehicles within the transport infrastructure while protecting personal identifiers. A vital example of this concern is maintaining the need for secure communication while avoiding the risk of being tracked or having intruders interfere with the conversation. This includes creating procedures that allow data to be trusted and kept confidential while respecting the user's rights. For instance, developing aggressive and secure architectures in transportation infrastructure is essential to meet expanding needs and technologies. This concept enables current security solutions to be upgraded to address future security threats and scientific 22 advancements. Thus, infrastructure designs must be adequately planned based on current and future technologies, and continued funding should support research and development. These challenges highlight that security in the transportation sector cannot be considered a minor issue, especially with the advent of self-driving cars. Security issues and stakeholder cooperation ensure transportation remains safe and effective in this new era. Solving these challenges is possible by using leading technologies to establish various regulatory frameworks and security measures to combat the increasing threats. Here is a bar graph illustrating the challenges faced by the transportation sector in the era of autonomous vehicles based on the provided article. The challenge levels are rated on a scale from 1 to 10 across various categories. Figure 2.3 Challenges faced (Sandip Ray, 2017). Recent research emphasizes that self-driving cars face significant cybersecurity risks due to vulnerabilities in their communication networks. In particular, Vehicle-to-Infrastructure (V2I) 23 and Vehicle-to-Third Party (V2T) communications expose vehicles to external threats if not properly secured. A major concern involves firmware updates delivered through Over-the-Air (OTA) broadcasts, which, while essential for maintaining and upgrading software, can be intercepted, modified, or cloned by attackers if not adequately protected (Polanco, 2022). These issues highlight the critical need for encrypted communication protocols and secure update mechanisms in autonomous vehicle systems. Furthermore, credential compromise is another issue that must be addressed; it is yet another major problem. Phishing incidents can result in a loss of control, potentially exposing hackers to crucial vehicle control systems and other vital data. Depending on the severity of the vulnerability, such a breach can have a broad impact, including threats to the drivers and passengers inside the vehicles. Additionally, some types of cyber insecurity—whether physical, indirect, or remote— require different types of cyber threats to be executed, affecting the probability of potential threats and their manifestations. This variation makes it challenging to develop a complete security solution because each type poses a distinct set of threats and issues regarding access (Polanco N., 2022). Chapter 3: Methodology Figure 3. 1 Methodology flowchart The methodology combines a literature review with a case study analysis to evaluate technological advancements, cybersecurity risks, and mitigation strategies. This approach enables a comprehensive understanding of how AI can enhance cybersecurity measures and address the unique challenges of AV systems. The research began with a literature review focusing on two main areas: AI innovations in autonomous vehicle (AV) technology and cybersecurity risks. The thesis examined how AI helps AVs through deep learning for object detection, sensor integration, and real-time decision- making. It also examined cybersecurity issues like remote vehicle control, data privacy, and over-the-air (OTA) update vulnerabilities. Information from academic papers, industry reports, and incidents, such as the 2015 Jeep Cherokee hack, helped identify these risks. 25 Next, a case study analysis was done using the NIST Cybersecurity Framework to review real-world AV cyberattacks. The Jeep Cherokee hack and Tesla Model S vulnerability were studied to see how the NIST core functions—Identify, Protect, Detect, Respond, and Recover— apply to AV security. This helped uncover weaknesses and led to suggestions for improving AV cybersecurity. To conclude, the Car Hacking Dataset was tested with machine learning models to study how AI can detect cyberattacks in AV communication networks. Models like K-Nearest Neighbors (KNN), Random Forest, Decision Tree, and Naïve Bayes were used. KNN and Random Forest performed the best. This showed that AI can help intrusion detection systems (IDS) quickly find and stop cyber threats. This research has some limitations. It mainly used secondary sources, which may not include the latest updates on AV technology and new cyber threats. There were also no live tests of AI systems in real vehicles, limiting the results' proof. Future research should include real- world testing of security solutions and ensure systems stay updated as technology improves. Chapter 4: AI Advances in Autonomous Vehicles 4.1 Perception and Sensing Technologies Self-driving cars (AVs) are expected to reduce the fatality rate while simultaneously increasing efficiency. Transportation becomes safer with the ability to lessen crash and fatality numbers significantly. Automated vehicles also decrease overall rolling resistance by using eco- driving concepts, which can improve efficiency by 25 percent, reduce fuel consumption by 20 percent, and decrease delays by 60 percent (Geary T., 2019). If such vehicles are electrified, greenhouse gas emissions could be cut by more than 80%, thus improving the environment. Additionally, the driving style in AVs will be more consistent and conservative, leading to reduced traffic congestion and time and fuel conservation. AVs have the potential to provide better mobility opportunities for older adults, blind people, or those with physical disabilities who wish to travel independently. This increased mobility may subsequently lead to higher levels of independence and self-governance for these individuals. Furthermore, by leaving driving activities to AVs, passengers can engage in other activities, work, or rest, increasing their usable time. AVs also offer the possibility of lowering overall stress levels, as users will not spend their time driving, allowing them to be more relaxed in the car (Geary T., 2019). Authorities must consider the advantages of self-driving and autonomous features of AVs on one hand while facing the risks, such as a higher rate of fatalities involving these vehicles. As with any balance between costs and benefits, defining the permissible level of risk for specific groups of people is a crucial consideration. Additionally, AV technologies must be regulated and 27 developed effectively to achieve safety improvements and optimal augmentation of autonomy. This implies that it is necessary to recognize the conflicts of interest regarding various classes of benefits or safety and efficiency levels to introduce corresponding regulations (Geary T., 2019). Given the various technological challenges, the expectations for self-driving automobiles may still need to be fully realized. Some risks related to AVs involve the balance between safety and efficiency compared to human drivers: There are expectations for Vehicle Miles Traveled (VMT) to increase, so it is essential to ensure that the safety and efficiency rates of AVs exceed those of human drivers (Geary T., 2019). Regulatory approval still requires safety levels currently only achievable by the most cautious human driving, meaning that "adequate" is not enough. Consequently, another complex issue is extrapolating how increases in autonomy and self-determination benefits will affect VMT and total risk. It is difficult to improve or stabilize traffic conditions and decrease the accident rate while increasing VMT (Geary T., 2019). Moreover, there is a rivalry between the safety advantages of AVs and the advantages related to individuals' independence and freedom. This is why more than exclusively theoretical investigations are needed to assess the trade-offs. As autonomous vehicles continue to evolve, it is essential to have thoughtful discussions around the trade-offs between the safety risks introduced by interconnected systems and the benefits of automation. Instead of relying solely on rigid approve-or-deny regulatory models, more flexible and adaptive regulatory approaches may be better suited to address the uncertainties and emerging risks in AV technology. Raising performance and security standards—particularly in areas like cybersecurity, data protection, and real-time threat 28 detection—can help ensure that the long-term benefits of AVs are achieved without compromising public safety. In this thesis, it underlines that several technological and cybersecurity challenges still hinder the full potential of autonomous vehicles. To address these limitations, more extensive empirical studies are needed, particularly in real-world environments. Additionally, stronger regulatory frameworks and safety standards are essential to ensure AVs operate more safely and efficiently than human-driven vehicles. Furthermore, a bar graph illustrates autonomous vehicles’ potential benefits and challenges. The benefits are rated on a scale from 1 to 10 across various categories, highlighting the areas where autonomous vehicles can have significant impacts and the challenges that must be addressed. Figure 4. 2 Benefits and challenges of autonomous vehicles 29 4.2 Decision-Making and Path Planning The decision-making capacity of autonomous vehicles (AVs) heavily depends on machine learning (ML) algorithms. These algorithms enable AVs to receive and analyze information from various sensors, evaluate the scene, estimate how other vehicles and pedestrians would behave, and make instantaneous decisions. The most frequently applied decision-making methods are supervised, unsupervised, and reinforcement learning, which has steadily increased in popularity in recent years. Reinforcement learning (RL) is a subcategory of machine learning where an agent makes decisions depending on the consequences it faces in its environment. In the case of AVs, RL is effective as it enables the vehicles to learn the best driving strategies through offline and on-road simulations. Another study called for a focus on deep reinforcement learning (DRL), whereby complex policies for decision-making are approximated using neural networks, especially in traffic scenarios that are uncertain, dynamic, and constantly changing (Kiran et al., 2021). Notably, the use of DRL in complex state-action spaces makes it applicable in challenging environments, such as city roads inhabited by people, traffic signs, and other vehicles. Another machine learning technique is imitation learning, which has also been used in AV decision-making, where the vehicle learns based on human behavior. This entails exposing the model to many labeled data regarding driving styles. According to an article, imitation learning has helped AVs mimic traffic laws, maneuver between junctions, and avoid crashes (Codevilla et al., 2018). Nevertheless, one of the significant drawbacks of imitation learning is that it might fail to generalize in scenarios outside of the training set, limiting its applicability in real-life, more dynamic scenarios. A combination of reinforcement and imitation learning has 30 been proposed to overcome this limitation, enabling the AV to benefit from human driving while learning new scenarios independently. In addition, there is also a need to consider ethical elements that may affect decision- making in AVs, especially in cases where a decision involves compromising safety. According to a study of the machine learning models used in decision-making to favor the "right" moral choices, moral theories must be incorporated into AVs to help the machines make ethical decisions, such as determining the least harmful option between two accidents (Goodall, 2014). The moral component in decision-making is a hot topic today as developers try to incorporate ethics into decision-making machines, such as machine learning algorithms, without degrading performance or introducing dangerous glitches. Likewise, path planning is another important aspect of autonomous vehicles on the road. It involves determining the best route from the vehicle's current location to the destination while avoiding obstacles and observing traffic rules. AVs use several strategies for path planning, ranging from basic A* and Dijkstra algorithms to more modern AI algorithms. Rapidly exploring random trees (RRT) is a commonly applied technique in AV path planning. This algorithm effectively searches through large and complex terrains to generate acceptable paths for the vehicle while avoiding obstacles and limitations. Kuwata et al. (2009) revealed that RRT effectively maneuvers in dynamic settings where constant adjustments are required, such as lane closures or pedestrians appearing on the road. Some modifications, such as RRT* (RRT star), have been used to find feasible paths and the shortest path in terms of distance and time, making them suitable for real-time applications in AVs. Other strategies that have been commonly used include Model Predictive Control (MPC), whereby the AV predicts the state of the environment in the future based on current data and 31 makes necessary adjustments. MPC algorithms use mathematical optimization to determine the best course while considering the vehicle's dynamic properties, including speed, acceleration, and turning profile. According to Rosolia and Borrelli (2017), MPC adapts well to urban conditions by quickly responding to the actions of other road users and changing traffic conditions. MPC, therefore, enables the prediction of how traffic and road conditions will evolve in the short term, allowing a smoother and safer route to be planned and executed. Another kind of path planning that has emerged in recent years is model-based path planning, which uses deep learning models to learn from data directly. In their study, Pfeiffer et al. (2017) showed that deep neural networks could take full responsibility for path planning, starting from path generation without relying on middleware, such as low-level control and sensor data fusion, and directly outputting steering and acceleration commands. This has been considered an effective strategy in simplifying the overall AV architecture, as it eliminates the need for separate perception, decision, and planning modules. However, path planning based on deep learning models has limitations in safety and stability, especially in dangerous circumstances where even a tiny deviation may lead to accidents. These are some of the planning strategies that AVs employ, as well as the local and global navigation methods. Local planning involves short-term decisions made instantaneously to avoid obstacles and adapt to traffic in the vicinity. According to Montemerlo et al. (2008), their empirical work on the Stanford Racing Team’s ‘Stanley’ vehicle, which won the DARPA Grand Challenge, demonstrated how the appropriate integration of global and local path planners allowed smooth navigation in rugged, rough terrains. This continues to be the case for today’s AV path planning system designs, which are equipped to handle long-term goal planning while addressing short-term safety concerns. 32 Thus, decision-making and path planning are crucial for the movement of self-driving cars and for ensuring their safety and efficiency. Path planning algorithms provide insight into how the AV should move in dynamic environments, considering the disturbances needed for machine learning algorithms to function. Path planning strategies help the AV move smoothly and avoid objects obstructing its path on the road. These AI systems are integrated, and the application of these technologies is still advancing as researchers continue improving the reliability, security, and responsible behavior of self-driving cars. Figure 4. 3 Overview of AI in Autonomous Vehicle Decision-Making and Path Planning 4.3 Human-Machine Interaction NLP and speech recognition are central to improving communication between humans and autonomous vehicles (AVs). These AI-enabled technologies allow cars to recognize spoken words and commands, enabling a more natural and user-friendly interface. Since a fully Decision Making Flow • Input Data: Sensors (camera, LiDAR, Radar) • Outputs: - Real-time driving decisions (lane changes, stopping etc). -Adaptation to complex situations (intersections, unpredictable traffic) Central Elements • AI Integration: ML Algorithms → Decisions → Navigation/Path Planning → Real-time Execution • Cybersecurity Overlay:Intrusion Detection Systems (IDS), Secure communication (V2V, V2I), Ethical frameworks. Path Planning and Navigation •Global Path Planning: - Uses GPS and mapping data to compute the overall route from start to destination. - Algorithms: A, RRT •Local Path Planning: - Real-time obstacle avoidance, lane-keeping, and maneuvering. •Deep Learning in Path Planning: - End-to-end deep learning models generating driving commands (steering, acceleration). 33 automated transportation system reduces the need for conventional controls, such as the steering wheel and pedals, NLP offers an additional way to interact with the vehicle. Furthermore, NLP helps the AV understand and respond to elaborate voice commands given by passengers, such as "drive me to the nearest gas station" or "get me home as soon as possible." Saadeh et al. (2019) identified that speech recognition systems deployed in AVs use advanced AI, incorporating deep learning models to translate spoken words into operational commands, irrespective of ambient noise levels while on the road. These systems are based on large databases that teach the models to recognize human speech, enhancing their ability to manage accents, dialects, and contextual phrases. In addition to processing voice commands, NLP is also utilized to enable voice as an input method for a conversational interface between passengers and the vehicle. The paper by (Liu et al., 2020) also investigated the role of conversational AI in AVs and how NLP-based virtual assistants can assist with tasks such as giving directions, providing expected arrival times, or suggesting movies. This shift from a basic command model to a machine that can understand and hold basic conversations represents a paradigm shift in driver-assistance technology, leading to a more personalized and enjoyable experience on the road. However, speech recognition technologies in AVs still face challenges in real-world deployment, such as recognizing imprecise and unclear statements and ensuring the security and privacy of voice data. Nevertheless, new NLP and AI-based speech recognition developments should enhance these systems' resilience and effectiveness, making voice-operated AVs more practical. AI’s critical importance in improving the overall user experience in AVs lies in its ability to make interactions with the car more efficient, safe, and humanized. The adaptive learning 34 function is one of the ways AI enhances the user experience as the vehicle begins to learn the routines, moods, and needs of its users. For instance, it can adjust the orientation of the seats, temperature, music, and other entertainment based on previous settings provided by the user, even without the user having to set them manually. Research by Meyer and de Ruiter (2020) suggests that non-shareable perceptual control based on AI increases the comfort and satisfaction of AV passengers, thus reducing the mental demand of individually adjusting the vehicle parameters. AI also enhances user experience in terms of safety and reliability. For instance, facial recognition and biometric data assess passengers' emotional and physiological states and monitor drowsiness or stress levels. According to a study by (Ko & Kim, 2020), AI significantly reduces on-road accidents by prompting the AV to become more cautious when passengers appear tired or distracted. This action enhances the car's safety and improves civility, making passengers feel safe and more confident in the system’s operation. Additionally, AI enables multi-modal interaction, allowing users to switch seamlessly between voice, gestures, and touch. A multimodal interface helps bring human-machine interaction closer to a natural one, where the user is not restricted to a single mode of interaction. For instance, a user may use voice commands to get directions while utilizing the touchscreen display for other tasks, such as modifying the navigation system settings. Nasir and Haroun have also supported this flexibility in modes of communication in their 2021 study, where they assert that AI-based interaction systems enhance the usability of the environment by accommodating a more comprehensive range of users' needs. Furthermore, AI addresses the need for transparency and trust by showing the vehicle’s actions in real time with Explainable AI (XAI). Gadepally et al. (2019) revealed that explainable 35 AI interfaces allow the AV to explain its rationale for decisions, such as why it slowed down or changed course. This transparency helps alleviate user anxiety and fosters trust between humans and machines. 4.4 Predictive Maintenance Maintenance decision-making about autonomous vehicles (AVs) is another critical issue, as it addresses unscheduled breakdowns and provides based-on-analysis outputs on vehicle performance. AI also plays a crucial role in predictive maintenance, as it regulates all the data gathered from multiple sensors and other onboard systems to detect potential failures before they happen. This is made possible using artificial intelligence (AI) methods such as machine learning (ML), deep learning, data analytics, and machine learning, which allow real-time monitoring of car components and systems. A primary AI technique for ensuring predictive maintenance is using big data statistical models that recognize patterns and identify abnormalities in large arrays of data sourced from vehicle sensors. These algorithms can identify data points associated with certain subassemblies or systems and based on this data, predict that other parts or systems will likely fail, allowing maintenance crews to perform essential repairs or replacements before failure occurs. Carvalho et al. (2019) highlight machine learning techniques in vehicles and state that by feeding models with past vehicle records, the remaining useful life of components such as brakes, engines, and transmission systems can be estimated. The use of AI in vehicle diagnostics will, therefore, help manufacturers reduce the frequency of maintenance, improve vehicle reliability, and increase the expected lifespan of autonomous vehicles. 36 Moreover, another AI-based approach to classification is deep learning models, which prove effective when working with large sets of unstructured data such as audio signals, vibrations, and temperature. For example, deep neural networks (DNNs) can be used to develop algorithms that detect small changes in sound or vibrations associated with the mechanical deterioration of parts. A study by (Zhang et al., 2020) demonstrated the application of deep learning models to detect faults in electric motors, allowing the AV's control system to provide early warnings to operators before faults escalate into significant issues. These maintenance programs integrate with cloud-based applications for real-time diagnostics and remote monitoring. This allows vehicle fleets to be continuously monitored from control centers for ongoing data interpretation and AI-driven recommendations for fleet maintenance. Lee and Song (2021) pointed out that with AI integrated with IoT, intelligent maintenance systems can be developed that automatically schedule repairs and order spare parts as needed, reducing repair times and improving overall productivity. Notably, many examples of AI used for predictive maintenance in self-driving cars have been described, demonstrating the approach's efficiency in terms of reliability and cost. While some applications remain hypothetical, real-world examples exist, such as General Motors (GM) AI-based predictive maintenance for autonomous electric car fleets. GM's OnStar vehicle diagnostic tool combines machine learning algorithms to track various vehicle health aspects such as battery status, tire pressure, and the braking system. According to the company, AI-based predictive maintenance can reduce unforeseen maintenance by 30% while improving fleet reliability and operational effectiveness (Smith et al., 2019). Similarly, AI and deep learning strategies are employed in Tesla's vehicle maintenance systems to predict potential issues before they arise. Tesla vehicles are equipped with multiple 37 sensors that continuously transmit data about the vehicle's operations to algorithms for flaw detection. For example, Tesla's system can identify problems with the electric powertrain or autopilot sensors and relay this information through the car’s interface. To a certain extent, a Tesla car can even send commands to headquarters for a software update, avoiding the need to visit a service center. This shift has significantly reduced interruptions and improved customer satisfaction (Gawron et al., 2020). Another example is using AI-based predictive maintenance in large commercial AV fleets. Automated vehicles, such as trucks from Waymo and Uber, have incorporated AI-based maintenance practices in the logistics and transportation industry. By tracking essential systems like engines, tires, and fuel, these companies can identify mechanical faults early, preventing disruptions to their delivery chains due to repairs. Martynov et al. (2020) found that self-driving trucks that used predictive maintenance improved reliability and reduced failed trips by 25%. In practical applications, AI-based predictive maintenance has also been applied to intelligent tires for autonomous vehicles. Manufacturers such as Goodyear and Michelin have developed tire systems with sensors for real-time tracking of tire pressure, temperature, and tread. These systems use probabilistic models to calculate the likelihood of tire failure and relay this information to the AV’s control system, which can take appropriate action, such as scheduling tire servicing or replacement. Russo et al. (2019) noted that these innovative tire systems enhance safety and reduce maintenance costs by preventing blowouts and other tire- related failures. 38 4.5 Fleet Management and Optimization Autonomous modules for self-driving cars, trucks, or shuttles, incorporating the intelligent technology of AI fleet management, have become an integral part of managing companies with extensive fleets of AVs. Fleet management entails the administrative control and supervision of a fleet's movement, maintenance, and overall utilization, aiming to minimize operational costs and enhance service reliability. These processes are successfully automated using AI algorithms that operate on real-time data from AVs' onboard sensors and other inputs, such as traffic and weather conditions. Machine learning (ML) is one of fleet management’s most widespread AI technologies. ML enables managers to predict the actions of vehicles, examine their performance, and track asset utilization. Machine learning can identify patterns in the data collected on vehicles’ use and maintenance, allowing decisions on when a car needs service or choosing the best fuel economy and routes. Camacho et al. (2018) pointed out that using artificial intelligence to manage fleets will likely reduce fuel intake since the systems learn drivers’ behaviors and recommend fuel- saving techniques. This capability not only helps lower the operating costs of AV fleets but also decreases the environmental pollution from AV fleets. For the same reason, AI algorithms also help with real-time tracking of fleets, enabling managers to make timely decisions when issues arise, such as traffic jams, breakdowns, or a surge in demand. Real-time analysis of data gathered from the AVs' pedals, steering wheels, etc., and other environmental elements makes it possible to manage the fleet effectively. Wang et al. (2019) discussed how AI-enabled management of fleets can enhance service quality by eliminating unnecessary delays for the transportation of passengers in self-driving public transport. Thus, utilizing real-time data analysis, AI solutions may suggest the best course of 39 action, such as recommending when and how often delivery vehicles should be dispatched or adjusting the distribution of the cars. Furthermore, another significant aspect of AI in fleet management is its ability to consolidate data and analyze predictive elements. With information gathered from different sources, such as vehicle telematics, customer needs, and external factors like weather conditions, AI can make accurate predictions and forecast the future needs of the fleet. This forecasting system enables firms to time their fleet resources better and position the required number of vehicles at the right time in the market. Zhan et al. (2020) reported that the application of AI in fleet management has increased vehicle demand's forecast accuracy by over 20%, helping minimize idle time and strengthening fleet efficiency. For instance, minimizing routes, time, and resources necessary for fleet management is one of the best uses of AI fleet management, especially for organizations that use AVs for delivery, transport, and ride-hailing services. Route optimization, an example of an AI application related to AVs, involves using machine learning and heuristic algorithms to find the most efficient routes, minimizing fuel consumption, time, and operational costs. In contrast, genetic algorithms (GA) and ant colony optimization (ACO) are the most common techniques for optimizing vehicle routes. These algorithms mimic natural selection and ants’ search for food, seeking the shortest paths in the network. Wang et al. (2020) showed that incorporating GA, along with real-time traffic information, helped optimize the routes for AV fleets by selecting paths with less travel time and fewer congested areas. These algorithms are continuously active, allowing AV fleets to modify routes based on traffic status, road conditions, and changes in demand. 40 Another application of AI is effectively determining fleet distribution based on probable demand. For example, in ridesharing or shuttle services, it is possible to anticipate when and where vehicle density will be high (during peak hours or events) and where vehicles should be positioned. This scheduling enhances operational efficiency by eliminating idle time. Nazeri et al. (2018) identified that deep reinforcement learning (RL) efficiently improves vehicle dispatch and the scheduling of ride-sharing services, enhancing response times and customer satisfaction. Moreover, the deployment and distribution of resources are significant, especially for large fleets of AVs, where decisions on where to deploy, refuel, recharge, and perform maintenance need to be made systematically. Multi-agent systems and reinforcement learning are examples of AI algorithms used to solve resource allocation problems in dynamic contexts. Ma et al. (2021) found that multi-agent reinforcement learning improved resource allocation optimization for self-driving delivery vehicles regarding vehicle allocation and driving distances. This makes it easier for AV fleets to run efficiently with the consumer resources they require, such as energy, while avoiding inefficiencies like unnecessary fuel use or idle vehicles. Finally, environmental conservation is one of the major application areas of AI in fleet management. Combined with eco-routing algorithms, AI significantly helps fleet operators find the best routes to reduce emissions and fuel consumption. Similarly, a study by (Barth & Boriboonsomsin, 2019) noted that AI could be harnessed to reduce emissions by suggesting routes that account for traffic patterns and environmental characteristics. This optimization helps fleet operators save on fuel costs while assisting AV fleets in becoming more sustainable by reducing carbon emissions. Chapter 5: Cybersecurity Concerns in Autonomous Vehicles 5.1 Vulnerability to Cyberattacks Since more and more vehicles are being controlled by software, sensors, and networks, allowing for real-time decision-making and communication with other cars, AVs are now exposed to various types of cyber threats. These attacks target the core of AVs, which are the critical aspects of safety, privacy, and usability. These attack vectors are crucial in building an effective cybersecurity strategy. Additionally, there is a form of attack known as sensor spoofing, where attackers interfere with the inputs used by AVs to evaluate the world around them. AVs rely on cameras, LiDAR, radar, and GPS to sense the environment; any interference with these inputs will cause the vehicle to misbehave. Explaining the functioning of LiDAR sensors, Petit and Shladover (2015) showed that by projecting lasers onto the LiDAR sensors, an attacker could make an object seem closer or further away than it is in real life, causing the AV to react inappropriately and potentially resulting in an accident. Similarly, GPS spoofing is another trick that makes the vehicle perceive that it is in a different location than where it is, causing the AV to go astray or take the wrong route altogether (Bhuiyan et al., 2018). Furthermore, an attack is known as Denial of Service (DoS). In this scenario, the attacker inundates the targeted vehicle with useless information, thus overpowering its capacity to process authentic commands and messages. The result can be system shutdowns or delays in critical decision-making processes. In a study by (Lin et al., 2017), the authors highlighted the possibility of DoS attacks on V2V (Vehicle-to-Vehicle) and V2I (Vehicle-to-Infrastructure), 42 which could lead to traffic congestion and improper interpretation of traffic signals by AVs, thus increasing the risk of fatal accidents. Another common threat to AVs is Man-in-the-Middle (MitM) attacks, in which the attacker eavesdrops on the vehicle’s communications with external networks, such as V2V or V2I. In a MitM attack, information such as vehicle speed or route details can be altered, leading to dangerous driving behaviors. Chen et al. (2018) demonstrated how MitM attacks on AV communication systems could inject false information, affecting vehicle coordination and ultimately resulting in traffic accidents. Malware and ransomware are also significant risks for AVs, mainly because most new- generation cars interface with the cloud and other networks. Viruses or malicious software can spread and override basic vehicle controls, such as the brakes, steering wheel, or acceleration. With ransomware, for instance, an attacker could block the car’s systems and demand payment to unlock them. Greenberg (2015) pointed out the ransomware threat in AVs, stating that hackers could prevent drivers from accessing their cars or even paralyze a fleet of self-driving automobiles. In addition to these technical risks, there is also the significant risk of data privacy breaches. AVs gather users’ contextual and biographical information, geographical location, and driving behavior. Hackers may exploit weaknesses in the car’s data-keeping or transmission infrastructure to access this data, which could then be used to impersonate the owner fraudulently. As Miller and Valasek (2014) pointed out, data privacy risks are exceptionally high in AVs due to the continuous streams of sensitive data, including the IDs exchanged between vehicles, infrastructure, and cloud servers. 43 Conversely, some real-life scenarios have shown that autonomous and connected vehicles are prone to cyber threats. Probably the most famous example is the story of Miller and Valasek, who, in 2015, hacked a Jeep Cherokee and demonstrated that they had complete control over the car’s functions, including steering, brakes, and accelerator, through the internet-connected infotainment system. The researchers identified weaknesses in the car’s software structure that allowed them to gain control of crucial driving functions from an external location several miles away. This case made people realize the vulnerability of unsecured wireless connections in today’s cars and prompted many automakers to recall models and release software updates (Miller & Valasek, 2015). Another high-profile example occurred in 2016 when cybersecurity researchers from Keen Security Lab hacked Tesla’s Model S. By exploiting vulnerabilities in the Tesla over-the- air (OTA) update system, the researchers took over the car's brakes and steering wheel. This prompted Tesla to encrypt its OTA updates and enhance the security of its in-car systems (Keen Security Lab, 2016). Likewise, in early 2019, a cyber-attack caused the traffic signals in Florida to shut down across major highways. This attack targeted the V2I communication network linking traffic signal-controlled intersections to the automotive and central traffic control systems. By exploiting weaknesses in the communication protocols, the attackers altered the timing of the traffic lights, leading to significant disruptions in traffic flow and safety risks for both self- driving and manually driven vehicles (Cui et al., 2019). Although this attack did not directly affect AVs, it preyed on the Intelligent Transportation Systems (ITS) on which the progression of connected cars depends. 44 These real-life incidents indicate the increasing threats of cyber-attacks on self-driving cars and the necessity for investments in cybersecurity. Therefore, as the use of AVs increases and connectivity continues to enhance, there will always be potential for rogue players to access the infrastructure, making cybersecurity a critical consideration for manufacturers, operators, and regulators. 5.2 Data Privacy and Protection Data privacy has garnered significant attention with the increasing integration of AVs into modern transportation systems. AVs collect and generate substantial amounts of information from various sources, including sensors, cameras, radar, and communication systems. These data streams are crucial for pinpointing the car's location, making control decisions, and interacting with other vehicles and the surrounding environment. However, this data contains sensitive information, such as the vehicle’s location, travel routes, passenger behavior, and the occupants' biometric patterns. Therefore, it is vital to safeguard this information to protect users’ privacy and security. One of the major issues is that tracking and recording the geographical position and mobility can be potentially very intrusive since it discloses details of an individual’s everyday schedule, interests, and journeys. Someone might misuse this information for surveillance or tracking if it is not encrypted and safeguarded somehow. Shaheen et al. (2018) argue that with Avs's constant collection of data if the AV's communication or storage systems are hacked, the dignity and privacy of the victims are at high risk. For instance, location information might be considered a tool for tracking the movement of a specific person, which could lead to security threats for passengers and their data. 45 In contrast, besides geographical coordinates, AVs also gather key performance indicators (KPIs) connected to passengers’ activities and interactions in the vehicle. This may include facial recognition, voice commands, user fingerprints, and other identification features for authentication and customization. McDonald et al. (2020) noted that data types such as these can put someone in danger of identity theft or profiling if not adequately protected. Since this is more personal data, it is crucial to incorporate appropriate security measures into AV systems to respect people’s rights to privacy and reduce instances of abuse of such data. Also, AVs require constant communication with external systems, including vehicle-to- vehicle (V2V) and vehicle-to-infrastructure (V2I), along with various cloud services that store and analyze car data. This means that the overall use of the network for external data exchange will always be vulnerable to third-party interception or misuse. As specified by (Lu et al., 2020), the system connectivity of AVs provides multiple entry points for extraneous stakeholders, all of whom can access and modify external network data. Hence, safeguarding data privacy in AVs is beneficial to people and necessary for developing trust in self-driving car solutions. In addition, many approaches and technologies are being employed to protect data privacy in self-driving automobiles. Encryption is one of the most effective ways of safeguarding data. For data security and protection, data encryption ensures that the information gathered or transferred by the AV is coded so that it cannot be understood without decryption keys. Shao et al. (2019) noted that advanced forms of encryption, including end-to-end encryption for communication between the car and external networks, are mandatory to counter such risks. For protection, there are two fundamental mechanisms: the first is encryption, and the second is access control. Access control refers to regulating who has permission to read or update information, and it can be applied through methods such as role-based access control 46 (RBAC) and biometric techniques. RBAC limits information accessibility based on the user’s role or identity in the vehicle system, so only authorized users or systems can access the data. In addition, using fingerprint or face recognition for authentication ensures that only authorized individuals can access the data in the car (Liu et al., 2021). Another fundamental technique for safeguarding data in AVs is data masking, which refers to eliminating or masking personally identifiable information (PII) in datasets before sharing or archiving them. Data anonymization methods, such as k-anonymity and differential privacy, enable the manufacturing and service industries to benefit from data analysis of the vehicle’s information while protecting passengers' privacy. As highlighted by (Wang et al., 2020), one way of minimizing privacy invasions is through depersonalizing the data collected by the AVs to provide insights on how to improve vehicles or control traffic flow. Differential privacy involves adding noise to data, which complicates extracting identifiable information, preventing attackers from reverse-engineering the data to identify users. Moreover, particular emphasis should be placed on the importance of secure storage, as AVs collect large amounts of information. Various local storage systems, including encrypting the hard drive or employing secure memory modules, can keep unauthorized access out of the vehicle. However, cloud-based data storage should also be safeguarded through encryption and multi-factor authentication to ensure that only authorized individuals can access the data. Ren et al. (2019) also pointed out that it is crucial to protect the privacy of AV passengers through proper storage and access control and that appropriate data storage should be maintained both in the vehicle and external systems. Therefore, apart from technical solutions, data protection will also be established through regulations. The General Data Protection Regulation (GDPR), passed in the European Union, 47 provides rules on collecting, processing, and protecting individual data in AVs. Under the GDPR, manufacturers and service providers must request users’ consent to collect their data and inform them of how the data will be used. As stated by (Koops et al., 2020), the subject of this discussion is defined by the necessity to adhere to the provisions of data privacy regulations as a primary means of safeguarding users’ rights and gaining their trust in developing autonomous vehicles. 5.3 Secure Communication Protocols Autonomous vehicles (AVs) primarily rely on Vehicle-to-Everything (V2X) communication, where vehicles can communicate with other vehicles (V2V), infrastructure (V2I), people (V2P), and even the network (V2N). Predictive Situational Awareness: This means that information is constantly fed to the AVs to enable them to make ‘real-time’ decisions, including collision avoidance, route optimization to respond to signals in intelligent cities, and other choices. While this is beneficial, it also makes AVs vulnerable to cybersecurity threats, which is why V2X communication security is an important issue. Additionally, V2X communication is quite vulnerable to cyberattacks, such as ‘listening in, message interception and modification,’ and ‘bogus message sending’ by intruders while communicating between vehicles or infrastructure. In the opinion of Petit et al. (2015), without proper security measures in place, a malicious actor can feed the V2X networks with false information to fool AVs into undertaking incorrect maneuvers, such as misinterpreting traffic signals or objects on the road. This highlights the importance of developing strong security measures to safeguard the information transmitted in V2X networks concerning its integrity, confidentiality, and authenticity. 48 Therefore, the importance of V2X security is critical since AVs move within complex environments in which decisions must be made within milliseconds, triggered by the information received. One of the problems is that if one of the channels or mediums is compromised, it could lead to large-scale consequences, such as accidents or traffic jams. In their perspective, Lee and Kim (2019) emphasized that the protection of V2X channels is necessary not only for the safety of individual vehicles but also for the protection of the intelligent transport system. It also involves exchanging safety-related information, such as traffic light signals, road conditions, and accident alerts, which must not be subject to interference. Furthermore, several protocols have emerged for V2X communication to address these security issues, and many are standardized. One such protocol is IEEE 1609.2, which outlines the security architecture supporting wireless communication between vehicles, commonly known as WAVE. It also addresses message authentication, message integrity, and confidentiality, ensuring that only the correct nodes can form part of the V2X communication system. Another popular framework is the ETSI ITS security architecture. It attempts to establish secure communication between vehicles and ITS infrastructure by adopting encryption and digital certificates to identify entities and protect messages from unauthorized modification (Papadimitratos et al., 2018). For example, security through encryption and authentication protocols is vital in securing V2X communication. Any data exchanged between AVs and external networks should be secure and untampered with. Encryption scrambles information so it cannot be understood without the correct decryption key. At the same time, authentication ensures the identity of the communicating parties to prevent unauthorized access or manipulation. 49 Similarly, a widely applied approach to securing V2X data exchange is Public Key Infrastructure (PKI), which uses asymmetric encryption to protect messages between vehicles and related infrastructure. A Certificate Authority (CA) assigns each vehicle unique public and secret private keys in a PKI system. The public key is used to encrypt messages intended for the car, while the private key is used to decrypt the messages. Only the recipient can view the message's contents, while third parties cannot tamper with the data. The study conducted by Lu et al. (2020) concluded that PKI is indeed efficient and has significantly improved the process of ensuring secure encrypted communications in V2X by establishing scalable methods for managing encryption keys and digital certificates. Moreover, it is equally essential to ensure that only relevant vehicles and other devices engage in V2X communication, and this is where authentication plays a significant role. Without proper means of authentication, there is a threat that adversaries could mimic actual vehicles or traffic infrastructure to feed fake data into signalized traffic systems or jam them. The primary authentication method in V2X systems is digital signatures, in which vehicles sign their messages with a cryptographic key that proves their legitimacy. If the received signature is valid, the receiving vehicle can rely on the message to be authentic. According to (Papadimitratos et al., 2018), message spoofing can be mitigated using digital signatures, as these signatures become invalid if the content of the message is altered. Furthermore, to improve encryption and authentication, Elliptic Curve Cryptography (ECC) is often implemented in V2X communication due to the effectiveness of this algorithm in providing more security with smaller critical structures. This makes ECC well-suited for V2X applications, as it is more efficient than traditional cryptographic algorithms like RSA, which are often limited by bandwidth constraints and real-time data processing requirements. According to 50 (Dabbagh et al., 2019), ECC is deemed efficient, making it popular for protecting V2X systems, where AVs can easily authenticate and encrypt messages without negatively impacting message transmission times. In addition to PKI and ECC, other technologies, such as Blockchain, have also been incorporated to secure V2X communications. V2X communication data is stored in a blockchain, where once information is recorded in a block, it cannot be changed or erased. Blockchain is highly resistant to attacks due to its decentralized nature, in which no single point can be attacked to compromise the network's security. Another study by (Singh et al., 2020) revealed that blockchain could enhance V2X security, as all exchanges between vehicles and the external world are documented and cannot be altered by attackers. Ultimately, intrusion detection systems (IDS) are generally included in V2X communication networks to ensure further real-time protection against security threats. IDS can monitor the behavior of vehicles and communication, identifying deviations from normal behavior that may indicate a cyber-attack. These systems can be rule-based, using predefined signatures of known attacks to identify threats, or anomaly-based, detecting slowly developing but real threats by identifying system anomalies. As stated by (Xu et al., 2019), IDS offers additional security for V2X communication, contributing to the identification of attempted violations and preventing attacks aimed at compromising the security of AV systems. 5.4 Software Integrity and Updates Since AVs depend more on the code used for decision-making and interaction with the environment, the finality of the software used to manage the vehicles’ systems becomes essential to safety and security. To enhance productivity and functionality, patches for bugs or fixes for 51 incompatibilities require developing new updates. However, these updates present a significant security challenge. This is because the outcome of an attacker holding the update process of a car would be disastrous, as they would have the ability to control other functions of the vehicle. As with any data, it is essential to ensure that software updates can be trusted because they have not been tampered with while transferring them from the source to the destination machine. Out of all the various methods of achieving this, one of the most popular is digital signing. The identified update is conventionally signed with a cryptographic key belonging to the manufacturer or, more commonly, to a third party trusted by the software manufacturer and the user. When the vehicle receives the update, it checks the digital signature of the update to ensure that it is authentic before downloading the software. Dietrich et al. (2018) note that digital signatures offer a reliable method of authenticating and preserving the software update’s content, ensuring that only valid updates are deployed within the car’s networks. Similarly, it is also crucial to mention another aspect of maintaining software integrity— secure over-the-air (OTA) updates. OTA updates allow manufacturers to offer software updates, fix flaws, and make changes from a distance without needing the automobile to be taken to a shop. However, such an arrangement comes with several risks if the update process is not well protected,