Large Language Models (LLMs) have demonstrated impressive capability in generating fluent text. LLMs have also shown a tendency to reproduce social biases such as stereotypical associations between gender and occupation. Like race and gender, morality is an important social variable. This work investigates whether LLMs reproduce the moral biases associated with political groups in the United States, an instance of a broader capability I refer to as "moral mimicry". I explore this hypothesis in the GPT-3/3.5 and OPT families of Transformer-based LLMs. Using tools from Moral Foundations Theory, I show that these LLMs are indeed "moral mimics". When prompted with a "liberal" or "conservative" political identity, the models generate text reflecting the moral biases associated with these groups. I investigate how moral mimicry relates to model scale. I hope that this work encourages further investigation of the moral mimicry capability, including how to leverage it for social good and minimize its risks.

Multiple disruptions in road networks have the potential for cascading effects that can cause significant degradations of network performance resulting in large increases in travel time. However, it is challenging to identify vulnerable combinations of links in road networks due to the complex interdependency of road segments and a prohibitively large number of possible combinations of links. In this paper, we present a deep reinforcement learning (DRL) framework to identify vulnerable combinations of links in road networks. We let a DRL agent select links to disrupt, and its policy is parameterized with deep neural networks (DNNs). The policy is directly learned from the consequences of disruptions of links (i.e., congestion incurred by the disruptions) in traffic simulations where multiple links are disrupted in a sequence. As a case study, we analyzed vulnerable combinations of links in the road network in the city of Davis in California, and compared the criticality of the disruptions of links selected by the proposed DRL-based method and heuristic-based methods that use betweenness centrality or traffic counts of links. In the results, we observed that disruptions by the DRL agent induced significantly larger congestion and increase in travel time of vehicles than heuristic-based methods. Furthermore, the links selected by the DRL agent reveal that the disrupted links are selected considering both the static properties such as the topology of the road network and the capacity of the links as well as dynamic properties imposed by the traffic demand.

The primary focus of this dissertation is on evaluating anomaly detection for Science DMZ networks. Compared to other anomaly detection problems, this introduces some unique challenges.First, if real-time threat detection and response is desired, there is a smaller time window for reaction in high speed networks such as Science DMZs. Anomaly detection, and real-time detection in particular, is already a challenging problem, but in the case of high speed networks it becomes even more difficult. Different approaches are required in order to keep up at high rates, such as either simplifying the detection methods, or finding creative means of optimizing existing methods, either through different hardware or modifying the algorithms. In some cases the best option is finding an entirely new detection technique, utilizing new or previously neglected metrics. Determining the specific techniques best suited for a given environment is another challenge. The ideal methods for anomaly detection will vary depending on the particular network and how it is used. Science Demilitarized Zone (DMZ) networks are generally made to perform a limited range of tasks, primarily facilitating high speed data transfers between research sites. Therefore, the behavior seen will be more predictable compared to more general purpose networks, which helps allow for more practical anomaly detection. This work contains research into different methods of anomaly detection on Science DMZs, providing evaluations of new real-time detection methods and tools for more effective monitoring.

The use of GPS navigation apps has been surging with increasing the use of smartphones. People used to use a map to find a way to their destination, but now the navigation apps provide us with several options that can be taken as the best path for drivers. Also Connected and Autonomous Vehicles (CAVs) have been attracting attention and have improved a lot in tandem with the popularization of electric vehicles. Also, it is expected that autonomous driving which means that the car is self-driven instead of being driven by a person will emerge in the near future. This autonomous driving will rely on algorithms that will determine the best path once the destination is set and will perform dynamic routing to adapt to road network congestion during the trip. These algorithms that suggests the best paths will affect road traffic flow. Furthermore, when more CAVs rely on the same best path algorithm in the future, it will lead to a significant impact on a city’s road network traffic if an adversary can intentionally cause real or virtual disruptions in an few strategic locations.This thesis presents the results of simulation experiments to determine vulnerability in road network in the city of Davis in California. Specifically, by developing a simulation model, this study models how disruption in one part of the network can spread and how this is related to or not related to some of the topological properties of the network. Lastly, this thesis discusses a way to minimize the effect of the vulnerability.

With the advanced wireless technologies and emerging machine learning techniques, there is an increasing trend of using both wireless communication and machine learning along with new platooning control as well as traffic signal control (TSC) algorithms to leverage and accommodate connected and autonomous vehicles (CAVs) for achieving high throughput and low latency traffic flow, lowering accidents, and reducing emissions in urban transportation. However, this development leads to highly connected wireless environments that would increase the potential for cyber-attacks on TSC. These attacks can undermine the benefits of new TSC algorithms. For example, an adversary can first perform reconnaissance to gain understanding of how the TSC operates under different traffic arrival rates, and then, choose a fixed number of vehicles to be attack vehicles to get scheduling priority and/or to create traffic congestion in one intersection which can spread to the entire network. These attacks can compromise TSC systems and significantly increase traffic delay and make TSC completely ineffective. Furthermore, machine learning based TSC approaches also lead to several issues: complexity, dimensionality, and convergence. They require a long period of time and a significant amount of training data to gain insights into improvement of traffic throughput and latency. They are not applicable to problems with a large state size. In addition, some of the approaches adopting neural networks (NNs) as a function approximator leads to a convergence issue due to the non-linearity of NNs. In this work, we have proposed the SecureTLC to address the security and machine learning issues caused by the novel technologies for modern intelligent transportation systems (ITS).

As our first work, we have investigated security vulnerabilities in four different backpressure-based (BP-based) traffic control algorithms: 1) Delay-based BackPressure Control (DBPC), 2) Queue-based BackPressure Control (QBPC), 3) Sum-of-delay-based BackPressure Control (SBPC), and 4) Hybrid-based BackPressure Control (HBPC) which is a combination of the delay-based and queue-based. Their performance has been compared when they are under two misinformation attacks: 1) time spoofing attack and 2) ghost vehicle attack. The time spoofing attack is a type of falsified data attack in which attack vehicles arriving at an intersection alter their arrival times. In the ghost vehicle attack, attack vehicles intentionally disconnect the wireless communication and thereby hide from the TSC. We have shown that the misinformation sent by attack vehicles can influence the signal phases determined by BP-based traffic control algorithms. We have considered an adversary that determines a set of arriving vehicles to be attack vehicles from many candidate sets (attack strategies) in order to maximize the number of disrupted signal phases. We have shown that by formulating the problem as a 0/1 Knapsack problem, the adversary can explore the space of attack strategies and determine the optimal strategy that maximally compromises the performance in terms of average delay and fairness. Through detailed simulation analyses we have also shown that while the DBPC has better fairness, it is more vulnerable to time spoofing attacks than the other schemes. On the other hand, the QBPC, which determines traffic signal phases using the aggregate queue information, suffers from a higher intensity of ghost vehicle attacks. We have examined the drawbacks of both the DBPC and QBPC under different attack scenarios for different traffic patterns including homogeneous and non-homogeneous vehicle arrivals at an isolated intersection. We have proposed two protection mechanisms against the attacks, namely, an auction-based protection algorithm (APA) and a hybrid-based protection algorithm (HPA). Using simulation analysis, we have shown that both protection schemes are able to mitigate the impacts of the time spoofing and ghost vehicle attacks.

In our second work, we have studied the impact of AI-enabled platooning control on fuel consumption. Nowadays, vehicular emissions and traffic congestion have been deteriorated by highly urbanization. The worsen traffic burdens drivers with a higher cost and longer time on driving, and exposures pedestrians to unhealthy emissions such as PM, $NO_{x}$, $SO_{2}$ and greenhouse gases. In response to these issues, CAV which enables information sharing between drivers and infrastructure was proposed. With advanced wireless technologies offering extremely low latency and CAV, platooning control can be realized to improve traffic efficiency and safety. However, conventional platooning control algorithms require complex computations and hence, are not a perfect candidate when applying to real-time operations. To overcome this issue, this chapter focuses on designing an innovative learning framework for platooning control capable of reducing the fuel consumption by the four basic platoon manipulations, e.g., split, acceleration, deceleration, and no-op. We integrated reinforcement learning (RL) with NNs to be able to model non-linear relationships between inputs and outputs for a complex application. The experimental results reveal a decreasing trend of the fuel usage and a growing trend of the reward. They demonstrate that the proposed DRL platooning control optimizes the fuel consumption by fine-tuning speeds and sizes of platoons.

In our third work, we have applied advanced machine learning techniques to TSC. A recent report has illustrated that an urban road network equipped with an advanced TSC system is capable of reducing average sojourn time of vehicles, traffic collisions, and traffic delay while at the same time decreasing energy consumption and improving parking. Modern TSC systems which leverage advanced machine learning techniques and scheduling algorithms have been a promising topic in recent years. RL has been broadly adopted in many areas, owing to its flexibility in combining an assortment of architectures and neural networks that assists RL agents in learning from a complex environment. Due to its compatibility with numerous existing deep learning techniques, RL can be integrated into a complicated deep learning framework which is extendable and powerful for many deterministic and nondeterministic applications. We proposed a deep RL (DRL) framework which has high extendability and flexibility by incorporating various RL-based algorithms and techniques. We shown that learning performance can be significantly improved by applying an on-policy temporal-difference (TD) learning method, SARSA, to the traffic signal control problem for a traffic network consisting of multiple intersections. The proposed traffic flow maps (TFMs) as input states are formulated by traffic flows at every time slot in the network. These states are fed into a non-linear function approximator to output the optimal action. In order to process TFMs, we combined a convolutional neural network, a dueling network architecture which enhances evaluations of action values when there is a huge action space, and a memory replay that breaks the strong temporal correlations among states. Moreover, we compared learning performance among DQN, 3DQN, DSARSA, and 2DSARSA as well as provide a thorough analysis on their weaknesses and strengths. To the best of our knowledge, this is the first work discussing and evaluating learning effects of different DRL-based agents.

In our fourth work, as an extended work, we have proposed a novel architecture exploiting both the global and local information to further improve the performance of TSC in the third work. We use the proposed TFMs as input states to represent the global information for a DRL agent and pressure-of-the-lanes (POL) to represent the local information as pressure metric for a number of backpressure (BP) controllers. We define a combinatorial reward function using the power metric which maximizes the overall throughput and minimizes the average end-to-end delay of a network. We proposed a hybrid-based two-level control (TLC) architecture which has high extendability and flexibility by enabling a collaboration between the centralized DRL agent using TFM and the decentralized BP controllers using POL in a hierarchical architecture to determine traffic signal phases that optimizes the network throughput and latency. We show that learning performance can be significantly improved by the collaboration between the two levels for a traffic network of multiple intersections. Moreover, we compare learning performance of the proposed TLC with BP, DQN, 3DQN, DSARSA, and 2DSARSA as well as provide a thorough analysis on their weaknesses and strengths.

As our future work, first, we plan to analyze the impacts of malicious attacks on TSC equipped with the DRL-based agent. From our experiments, we have shown that DRL performs well for a traffic network with multiple intersections. We expect to explore the tolerance of DRL-based agent when it is under adversarial attacks. How many attacks can it tolerate? We want to examine the resilience of DRL-based agent as well. How well can it recover after encountering attacks? Second, as we start to consider TSC problems for larger areas, increases in dimensions of states and actions are foreseeable. High dimensional states and actions causing difficulty in learning will be a worthy research topic. We plan to design a two-level control architecture which is capable of reducing the dimension of actions by transferring some workload from a centralized agent to local controllers.