PhD Oral Exam - Nadia Abdolkhani, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The rapid proliferation of the Internet of Things (IoT) has intensified the need for communication systems that are spectrum-efficient, energy-aware, and resilient. Yet a persistent mismatch remains, licensed bands are often underutilized while unlicensed bands suffer congestion from IoT, vehicular, and Unmanned Aerial Vehicle (UAV) devices crowding a few public frequencies. Cognitive Internet of Things (CIoT) networks, built on Cognitive Radio (CR), offer a path to address spectrum scarcity. This tension calls for a rethink of spectrum allocation and management toward intelligent radio that unites cognition with learning. In this thesis, we advance data-driven, Reinforcement Learning (RL) approaches that learn directly from interactions with the environment to enable agile spectrum borrowing and sharing on licensed bands. We develop Deep Reinforcement Learning (DRL) frameworks that optimize power control, time allocation, access coordination, channel selection, and cooperative caching under realistic constraints such as interference, fading, jamming. Furthermore, we design Energy Harvesting (EH)-aware algorithms for self-sustaining, green devices. Across the work, we emphasize computational efficiency suited to resource-constrained nodes without sacrificing performance or security.

To drive sustainable spectrum access in CIoT networks, we develop a DRL approach for optimizing the joint power control and access coordination in a Wireless Power Transfer (WPT)-EH CIoT. Our framework formulates the problem of joint power allocation and channel access as a model-free Markov Decision Process (MDP). A Deep Q-Network (DQN)-based CIoT agent learns when to harvest energy, when to transmit, and how much power to allocate, all while satisfying interference constraints with the primary network. Unlike traditional approaches that rely on prior knowledge or channel models, the proposed DRL agent autonomously adapts to varying channel occupancy, interference dynamics, and EH conditions. We extend the DRL framework to a Simultaneous Wireless Information and Power Transfer (SWIPT)-enabled CIoT systems. Here, a Double Deep Q-Network (DDQN) equipped with an Upper Confidence Bound (UCB) exploration strategy is proposed to jointly optimize the Time Switching (TS) factor and transmit power under fading and interference constraints. The CIoT transmitter dynamically balances between energy harvesting and information transmission, learning to allocate time and power efficiently to enhance throughput and lifetime.

To tackle secure spectrum access in CIoT networks, we develop a DRL framework for EH-enabled CIoT under jamming attacks. We focus on the security challenges of CIoT systems exposed to intentional radio jamming. To counteract these adversarial conditions, a robust DDQN-based framework enhanced with a novel Interference-Aware Upper Confidence Bound (UCB-IA) is introduced. This approach enables the CIoT device to recognize jamming activity, adjust transmission strategies, and maintain reliable communication without prior knowledge of the jammer’s behaviour. The model leverages locally observable features such as energy arrivals, channel gain, and jamming intensity to make adaptive decisions. We then extend our DRL framework into a hierarchical DRL framework for robust access in CIoT networks under smart jamming attacks. This is to address the growing complexity of decision-making in hybrid action spaces that combine discrete (mode/channel) and continuous (power) decisions. A Hierarchical Deep Deterministic Policy Gradient (H-DDPG) framework is proposed, decomposing the decision process into three levels: high-level mode selection (transmit or harvest), mid-level channel selection, and low-level continuous power control. Meanwhile, the jammer is modeled as a learning-based agent to simulate adaptive attack patterns. This hierarchical design improves scalability and exploration efficiency, allowing the CIoT agent to make robust and fine-grained decisions.

To improve collaboration towards efficient spectrum utilization in CIoT networks, we explore caching-based cooperation for spectrum sharing in the overlay spectrum access paradigm. In contrast to conventional CR strategies where secondary users vacate the channel when primary users return, the proposed DRL framework encourages collaborative caching, enabling CIoT agents to store and deliver Primary User (PU) content. This cooperative behavior reduces latency and improves throughput by minimizing data retrieval distance, thus benefiting both primary and secondary networks. The DRL-based caching policy optimizes storage and access coordination under energy, interference, and capacity constraints.

Finally, to develop a dynamic strategy for the complex hybrid CIoT spectrum access environment, we propose a hierarchical DRL framework to optimize cooperative caching and SWIPT-EH. This framework presents a comprehensive three-level Hierarchical Soft Actor-Critic (H-SAC) model that unifies EH, hybrid underlay–overlay access, power allocation, and cooperative caching into a single framework. The high-level policy governs the TS factor for EH, the mid-level policy manages hybrid access and caching collaboration, and the low-level policy optimizes transmit power and cache placement decisions. The joint optimization is formulated as a weighted-sum multi-objective task to maximize throughput and cache hit ratio while minimizing delay and energy consumption.

Collectively, this thesis presents a cohesive body of work that advances Artificial Intelligence (AI)-driven, energy-efficient, and secure spectrum management for next-generation CIoT networks. Through the integration of RL, hierarchical control, and cooperative caching, the proposed frameworks demonstrate the potential of self-sustaining and intelligent wireless systems capable of optimizing communication performance under dynamic, uncertain, and adversarial conditions.