UC Irvine

Human-in-the-loop (HITL) IoT systems merge artificial intelligence, automation, and user interaction, enabling IoT environments to adapt dynamically based on human feedback. These systems are pivotal in applications requiring real-time responsiveness, user-specific customiza- tion, and privacy preservation, such as smart homes, healthcare monitoring, and advanced driver assistance systems. This thesis investigates the design principles, challenges, and privacy implications of HITL IoT systems, with a focus on enhancing personalization, efficiency, and trust. We present four key contributions addressing privacy challenges in HITL IoT systems. First, ERUDITE, a wearable neurotechnology platform, uses EEG data to personalize adaptive learning environments, demonstrating a 26% improvement in learning outcomes. Second, ada- PARL introduces adaptive privacy-aware reinforcement learning, reducing privacy leakage by 23% while enhancing utility by 57% in smart home simulations. Third, adaEXPARL proposes a novel early-exit strategy for Deep Q-Networks, achieving a 31% improvement in privacy with minimal utility trade-offs. Finally, we develop a privacy control mechanism for Microsoft HoloLens, incorporating real-time data masking and anonymization, establishing a benchmark for privacy-preserving IoT systems. Through these contributions, this thesis provides a compre- hensive framework for designing HITL IoT systems that balance user adaptability, privacy, and utility, laying the foundation for their application in sensitive and user-driven environments.