Trustworthy AI for Smart Systems: Ensuring Resilience, Sustainability, and Responsibility in Autonomous Decision-Making

Abstract

The rapid growth of smart systems, such as those deployed in agriculture and healthcare, has underscored their transformative potential in enhancing efficiency, sustainability, and decision-making. However, this proliferation also raises pressing concerns around trustworthiness, privacy, resilience, and ethical use of AI technologies. This dissertation investigates the foundational principles and practical implementations of trustworthy AI in smart environments, with a multidisciplinary focus on sustainability, fairness, and transparency. To this end, the research explores three key tasks. First, it develops sustainable and resilient AI for smart farm systems using techniques such as transfer learning, deep reinforcement learning, and federated learning to address energy constraints, cyber threats, and operational uncertainty in real-world deployments. Second, it designs fair and privacy-preserving AI frameworks for smart healthcare systems by incorporating uncertainty-aware and bias-mitigating mechanisms into federated learning models for equitable disease detection, particularly Alzheimer's disease. Third, it creates explainable AI solutions for smart animal welfare systems through interpretable Bayesian models that provide causal reasoning and robustness against sensor noise and adversarial disruptions in complex environments. The key contributions include the development of a solar-powered, energy-adaptive monitoring framework; a novel TL-DRL approach to enhance farm system efficiency; and the first integration of evidential neural networks into federated learning for fair, uncertainty-aware medical diagnosis. The proposed methods are evaluated on real-world datasets, demonstrating improvements in system reliability, fairness, and predictive accuracy; an uncertainty-aware Bayesian network that unifies deep and interpretable features with feature-level uncertainty modeling and propagation for structured inference and explanation. This dissertation advances the field of trustworthy AI by addressing critical gaps in resilience, fairness, and interpretability, laying the foundation for ethical and robust deployment of AI in high-stakes, resource-constrained smart systems.