<p>Urban sustainability systems depend on distributed Internet of Things (IoT) networks for real-time monitoring and control. These networks support energy management, traffic optimization, and environmental sensing. However, the data they collect often include sensitive user information. This raises serious concerns about privacy, security, and centralized data handling. Existing solutions such as cloud-based analytics, encryption schemes, and access-control mechanisms have been proposed to enhance data security in IoT-enabled smart cities. However, these methods typically rely on centralized servers, incur high communication costs, and remain vulnerable to data leakage and single points of failure. In this work, we present a privacy-preserving artificial intelligence (AI) framework for smart urban environments. The system combines federated learning with local differential privacy. Raw data stay on edge devices. Only encrypted and noise-added model updates are shared. We use field programmable gate array (FPGA)-based edge hardware to ensure low latency and energy efficiency. Our framework supports real-time AI tasks in resource-constrained networks. It is scalable and adapts to heterogeneous device configurations. Experimental results show strong performance in accuracy, privacy, and energy use. Compared to centralized training, our approach reduces communication and power consumption, while maintaining model quality. The proposed system achieved only a minor accuracy drop of 1.2–1.5% compared to centralized training, while the success rate of membership inference attacks decreased by over 50%. These results confirm that the framework provides a strong balance between model performance and privacy protection. This work demonstrates that privacy-preserving AI can be practically deployed in smart cities.</p>