Abstract
The growing demand for real-time data processing on light-weight devices, without compromising data privacy, necessitates innovative approaches to balance these constraints. This study presents a novel hybrid edge-cloud federated learning framework aimed at developing an automated smoking detection system as a case study. Unlike traditional AI solutions that often face scalability and cost challenges due to reliance on high-performance computing resources, our decentralized approach distributes processing tasks across fog-edge layers, significantly reducing the dependence on centralized infrastructure. Additionally, while existing solutions typically focus on object detection (such as cigarettes), our framework specifically targets smoking behavior, enhancing detection accuracy in public, also, we collected a fully balanced dataset for smoking classification. Moreover, we address data privacy concerns by using federated learning, enabling local data processing on edge devices while securely sharing the model with the server. The suggested framework not only represents over 80\% accuracy in detecting smoking activities but also addresses architectural challenges in integrating cloud and edge computing. By leveraging federated learning, our solution ensures real-time responsiveness, and operational efficiency, and offers a scalable, accessible method for applying smoke-free zones. This study highlights the potential of combining lightweight devices with federated learning and edge computing to advance public health initiatives and provides a foundation for future research in decentralizing high-performance computing tasks.