Federated learning enables collaborative model training across distributed devices without sharing raw data, offering strong privacy guarantees. However, deploying such systems on edge devices faces limitations in computation and energy efficiency. This paper presents a hybrid edge–cloud federated learning framework applied to the problem of lightweight smoking detection. The proposed approach leverages edge devices for preliminary local training and offloads computationally intensive tasks to the cloud, achieving a balance between low-latency inference, privacy, and resource efficiency. Experimental results show that the hybrid strategy outperforms purely edge-based or cloud-based alternatives, making it a promising solution for healthcare monitoring applications.