Internet of Things (IoT) technology has created a new dimension for data collection, transmission, processing, storage, and service delivery. With the advantages offered by IoT technologies, interest in smart home automation has increased over the years. Nevertheless, smart connected homes are characterized with the security and privacy problems that are associated with aggregating multiple sensors' data and exposing them to the Internet. In this paper, we propose LPM, a lightweight privacy-aware model that leverages information theoretic correlation analysis and gradient boosting to fuse multiple sensors' data at the edge nodes of smart connected homes. LPM employs federated learning, edge and cloud computing to reduce privacy leakages of sensitive data. To demonstrate its applicability, two services, commonly provided by smart homes, i.e., occupancy detection and people count estimation, were experimentally investigated. The results show that LPM can achieve accuracy, F1 score and AUC-ROC of 99.98%, 99.13%, and 99.98% respectively for occupancy detection as well as Mean Squared Error (MSE), Mean Absolute Error (MAE), and R2 of 0.0011,0.0175, and 98.39% respectively for people count estimation. LPM offers the opportunity to each smart connected home to participate in collaborative learning that is achieved through the federated machine learning component of the proposed model.