The rapid growth of global energy consumption and increasing concerns about data privacy in smart buildings have created an urgent need for secure and scalable AI solutions for energy efficiency optimization. Traditional centralized approaches face critical limitations, including data privacy risks, communication bottlenecks, and inability to handle distributed non-IID data from diverse building systems. To address these challenges, this study proposes a novel Attention-based Federated Deep Neural Network (Fed-DNN) framework that combines federated learning with homomorphic encryption and differential privacy to enable collaborative, privacy-preserving energy efficiency classification. By integrating attention mechanisms, our framework dynamically identifies and prioritizes critical building features while maintaining data locality across distributed edge devices. Experimental results on the UCI-ENB2012 dataset demonstrate 98.05% classification accuracy with perfect precision and recall for Low-efficiency buildings and 4% misclassification between Medium and High categories. The privacy-preserving design ensures secure aggregation of model updates without raw data sharing, addressing key vulnerabilities in existing systems. This research is particularly timely as it directly supports UN Sustainable Development Goals SDG 7 and 11 by providing a practical solution for energy-efficient building management that balances AI performance with data security.