Abstract
The rapid proliferation of smart home Internet of Things (IoT) systems introduces significant challenges in preserving data privacy while maintaining efficient and sustainable model training under heterogeneous and resource-constrained wireless environments. Conventional centralized learning and static federated learning approaches incur excessive communication overhead and fail to adapt to dynamic bandwidth, latency, and energy constraints. This paper proposes FLASH, a Federated Learning Framework for Sustainable Smart Home environments that enables privacy-preserving model aggregation through Flexible Latency-Aware Scheduling and Heuristic Compression. Specifically, FLASH dynamically selects a client-specific model compression ratio to ensure each training round satisfies a server-defined latency deadline while maximizing update fidelity. These decisions are guided by historical client resource profiles, including computation capacity and channel conditions, allowing efficient utilization of edge resources and avoidance of redundant transmissions. FLASH is evaluated through real-world experiments on a heterogeneous smart home deployment with diverse device capabilities and network conditions. Compared with standard hierarchical FL baselines, FLASH reduces per-round communication overhead by up to 74.7%, shortens training round latency by 75.6%, and improves convergence efficacy for up to 2.41 times under fluctuating network conditions, demonstrating its effectiveness for scalable, energy-efficient, and privacy-preserving sustainable smart home intelligence.