LiteFL: A Lightweight Privacy-Preserving Federated Learning Framework for Non-IID Distributed Wearable IoT Devices

The rapid adoption of Wearable Internet of Things (WIoT) devices in smart cities has enabled a wide spectrum of applications. Although Federated Learning (FL) was once considered the silver bullet for privacy-preserving systems, it falls short in WIoT settings. This is primarily because deep learning (DL) models typically require extensive computation, which exceeds the limited processing capabilities of resource-constrained WIoT devices. Additionally, the non-independent and identically distributed (non-IID) nature of WIoT data hinders model convergence and generalization. To this end, we propose LiteFL, a lightweight privacy-preserving FL framework tailored for non-IID WIoT environments. LiteFL partitions the model into extractor, processor, and classifier modules, offloading the computation-intensive processor to the server while retaining lightweight components on clients. It integrates differential privacy (DP) for data perturbation and fully homomorphic encryption (FHE) for the proposed secure aggregation mechanism. Experiments on a real-world WIoT dataset demonstrate LiteFL achieves an AUC of 0.939 and F1-score of 0.845, outperforming baselines by over 5%. It also mitigates the cold-start problem in real-world applications by enabling knowledge transfer across heterogeneous distributions, advancing scalable, secure FL for resource-constrained WIoT systems.