Privacy-Preserving Federated Learning for Smart Walking Sticks Using IoT

Assistive mobility devices for the elderly and visually impaired increasingly use IoT sensors for real-time obstacle detection, gait monitoring, and fall prediction, but centralized approaches risk privacy breaches and fail to personalize across non-IID user patterns. This paper examines federated learning (FL) within hybrid edge-cloud IoT architectures to enable privacy-preserving personalization. The proposed framework combines low-latency edge inference on resource-constrained devices with secure cloud aggregation of model updates. A prototype smart walking stick—equipped with ultrasonic sensors, IMU, and ESP32—serves as the case study. Trials demonstrate ~96% fall-prediction accuracy, 86–89% personalization gain in obstacle/gait adaptation after 10 FL rounds,