Advances in Federated Learning: Combining Local Preprocessing with Adaptive Uncertainty Symmetry to Reduce Irrelevant Features and Address Imbalanced Data

Abstract

Federated learning is increasingly being considered for sensor-driven human activity recognition, offering advantages in terms of privacy and scalability compared to centralized methods. However, challenges such as feature selection and client imbalanced data persist. In this study, FLP-DS2MOTE-USA is suggested, a system that integrates federated local preprocessing, adaptive thresholding based on uncertainty symmetry, and a density- sensitive synthetic minority over-sampling approach. Each client preprocesses data locally and employs DS2MOTE for class balancing. On the server side, adaptive thresholding based on uncertainty symmetry is utilized to identify the optimal client for training the global mode. Evaluation on two distinct datasets—Human Activity Recognition with Smartphones and Human Activity Recognition (OpenPose) —reveals that our model outperforms FedAvg, FedSgd, FedSmote, and FedNova, achieving accuracies of 90.57% and 96.58%, respectively. In addition, FLP-DS2MOTE-USA minimizes update size and network overhead on the Human Activity Recognition with Smartphones, while achieving improvements on the OpenPose dataset. Overall, the proposed method not only addresses issues of imbalanced data but also reduces computational complexity via streamlined local preprocessing, and server-side mechanisms ensure client privacy. It outperforms traditional federated learning techniques in both accuracy and efficiency.