Federated learning for human activity recognition: Balancing privacy, efficiency, and accuracy through innovative aggregation techniques
Muhammad Rafsan Kabir, Rashidul Hassan Borshon, Riasat Khan · Array · 2025
Recognizing human activities using machine learning is a widespread application but comes with significant concerns regarding data privacy. To mitigate this challenge, we explore the efficiency of federated learning (FL) strategies in a privacy-preserving manner for identifying human activities from smartphone sensor data. We implemented five distinct FL aggregation techniques, Vanilla Averaging, Weighted Averaging, Selective Averaging, Best-Model Update, and Continual Federated Learning, across two model architectures, namely Multi-layer Perceptron (MLP) and Convolutional Neural Network (CNN). These experiments shed light on the balance between model architectures, inference time, and accuracy within FL frameworks. Specifically, our CNN model utilizing the Selective Averaging technique achieves an accuracy of 92.30% and an F1 score of 92.37%, thus surpassing the performance of other configurations tested. Due to the fewer parameters and shorter inference time, the MLP model is suitable for memory-constrained edge devices compared to the CNN model despite a slight reduction in classification accuracy. This work illustrates the importance of carefully selecting both the model architecture and the FL technique to achieve optimal results in real-world applications prioritizing privacy.