A Comparative Analysis of FedAvg, FedProx, and Scaffold in Gait-Based Activity Recognition by Evaluating Accuracy, Privacy, and Explainability
Dr. P. Prathusha, Kommala Aparna
Gait-based activity recognition plays a crucial role in various applications, including healthcare
monitoring, security surveillance, and biometric authentication. Traditional machine learning models
for human activity recognition (HAR) rely on centralized data collection, raising concerns about privacy
and data security. Federated Learning (FL) offers a promising solution by enabling distributed model
training while preserving user data privacy. In this study, we conduct a comparative analysis of three FL
algorithms—FedAvg, FedProx, and SCAFFOLD—evaluating their performance in terms of accuracy,
privacy, and explainability. Using publicly available gait datasets (e.g., MobiAct, WISDM, OU-ISIR
Gait), we simulate a federated environment where decentralized clients participate in model training.
Our results indicate that SCAFFOLD achieves the highest accuracy (89.1%) and fastest convergence (70
rounds) due to variance reduction techniques, while FedProx improves stability across heterogeneous
data. FedAvg, although less accurate, minimizes communication overhead (8.5 MB), making it suitable
for resource-constrained devices. Furthermore, SCAFFOLD demonstrates superior privacy preservation
(0.9 privacy score) and explainability (79.4), making it a strong candidate for interpretable and secure
gait-based HAR applications. This study provides insights into selecting the most suitable FL algorithm
based on the trade-offs between model performance, privacy, and interpretability.
