Gait-based activity recognition has gained significant attention in applications such as healthcare,
security, and rehabilitation. However, traditional centralized machine learning models pose challenges
related to data privacy, scalability, and interpretability. Federated Learning (FL) addresses these
concerns by enabling distributed model training without sharing raw data, ensuring privacy preservation.
Simultaneously, Explainable AI (XAI) techniques enhance model transparency, making gait recognition
systems more interpretable. This paper presents a comprehensive survey on FL and XAI techniques for
gait-based activity recognition, focusing on recent advancements, benchmark datasets, and existing
challenges. A comparative study of different FL methods (FedAvg, FedProx, FedBN, Personalized
FL, and Hierarchical FL) and XAI techniques (SHAP, LIME, Grad-CAM, Attention-based XAI, and
Hybrid Neuro-Fuzzy models) demonstrates their effectiveness. Our findings show that Hierarchical FL
combined with Hybrid XAI achieves the highest accuracy (93.1%) while maintaining strong privacy
(0.91 score), albeit at a higher computational cost. Despite significant advancements, challenges such
as communication efficiency, model personalization, and computational overhead persist. This study
highlights the need for standardized benchmarks and optimized FL-XAI frameworks to enhance realworld deployment in resource-constrained environments.
