Application of Deep Learning Method under Federated Learning for Human Action Recognition

Human Action Recognition (HAR) is growing in importance in various fields such as security, healthcare, and sports, and deep learning methods such as Convolutional Neural Networks (CNNs) have reached promising results. This study introduces federated learning (FL) techniques to integrate CNN models for multi-device action recognition to improve model performance while ensuring privacy. Specifically, the Residual Network (ResNet) 50 model is used as the backbone for feature extraction and human action classification. Second, FL is used to train classification models across multiple devices. This study aggregates model updates using the Federated Averaging (FedAvg) algorithm. The method uses data augmentation techniques to improve generalization and prevent overfitting. The model is trained on a public HAR dataset from Hugging Face. Experimental results present that the FL-based model achieves competitive performance in action recognition while maintaining data privacy. However, the model also faces the challenge of overfitting due to the significant gap between training and validation accuracies. Future work will explore optimization strategies such as Federalized Proximal (FedProx) and other regularization methods to enhance the model's generalization and address the overfitting problem.