Human action recognition (HAR) has attracted significant attention in recent years, coinciding with the rapid development of deep learning techniques. Since HAR applications often involve sensitive data, privacy concerns have become a major research focus. To solve this problem, various approaches have been proposed to ensure data security of participating parties. In this work, we use federated learning, a decentralized framework, to preserve private data exchange between parties. We leverage the MoViNets model, a state-of-the-art architecture that has demonstrated superior performance in HAR tasks, to investigate the effectiveness of federated learning in this context. Additionally, this study addresses the problem of non-IID data, a common problem in federated learning settings.