In recent years, Federated Learning (FL) has gained significant traction as an effective solution for various computer vision applications, particularly due to its strengths in preserving data privacy and minimizing communication overhead. However, its application to advanced tasks like video-based action recognition introduces unique complications. Additionally, the client drift caused by data heterogeneity remains a significant challenge. To address this challenge, we introduce a novel framework, Synergized Twin Layer for Federated Action Recognition (STL-FAR), which leverages both local and global layers to harmonize client-specific patterns with a global model representation. Specifically, the STL-FAR framework is composed of two main components in Synergized Twin Layer (STL): a local classifier that adapts to local client data, enhancing the robustness and accuracy of action recognition on individual clients, and a global classifier that unifies the local models into a consolidated global model through federated averaging, thereby reducing inter-client discrepancies and improving overall model generalization. Moreover, we incorporate a Cloud-to-Client Knowledge Distillation (CCKD) mechanism within the framework, where the server supervises the client, ensuring consistent and robust performance across clients. We demonstrate the efficacy of STL-FAR through extensive experiments on two benchmark action recognition datasets. Our findings indicate that STL-FAR outperforms existing federated learning methods. This work advances federated action recognition and provides a promising solution to the issue of data heterogeneity in federated learning scenarios.