Federated Learning (FL) and Kolmogorov-Arnold Network (KAN) have emerged as popular paradigms in machine learning. Our study explores their integration for handling Non-Independent and Identically Distributed (Non-IID) data in FL scenarios. We transform KAN into a parameter optimization problem that is well-suited for resolution through FL, and formulate a general FedKAN (Federated Kolmogorov-Arnold Network) optimization model. We introduce FedCKAN (Federated Convolutional Kolmogorov-Arnold Network) for vision tasks by integrating Convolutional KAN. Furthermore, we propose FedMKAN (Federated Meta Kolmogorov-Arnold Network) to address Non-IID data in FL by incorporating meta-Learning. Extensive experiments on FMNIST (Fashion-MNIST) and MNIST demonstrate that FedMKAN maintains high performance across different Non-IID levels, despite FL scales and batch sizes. Notably, on highly Non-IID FMNIST ( \(\alpha\)  = 0.05), the proposed method achieves nearly a 25% increase in accuracy compared to FedKAN, boasting 94.73% versus 70.00%.