Data heterogeneity has become a major concern in federated learning, where the client data may sample from different distributions and be biased regarding features and labels. Such Non-IID data degrades the performance of federated learning. Previous research attributed this problem to weight divergence between the global optimal solution and local ones and proposed regularization as a solution. However, we note that weight divergence does not affect each weight equally, making universal strength regularization unnecessary. To address this challenge, we introduce a method called FedAR that regularizes the local model based on data heterogeneity with varying strengths for each dimension. Empirical evaluations conducted on MNIST, FEMNIST, and CIFAR-10 indicate that FedAR facilitates personalized federated learning with better generalization performance in both IID and Non-IID settings.