Graph Neural Networks (GNNs) have demonstrated remarkable potential in processing graph-structured data and have been widely adopted for various graph-related tasks. In recent years, graph knowledge distillation has emerged as an effective model optimization technique, achieving outstanding performance in enhancing GNN capabilities. However, it remains constrained by the inherent sparsity and dependency of graph data. To address this, we propose Mixup-driven Distillation for Graph Neural Networks (MD-GNN), an innovative framework that combines Mixup-based data augmentation with knowledge distillation to boost GNN performance. Specifically, we first leverages Mixup to perform linear interpolation on both node features and labels, generating diversified training samples while preserving original graph topology. Subsequently, based on these generated data, we extract knowledge from a pre-trained GNN teacher through output logit alignment to guide the training of student GNN models. Extensive experiments demonstrate that MD-GNN significantly outperforms existing baseline GNN models, achieving performance gains of 2.54%–3.73% on three benchmark datasets (Cora, CiteSeer, and PubMed). Notably, MD-GNN exhibits superior generalization capabilities and enhanced robustness, particularly showing notable advantages in noisy scenarios.