Enhancing the Resilience of Graph Neural Networks to Topological Perturbations in Sparse Graphs