Abstract Graph Neural Networks (GNNs) exhibit potential in predicting the properties of molecules, but computational analyses with the GNNs often encounter the problem of data imbalance or overfitting. Augmentation techniques have emerged as a popular solution, and adversarial perturbation to node features achieves a significant improvement in enhancing the model's generalization capacity. Despite remarkable advancement, there is scarce research about systematically tuning the adversarial augmentation. We propose a new framework for an adversarial perturbation with influential graph features. Our method selects the data to apply adversarial augmentation based on the one-step influence function that measures the influence of each training sample on prediction in each iteration. In particular, the approximation of the one-step influence function has wide applicability to evaluate a model's validity on the observation level for a large-scale neural network. Selected data using the one-step influence function are likely to be located near the decision boundary, and experimental results demonstrated that augmentation of such data has improved the model's performance.