A Data Augmentation-Based Graph Model Defense Method

Graph Neural Networks (GNNs) have shown great potential in various graph - related tasks in recent years. However, they are susceptible to adversarial attacks, which can degrade their performance. To counter these attacks, numerous defense models have been proposed, with data augmentation methods being widely used by researchers. Despite this, the process of data augmentation can lead to attack exacerbation, where the performance of GNN s deteriorates significantly under stronger adversarial attacks. Moreover, data augmentation often sacrifices performance on clean samples to improve that on poisoned ones. To address these issues, we propose a robust defense model based on data augmentation, named DABD. This model first preliminarily purifies the graph through preprocessing. Then, a structural purification module is introduced to detect suspicious nodes and reduce their impact, resulting in a purified graph. A structural restoration module is used to obtain a restored graph to prevent performance decline on clean samples. Finally, contrastive learning is applied to the purified and restored graphs to obtain effective node representations. Experimental results on the Cora and Citeseer datasets show that, under Mettack attacks, the model not only achieves an average classification accuracy more than 1.1 % higher than state - of - the - art baselines but also alleviates the attack exacerbation phenomenon.