GraphGallery is a library for training and testing graph neural networks. GraphGallery implements adversarial attacks on graph neural networks. The library is compatible with PyTorch, TensorFlow 2.x, Pytorch Geometric (PyG), and Deep Graph Library (DGL). In addition, 17 datasets with graph data are available in the library.
Adversarial attacks on graph neural networks
Adversarial attacks on a model are used to test models for noise immunity. The attack aims to force the model to produce an incorrect prediction with minor modifications to the input data or model components. GraphGallery implements both targeted and untargeted attacks. Targeted attacks are aimed at the fact that the input object would be classified as some kind of predefined class. Non-targeted attacks, in turn, focus on the fact that the input object would simply be misclassified. In addition to the attacks themselves, the libraries have implemented models that are resistant to attacks. Among them are RobustGCN, SBVAT, SimPGCN, and GCN-Jaccard.
More about available datasets
GraphGallery contains popular single graph datasets of various sizes. 8 out of 17 datasets are directional. In CORA, CiteSeer, and PubMed, nodes are documents and links are citations. Amazon Computers and Amazon Photo are parts of the Amazon co-purchase graph, where nodes represent products and links show that two products are often bought together. Node tags are reviews on the platform in bag-of-words format, and class labels are the product category.