Adversarial network mining

Abstract

Recent advancements in the field of graph representation learning and the introduction of algorithms like DeepWalk, LINE and Graph Convolutional Networks (GCNs) have helped achieve state-of-the-art results for tasks like node classification and link prediction. However, research indicates that these machine learning algorithms are susceptible to attacks and slight perturbations in the graph data can result in poor results. This shows the need to build reliable models resilient to such attacks. In our thesis, we aim to develop novel attack strategies for both unsupervised and supervised graph embedding algorithms.