Performance evaluation of adversarial attacks on whole-graph embedding models

Graph embedding techniques are becoming increasingly common in many fields ranging from scientific computing to biomedical applications and finance. These techniques aim to automatically learn low-dimensional representations for a variety of network analysis tasks. In literature, several methods (e.g., random walk-based, factorization-based, and neural network-based) show very promising results in terms of their usability and potential. Despite their spreading diffusion, little is known about their reliability and robustness, particularly when applied to the real world of data, where adversaries or malfunctioning/noisy data sources may supply deceptive data. The vulnerability emerges mainly by inserting limited perturbations in the input data when these lead to a dramatic deterioration in performance. In this work, we propose an analysis of different adversarial attacks in the context of whole-graph embedding. The attack strategies involve a limited number of nodes based on the role they playin the graph. The study aims to measure the robustness of different whole-graph embedding approaches to those types of attacks, when the network analysis task consists in the supervised classification of whole-graphs. Extensive experiments carried out on synthetic and real data provide empirical insights on the vulnerability of whole-graph embedding models to node-level attacks in supervised classification tasks.