Universal adversarial attacks on graph neural networks

• المؤلف الرئيسي: Liao, Chang
• مؤلفون آخرون: Tay, Wee Peng
• التنسيق: Final Year Project
• اللغة: English
• منشور في: Nanyang Technological University 2021
• الموضوعات: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Electrical and electronic engineering
• الوصول للمادة أونلاين: https://hdl.handle.net/10356/149401

This project aims to study the robustness of graph-level graph neural networks (GNNs) against universal adversarial attacks in white-box and grey-box scenarios. Graph-level GNNs are widely used in critical domains such as quantum chemistry, drug discovery, while the robustness of them against universal adversarial attacks is unknown. In particular, there is currently no universal adversarial attack algorithm applicable to graph-level GNNs. In this project, a data-free universal adversarial attack for graph-level GNNs is proposed. We conducted several experiments to show that: 1. Graph-level GNNs are not robust against universal adversarial attacks even under grey-box scenarios. 2. Node features can be used as node embeddings with better performance than using the difference between graphs as node embedding. 3. Betweenness centrality may better reflect the importance of a node in an adversarial attack context.