Cyber Security Seminar by Jing Xu MSc - Backdoor Attacks in Graph Neural Networks

Abstract:

Many real-world data can be modeled as graphs, such as social relations, transportation networks, and protein structures. And Graph Neural Networks (GNNs) have emerged as state-of-the-art for machine learning on graphs due to their superior ability to incorporate information from neighboring nodes in the graph. However, similar to Convolutional Neural Networks(CNNs), GNNs are also vulnerable to adversarial attacks, one of which is the backdoor attack. In this talk, I will discuss the difference between backdoor attacks in CNNs and GNNs, and how we can implement backdoor attacks in GNNs. Also, I will discuss how we apply backdoor attacks for defense purpose.

Short bio:

Jing Xu is currently a Ph.D. candidate in the Cybersecurity group, INSY department/TU Delft. Her research interest includes the security concern in GNNs, applications of GNNs, and explainable artificial intelligence.