Adversarial Attack and Defense on Graph Neural Networks

PhD Qualifying Examination


Title: "Adversarial Attack and Defense on Graph Neural Networks"

by

Mr. Haoyang LI


Abstract:

Graph Neural Networks (GNNs) have achieved great success in various graph 
tasks, such as node classification and online recommendation. Despite 
these successes, recent studies have revealed that GNNs are vulnerable to 
adversarial attacks on graph data, including topology modifications and 
feature perturbations. The attackers can slightly manipulate graph data to 
mislead GNNs into making wrong predictions. Moreover, since the attacker 
behaviors will degrade the performance of GNNs and lead to economic loss 
in real-world applications, existing researchers propose how to defend 
against such adversarial attacks, i.e., GNNs can make predictions 
correctly under attacks. Given the importance of graph analysis in 
real-world applications, it is necessary to provide a comprehensive survey 
of existing adversarial attacks and defenses on GNNs. In this survey, we 
first categorize existing adversarial attacks and defenses, and review the 
corresponding state-of-the-art methods. Then, we provide future directions 
on attacks and defenses.


Date:			Monday, 27 March 2023

Time:                  	2:00pm - 4:00pm

Venue:                  Room 5501
                         Lifts 25/26

Committee Members:	Prof. Lei Chen (Supervisor)
 			Prof. Bo Li (Chairperson)
 			Dr. Dan Xu
 			Prof. Ke Yi


**** ALL are Welcome ****