A Survey on Textual Adversarial Attack and Defense with Combinatorial Optimization

PhD Qualifying Examination


Title: "A Survey on Textual Adversarial Attack and Defense with 
Combinatorial Optimization"

by

Mr. Ning LU


Abstract:

With the development of computing power and the availability of large 
amounts of data, deep neural networks (DNNs) have seen remarkable 
achievements in various tasks (e.g., image classification, translation, 
and audio recognition). However, researchers have found that DNNs are 
vulnerable to adversarial examples (AEs), crafted by adding imperceptible 
perturbations to the original input and misleading DNNs to make wrong 
predictions. Recent studies have shown that textual AEs can effectively 
disturb the natural language processing (NLP) models, which may be 
employed to evade toxic comment detection or make AI generate offensive 
speech. Nevertheless, generating AEs is challenging in the NLP domain due 
to the discrete property of symbolic texts. The combinatorial optimization 
(CO) algorithm fits the problem settings naturally, the goal of which is 
to find an optimal or approximation solution in a discrete search space 
efficiently. In this survey, we present a comprehensive review of the 
recent research on the techniques of adversarial attack and defense in 
NLP. We build a taxonomy from different perspectives and discuss their 
properties. Furthermore, we investigate the CO applications in this field. 
Finally, we discuss the existing challenges and future research 
directions.


Date:			Wednesday, 30 November 2022

Time:                  	2:30pm - 4:30pm

Venue:                  Room 5566
                         Lifts 27/28

Committee Members:	Prof. Cunsheng Ding (Supervisor)
 			Dr. Qi Wang (Supervisor, SUSTECH)
  			Dr. Dimitris Papadopoulos (Chairperson)
 			Dr. Amir Goharshady
 			Dr. Maosheng Xiong (MATH)


**** ALL are Welcome ****