A Survey of Safety of Deep Neural Networks: Verification, Testing, Adversarial Attack and Defense, and Interpretability

PhD Qualifying Examination


Title: "A Survey of Safety of Deep Neural Networks: Verification, Testing, 
Adversarial Attack and Defense, and Interpretability"

by

Mr. Jaewoo SONG


Abstract:

Deep neural networks (DNNs) have been showing high level of performance 
previously unreachable by other machine learning techniques. Powered by high 
computation power and abundance of data these days, DNNs are used in many 
different areas. Nonetheless, the black-box nature of DNNs makes it hard for a 
human to understand why and how they work so well. Consequently, many questions 
have arisen regarding the safety of DNNs.

Safety of DNNs can be analyzed in various aspects. Among those are four main 
aspects: verification, testing, adversarial attack and defense, and 
interpretability. Verification is a process of checking whether certain 
properties hold on the DNNs with provable guarantees. Testing is about 
generating and running test cases on DNNs under suitable coverage criteria. 
Adversarial attack creates perturbed inputs in order to show the lack of 
robustness of DNNs while adversarial defense aims to detect the attack and make 
the DNNs more robust. Interpretability is achieved by making DNNs more 
understandable to humans. This survey paper describes and summarizes researches 
on these four main aspects.


Date:			Tuesday, 26 November 2019

Time:                  	10:00am - 12:00noon

Venue:                  Room 4475
                         Lifts 25/26

Committee Members:	Prof. Fangzhen Lin (Supervisor)
 			Dr. Qiong Luo (Chairperson)
 			Prof. Shing-Chi Cheung
 			Dr. Shuai Wang


**** ALL are Welcome ****