ORCID Identifier(s)

0000-0003-0277-6315

Graduation Semester and Year

2022

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

Department

Computer Science and Engineering

First Advisor

Shirin Nilizadeh

Abstract

The massive advances on the field of deep neural networks in the 2000 and 2010 decades led to an overwhelming adoption of these algorithms on all sorts of domains and applications. Under this widespread adoption scenario, it is natural that these neural networks have also been employed on safety-related use cases, bringing substantial improvements to the performance of existing as well as novel systems. Examples of these safety-inclined applications include scene recognition, object detection and tracking, speech recognition, audio event detection and classification, just to cite a few ones. Unfortunately, these neural network algorithms have been shown to be vulnerable to different forms of attacks that can prevent them from performing as intended and as designed. These attacks have also, so far, been shown to be impossible to be fully eliminated or even dealt with to a definitive degree of satisfaction. This is because these attacks exploit the very fundamental way these algorithms are conceived in the first place, deriving their malicious efficacy from the very intrinsic neural networks properties. The focus of this dissertation is on audio event detection (AED) systems and on to seek to contribute for the advance of neural network safe use on the AED domain. Existing real AED systems are tested to exhaustion to evaluate the state-of-the-art. Research and implementation efforts are then switched to neural networks (NN), the main component behind the AED capabilities by several of these modern systems. Throughout this doctoral research, different state-of-the-art AED devices are field tested, several AED classifiers are implemented, attacked, as well as defended, and a full End-to-end AED system is proposed. These experiments are done under the objective to generate new knowledge to contribute to the mitigation and bridging of the existing gaps in practical AED systems.

Keywords

Audio event detection, Deep learning, Adversarial attack

Disciplines

Computer Sciences | Physical Sciences and Mathematics

Comments

Degree granted by The University of Texas at Arlington

30369-2.zip (6563 kB)

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.