Back to search

University of Central Florida

The Susceptibility of Deep Neural Networks to Natural Perturbations

Abstract

dc:description.abstract

Deep learning systems have achieved great success in various types of applications in recent years. They are increasingly being adopted for safety-critical tasks such as face recognition, surveillance systems, speech recognition, and autonomous driving. On the other hand, it has been found that deep neural networks (DNNs) can easily be fooled by adversarial input samples. These imperceptible perturbations on images can lead any machine learning system to misclassify the objects with high confidence. Furthermore, they can be almost indistinguishable to a human observer. These systems can also be exposed to adverse weather conditions such as fog, rain, and snow. This vulnerability raises major concerns in security-sensitive environments. Therefore, vulnerability of deep learning systems to synthetic adversarial attacks has been extensively studied and demonstrated, but the impact of natural weather conditions on these systems has not been studied in detail. The main contribution of this thesis is exploring the effects of fog on classification accuracy of the popular Inception deep learning model. We use stereo images from the Cityscapes dataset and computer graphics techniques to mimic realistic naturally occurring fog. We show that the Inception deep learning model is vulnerable to the addition of fog in images. We also review the types of adversarial attacks and defenses, describe the state-of-the-art methods for each group, and compare their results. Adversarial images can be used to generate targeted attacks or non-targeted attacks. Targeted attacks misguide the deep learning networks to produce responses from a specific a priori determined class. In non-targeted attacks, all images in the dataset are not assigned to a specific class; instead, the output of the deep neural network is arbitrarily wrong. In this thesis, we create non-targeted, iterative, and physical attacks.

Degree

thesis:*
Grantor dc:publisher
University of Central Florida

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Ozdag, Mesut
Contributors dc:contributor
  • Jha, Sumit Kumar

Subjects

dc:subject × 3

Rights

Language dc:language
English

Identifiers

dc:identifier.*
Identifier
CFE0008018; DP0023158

Chain of custody

source
Harvested from
Central Florida
Base URL
stars.library.ucf.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Ozdag, Mesut. The Susceptibility of Deep Neural Networks to Natural Perturbations. University of Central Florida, https://stars.library.ucf.edu/etd2020/112