University of Central Florida
Methods For Defending Neural Networks Against Adversarial Attacks
Abstract
dc:description.abstractConvolutional Neural Networks (CNNs) have been at the frontier of the revolution within the field of computer vision. Since the advent of AlexNet in 2012, neural networks with CNN architectures have surpassed human-level capabilities for many cognitive tasks. As the neural networks are integrated in many safety critical applications such as autonomous vehicles, it is critical that they are robust and resilient to errors. Unfortunately, it has recently been observed that deep neural network models are susceptible to adversarial perturbations which are imperceptible to human vision. In this thesis, we propose a solution to defend neural networks against white box adversarial attacks. The proposed defense is based on activation pattern analysis in the frequency domain. The technique is evaluated and compared with state-of-the-art techniques on the CIFAR-10 dataset.
Degree
thesis:*- Grantor dc:publisher
- University of Central Florida
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shah, Sharvil
- Contributors dc:contributor
-
- Ewetz, Rickard
Subjects
dc:subject × 3- Adversarial attacks; Defense mechanisms; Activation pattern analysis; CIFAR-10 dataset; Deep learning
- Computer Sciences
- Neural networks (Computer science)--Statistical methods; Neural networks (Computer science)--Research; Neural networks (Computer science)--Design and construction; Computer vision--Research; Deep learning (Machine learning)
Rights
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
- CFE0009258; DP0026862