Deep neural networks have demonstrated remarkable accuracy for most image classification machine learning tasks. However, these networks remain susceptible to adversarial attacks, where slight perturbations in input data produces a misclassification. Without effective defense, this vulnerability creates a significant obstacle to the practical applications of neural networks. Therefore, in this paper we propose four unique interpretations of adversarial attacks designed to test the limits of adversarial defenses. To conclude the paper we assess the strengths and weaknesses of the four defenses we designed and recommend an approach to ensure the safety and security of neural networks in the public domain.

• This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator

• Campbell, Alasdair
• Lasselle, Jared

Publisher

• Worcester Polytechnic Institute

Identifier

• 121382
• E-project-042324-130037

Palavra-chave

• Machine Learning
• Adversarial Defense
• Neural Networks
• Adversarial Attack

Advisor

• Li, Yanhua

Year

• 2024

Date created

• 2024-04-23

Resource type

• Major Qualifying Project

Major

• Computer Science

Source

• E-project-042324-130037

Rights statement

• In Copyright