Graduate Defense: Chandler Underwood

Thesis Information

Title: Transformer Reinforcement Learning Approach to Attack Automatic Fake News Detectors

Program: Master of Science in Computer Science

Advisor: Dr. Francesca Spezzano, Computer Science

Committee Members: Dr. Edoardo Serra, Computer Science; and Dr. Yantian Hou, Computer Science

Abstract

“Misinformation and disinformation disguised in the form of fake news stories have garnered a lot of attention as of late largely because they confuse and often anger the public, leading to a less cohesive society for us all. In response to the ever-growing issue of fake news stories circulating on social media, researchers have crafted various solutions to predict the veracity of stories in hopes of catching illegitimate ones before they can spread. In response to this research area, a newer research area focused on attacking fake news detectors is forming. In this work we have built and adversarial text generator using T5 and reinforcement learning capable of attacking a vast array of fake news detection systems. This work focuses on attacking the state-of-the-art (SOTA) text-based fake news classifier and compare our performance directly with the SOTA text-based adversary. Our adversary is more effective than the current SOTA, but we continue to struggle to produce adversarial text that could avoid detection by some other system or human readers.”