[Defense] Linguistic Deception Detection - Models, Domains, Behaviors, Stylistic Patterns to Large Language Models (LLMs)
Wednesday, April 16, 2025
3:00 pm - 4:30 pm
In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
Sadat Shahriar
will defend his dissertation
Linguistic Deception Detection - Models, Domains, Behaviors, Stylistic Patterns to Large Language Models (LLMs)
Abstract
Deception in language—ranging from fake news and spam to phishing and rumor—has long been a tool for manipulation, exploiting linguistic ambiguity and psychological triggers to mislead readers. Deception spanned varied domains, yet shared common traits, which enabled the development of domain-independent detection methods that transferred knowledge across tasks using feature augmentation and multi-task learning. Psychological modeling further revealed how deception often plays on urgency, fear, and enticement. However, with the advent of Large Language Models (LLMs), the landscape of deception has shifted dramatically. These models can generate fluent, context-aware, and human-like text that often evades even SoTA detectors, blurring once-reliable cues of manipulation. Beyond mundane misuse to aid in fake reviews or partisan journalism, LLMs exhibit a more profound ability: generating scientifically coherent, logically sound ideas that closely resemble human reasoning. While this raises serious concerns around idea attribution and originality within the broader deception landscape, it also opens an opportunity to understand the underlying thought patterns of LLMs—moving beyond shallow stylistic rephrasings to deeper cognitive structures. This dissertation unifies classical deception paradigms with emerging LLM-centric challenges, offering a comprehensive framework to detect and reason about deception in its many evolving forms.
Wednesday, April 16, 2025
3:00 PM - 4:30 PM
PGH 501B
Dr. Arjun Mukherjee, dissertation advisor
Faculty, students, and the general public are invited.