Tools
Tools
Deva, Abhinav (2025) Enhancing Fake News Detection Through GPT – 4o Generated Synthetic Data. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (758kB) | Preview |
Preview |
PDF (Configuration Manual)
Download (4MB) | Preview |
Abstract
This research proposes integrating GPT-4o's generative ability with ALBERT's parameter-efficient classifier framework to create adaptive detection systems. We build an improved prompting framework that uses Chain-of-Thought reasoning and limited-shot learning to make 15,000 fake news samples in five main categories. Using six detection model types (traditional ML and transformer), the framework performs extensive cross-evaluation experiments on both synthetic and real data. Using detection performance metrics, our solution provides a novel evaluation framework for evaluating the quality of synthetic data in adversarial settings.
The findings demonstrated an unequal relationship between generation and detection: models trained on synthetic data only correctly identified real fake news 53.3% of the time, while synthetic samples were not detected, resulting in an average performance drop of 27.6%. By maintaining 99.5% accuracy, ALBERT demonstrated remarkable strength, indicating that parameter-efficient architectures provide more robust defences against attacks than merely computational advantages.
Actions (login required)
 |
View Item |