Can ChatGPT Hack Your Computer?
Exploring myths and realities about the capabilities and limitations of language models like ChatGPT in cybersecurity.
Thesis Proposal
Fake News Detection
Combine LLMs and computer vision approaches to tackle misinformation. For example, integrate text-based analysis with image verification to identify and classify fake news, leveraging the strengths of both modalities.
Hallucination Detection
Develop techniques to identify when LLMs generate false or nonsensical information (“hallucinations”). Use metrics or methods to evaluate output reliability and explore corrective strategies to improve accuracy in various applications.
Mental Health Analysis
Analyze Reddit posts to identify typical writing patterns linked to mental health conditions. Use embeddings from Large Language Models (LLMs) to represent these patterns and compare across different models. Build a network of these embeddings and study its structure to uncover insights into mental health trends or behaviors.
Vision Distillation
In a Vision Transformer, distill the information from the query matrix (Q) onto a blank image to visualize how the model identifies and extracts key details, providing insights into its behavior when processing visual inputs.
Vision Explainability
Calculate the gradient of the attention matrix in a Vision Transformer and construct a multilayer gradient network. Create interpretive masks based on centrality, token importance, or other metrics to enhance the understanding of the model’s decision-making process.