Publications

This work examines whether machine unlearning transfers between text and vision across LLaVA-1.5, InstructBLIP, and IDEFICS. It finds that transfer is asymmetric and incomplete, with typographic attacks recovering knowledge that appeared to be forgotten. The proposed influence-guided method, CrossInf, narrows the transfer gap and improves attack robustness while preserving model utility.

While coronary imaging is widely used for anatomical assessment, CCTA reports play a distinct last-mile role in clinical care by providing CAD-RADS scores that guide patient management. We introduce CCTA-RADS, the largest publicly available dataset of 940 real-world CCTA reports from Tampa General Hospital, each annotated with CAD-RADS scores. Direct approaches, including state-of-the-art LLMs and fine-tuned BERT models, underperform on diverse real-world clinical data. To address this, we propose a two-stage pipeline that normalizes heterogeneous reports into structured JSON format with an LLM-based parser, then applies fine-tuned BERT classification, improving F1-score by 6%-13% over direct methods.

Coronary CT angiography (CCTA) is a valuable clinical tool for evaluation of coronary artery disease. However, the standardized framework, CAD-RADS, is not consistently applied due to variability in terminology and narrative format. We developed an AI-based system leveraging large language models (LLM) to automate the interpretation of CCTA reports. The training dataset comprised 940 reports from 2020-2024. BioBERT demonstrated the highest performance and was selected for deployment. External validation using 500 independent reports showed 87.6% overall accuracy in AI-interpreted CAD-RADS classification.

Network Intrusion Detection Systems (NIDS) play a critical role in identifying and mitigating malicious activities within computer networks. This thesis explores the application of advanced NLP techniques, particularly LLMs, to enhance NIDS performance. We investigate multiple approaches, including Masked Language Models (MLMs) such as BERT, RoBERTa, and DistilBERT, as well as large-scale generative models like Gemma (2B, 9B, and 27B parameter versions) for intrusion detection tasks. Our study implements standard ML models on NSL-KDD and CICIoT2023 datasets to establish baselines, then applies MLMs both as classifiers and feature extractors. We conduct experiments with various prompting strategies including Zero-Shot, One-Shot, In-Context Learning, and Chain-of-Thought reasoning.