Poisoning Attacks, LLMs, Machine Learning, Privacy
This work examines whether large language models can be poisoned to produce recommendations that lead users to install malware, and evaluates whether such attacks can circumvent existing alignment and safety techniques.
Supervisor: Dr. Bailey Kacsmar
Collaborators: Miriam Bakija, Samuel Feldman

Learning Analytics, Educational Data Mining, Natural Language Processing
Guided by the Simple View of Reading, which conceptualizes reading comprehension as the interaction between decoding and language comprehension, the analysis focuses on linguistic features related to word recognition and semantic understanding.
Supervisor: Dr. Carrie Demmans Epp

Natural Language Processing, Privacy, and Cryptography
She completed her undergraduate thesis under the supervision of Dr. Farig Yousuf Sadeque and Md Faisal Ahmed. The proposed protocol uses encryption to protect sender-receiver interactions, using LSTM and GRU models to assess emotional states by utilizing both symmetric and asymmetric encryption. Training the Word2Vec embedding model with encrypted words not only preserves vital syntactic and semantic relationships but also maintains data privacy. The research delves into potential vulnerabilities and strategies for their management at every stage of the protocol.

Natural Language Processing, Fine-tuning Transformers
Under the supervision of Annajiat Alim Rasel, contributed to a research project. Developed a dataset focused on gender-biased defamation in Bengali, a low-resourced language, by extracting data from various datasets. The models included state-of-the-art transformer-based NLP models, BanglaBERT, XLM-RoBERTa, m-BERT, and DistilBERT, to detect gender-specific defamation on Social Media.

Natural Language Processing, Image Captioning, Computer Vision
Under the supervision of Annajiat Alim Rasel, contributed to a research project. The VGG-16 model, a Convolutional Neural Network (CNN), is employed to extract features from images. Long Short-Term Memory (LSTM) RNN is then utilized to generate descriptive captions. The entire system is trained on the Flickr8k dataset, ensuring standard performance in generating contextually relevant captions for diverse images.