2025 Projects
The Effectiveness of using rs-fMRI for Prediction of Cognitive Impairment Stage Conversion
Advisor: Dr. Siadat
Alzheimer's Disease (AD) is a neurodegenerative disease that affects over 7 million individuals in America alone. Alzheimer's Disease currently has no cure but there are ways to delay the progression of it if caught early enough. Our goal is to show whether resting-state fMRI images can be used as an effective tool in order to predict the progression of the disease. For this we took 303 patients with rs-fMRI images and preprocessed them in fMRIPrep. We then converted the preprocessed fMRI images into Functional Connectivity (FC) Matrices to get the correlation of BOLD signals from the patients fMRI. Finally the FC Matrices are turned into graphs to be passed through our Graph Neural Network (GNN) model, where it predicts if the patient will remain stable or convert to the next stage. The stages we are considering are Cognitively Normal (CN), Mildly Cognitively Impaired (MCI), and Alzheimer's Disease (AD). From our current testing we reported an accuracy score of about 68% for the prediction of stable versus converter patients. With our results it is clear that fMRI scans can be used effectively in the prediction of Alzheimer's Disease progression and may provide greater results when combined with other modalities of medical data.
Lecture Notes |
Can a Large Language Model Win a Kaggle Competition?
Advisor: Dr. Wardat
Large Language Models (LLMs) have demonstrated exceptional capabilities across diverse application domains. However, their potential in solving intricate problems within competitive environments, such as Kaggle competitions, remains relatively unexplored. This paper presents the first investigation into leveraging LLMs to autonomously build deep learning solutions specifically for Kaggle competitions. Our proposed method, CompeteRAG, employs a Retrieval-Augmented Generation (RAG) framework, utilizing top-voted Jupyter notebooks as exemplary templates to construct targeted solutions. If initial code-generation attempts result in incomplete or erroneous outputs, CompeteRAG dynamically analyzes these errors and iteratively refines the generated code. We rigorously evaluate CompeteRAG's efficacy using two state-of-the-art (SoTA) LLMs specialized in code generation: GPT-04-mini-high and DeepSeek-R1. Additionally, we introduce a comprehensive dataset comprising highly-rated Jupyter notebooks that build DNN solutions for Kaggle competitions. Our study provides critical insights into the strengths and inherent limitations of current LLM methodologies when confronting complex structured-data tasks common in competitive environments. Experimental results underscore CompeteRAG's robust capability in generating effective deep neural network (DNN) solutions within Kaggle's competitive landscape. Additionally, we introduce a comprehensive dataset comprising highly-rated Jupyter notebooks that build DNN solutions for Kaggle competitions. This dataset serves as a valuable resource for enhancing future code-generation techniques. Our study provides critical insights into the strengths and inherent limitations of current LLM methodologies when confronting complex structured-data tasks common in competitive environments.
Lecture Notes |
Optimizing Cryptographic Embeddings in LLM Text
Advisor: Dr. Hanna
What if a dictator bans encryption and forces all communication to be public? Is private communication still possible? We explore a cryptographic framework that enables secure messaging by hiding in plain sight, embedding encrypted messages into natural-sounding text generated by large language models (LLMs). To casual observers, these outputs appear indistinguishable from typical LLM generations. Building on the work of Gligoroski et al., we extend their cryptographic embedding scheme by improving its resistance to known-position attacks—threat scenarios where an adversary knows the exact locations of embedded characters. Our modified encoding strategy ensures that even with positional knowledge, the adversary gains minimal advantage in distinguishing ciphertext from innocuous text. To improve coherence and embedding success rates, we introduce a novel tree-based decoding framework that recursively backtracks through LLM generations, selecting token paths that best align with the encrypted message while preserving fluency and meaning. This method significantly enhances the quality of the cover text without compromising the stealth of the embedding. Finally, we present an early investigation into context-based attacks, an underexplored threat model wherein adversaries exploit thematic or topical patterns of LLMs to detect hidden content. We assess how context may unintentionally betray the presence of cryptographic embeddings and propose mitigation strategies within our framework. Our work advances the field of steganography in generative models, offering a path toward resilient, high-quality covert communication in adversarial environments.
Lecture Notes |
PubMedBERT: Improving Retrieval-Augmented Generation in Medical Question
Advisor: Dr. Qiang
Large Language Models (LLMs) are advanced AI models trained on extensive datasets of text that have strong capabilities in determining semantic relationships between words to answer general questions. While they perform well across various domains, these models have a few common issues, particularly in the medical field. LLMs struggle with hallucinations, incorrect or misleading responses, and having outdated knowledge due to static training data. This project explores Retrieval Augmented Generation (RAG), which is a framework designed to tackle these issues. RAG enhances the performance and accuracy of LLMs by retrieving up-to-date data from external sources relevant to the user's question. It incorporates this information along with the query as input to the LLM, which generates a response based on its training set along with this additional information. Our focus is to apply this framework in the biomedical domain by utilizing RAG with datasets such as PubMedQA, which consists of biomedical question answering compiled from PubMed abstracts. This helps us achieve our goal of evaluating and improving the model's ability to answer complex biomedical questions accurately, making LLMs more practical for clinical and scientific scenarios.
Lecture Notes |
Adversarial Attacks and Defenses for Panoptic Perception Models in Autonomous Driving
Advisor: Dr. Xu
Panoptic perception models in autonomous driving use deep learning models to interpret their surroundings and make real-time decisions. However, these models are susceptible, carefully designed noise can fool models all while being imperceptible to humans. In this project, we investigate the impact of black-box adversarial noise attacks on three core perception tasks: drivable area recognition, lane line segmentation, and object detection. Unlike white-box attacks, black-box attacks assume no knowledge of the model's internal parameters making them a more realistic and challenging threat scenario. Our goal is to evaluate how such an attack affects the model's predictions and explore countermeasures towards such attacks. In response to our implemented attack, we have tested various defense methods. With each defense method we have assessed the recovery on prediction accuracy. This research aims to provide valuable insights into the vulnerabilities of panoptic perception models and highlights strategies for enhancing their resilience against adversarial manipulation within real-world scenarios.
Lecture Notes |
BBOB: Open, Adaptive Object Detection with a Vision-Language TinyLLM
Advisor: Dr. Hanna
Lecture Notes |
PA-JJAMA: An LLM Based Intrusion Detection System for CAN Bus Networks
Advisor: Dr. Qiang
Recent studies have demonstrated significant success in detecting attacks on the Controller Area Network (CAN) bus network using machine learning and deep learning models, including convolutional neural networks and transformer-based architectures. Building on this foundation, our work investigates the use of large language models (LLMs) not only for intrusion detection but also for providing interpretable explanations of their decisions. We fine-tuned three LLMs, i.e., SecureBERT, LLaMA-2, and LLaMA-3, for intrusion detection on CAN bus data. Among them, LLaMA-3 delivered the best results, achieving SOTA performance on the Car-Hacking dataset. Beyond attack classification, we evaluated LLaMA-3’s ability to generate reasoning for its decisions through zero-shot prompting. The model successfully articulated its rationale, particularly for Denial-of-Service (DoS) attacks, demonstrating strong potential for explain-ability in intrusion detection systems. These findings highlight the potential of LLMs to serve as a highly accurate intrusion detection system while simultaneously providing interpretable explanations, thereby enhancing the investigative capabilities of cybersecurity professionals.
Lecture Notes |