Prerequisites:  Students are expected to have basic software development experience and an interest in cybersecurity or LLMs

Courses:  ECE 496 (or lower with instructor approval)

Focus:   W (Writing) O (Oral)

Description:  

Advisor:  Dr. Danielle Cummings, Visiting Professor

Project Overview:

The VENOM-RAG project investigates the vulnerability of Retrieval Augmented Generation (RAG) models to adversarial data poisoning, a critical threat to the integrity of AI-driven knowledge search. The project proposes the exploration of tailored attack methodologies against DoD-specific data, such as intelligence reports, sensor data, or technical manuals, aiming to manipulate the outputs of RAG models. Through rigorous vulnerability exposure and the development of novel neutralization strategies, this research will provide actionable insights and a framework to understand vulnerabilities in DoD's knowledge core, ensuring reliable and trustworthy information access for enhanced decision-making across all branches. Students will work on setting up local open source LLMs such as ollama and work with python and jupyter notebooks to expose methods for poisoning RAG model datasets along with developing mitigation tactics for prevention of dataset poisoning.

Related Paper:   PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models

Other Details:

• This capstone is offered in tandem with the ICS and ECE departments who will work together on project material
• Project is sponsored by NIWC (Naval Information Warfare Center) who students will work with providing guidance and developing project milestones