Title:
Investigating The Security of Modern AI and Cloud Infrastructure
Abstract:
This defense examines security assumptions in AI and cloud infrastructure across three threat levels: shared memory attacks (leaking LLM user tokens via side channels, faulting GGUF models), shared hardware attacks (targeting stack/register data to bypass TLS, PQC schemes, and machine learning algorithms), and remote service attacks (bypassing LLM alignment and Guard models with adversarial suffixes).
This research reveals that isolation assumptions underlying modern AI deployment may be more fragile than commonly assumed across multiple levels of the interaction hierarchy.
Advisor:
Prof. Berk Sunar
ECE Department, WPI
Committee Members:
Prof. Fatemeh Ganji
ECE Department, WPI
Dr. Yarkin Doroz
NVIDIA
Jeff Hamalainen
MITRE Corp.