TranScope: What the Software Hides About LLM Training Data, the Hardware Reveals at Scale, and Accelerators Magnify

This project looks at micro‑architectural side‑channel weaknesses in large language models (LLMs) and vision transformers when they run inference. It studies how knowing whether a piece of data was used in training changes the hardware’s execution pattern. Even if the software output is constant‑time and masks confidence, the hardware still shows different behavior. We examine how transformer models interact with hardware at the cycle level, including Translation Lookaside Buffers (TLBs), page tables, and on‑core accelerators. The key idea is that tokenization done during training moves vocabulary tokens to different places in memory. That shift changes page‑table accesses during later inference, creating a measurable side‑channel. The project supplies code to model

these micro‑architectural differences, builds simulation frameworks that record cycle‑level traces, and tests whether membership‑inference attacks work without using input‑dependent branches or dynamic optimizations. By looking at these hardware signals, we develop a way to tell if data is in‑distribution or out‑of‑distribution, which helps design secure hardware accelerators and privacy‑preserving machine‑learning deployment pipelines.

Message: Hi FE, I need help with “TranScope: What the Software Hides About LLM Training Data, the Hardware Reveals at Scale, and Accelerators Magnify” in “Computer Science & Engineering / Hardware Security”