Date: Tuesday, November 28, 2023

Time: 3:30 PM - 5:30 PM ET

Location:

      In-Person: KACB 1315

      Virtual: Click here to join the meeting

Jaewon Lee

Ph.D. Student

School of Computer Science

College of Computing

Georgia Institute of Technology

Committee:

Dr. Hyesoon Kim (advisor), School of Computer Science, Georgia Institute of Technology

Dr. Moinuddin Qureshi, School of Computer Science, Georgia Institute of Technology

Dr. Tushar Krishna, School of Electrical and Computer Engineering & School of Computer Science, Georgia Institute of Technology

Abstract

Graphic Processing Units (GPUs) had been considered acceptable even if they were insecure; however, the surge in the usage of Artificial Intelligence (AI) applications now involves GPUs in critical life and financial decision-making. Recent studies successfully demonstrate that attackers can induce failures in AI models by exploiting vulnerabilities in GPU memory.

We propose GPUShield, an efficient GPU memory bounds checking scheme that utilizes the characteristics of GPU programs. GPUShield minimizes metadata access by leveraging GPU's region-based memory access, metadata caching, and memory coalescing.

Additionally, we introduce our ongoing work, BNPL, a practical fine-grain GPU memory safety scheme. We implement efficient all-time bounds checking with pointer alignment so that we can eliminate the need to access metadata bounds in memory, thanks to GPU characteristics.

Finally, we address another type of threat: power side-channel attacks, and propose mitigation with a frequency binning strategy.