Rochester System Security Lab
University of Rochester
Rochester System Security Lab

(GPU) System Security and Performance

We investigate practical security vulnerabilities in modern systems—particularly GPUs—and design performance-conscious defenses. Our research focuses on the following areas:

  • GPU memory safety
  • CPU/GPU side channels
  • GPU microarchitectural optimization

Lab PI

Yanan Guo

Yanan Guo

Email: yanan.guo@rochester.edu

Website: yananguo.com

Assistant Professor

Department of Computer Science

University of Rochester

GPU security Side channels Memory safety
View full members page →

Research Highlights

GPU Side-Channel Attacks

Work 1 figure

Side-Channel Attacks Across GPU Instances (USENIX Security'26)

  • NVIDIA MIG is designed to provide cache and memory isolation between users.
  • This work demonstrates, for the first time, that MIG’s isolation is not perfect.
  • We introduce a cross-MIG cache side-channel attack that reveals input and output characteristics of LLM inference.
Paper Code
Work 1 figure

Cross-VM GPU TLB Side-Channel Attack (NDSS'26)

  • GPU virtualization is widely used on public clouds.
  • This work demonstrates the first cross-VM side-channel attack on virtualized GPUs.
  • It shows practical exploits in cloud gaming and virtual desktop workloads.
Paper Code

GPU Memory Safety

Work 1 figure

Attack on CUDA ASLR (IEEE S&P'26)

  • This work conducts a thorough analysis of the virtual memory layout of CUDA programs.
  • We identify several significant vulnerabilities in CUDA ASLR.
  • One of the identified vulnerabilities enables a GPU-to-CPU attack.
  • NVIDIA has fixed one of the identified vulnerabilities in driver version 570.
Paper Code Project
Work 1 figure

Memory Safety Tool for CUDA (USENIX Security'26)

  • This work develops CuSafe, the first compiler-level framework for detecting spatial and temporal memory bugs in CUDA programs.
  • Cusafe is open-source and runs on commodity NVIDIA GPUs without hardware or driver modifications.
  • On average, CuSafe incurs 13% runtime slowdown and 0.3% memory overhead.
Paper Code Demo
Work 1 figure

GPU Memory Exploitation (USENIX Security'24)

  • This work thoroughly analyzes buffer overflow vulnerabilities in CUDA on NVIDIA GPUs.
  • We show that buffer overflows can enable code injection and ROP attacks.
  • We uncover powerful ROP gadgets in CUDA libraries, including arbitrary memory read/write primitives.
Paper Code Demo

GPU Performance Characterization

Work 1 figure

Characterization of Cache Eviction Priority Hints on NVIDIA GPUs (MICRO'25)

  • This work reverse engineers the microarchitectural behavior of eviction priority hints on NVIDIA GPUs.
  • We show that the current implementation of these hints can lead to unexpected performance degradation.
  • For example, using evict_last to pin data in cache can instead trigger severe cache thrashing.
Paper Code

Sponsors