InStorm: An Instruction-Level Vulnerability Testing Framework for Graphics Processing Units

View at publisher Summarise with VIRAT
Authors: Yang Zhang, Jian Wang, Z. Cheng, Z. Zhang
Cited by: 0
References: 23

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In the last decade, an increasing number of processor vulnerabilities have been disclosed. However, existing methods have not tested the integrity and security of the GPU instruction set and microarchitecture. In this paper, we propose an instruction-level vulnerability testing frameworkInStormfor modern GPUs. Due to manufacturers' closed-source strategies, we reverse engineer the target ISA and generate efficient testing instructions by skipping redundant ones. Then, the generated instructions are executed on the target GPU to identify the potential vulnerabilities. Unfortunately, we discovered dozens of undocumented instructions with unexpected functions across six NVIDIA GPUs. Even worse, these instructions can bypass the GPU exception handling mechanism. In addition, we study a vulnerability exploitation method. An attacker can inject undocumented instructions into the victim GPU application by calling a malicious kernel loading function. Finally, we evaluate our exploitation method on a generative adversarial network accelerated by the NVIDIA GPU. The attack result demonstrates that instruction vulnerabilities can pose a significant security risk to GPU applications. We have responsibly disclosed these vulnerabilities and obtained an identifier CNVD-2025-10321.

Fields of Study & Subjects

Security and Verification in Computing Cryptographic Implementations and Security Parallel Computing and Optimization Techniques

Publication details

Journal / source
IEEE Transactions on Dependable and Secure Computing
Volume
23
Issue
4
Pages
7945–7957
Publication date
31 March 2026
Language
EN