A Survey of Techniques for Improving Security of GPUs

Sparsh Mittal, S. B. Abhinaya, Manish Reddy, Irfan Ali

Abstract

As Graphics Processing Units (GPUs) transition from specialized gaming hardware to general-purpose accelerators in finance, healthcare, and defense, their security has become a critical concern. This survey systematically categorizes the landscape of GPU vulnerabilities—ranging from passive data-leaking side channels to active malware and Denial-of-Service (DoS) attacks—and evaluates the emerging countermeasures intended to secure the massively parallel future of computing.

Introduction

The computing industry is at an "inflection point" where GPUs are central to cryptography, cloud computing, and autonomous systems. Despite this, GPU security research lags significantly behind CPU security. Three core challenges define this gap:

  1. Isolation Gaps: The host CPU often lacks visibility into GPU execution, allowing malware to operate undetected.
  2. Security through Obscurity: Proprietary microarchitectures and undocumented memory management force researchers to rely on reverse engineering.
  3. Data Remanence: Unlike CPUs, GPUs frequently fail to clear memory after process termination, enabling malicious actors to "scrape" sensitive data left by previous users.

Taxonomy of GPU Attacks

The survey classifies GPU security research into two primary categories: passive and active attacks. Passive attacks, such as information leakage and timing side-channels, exploit system characteristics to extract data. Active attacks, including malware and buffer overflows, involve malicious code execution or service disruption.

Taxonomy of GPU security vulnerabilities, affected components, and application domains.

Key Insights and Vulnerabilities

The Coalescing Paradox

Memory coalescing, designed to boost performance by merging memory requests, creates a deterministic timing signature. Attackers can exploit this to count coalesced requests and mathematically derive secret keys, such as those used in AES-128.

Browser "Ghosting"

Rendered images from closed browser tabs often persist in GPU global memory. Researchers have demonstrated the ability to reconstruct images of visited websites, including personal information, by simply allocating the same memory space in a subsequent process.

Malware Safe-Havens

Because antivirus software typically runs on the CPU, it struggles to inspect the "black box" of GPU execution. This allows malware to perform polymorphic unpacking or run stealthy keyloggers that bypass traditional OS hooks.

Key Results and Defensive Measures

The survey highlights several critical findings from the literature regarding the efficacy of attacks and the cost of defense:

  • AES Cracking: Timing side-channels can recover a full 16-byte AES-128 key in approximately 30 minutes.
  • Memory Scraping: In shared environments, malicious kernels can leak 100% of data in shared and global memory under specific conditions.
  • DoS Effectiveness: Flooding specific GPU functions can cause total system freezes on major operating systems, requiring hard reboots.
Conditions required for complete or partial information leakage across different GPU memory types.

Defensive strategies like Random-sized Subwarps (RSS) and Randomized Coalescing aim to mitigate timing attacks. However, software-based detection tools like "Oclgrind" can incur performance overheads of up to 300x, highlighting the need for hardware-level solutions.

Conclusion and Future Outlook

Future research must address the unique challenges of "fused" chips where CPUs and GPUs share silicon and memory. Furthermore, the industry must move away from "security through obscurity" toward standardized documentation and open-source tools. Achieving robust GPU security without sacrificing performance remains the primary hurdle for the next generation of parallel computing.

Summary updated

BibTeX
@article{mittal2018survey,
  author = {Sparsh Mittal and S. B. Abhinaya and Manish Reddy and Irfan Ali},
  title = {{A Survey of Techniques for Improving Security of GPUs}},
  journal = {J. Hardw. Syst. Secur.},
  year = {2018},
  url = {https://openreview.net/forum?id=YI0I8gm6H0},
}
Figure