Georgia Institute of Technology
Hardware-Software Co-Designed GPU Security Mechanism for Memory Safety
Abstract
dc:description.abstractMemory safety has long been a critical challenge in low-level programming languages such as C and C++. Similar concerns have recently emerged in GPU programming, which relies on C/C++-style languages and now plays a vital role in critical systems such as autonomous driving and medical diagnostics. Recent studies have identified vulnerabilities in GPU programs, including buffer overflow attacks that adversaries can exploit to execute malicious code. These exploits pose a significant threat, potentially compromising the integrity of other users' deep learning applications on shared cloud servers. To address this pressing issue, this dissertation proposes hardware-software co-designed mechanisms to enhance GPU memory safety. By leveraging the massively parallel architecture and unique programming model of GPUs, these mechanisms integrate hardware and software collaboration to deliver highly optimized solutions that surpass the efficiency of directly applying CPU-centric approaches to GPUs. The first proposed mechanism, GPUShield, introduces region-based bounds checking using pointer tagging, tags stored in the upper bits of a pointer are utilized to retrieve bounds metadata in memory. By exploiting the GPU's specialized load/store instructions and employing static-time filtering, GPUShield achieves significant performance advantages over existing software-based GPU memory safety mechanisms. The second mechanism, LMI, is a fine-grained memory safety solution that employs in-pointer metadata for runtime bounds checking. To mitigate the reduced availability of upper bits for tagging caused by virtual memory expansion and multi-GPU support, LMI efficiently utilizes power-of-two-aligned pointers. This design enables fine-grained memory safety with minimal hardware overhead compared to earlier mechanisms
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Computer Science
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lee, Jaewon
- Advisor dc:contributor.advisor
-
- Kim, Hyesoon
- Committee members dc:contributor.committeemember
-
- Qureshi, Moinuddin
- Krishna, Tushar
- Mukhopadhyay, Saibal
- Lee, Jaekyu
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/77792