GPUBreach: Privilege Escalation Attacks on GPUs Using Rowhammer (arxiv.org)

Researchers have unveiled a significant security vulnerability, dubbed GPUBreach, highlighting the susceptibility of NVIDIA GPUs with GDDR memory to targeted Rowhammer-based privilege escalation attacks. Unlike previous exploits that only caused untargeted bit-flips in machine learning model data, this new approach allows unprivileged users to manipulate GPU memory across processes. By ingeniously leveraging GPU page table management, attackers can introduce targeted bit-flips, leading to unauthorized access and the potential theft of sensitive information, including cryptographic keys. The implications for the AI/ML community are profound, as this vulnerability could undermine the integrity of machine learning models and the confidentiality of data processed on GPUs. The ability to tamper with GPU memory not only enables direct attacks on models but can also result in escalation to root privileges on the host CPU, thereby bypassing existing security measures like the Input-Output Memory Management Unit (IOMMU). As the use of GPUs in machine learning continues to expand, addressing these vulnerabilities will be crucial to maintaining system security and trustworthiness in AI applications.