Vulnerable Intel GPU Context: Prohibit Complete Context Restore by Modifying Kernel Driver

GPUs are increasingly utilized in modern virtual computing platforms, as they not only specialize in graphics-related tasks but also accelerate the computation of various workloads. Consequently, vendors are developing new GPU virtualization technologies to enable the efficient sharing of a single GPU in virtualized environments. However, these advancements also introduce the potential for novel attack scenarios. GPUs employ contexts to manage multiple processes, executing context save/restore operations during context switches. We devised an attack primitive that prematurely terminate context restoration by exploiting mechanisms within the context-switching process. This primitive allows other contexts to observe the state of previously executed contexts on the GPU.In virtual computing environments, where multiple virtual machines share a single GPU through GPU virtualization, an attacker with root privileges on one virtual machine can leverage this attack primitive to circumvent security boundaries between virtual machines. In this paper, we successfully established a covert channel between two distinct virtual machines and conducted a website fingerprinting attack. Finally, we propose mitigation strategies to prevent the described attack.