Demystifying and Exploiting ASLR on NVIDIA GPUs

GPUs are foundational to modern AI workloads, powering deep learning training and inference. As their deployment becomes increasingly widespread, GPUs have also emerged as attractive targets for attackers. To strengthen their defenses, security measures, such as Address Space Layout Randomization (ASLR), are deployed. However, in contrast to the extensive research on CPU ASLR, in-depth studies of GPU ASLR are still missing.

This paper presents the first comprehensive examination of ASLR on NVIDIA GPUs. We propose two novel techniques to thoroughly inspect memory mappings and collect randomized GPU addresses at scale. Leveraging these techniques, we construct a fine-grained GPU memory map and introduce entropybased metrics to quantify the strength of randomization. Our study uncovers multiple previously unknown weaknesses of ASLR on NVIDIA GPUs, including an unrandomized GPU heap and correlated ASLR offsets between GPU and CPU regions, which undermine the security of both GPU and CPU ASLR. These findings have been confirmed by NVIDIA. Furthermore, we conduct a practical case study demonstrating how these weaknesses can be exploited to infer CPU ASLR offsets from the GPU. Finally, we give mitigations to enhance GPU ASLR security.