Research on security of GPU memory in deep learning systems

Abstract

Modern deep learning frameworks rely heavily on GPUs to accelerate the computation. However, the security implication of GPU device memory exploitation on deep learning frameworks has been largely neglected. In this dissertation, we argue that GPU device memory manipulation is a novel attack vector against deep learning systems. We present a novel attack method leveraging the attack vector, which makes deep learning predictions no longer different from random guessing by degrading the accuracy of the predictions. To manipulate GPU device memory, we also propose a novel exploit technique that can execute arbitrary code by exploiting a vulnerability of a GPU function. To the best of our knowledge, we are the first to show a practical attack that directly exploits deep learning frameworks through GPU memory manipulation. We demonstrate the attack against DjiNN, a DL inference server, in a realistic environment to present the practicality. We confirmed that our attack works on three popular deep learning frameworks. TensorFlow, CNTK, and Caffe, running on CUDA. Finally, we propose defense mechanisms against our attack, and discuss concerns of GPU memory safety.