Virt-ML: A Virtualization-based Isolation Strategy for Privacy-Preserving Machine Learning on Mobile Devices

Nowadays machine learning models, particularly deep learning models, are extensively used on mobile devices. Performing such AI tasks in mobile application usually call for preserving privacy or protecting trusted computing. To simplify the solution, many relevant researches take advantage of TrustZone to provide a secure executing environment for machine learning models. However, the Trusted Computing Base (TCB) may be bloated as there should be more and more secure applications running in the same isolated executing environment provided by TrustZone. In this work, we proposed Virt-ML, which fully utilizes the virtualization technology provided by the arm cpu to provide similar isolated computing environment for machine learning models on mobile devices. With Virt-ML, we can avoid TCB bloat and reduce the risk of a compromised isolated execution environment in a virtual machine or host operating system. Virt-ML utilizes a DICE-based protection method to implement our strategy and its performance is stable in our experiment.