An Access Control Mechanism for Pre-trained ML Models in Mobile Apps

Abstract

Utilizing machine learning inference in mobile platforms have been made realistic since recent smartphones may have high-performance processors and hardware accelerators of signal and image processing. However, it could bring issues. One of the most significant issues is that any machine learning inferences can be unlimitedly applied to personal data which once the application got permitted to. Furthermore, it could be quite hard to understand the behavior of a ‘pre-trained’ machine learning model and the behavior of an installed application can be transparently modified afterword by replacing a used model. This research aims to implement access control against such functionalities of machine learning inference for mobile platforms. In this paper, we describe how to verify used pre-trained models and how to implement permission control at Android Neural Networks API for Android as a target platform. And we also show an evaluation result with a prototype implementation.