Trusted Execution Environments in Protecting Machine Learning Models

Abstract

The adaptation and application of machine learning (ML) has grown extensively in recent years, and has awakened concern about the safety of intellectual property (IP) related to the machine learning models. The training of machine learning models is a time-consuming and expensive task, that has increased the demand of better solutions to protect the intellectual property of the machine learning models. This thesis explores the promising potential of Trusted Execution Environments (TEE) like Intel's Software Guard Extensions (Intel SGX), in protecting intellectual property related to machine learning models. The concern of ML model safety arises especially when the software solution needs to be distributed to clients or machine learning operations needs to be done in an untrusted environment. The main focus of this thesis is on Intel's SGX, which is one of the most used TEE implementations. This thesis tries to answer to the questions on how TEEs can be used to protect IP of the ML models, what aspects need to be considered and what limitations may arise.