Private Evaluation of a Decision Tree Based on Secret Sharing

Abstract

There has been increasing interest in developing privacy-preserving algorithms for evaluating machine learning (ML) models. With the advancement of cloud computing, it is now possible for model owners to host their trained ML models on a cloud server and offer cloud computing solutions on different ML tasks to users (clients). Thus private evaluation of ML models is an attractive area of research as it allows solution providers to protect their propriety ML models and users to protect their sensitive data while using cloud computing solutions. In this work, we propose an algorithm to privately evaluate a decision tree. We examine current state-of-the-art private evaluation protocols and present a solution that is sublinear in tree size and linear in tree depth. The key feature of our proposal is that it is entirely based on secret sharing and thus there are no computational costs associated with heavy cryptographic primitives such as modular exponentiation. We propose a new method to privately index arrays that avoids the use of public/symmetric key cryptosystem, typically associated with private array indexing protocols. The results of our experiments show that our solution has a low communication cost compared to existing methods (lower by a factor of $\approx$10 in the online phase), and demonstrate a faster runtime at low network latency (such as LAN network). We conclude by suggesting improvement to our protocol and proposing potential areas of future research.