Efficient and fully outsourced privacy-preserving decision tree training and prediction based on homomorphic encryption

Abstract

Outsourcing machine learning models to cloud servers allows data owners to train and utilize models without investing in dedicated hardware. However, this approach raises significant concerns regarding the proprietary nature of the models and the data privacy, including the confidentiality of training data, intermediate computations, input queries, and prediction results. In this paper, we propose Secure Decision Tree (SDT), a secure and efficient framework for outsourcing decision tree training and inference. The proposed solution leverages homomorphic encryption and introduces a novel structure called the encrypted decimal matrix to enable computations on encrypted data without disclosing sensitive information. Unlike existing solutions, SDT ensures data privacy without involving the data owner during training or inference, avoids reliance on secure multi-party computation, and prevents exposure of secret keys to external parties. Furthermore, SDT protects the proprietary rights of trained models and conceals statistical properties of the data and model from the cloud. Experimental evaluations on benchmark datasets from the UCI data repository demonstrate that SDT achieves classification accuracy comparable to standard (unencrypted) approach while maintaining strong privacy guarantees and incurring minimal computational overhead.