SecureRC: A system for privacy-preserving relation classification using secure multi-party computation

Abstract

Natural Language Processing (NLP) transforms human language into machine language that can be understood by machines through computer technology. Relation classification is an important semantic extraction task in NLP, which can accurately obtain the semantic relationship between two entities in a text. It has been used extensively in many NLP applications, such as information extraction and question answering. Relation classification can extract accurately relationships between two entities from large amounts of linguistic data. However, collecting such vast amounts of data in relation classification poses serious privacy issues. As far as we know, there is no existing work that implements privacy-preserving relation classification task. In this paper, we consider how to implement privacy-preserving relation classification using attention-based gated recurrent unit (GRU) network. Specifically, we first design three basic privacy-preserving protocols for the non-linear functions (sigmoid and tanh) using secure multi-party computation (MPC). Then, we propose a secure computation protocol SecureGRU for the GRU network based on these three basic protocols. Finally, based on the SecureGRU and the attention mechanism, we obtain the privacy-preserving relation classification system SecureRC. In the semi-honest adversary model, we prove the security of these protocols. Any honest-but-curious adversary is not able to obtain anything beyond what he is allowed to learn. The proposed system is implemented in Python. Experimental results demonstrate the performance of our proposed protocols.