Privacy-preserving word vectors learning using partially homomorphic encryption

Abstract

This paper introduces a privacy-preserving scheme for learning GloVe word vectors on encrypted data. Users first encrypt their private data using a partially homomorphic encryption algorithm and then send the ciphertext to a cloud server to execute the proposed scheme. The cloud server generates high-quality word vectors for subsequent machine learning tasks by filtering out disturbances. We conduct a theoretical analysis of the security and efficiency of the proposed approach. Experimental results on real-world datasets demonstrate that our scheme effectively trains word vectors without compromising user privacy or the integrity of the word vector model, while keeping the user-side implementation lightweight and offline.