Employing Brakerski/Fan-Vercauteren's for secure data classification in GRU networks for sentiment analysis

Abstract

Due to the widespread use of technology and the resulting vast amounts of big data, which are beneficial for product and application development, sentiment analysis is one such benefit. It allows users to understand their opinions and guide development and growth in the right direction. However, the use of sensitive textual data remains a matter of privacy and requires careful handling. Therefore, users must be provided with the necessary guarantees regarding the security of their privacy. This work introduces a privacy-preserving binary sentiment analysis system that categorizes text as positive or negative through comprehensive contextual understanding while safeguarding user data using homomorphic encryption. The Gated Recurrent Unit (GRU) model is trained on unencrypted data of type IMDB and performs inference on encrypted inputs with the BFV graded homomorphic encryption technique. To facilitate efficient encrypted inference, it was replaced non-polynomial activations with low-degree polynomial approximations, reduce the depth of multiplication. The accuracy attained by the proposed technique is 90.1%.