Efficient and Privacy-Preserving Feature Selection Based on Multiparty Computation

Feature selection is a critical data preprocessing stage that has been proven beneficial in data mining and machine learning applications. As most current works focus on privacy during the training and inference tasks in machine learning, implementing privacy preservation in preprocessing is a powerful complement. In this paper, we present an <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</u>fficient and <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</u>rivacy-preserving <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">f</u>eature <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</u>election protocol (EPFS) based on secure multiparty computation (MPC). We customize a novel method called approximate fixed-point representation to reduce the bitwidth of the sample distribution probability, thereby decreasing communication overhead. We optimize the comparison protocol by reducing the high-order bits of values according to the characteristics of the datasets and design the feature score calculation protocol together with several other MPC-based sub-protocols. We also construct an efficient feature selection workflow to obtain the reduced feature matrix, which avoids the numerous calls of secure comparison and equality test protocols in loops. Experiments on several real-world datasets show that the improved comparison protocol achieves a 29%-53% improvement in runtime and a 6%-32% reduction in communication compared to the general comparison protocol. The optimized feature selection workflow exhibits an upper performance bound, achieving a 38% improvement in runtime compared to prior work. Besides, we implement secure logistic regression training based on the selection features, where the accuracy has improved by an average of 8% compared to training on raw features.