Depthwise Separable Convolutional Neural Network for Confidential Information Analysis

Abstract

Confidential information analysis can identify the text containing confidential information, thereby protecting organizations from the threat posed by leakage of confidential information. It is effective to build a confidential information analyzer based on a neural network. Most of the existing studies pursue high accuracy to design complex networks, ignoring speed and consumption. The optimal defense is to automatically analyze confidential information without compromising routine services. In this paper, we introduce a lightweight network, DSCNN, that can be adapted to low-resource devices. We also introduce two hyper-parameters to balance accuracy and speed. Our motivation is to simplify convolutions by breaking them down because the space dimension and channel dimension are not closely related in the convolutions. Experimental results on real-world data from WikiLeaks show that our proposed DSCNN performs well for confidential information analysis.