Text Semantic Steganalysis Based on Word Embedding

Abstract

Most state-of-the-art detection methods against synonym substitution based steganography extract features based on statistical distortion. However, synonym substitution will cause not only statistical distortion but also semantic distortion. In this paper, we propose word embedding feature (WEF) to detect the semantic distortion. Furthermore, a fused feature called word embedding and statistical feature set (WESF) which consists of WEF and statistical feature based on word frequency is designed to improve detection performance. Experiments show that WESF can achieve lower detection error rates compared with prmethods.