A Word-Embedding-Based Steganalysis Method for Linguistic Steganography via Synonym Substitution

Abstract

The development of steganography technology threatens the security of privacy information in smart campus. To prevent privacy disclosure, a linguistic steganalysis method based on word embedding is proposed to detect the privacy information hidden in synonyms in the texts. With the continuous Skip-gram language model, each synonym and words in its context are represented as word embeddings, which aims to encode semantic meanings of words into low-dimensional dense vectors. The context fitness, which characterizes the suitability of a synonym by its semantic correlations with context words, is effectively estimated by their corresponding word embeddings and weighted by TF-IDF values of context words. By analyzing the differences of context fitness values of synonyms in the same synonym set and the differences of those in the cover and stego text, three features are extracted and fed into a support vector machine classifier for steganalysis task. The experimental results show that the proposed steganalysis improves the average F-value at least 4.8% over two baselines. In addition, the detection performance can be further improved by learning better word embeddings.