Text Steganalysis Based on Capsule Network with Dynamic Routing

Abstract

With the growth of natural language processing technology, coverless text steganography has attracted the attention of a large number of researchers. Most existing text steganalysis methods are based on traditional neural network to extract and analyze the semantic features of automatically generated steganographic text. However, due to the limitation of traditional neural networks to preserve subtle features, these methods cannot obtain satisfactory results when detecting the differences between steganographic text with low embedding rate and natural text. This paper demonstrates that using a capsule network to detect whether the natural text contains secret information and gets robust and accurate performance. The capsule network extracts and preserves the sematic features of text, analyzes the subtle differences between steganographic text and natural text. To strengthen the generalization of the method, we choose word2vec to vectorize text and use steganographic text generated based on RNN and variable-length coding as the data set for experiments. Experimental results show that detection accuracy of our method can achieve 92% in steganographic text with the low embedding rate (1–3 bit/word), which is about 7% higher than that based on other neural networks; in high embedding rate (4–5 bit/word), the detection accuracy can reach more than 94%.