Generating steganographic image description by dynamic synonym substitution

With the popularity of photo-sharing websites, it is wise to select these websites on which people usually post images and semantically related descriptive texts as channels of steganography. How to preserve the semantic correlation between the descriptive text and corresponding image for the security of text steganography on these websites is a challenging task. In this paper, we propose a novel steganographic image caption model (SIC) based on a typical image caption model Neural Image Caption (NIC) to automatically generate stego descriptions for given images. In the process of generation, secret information is embedded in the description by an elegant embedding method called dynamic synonym substitution (DSS) which can overcome the main drawback of normal synonym substitution based methods that the statistical characteristics of synonyms frequencies are changed after embedding. Our model takes the advantages of both generation based and synonym substitution based steganography and the good performance is evaluated by subjective and objective experiments. Experimental results show that stego descriptions generated by our model are almost as natural as descriptions generated by NIC and achieve high security against statistical analysis of steganalysis tools. (C) 2019 Elsevier B.V. All rights reserved.