MM-Stega: Multi-modal Steganography Based on Text-Image Matching

Abstract

This paper proposes a multi-modal steganography (MM-Stega) scheme based on text-image matching. Currently, most steganographic methods embed secret information into a cover by modifying its content. However, the distortion of the cover caused by the modification may be detected by steganalysis methods. Other steganographic methods hide secret information by generating covers, but the imperceptibility of this kind of steganographic methods is limited by the quality of the generated covers. Our method is different from these steganographic methods in two aspects. First, our method utilizes multi-modal covers, i.e., texts and images, while most steganographic methods use single-modal covers. Second, our method embeds secret information in the relevance between the texts and images without modifying or generating a cover, thus our method has strong resistance to steganalysis. Our method is based on a text-image matching model which can measure the similarity between a text and an image. The text-image matching model utilizes a visual semantic embedding (VSE) model, which can project texts and images into a common subspace. After choosing a text from the text database randomly, several images relevant to the text are selected with the text-image matching model on the basis of the secret information that needs to be hidden. Experimental results and analysis prove that our method has adjustable hiding capacity and desirable security.