Improving Text-Image Matching with Adversarial Learning and Circle Loss for Multi-modal Steganography

摘要

This paper proposes a multi-modal steganography method based on an improved text-image matching algorithm. At present, most of the steganography methods are based on single modality of carriers and embed confidential information into the carriers by cover modification or cover synthesis. Since the distortions between the covers after embedding and the original covers are inevitable, these steganography methods can be detected by the existing steganalysis methods. To solve this problem, we propose multi-modal steganography which hides the confidential information in the semantic relevancy between two modalities of original carriers. The semantic relevancy between the two modalities is measured by text-image matching, which affects the impercep-tibility of the proposed method to a large extent. In order to increase the security of multi-modal steganography, we improve the current text-image matching algorithm with adversarial learning and circle loss. By selecting and transmitting the original multi-modal carriers with high relevancy, the proposed method can escape from the detection of current steganalysis methods. It is also illustrated by the theoretical analysis and experimental results that the semantic relevancy between the selected multi-modal carriers is enhanced.