Simultaneous convolutional neural network for highly efficient image steganography

Over the past decade, steganography has attracted a large number of researchers' attention because of its practical value to information security. Unlike cryptography which aims to protect the content of the message, steganography involves hiding the message within a transport layer. Most traditional methods directly hide information in one or two least significant bits. Recently, miscellaneous effort with deep learning approach has been devoted to hiding more information without losing the integrity of the transportation layer. In this paper, our work focuses on image steganography which means hiding an image (secret image) inside another image of the same size (cover image). A deep convolutional neural network with similar architecture to U-Net is employed as the hiding network, and a new training scheme is proposed to speed up the training phase. Through extensive experiments, it has been verified that the new network architecture, combined with the new training strategy, can result in lower mean square error of pixel difference whereas the training time is reduced by half.