A New Steganalysis Method Using Densely Connected ConvNets

摘要

Steganography is an ancient art of communicating a secret message through an innocent-looking image. On the other hand, steganalysis is the counter process of the steganography, which targets to detect hidden trace within a given image. In this paper, a new approach to steganalysis is presented to learn prominent features and avoid loss of stego signals. The proposed model uses diverse sized filters to capture all useful steganalytic features through a densely connected convolutional network. Moreover, there is no fully connected network in the proposed model, which allows testing any size of images regardless of the image size used for training. To justify the applicability of the proposed scheme, it has been shown experimentally that the proposed scheme outperforms most of the related state-of-the-art methods.