Universal Deep Network Steganalysis of Color Based on Channel Representation

Abstract

Up to now, utmost being steganalytic styles are designed for grayscale images, and they aren't suitable for color images that are extensively used in current social networks. To begin with, we divide the input image based on the different embedding spaces of its colors, resulting in three separate channels.. The proposed system includes preprocessing, convolutional, and bracket modules. To save the steganographic vestiges in each color channel, in preprocessing module, we originally separate the input image into three channels according to the corresponding embedding spaces(i.e RGB for spatial steganography and YCbCr for JPEG steganography), and also prize the image residuals with 62 fixed high- pass pollutants, eventually concatenated all abbreviated residuals for posterior analysis analysis rather than adding them together with normal complication like being CNN- grounded steganalyzers. To accelerate the network confluence and effectively reduce the number of parameters, in convolutional module, we precisely design three types of layers with different roadway connections and group complication structures to further learn high- position steganalytic features. In bracket module, we employ a global normal pooling and completely connected subcaste for bracket