Convolutional Neural Network for Larger JPEG Images Steganalysis

Abstract

This paper proposes an effective steganalytic scheme based on CNN in order to detect steganography on larger JPEG images. Most of the CNN schemes were designed very deep to achieve high accuracy, resulting in inability to train large size images due to the limitation of GPUs’ memory. Most existing network architectures use small images of 256 $\times$ 256 or 512 $\times$ 512 pixels as their detection objects which are far from meeting the needs of practical applications. Meanwhile, the resizing operation on stegos will make the slight noise signal caused by steganography become difficult to detect. In our proposed network architecture, we try to solve the problem by compressing the depth of network structure. And in order to reduce the data dimension, we apply a histogram layer to transform the feature maps to feature vectors before the fully connected layer. We test our network on images of size 512 $\times$ 512, 1024 $\times$ 1024 and 2048 $\times$ 2048. For different application scenes, we take two methods to generate large samples. The result demonstrates that the proposed scheme can make directly training the steganalysis detectors on large images feasible.