Steganalysis of Adaptive Image Steganography using Convolution Neural Network and Blocks Selection

The method of detecting secret information in images in order to discriminate between a suspect and carrier images are known as image steganalysis. Since content-adapted steganographic methods adaptively integrate knowledge into regions with distortion-based rich textures, state-of-the-art steganalysis approaches with the development of steganography technology cannot achieve the required detection performance. This makes it difficult to extract effective steganalysis features. Due to the impressive performance of Convolution Neural Network (CNN) in the field of image processing, a growing number of research papers are focusing on developing steganalysis methods based on the CNN. Moreover, most research works rely on the entire image to extract the steganalysis features. In this paper, we introduce a new image steganalysis method by designing an entropy-based regional selecting method, and a new CNN framework to discriminate between a stego and cover image. We design a new approach for image block selection that finds the blocks with the highest entropy, allowing the CNN to concentrate on complex textures in image regions while reducing computational complexity. To allow feature diversity, we built a new CNN framework composed of two subnets of different kernel sizes, which we then repeatedly combined and separated. Those two aspects are consequently improved the performance with a reasonable number of epochs for training. When the results of the experiments are compared, our proposed method can lead to improving the detection accuracy of content-adaptive steganographic methods.