Deep Learning Techniques for Automation and Industrial Applications

8A Comparative Analysis of Different CNN Models for Spatial Domain Steganalysis

Ankita Gupta^*, Rita Chhikara and Prabha Sharma

Department of Computer Science and Engineering, The NorthCap University, Gurugram, Haryana, India

Abstract

Universal steganalysis detects whether an image is a cover image or a stego image with some hidden information embedded through various steganographic methods. Steganalysis started with simple machine learning techniques of feature extraction and then classification. The focus on steganalysis has now shifted towards exploring deep learning methods. The most popular and successful deep learning models for image classification are based on convolutional neural networks (CNNs). CNN networks also show a resemblance to traditional steganalysis of using filters for feature extraction. Due to the use of content-adaptive steganographic methods, the stego message is hidden more in the complex areas of the image, and thus cannot be detected with a simple statistical analysis of the image. The stego information in these steganographic methods basically affects the dependencies between the pixels introduced through various kinds of noise present in the image. Thus, the difference between the cover and stego image is identified through the noise part rather than the image content. Different researchers have used various preprocessing filters to calculate the noise residuals, passing them to the CNN network instead of images directly. This work uses a content-adaptive ...