Towards a Universal Steganalyser Using Convolutional Neural Networks.

A universal steganalyser has been the goal of many research leading to some good trials. Such steganalysers relied on machine learning and a wide range of features that can be extracted from images. However, increasing the dimensionality of the extracted features leads to the rapid rise in the complexity of algorithms. In recent years, some studies have indicated that well-designed convolutional neural networks (CNN) can achieve comparable performance to the two-step machine learning approaches. This paper aims to investigate different CNN architectures and diverse training strategies to propose a universal steganalysis model that can detect the presence of secret data in a colour stego-image. Since the detection of a stego-image can be considered as a classification problem, a CNN-based classifier has been proposed here. The experimental results of the proposed approach proved the efficiency in the main aspects of image steganography compared with the current state-of-the-art methods. However, a universal steganalysis is still unachievable, and more work should be done in this field.