Steganalysis of digital images by means of shallow and deep machine learning: existing approaches and new solutions

Abstract

The article considers the current state of the problem of steganalysis of digital images in order to develop and study effective methods of revealing hidden (invisible) messages in container images. In the first part of the article, we provide a classification of the existing approaches and detail the previously obtained results of steganalysis performed using shallow and deep machine learning methods. We also describe the indicator systems used in shallow machine learning today and classifiers based on them (ensemble methods, support vector machines, etc.). An alternative method is based on using deep convolutional neural networks with various modifications (additional layers, special activation functions, etc.). The paper presents the results of the comparative analysis of the effectiveness of different ap proaches and different neural network archi tectures used in steganalysis. The analysis was performed using standard image sets. Hidden messages were embedded using adaptive spatial steganography algorithms: WOW, HUGO, and S-UNIWARD. The study demonstrated the universality and effectiveness of deep machine learning and showed that it is a promising method that can be used in steganalysis. In the second part of the article, we suggest a new architecture for a deep neural network and describe its performance when applied in the steganalysis of colour images. The key idea of the suggested approach is to use relatively simple convolutional networks for subsequent analysis of small fragments (blocks) of initial large images. The obtained classification results are then fused in a sequence of binary features using a Naive Bayes classifier. The experiments were performed using the PPG-LIRMM-COLOR database. WOW and S-UNIWARD algorithms were used to embed steganographic messages with various payload. The precision of the steganalysis of large images is compatible with and, in some cases, even better than the results obtained by other authors.