Image stegoanalysis using deep neural networks and heteroassociative integral transformations

The problem of steganalysis of digital images is considered. The proposed approach is based on the use of deep convolutional neural networks with a relatively simple architecture, distinguished by the use of additional layers of special processing. These networks are trained and used for steganalysis of small fragments of the original large images. For the analysis of full sized images, it is proposed to carry out secondary post-processing, which involves combining the obtained classification results in blocks as a sequence of binary features according to the scheme of a naive Bayesian classifier. We propose to use integral heteroassociative transformations that provide the selection of the estimated and stochastic (masking) components on the processed image fragment based on the forecast model of one part of the fragment in relation to another to identify violations of the structural and statistical image properties after message embedding. Such transformations are included in the architecture of trained neural networks as an additional layer. Alternative versions of deep neural network architectures (with and without an integral layer of heteroassociative transformation) are considered. The PPG-LIRMM-COLOR images base was used to create data sets. Experiments have been carried out for several well-known stego algorithms (including the classic block and block-spectral algorithms of Kutter, Koha — Zhao, modern algorithms EMD, MBEP and algorithms for adaptive spatial steganography WOW and S-UNIWARD) and for the stego algorithms based on the use of heteroassociative compression transformations. It is shown that the accuracy of steganalysis obtained when implementing the proposed information processing schemes for large images with relatively low computational costs is comparable to the results obtained by other authors, and in some cases even exceeds them.