RUS  ENG
Full version
JOURNALS // Proceedings of the Institute for System Programming of the RAS // Archive

Proceedings of ISP RAS, 2020 Volume 32, Issue 4, Pages 155–164 (Mi tisp531)

This article is cited in 6 papers

Using of convolutional neural networks for steganalysis of digital images

A. A. Polunin, E. A. Yandashevskaya

Russian Federation Security Guard Service Federal Academy

Abstract: The article substantiates the relevance of steganalysis, as a determination of the presence of a hidden channel in telecommunication systems, whose nodes exchange digital images. The article deals with the application of convolutional neural networks to solve this problem. It is assumed that the probability of correct image classification using a well-trained convolutional neural network will be comparable or even better than characteristics of statistical algorithms or the RM model. We introduce principles of construction and capabilities of convolutional neural networks in the framework of their applicability to solving the problem of steganalysis. To improve the efficiency and effectiveness of the stegocontainer recognition process, a version of the image classification model for a convolutional neural network is proposed. It is based on combination of several convolutional and fully connected layers. We have developed software for this model version with the ability to train a neural network and evaluate the quality of classification. The analysis of existing software products designed for the task of determining the fact of using steganography in digital images is carried out. The advantage of classifiers based on neural networks in comparison with statistical ones is proved. Using the developed software, an experimental study of classification model on sets of digital images contained in open sources has been carried out. The article presents the results of neural network training, as well as an analysis of the strengths and weaknesses of the selected model.

Keywords: steganography, steganalysis, digital images, neural network, convolutional layer, fully connected layer, machine learning.

DOI: 10.15514/ISPRAS-2020-32(4)-11



© Steklov Math. Inst. of RAS, 2024