Application of Convolutional Neural Networks to Increase the Security Level of Steganographic Methods

Abstract

This paper proposes an approach to increase the stealthiness of image steganography methods using convolutional neural networks (CNNs). CNNs are integrated into the data embedding process and help minimize the traces of hidden information that can be detected by steganalyzers. Two implementation variants are considered: based on the least significant bit (LSB) and discrete cosine transform (DCT) methods, as well as their modifications using CNNs. The task of ensuring the stealth and reliability of data embedding is solved in stages: the visual quality of the stego images, the reliability of message extraction, and resistance to detection by classical steganalysis methods are analyzed. The results of experiments on assessing quality and stealthiness confirm the effectiveness of the proposed approach.