A Review of modern methods for steganalysis and localization of embedded data in digital images

Abstract

The article provides a systematic review of modern steganalysis methods for digital images based on artificial neural networks. The primary stages of development of advanced cover-image models, from widely used artificial neural networks to contemporary hybrid models, are considered. Advantages and limitations of various types of neural networks for constructing stegodetectors for digital images are investigated. Based on comparative analysis of steganalysis accuracy, it is established that the use of advanced artificial neural networks achieves a detection accuracy of steganograms exceeding 90%, even at low embedding rates (less than 20%). Additionally, applying complex methods of processing both examined images, and feature vectors in multidimensional spaces with studied neural networks allows reducing the computational complexity of configuring stegodetectors without significant losses in stego images detection accuracy.