Influence of SRM filters preprocessing on stego data localization in digital images

Abstract

Early detection and counteraction to unauthorized transmission of sensitive information via publicly available networks are topical tasks today. Of special interest are steganalysis methods aimed for effective destruction of hidden messages embedded into innocuous media files, like digital images. However, practical usage of such methods introduces significant changes into statistical and spectral parameters of processed images, thus revealing the intrusion into stego channels. There are proposed novel methods for localization positions of embedded stego bits into cover images and pointwise processing only these positions. The article quantifies the impact of cover images preprocessing on accuracy of stego bits localization. The case of Spatial Rich Model (SRM) filters usage is considered, while stego bits position detection is performed using novel deep neural networks, such as Unet, LinkNet, PSPNet and FPN models. The results of comparative analysis of localization accuracy proved effectiveness of SRM filters usage, namely to increase of localization accuracy up to five times (from 2.01% to 10.9% of Intersection-over-Union metric values) even for modern adaptive embedding (like MG and MiPOD) and low cover image payload values (about of 3%-5%). Obtained results create preconditions for development of high-accuracy methods for localization positions of stego bits embedded into cover images according to novel embedding methods.