Context-aware data hiding ARM-3 method and context MSE metrics

Abstract

The paper presents a new context-aware adaptive data hiding (DH) method, ARM-3, for spatial domain of still cover images using Remainder Modulo Three with optimal pixel adjustment process (RM-3 OPAP) DH method. The method efficiency is estimated theoretically and experimentally by average mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM). It is shown that ARM-3 and RM-3 OPAP DH methods having the same average MSE and SSIM have quite different stego/cover image detection error (DE). A new metrics, context MSE (CMSE), discriminating images with the same MSE and SSIM but different DE is proposed. Experiments were conducted on BOSSbase 1.01 dataset with 10000 gray-scale images. ARM-3 and S-UNIWARD DE ranges overlap, for 0.0625-0.195\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.0625-0.195$$\end{document} bit per pixel (bpp) embedding rates (ER), and, thus, are comparable but ARM-3 is about 200 times faster than S-UNIWARD.