A Local Adaptive Model of Natural Images for Almost Optimal Detection of Hidden Data
This paper proposes a novel methodology to detect data hidden in the least significant bits of a natural image. The goal is twofold: first, the methodology aims at proposing a test specifically designed for natural images, to this end an original model of images is presented, and, second, the statistical properties of the designed test, probability of false alarm and power function, should be predictable. The problem of hidden data detection is set in the framework of hypothesis testing theory. When inspected image parameters are known, the Likelihood Ratio Test (LRT) is designed and its statistical performance is analytically established. In practice, unknown image parameters have to be estimated. The proposed model of natural images is used to estimate unknown parameters accurately and to design a Generalized Likelihood Ratio Test (GLRT). Finally, the statistical properties of the proposed GLRT are analytically established which permits us, first, to guarantee a prescribed false-alarm probability and, second, to show that the GLRT is almost as powerful as the optimal LRT. Numerical results on natural images databases and comparison with prior art steganalyzers show the relevance of theoretical findings.