Reliable detection of hidden information based on a non-linear local model

Abstract : This paper investigates the reliable detection of information hidden in natural images. It is aimed to design a test with analytically predictable probabilities of error. To this end, the problem of hidden information detection is cast in the framework of hypothesis testing. The optimal test solving the decision problem of steganalysis requires image parameters which are not available in practice. To design a feasible test, a non-linear locally-adapted model of natural images is proposed. This model is linearized to allow an efficient and simple estimation of image parameters which leads to the design of an almost optimal test. Numerical results on a large number of natural images show the relevance of the theoretical findings.