Quantitative steganalysis using binary classifier

In this paper, the problem of determining secret message length using binary steganalytic classifier is researched. It is assumed that a steganalyst is able to cut a large stego image into $k$ smaller images and to apply the binary classification to every one of them. According to the information-theoretic approach to the steganographic security, a steganalyst's expected error calculation formula is derived. Determining the optimal choice of $k$ depended on the properties of a binary classifier and a given image is formulated as a minimization problem. Presented approach can be used to estimate impact of various parameters on stegosystem security against quantitative steganalysis.