In this paper, an unsupervised steganalysis method that combines artificial training sets and supervised classification is proposed. We provide a formal framework for unsupervised classification of stego and cover images in the typical situation of targeted steganalysis (i.e., for a known algorithm and approximate embedding bit rate). We also present a complete set of experiments using (1) eight different image databases, (2) image features based on Rich Models, and (3) three different embedding algorithms: Least Significant Bit (LSB) matching, Highly undetectable steganography (HUGO) and Wavelet Obtained Weights (WOW). We show that the experimental results outperform previous methods based on Rich Models in the majority of the tested cases. At the same time, the proposed approach bypasses the problem of Cover Source Mismatch -when the embedding algorithm and bit rate are known- since it removes the need of a training database when we have a large enough testing set. Furthermore, we provide a generic proof of the proposed framework in the machine learning context. Hence, the results of this paper could be extended to other classification problems similar to steganalysis.