On dangers of overtraining steganography to incomplete cover model

This is the authors' abstract. We don't add key points for this paper.

A modern direction in steganography calls for embedding while minimizing a distortion function defined in a sufficiently complex model space. In this paper we show that, quite surprisingly, even a high-dimensional cover model does not automatically guarantee immunity to simple attacks. Moreover, the security can be compromised if the distortion is optimized to an incomplete cover model. We demonstrate these pitfalls with two recently proposed steganographic schemes and support our arguments experimentally. Finally, we discuss how the corresponding models might be modified to eliminate the security flaws.