Grey-Box Steganography

In Steganography secret messages are encoded into unsuspicious covertexts such that an
adversary cannot distinguish the resulting stegotexts from original covertexts. To accomplish their respective tasks, encoder and adversary need information about the covertext distribution. In previous analyses, the knowledge about the covertext channel was unbalanced: while the adversary had full knowledge, the encoder could only query a black-box sampling oracle. In such a situation, the only general steganographic method known is rejection sampling, for which the sampling complexity has been shown to be exponential in the message length. To overcome this somewhat unrealistic scenario
and to get finer-grained security analyses, we propose a new model, called grey-box steganography. Here, the encoder starts with at least some partial knowledge about the type of covertext channel. Using the sampling oracle, he first uses machine learning techniques to learn the covertext distribution and then tries to actively construct a suitable stegotext – either by modifying a covertext or by creating a new one. The efficiency of grey-box steganography depends on the learning complexity of the concept class for the covertext channel, the efficiency of membership tests, and the complexity of the modification
procedure. We will give examples showing that this finer-grained distinction provides more insight and helps constructing efficient stegosystems that are provably secure against chosen hiddentext attacks. This way we can also easily model semantic steganography.