Low volume generic steganalysis with improved generalization

Abstract

Practical Steganalysis needs to be carried blind as the Steganalyzer won't have access to any other data except for the suspicion of a covert channel. The process needs to be carried out universal with a bunch of statistical features as the Steganalyzer needs to identify stego images created with even new steganographic tools. This poses a need for the generalization of the Steganalyzer to be enhanced. The proposed approach improves the generalization of the developed Steganalyzer by training it with a group of statistical features carefully coined to characterize embedding distortions that can disturb different features of an image. The designed Steganalyzer uses mixed blind generic classification for identifying unfamiliar tools i.e., not introduced during training phase.