Spatial domain universal steganalysis based on singular value decomposition

Increasing number of universal steganalysis schemes provide more reliable detection performance since neither of them works flawless for varying type and complexity of images and fusion techniqes applied to decision, feature and score levels increases the overall detection accuracy. In this paper, we propose singular value decomposition based features some of which we have previously tested under non-blind steganalysis concept with the following feature selection process under the blind-stegnalysis concept. The idea of using SVD is that it provides modeling absolute and relative linear dependencies among image rows and columns. Thus any image which do not exhibit such kind of dependencies would be classiffied as a suspicious medium. Since row images exhibit strong linear dependencies, we test the proposed features for two popular and publicly available spatial domain steganographic algorithms, namely, Steghide and Hide4pgp. Finally we apply our steganalyzer on Perturbed Quantization (PQ) data hiding which hasnpsilat been broken unconditionally with any kind of steganalysis attact. Experimental results indicate that proposed features have strong capability of detecting hidden messages.