EFFECTIVENESS OF STEGO IMAGE CALIBRATION VIA FEATURE VECTORS RE-PROJECTION INTO HIGH-DIMENSIONAL SPACES

Context. The topical problem of sensitive information protection during data transmission in local and global communication systems was considered. The case of detection of stego images formed according to novel steganographic (embedding) methods was analyzed. The object of research is special methods of stego images features pre-processing (calibration) that are used for improving detection accuracy of modern statistical stegdetectors. Objective. The purpose of the work is performance analysis of applying special types of image calibration methods, namely divergent reference techniques, for revealing stego images formed according to adaptive embedding methods. Method. The considered divergent reference methods are aimed at search an appropriate transformation for cover and stego images features that allows increasing Euclidean distance between them. This can be achieved by re-projection of estimated features into a high-dimensional space where cover and stego features may have higher inter-cluster distances. The work is devoted to analysis of such methods, namely by applying the inverse Fast Johnson-Lindenstrauss transform for estimation preimages of cover and stego images features. The transform allows considerably decreasing computation complexity of features calibration procedure while providing a fixed level of relative positions changes for cover and stego images features vectors, which is of particular interest in steganalysis. Results. The dependencies of detection accuracy, namely Matthews correlation coefficient, on cover image payload and dimensionality of estimated preimages for feature vector were obtained. The case of usage state-of-the-art HUGO, S-UNIWARD, MG and MiPOD embedding methods for message hiding into a cover image was considered. Also, the variants of stego image features preprocessing by full access to stego encoder for a steganalytic as well as limited a prior information about used embedding method were analyzed. Conclusions. The obtained experimental results proved effectiveness of proposed approach in the most difficult case of limited a prior information about used embedding method and low cover image payload (less than 10%). The prospects for further research may include investigation of applying special methods for features preimages estimation in a high-dimensional space for improving detection accuracy for advanced embedding methods.