Detecting Hidden Messages Using Image Power Spectrum

Abstract

In this paper we present a study of the effects of data hiding on the power spectra of digital images. Several imperceptible data hiding techniques have been proposed that provide strong visual security and robustness. Although imperceptible to the human visual system, the hidden data affects the natural qualities of the image, such as the image power spectrum. In this study, we classify a large image database into a number of categories. For each category, we calculate the slope of the power spectra for the marked and unmarked images. We note that in the case of spatial data hiding the average slope of the power spectra of marked images is 54.93% higher compared to that of the unmarked images. Also in the cases of transform domain data hiding we note that the average slope of the power spectra of the images marked using a discrete cosine (wavelet) transform (DC(W)T) based technique is higher by 9.12% (38.39%). We also test a commercially available data hiding software namely Digimarc Corp.'s MyPictureMarc 2005 VI.0. In this case the average power spectra of the marked images is 35.99% higher. Hence we see that the proposed scheme is a tool for universal steganalysis with varying degrees of success depending on the type of embedding.