Statistical detection of data hidden in least significant bits of clipped images

This paper studies the statistical detection of data hidden in the Least Significant Bits (LSB) plan of natural clipped images using the hypothesis testing theory. The main contributions are the following. First, this paper proposes to exploit the heteroscedastic noise model. This model, characterized by only two parameters, explicitly provides the noise variance as a function of pixel expectation. Using this model enhances the noise variance estimation and hence, allows the improving of detection performance of the ensuing test. Second, this paper introduces the clipped phenomenon caused by the limited dynamic range of the imaging device. Overexposed and underexposed pixels are statistically modeled and specifically taken into account to allow the inspecting of images with clipped pixels. While existing methods in the literature fail when the data is embedded in clipped images, the proposed detector still ensures a high detection performance. The statistical properties of the proposed GLRT are analytically established showing that this test is a Constant False Alarm Rate detector: it guarantees a prescribed false alarm probability.