Dataset, Noise Analysis, and Automated Parameter Estimation for Natural Steganography

Natural steganography concerns embedding a secret message in a cover-source following some distribution S_1, such that after embedding the distribution of the stego-media mimics another cover-source with distribution S_2 (without embedding). Prior works have studied natural steganography over image files, where S_1 and S_2 correspond to the light intensity distribution of a photo taken respectively at some ISO_1 and ISO_2, and much effort has been dedicated to various embedding methods. On the other hand, while the nature of mimicking the distribution S_2 by embedding messages into S_1 sources means that accurate estimations of such distributions are crucial, relatively little attention has been given to this aspect. Furthermore, deploying these stegosystems in practice requires users to estimate the noise distributions of their cameras, which poses a challenging technological barrier for average users and limits the utility of the stegosystems. An objective of this work is to verify the existing claim that, for each fixed ISO value, the pixel values follow a family of Gaussian distributions where the variance is an affine function of the mean. Towards estimating and verifying the concerned distributions, we have created a comprehensive image dataset with the mainstream Sony A6400 camera in a professional photo-shooting environment. Analyses over our dataset reveal that parameters of the light intensity distributions appear to have more complicated behaviour than reported in prior works - they seem to depend on the overall exposure level induced by the camera settings. For the ease of analysis, we have also developed a set of tools for automating the parameter estimation process. We believe that these tools will eventually improve the accessibility of natural steganography.