Hiding, detecting, and removing steganographic noise

Unlike cryptography, steganography in natural images cannot currently be made computationally secure. It is not a matter of how much security is "enough". Instead, steganography cannot have provable security due to the lack of a model of the content of an image, or more precisely, the lack of specificity in the models that do exist. This problem affects both sides of the steganography-steganalysis competition. Despite this seemingly fatal flaw, the desire to hide information still exists. The need to communicate secretly still exists. So, without perfect security, how can a steganographer and a steganalyst best pursue their respective tasks?; This dissertation investigates the dual problems of hiding and detecting steganography. The general premise treats steganography as noise, stego-noise. The contributions of this work include the development of new approaches to both steganography and steganalysis. Along the way, new methods of non-linear denoising, edge detection, and a new form of parametric wavelets were also created.; For hiding stego-noise, a technique is presented for hiding within the most visually significant areas of the image, the edges. A mild distortion channel is developed which effectively detects image edges and textures. This distortion channel generates visually similar images that serve as bounding structures for embedding hidden data. For detecting stego-noise, a formal steganalysis method is presented, based on estimating a clean image. Estimating a clean image typically relies on a conventional noise model that does not apply to steganography. A new method is developed which does not rely on conventional noise models. This work introduces a new steganalysis method based on estimating statistics of image noise. For removing stego-noise, a number of denoising methods are investigated for their ability to preserve an image while removing any hidden information.