Research On The Theoretical Models And Some Key Techniques Of Image Information Hiding

Information hiding, which has been proposed as a solution to the problem of covert communication on the Internet and copyright protection for multimedia data, is a process of embedding messages directly into the media data by making small modifications to them. This thesis discusses the theoretical models and some key techniques of information hiding in digital images.In this thesis, we attempt to establish a theoretical framework for image information hiding. At fist, a modification of a Simmon's "prisoners" problem is proposed to admit typcal image information hiding scenarios. Then based on information theory, we discuss the security of informatin hiding system, including the security of plain messages, secure messages and host images. After comparing information hiding with communication, we construct some communication models to describe different image information hiding schemes. The capacity of host channel and attack channel is investigated, and the influence of side information is studied.As the basis of image information hiding technologies, typical schemes on host signal selection in images are discussed in the light of perceptual importance viewpoint. After reviewing some statistical models proposed to better characterize the host signal in DCT, DWT and spatial domain, we describe the statistical property of host signal using Gaussian scale mixture distribution, which is more general and flexible. We also present a new method for estimating the parameters of generalized Gaussian distribution rapidly.Message embedding algorithm is the important factor for the performance of information hiding system. After analyzing the relationship among embedding parameters adjustment, attack elusion and performance evaluation, we construct an "evaluation -automatic adjustment" message embedding framework. Some key issues of performance evaluation are investigated in this thesis:1) Evaluation of the image degradation. Based on the Laplacian distribution model of DCT coefficients, we deduce a theoretical relationship between the scaling parameter in some typical watermarking algorithms and the degradation of watermarked images. Experimental results show that the estimation error of SNR and PSNR is less than 1 dB. According to this relationship, we design an adaptive insertion strategy that can be employed to embed watermarks in natural images with assigned SNR or PSNR.2) Evaluation of robustness against noise removal attack. By analyzing the mathematical models of noise removal attack, we deduce some indexes for robustness evaluation defined in frequency domain, power spectrum and locally variances of difference images.3) Evaluation of the detectability of LSB information hiding algorithm. We develop some methods for evaluating the security of LSB steganography using randomness testand two-dimension self-correlation function.4) Evaluation of the statistical visibility of information hiding algorithms. We evaluate the statistical visibility of information hiding algorithms based on the coefficient of kurtosis, differential entropy, goodness of fit test and Kullback entropy. On these bases some methodologies are suggested to defend against statistical steganalysis.With the ability to detect or decode message without the original image, blind information hiding is very useful in digital watermarking and steganography. By posing the difference on mathematical models between blind detection and non-blind detection, we devise the sign correlation detector (SCD) and the normalized liner correlation detector (NLCD) based on the Laplacian distribution model and the local Gaussian distribution model of the host signal respectively. A series of experiments in DCT domain and DWT domain show that these two detectors are more robust than linear correlation detector (LCD). On the basis of locally optimum detection theory and the generalized Gaussian distribution model of the host signal, a general blind detector -amplitude-sign correlation detector (ASCD) is derived. By deducing asymptot...