Imgae Content Based Steganalysis And Tampering Detection

As two important technical tools for defending the security of image information, image steganalysis and tampering detection have evolved into important subjects of interest in the field of multimedia information security. Image steganalysis, as the opposite technology against steganography, aims to expose the existence of secret message and can also serve as an effective way to test the security of steganographic algorithms. Tampering detection, as a major branch of digital image blind forensics, is to identify the authenticity of digital images by analyzing the inherent features of images in the absence of watermarking or signature.Image steganalysis and tampering detection are fulfilled based on the salient and statistically distinct features caused by embedded secret message or tampering manipulation. Natural images represent the vision information of sceneries with different spacial structure in the external world and have local stationary characteristics. Image regions with different content have different statistical characteristics. Furthermore, the steganalytic features and tampering detection features have a close relationship with image content. Consequently, it is of important significance to make researches on image steganlysis and tampering detection based on image content.In this dissertation, the image source is modeled as a local stationary Markov source. Based on the analysis of principle of LSB matching steganography and composite JPEG images, the thesis focuses on developing the alogorithms of image steganalysis and tampering detection based on image content. The main contributions of this thesis are summarized as follows:1. Statistical analysis of natural images. Combining with the advances in statistical modeling of natural images and applying probability and statistics theory and information theory, we analyze the distributions of pixel values and detail components of natural images and the relationship between the distributions and image complexity. The analytic results demonstrate that natural images have strong spatial dependences among adjacent pixels, and the statistical characteristics are similar for the same scenery and apparently different for diverse sceneries, that is, regions with different content present different statistical characteristics.2. Statistical analysis of cover and stego images. On the self-built database with single image conten, we investigate the impact of LSB matching steganography on statistical characteristics of natural images. In addition, taking some representative steganalytic features for example, we analyze the relationship between the steganalytic features and image content and come to the conclusion that the statistical changes are more evident in flat regions after embedding, which provides a theoretic basis for researching image content based steganalysis techniques and proposing new reliable steganalytic algorithms.3. Steganalysis of LSB matching based on local variance histogram. We model the LSB matching embedding as additive noise, analyze the difference of local variances between the cover and stego images, and propose a method for steganalysis of LSB matching based on local variance histogram. The"signal to noise ratio"of stego noise is increased by difference preprocessing. The weighted features of the local variance histogram are extracted to characterize changes in regions with different complexity between the cover and stego images. Features extracted from a given image and its downsampled version are combined as classification features. The experimental results demonstrate that this approach has good performance.4. Steganalysis of LSB matching based on local linear transform and weighted features of characteristic functions. We analyze the largest difference between the characteristic functions of detail components before and after embedding and porpose a method for steganalysis of LSB matching based on local linear transform and weighted features of histogram characteristic functions. The embedded secret information is regarded as a kind of stochastic texture in a fine scale. Images are decomposed into a group of detail subbands with a bank of local linear transform masks which are very sensitive to textures. New weighted features of characteristic functions are constructed to capture the largest differences between the characteristic functions of cover and stego images. A feature selection algorithm is adopted to find a suboptimal feature set. In comparison with the state-of-the-art steganalysis algorithms, this approach performs the best, on the whole.5. Blockwise joint judegement steganalysis based on image content. Based on the relationship between the steganalytic features and image content, we propose a blockwise joint judgement steganalysis algorithm. We decompose images into small sub-images, categorize these sub-images based on image content, train a classifier for each category and assign different weights to each category. The detection result of a whole image is obtained by weighted fusion of the results of its sub-images. Experimental results for steganalysis of LSB matching indicate that this approach outperforms representative steganalysis algorithms.6. Detecting composite JPEG images based on inconsistencies of blocking artifacts. For a class of JPEG image forgery, a simple and effective detecting algorithm is proposed on the basis of inconsistencies of blocking artifacts between tampered region and non-tampered region and relationship between blocking artifacts and image content. A testing image is cropped and recompressed with the estimated primariy quality factor. The blocking artifacts factor mapping is extracted via computing the blocking artifacts with different block-dividing manners at primary quality factor. Automatic tampering detection and location of tampered region are fulfilled by image segmentation. Experimental results demonstrate that this approach is robust to underlying image content and applicable to images of various qualities and small tampered regions. With the false positive rate of 1%, the detection accuracy for composite images is above 90% for differences of quality factors before/after tampering larger than 15.