Research On Steganalysis For JPEG Images And Decompressed Images

As one of the most important technical tools for defending the security of imageinformation, image steganalysis has become an attractive research topic in the field ofmultimedia information security. As the counterpart of steganography, image steganalysis triesto detect, extract, restore or destroy the secret messages embedded into the digital images. TheJPEG image has become one of the most widely used image types because of its highcompression ratio, good quality and small storage. It will be a very good choice to use JPEGimages as covers to hide the secret messages without arousing any suspicion. Moreover, it isquite reasonable to expect that most casual users of steganographic programs will employ JPEGdecompressed images as covers, and then perform spatial steganographic algorithms on themdirectly. Therefore, it is of great significance to make researches on the steganalysis for JPEGimages and decompressed images.In this dissertation, the DCT coefficients of JPEG images are modeled as the GeneralizedGaussian Distribution. Based on the analysis of the influences of the hidden message and JPEGrecompression on the statistical characteristic of the coefficients, the thesis focuses on the studyof steganographic algorithm identification for JPEG images and universal blind detection,embedding rate estimation method for decompressed images. The main contributions of thisthesis are summarized as follows:1. The statistical modeling methods and steganalytic techniques for JPEG images arediscussed. First, the statistical model of DCT coefficients and the description method of commonstatistics are presented. Then, the Generalized Gaussian Distribution model and its parameterestimation method are discussed in detail. Finally, the state-of-the-art and the development trendof JPEG image steganalysis are outlined.2. A novel multiclass steganalytic method with recompression detection for JPEG images isproposed to classify single-and double-compressed stego images to selected steganographicmethods. First, based on the statistical distribution of the first digits of DCT coefficients, a JPEGrecompression detection method is proposed. Then, the blind detection features are extractedfrom histogram, intrablock correlation, interblock correlation and spatial blockiness. Finally, amulticlass detector of current steganographic methods is constructed with Support VectorMachine. The experimental results show that the proposed recompression detection schemeoutperforms the existing methods significantly and is robust to the embedding changes, the lowerdimensional steganalytic feature provides a better performance and the multiclass steganalyzercan identify the current JPEG steganographic methods reliably.3. Based on the variance analysis of the noise residuals in the DCT domain, a noveluniversal steganalyzer is proposed for detecting additive noise steganography in JPEGdecompressed images. On the basis of the influence of the hidden message on the statisticaldistribution of ac DCT coefficients, we first develop a new steganalytic feature which is definedas the ratio between different ranges of the normalized ac coefficients histogram. Then theeffectiveness of the proposed feature is investigated by analyzing the steganographic algorithms such as LSB matching,±K embedding and stochastic modulation steganography theoretically.Extensive experimental results demonstrate that the proposed approach outperforms the existingstate-of-the-art schemes significantly, is less sensitive to the image content and can detect thesteganography effectively even at a very low embedding rate. In addition, the method using aone-dimensional feature is not only practical and real-time, but also can provide a better controlof the false positive rate and the false negative rate by adjusting the detection threshold.Moreover, the proposed feature can also be used to identify JPEG compression besidessteganalysis, which indicates that the proposed method has a great promise in practicalapplications.4. Based on the analysis of JPEG error and stegonoise, a novel quantitative steganalyzer isproposed for LSB matching steganography in JPEG decompressed images. First, a particulartheoretical argument is presented to prove that the cover images which are originally stored inJPEG format can be approximately estimated through JPEG recompression with the detectedquantization table. Then, based on the relationship between the message embedding rate and thevariance of the stegonoise in the DCT domain, a polynomial regression model is constructed toestimate the secret message length. The extensive experimental results show that the proposedscheme is not only computationally feasible but also significantly outperforms the existing state-of-the-art estimators especially for the images with high quality factors and embedding rates.The order of magnitude of prediction error can remain around10-3measured by the MeanAbsolute Error, Interquartile Range and Standard Deviation. The order of magnitude of medianabsolute difference can even remain at10-4. Moreover, our estimator is stable and robust withrespect to the embedding rate and the quality factor.Finally, the research work in this thesis is summarized and the further research directions ofimage steganalysis are discussed.