Research On Forensics Aided Steganalysis Of Heterogeneous Images

As a new research direction of information security field, modern information hidingtechniques have attracted extensive attention in academe once brought forward. In the past tenyears, the battle between steganography and its counterpart steganalysis became more and moredrastic. From the literature released in recent years, we can see that the detection techniques forimage steganography have achieved fruitful research results, which exhibit excellent detectionperformance under the laboratory environment. However, the existing steganalysis algorithmsare difficult to obtain high detection accuracy when applied to the heterogeneous image sourcesunder the practical network environment, which consist of images generated by various imageacquisition devices, equipped with multiplicate image quality, image content and undergoingdiverse complex image processing. Therefore, it is of great significance to make researches onthe forensics aided steganalysis of heterogeneous images.Based on the analysis of statistical properties and models of heterogeneous images, thethesis focuses on the study of image forensics techniques and forensics aided steganalysis ofheterogeneous images. The main contributions of this thesis are summarized as follows:1. Statistical properties and models of heterogeneous images are analyzed. First, thequantization error and rounding error produced in JPEG compression and decompression andthe statistical model of DCT coefficients are carefully investigated. Then, the correlation andperiodicity presented between resampling image pixels are analyzed. Finally, textural featureand texture classification model aiming at the images with different texture complexity arestudied.2. Two practical forensics aided steganalyzers (FA-steganalyzer) for heterogeneous bitmapimages are constructed, which can properly handle steganalysis problems for mixed imagesources consisting of raw uncompressed images and JPEG decompressed images with differentquality factors. First, based on the JPEG error analysis and additive noise steganography model,we propose two image forensics techniques including identifying JPEG decompressed images,and further detecting quantization table of a JPEG decompressed image under the assumptionthat the images under investigation are mixed covers and stegos. Then, combining theaforementioned two image forensics techniques, two FA-steganalyzers are devised. The firstFA-steganalyzer consists of a JPEG decompressed image identifier followed by twocorresponding steganalyzers, one of which is used to deal with uncompressed images and theother is used for mixed JPEG decompressed images with different quality factors. In the secondFA-steganalyzer scheme, we further estimate the quality factors for JPEG decompressed images,and then steganalyzers trained on the corresponding quality factors are used. Moreover, by designing three setups, how to set the decision threshold of image forensics classifier to balanceeach homogeneous steganalyzer is analyzed. Finally, extensive experimental results show thatthe proposed two FA-steganalyzers outperform the existing steganalyzer that is trained on amixed dataset. Additionally, in our proposed FA-steganalyzer scheme, we can select thesteganalysis methods specially designed for raw uncompressed images and JPEG decompressedimages respectively, which can achieve much more reliable detection accuracy than adoptingthe identical steganalysis method regardless of the type of cover source.3. An efficient resampling detecting aided steganalyzer of heterogeneous images includingraw single-sampled images and resampled images at different scaling factors is proposed. First,combining texture analysis, we explore the relationship between texture and resamplingoperations, and present a reliable resampling detection method based on image texture. Theimportance of the proposed method is its ability to detect resampling not only in cover imagesbut also in stego images. Then, combining the resampling detection technique, a resamplingdetection aided steganalyzer is constructed. We first employ the resampling detector to decidethe image source, and then send the image to the steganalyzer specially designed to work withimages of that class. Moreover, the effect of resampling detection on steganalysis is analyzed bydetailed deduction from error probability. Finally, experimental results show that the detectionperformance of the proposed FA-steganalyzer outperforms the existing steganalyzer that istrained on a mixed dataset especially at low embedding rates.4. A steganalyzer based on Gaussian mixture model (GMM) clustering is proposed,targeting at the images with different texture complexity. Two changes are made compared tothe existing steganalyzer that is trained on a mixed dataset. In the training stage, the GMMclustering algorithm is exploited to classify the training samples into limited categories; in thetesting stage, the posterior probability of testing samples belonging to each category iscalculated, and the samples are submitted to the steganalyzers corresponding to the maximumposterior probability for test. Moreover, the effect of GMM clustering on the performanceimprovement of steganalysis is analyzed form the viewpoint of image information measurement.Finally, experimental results aiming at LSB matching steganography and two adaptivesteganographic algorithms have shown that the proposed steganalyzer outperforms the existingsteganalyzer that is trained on a mixed dataset.Finally, the research work in this thesis is summarized and the further research directionsof forensics aided steganalysis of heterogeneous images are discussed.