Research On Fundamental Problems Of Universal Image Steganalysis

Media data in internet may contain secret message embedded by steganography. Steganography can disguise secret message in normal media data for secret communication. Due to invisibility of message in media data, human eye is unable to notice any presence of it, thus computer algorithm is needed for steganalysis. Digital image is one of the most popular media formats and many steganography for digital image have been developed. For this point, it is necessary to develop universal steganalysis method which can detect universal kinds of steganography, which is most difficult.In this dissertation, we focus on universal steganalysis for digital image. We detect stego image by feature extracted from digital image and statistical learning method of classification and regression.The main contribution of this dissertation:1. We proposed universal steganalysis feature based on neighbor information. Extracting feature is a key part in universal steganalysis. Feature provides discriminative information so that classifier can detect stego image from normal image. In this method, we first extract multiple neighbor information from pixel neighbor and then process them. Finally we calculate histogram of processed neighbor information as feature. Experimental results demonstrate that neighbor information feature is effective for detecting steganography of different kinds.2. We proposed a quantitative steganalysis method based on subspace learning and gradient boosting.quantitative steganalysis aims to predict embedding density of secret message in stego image. Universal quantitative steganalysis is done by feature extracted from image and regression method. In our method, we first combine several different feature sets, and then fuse them by subspace learning. Finally fused feature is input into estimator trained by gradient boosting for prediction result. Experimental results demonstrate that our method achieves high accuracy and meanwhile with low time complexity.3. We proposed a pratical steganalysis method based on clustering and support vector machine. In image steganalysis, image content greatly impacts accuracy. In order to obtain generalization ability, it is indispensable to collect more image samples for training. However, large number of training samples bring great burden for training and testing. In our method, training sample set are divided into several non-overlapping subsets by clustering. For each subset, a classifier is trained on samples in it. In testing procedure, candidate sample is assigned to a subset according to distance between its feature and center of subset, and then detected by classifier trained on this subset. Experimental results demonstrate that this method is practical for its lower time complexity and high accuracy.