Research On Some Key Issues In Image Steganalysis

Steganography is a technique of covert communication which aims at embedding secret messages in innocuous cover objects so as not to arouse suspicion. By contrast, steganalysis, the counter problem of steganography, is intent on identifying the existence of secret messages in a given medium, and then extracting and breaking the hidden secret messages. Since 1990 s, image steganography and image steganalysis have evolved into an important subject in the field of information security. This dissertation mainly focuses on the research of image steganalysis.Image steganalysis has made great progress in the last 20 years. A lot of steganalysis methods have been proposed and some of them could get very good detection performance. However, there are still some issues remaining unsolved. For example, there is no efficient steganalysis method for some typical steganography methods. Moreover, most existing blind steganalysis methods did not consider the difference in statistical properties among different images and the impact of embedding algorithm mismatch. This thesis concentrates on dealing with these issues and the contributions of the thesis are as follows:1. The histories, basic concepts, models and evaluating criteria of steganography and steganalysis are introduced briefly. Two typical categories of image steganography are introduced: spatial domain steganography and JPEG steganography. Special steganalysis and universal steganalysis are described in detail and the typical steganalytic features and classifiers used in universal steganalysis are discussed. Some new steganalysis methods using the difference in statistical properties among different images are introduced briefly.2. Quantitative steganalysis of least significant bit matching revisited(LSBMR) for consecutive pixels. LSBMR steganography uses two adjacent pixels as an embedding unit to conceal secret messages. In each unit, the modification probability of the first pixel is twice the modification probability of the second pixel, thereby causing an “asymmetric” effect on the statistical distribution of pixel difference in LSBMR for consecutive pixels(LSBMRCP). O n the basis of the analysis of this “asymmetric” effect, a fast steganalytic feature is deduced, and an embedding rate estimating method that employs an iteration strategy is proposed. Extensive experimental results show that the steganalytic feature can effectively detect LSBMRCP steganography and that the detection performance is superior to previous proposed methods. Moreover, the embedding rate estimating method is quite accurate that the order of magnitude of prediction error is maintained at 10-2 measured by the mean absolute error, median absolute difference and interquartile range, thus indicating that the proposed method significantly outperforms the method proposed by Tan.3. Fast detection of random LSBMR steganography. By analyzing the weighted-smoothing effect of LSBMR steganography on the distribution of pixel difference, an equation is deduced to estimate the frequency of difference value 0 using the frequencies of difference values 1 and 2. The sum of the ratio of the estimated value to the actual value as well as the ratio of the frequency of difference value 1 to difference value 0 is used as the steganalytic detector. Experimental results show that the proposed method can effectively detect LSBMR steganography and can outperform previous proposed methods in most circumstances.4. Blind detection of JPEG image steganography using K-means clustering. Most existing blind image steganalysis methods did not consider the impact of the differences in statistical properties among different images. To address this problem, a “cluster and classify” strategy is proposed. In the training phase, the training images are categorized into k clusters using K-means clustering according to the clustering features which are deduced from the horizontal and vertical intra-block co-occurrence matrices of the absolute value of DCT coefficients. After that, steganalytic features are extracted from each cluster and the training process is specialized for each cluster separately. Given a test image, the clustering features are extracted and the distances from the test image to each cluster center obtained in the training phase are calculated, and then the test image is pre-classified as the cluster with the minimum distance. After that, the steganalytic features of the test image are extracted and sent to the corresponding classifier and the final result(cover or stego) is outputted. Experimental results on typical JPEG steganographic algorithms show that the proposed method can significantly enhance the detection performance of existing steganalytic features, especially at low payloads.5. Ways to mitigating the impact of embedding algorithm mismatch in image steganalysis. This thesis point out that, in partial mismatch condition, auxiliary samples can be used to improve the performance of classifiers. Two methods for selecting auxiliary samples are proposed:(1) sample selection based on distance and(2) sample selection based on classifiers. Experimental results show that both two methods can enhance the performance of classifiers under partial mismatch condition, and the method based on classifiers can get better performance. Since the performance of different auxiliary embedding algorithms are different, a method is proposed to search the most appropriate embedding algorithm. By employing a bank of classifiers trained on some embedding algorithms and tested on the target training samples, those samples in the classifier that get the best performance are considered as the most appropriate samples. Experimental results demonstrate the effectness of the proposed method.Finally, the conclusion of the thesis and a discussion of the direction for future research are presented.