Steganalysis of Digital Images Using Rich Image Representations and Ensemble Classifiers By

Modern detectors of steganographic communication in digital images are implemented as supervised classifiers trained in pre-defined feature spaces also called image models. Currently, the Support Vector Machine (SVM) is the machine-learning tool of choice in the steganalysis community due to its accuracy and a well-founded theory. However, in order to keep the SVM training computationally feasible, feature spaces need to be designed to be low-dimensional. Consequently, their construction often consists of a series of clever tricks and heuristic dimensionality-reduction techniques. Recent trends in steganalysis, however, have shown that more complex and higher-dimensional image models could deliver substantially better performance. In this dissertation, we propose a novel framework for steganalysis of digital images in which we replace SVMs with the ensemble classifier, a scalable machine-learning alternative offering comparable accuracy at a fraction of a computational cost. This allows us to approach the feature-space building process in a more systematic and exhaustive way. In particular, we propose to construct feature spaces as collections of a large number of simpler submodels, each of them capturing different types of dependencies among image coefficients. As a result, we obtain a high-dimensional rich statistical descriptor of images, the so-called rich model, which is then used for the classifier training. To demonstrate the power of the proposed methodology, we construct rich models for images in raster formats and JPEGs, the two most commonly used image representations. The 34, 671-dimensional spatial domain rich model consists of co-occurrences of neighboring samples of noise residuals obtained by various filters, and the 22, 510-dimensional JPEG domain rich model consists of submodels capturing different types of dependencies among DCT coefficients of JPEG images. Both rich models, combined with the ensemble classifier, are shown to significantly outperform previous art across a wide range of steganographic schemes hiding data in both domains. This work is presented as a self-contained text, covering all technical details of the ensemble classifier and the rich-model construction, their implementation, and experimental performance evaluation.