Date of Award

8-12-2008

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Electrical and Computer Engineering

First Advisor

Gilbert L. Peterson, PhD

Abstract

Over 725 steganography tools are available over the Internet, each providing a method for covert transmission of secret messages. This research presents four steganalysis advancements that result in an algorithm that identifies the steganalysis tool used to embed a secret message in a JPEG image file. The algorithm includes feature generation, feature preprocessing, multi-class classification and classifier fusion. The first contribution is a new feature generation method which is based on the decomposition of discrete cosine transform (DCT) coefficients used in the JPEG image encoder. The generated features are better suited to identifying discrepancies in each area of the decomposed DCT coefficients. Second, the classification accuracy is further improved with the development of a feature ranking technique in the preprocessing stage for the kernel Fisher s discriminant (KFD) and support vector machines (SVM) classifiers in the kernel space during the training process. Third, for the KFD and SVM two-class classifiers a classification tree is designed from the kernel space to provide a multi-class classification solution for both methods. Fourth, by analyzing a set of classifiers, signature detectors, and multi-class classification methods a classifier fusion system is developed to increase the detection accuracy of identifying the embedding method used in generating the steganography images. Based on classifying stego images created from research and commercial JPEG steganography techniques, F5, JP Hide, JSteg, Model-based, Model-based Version 1.2, OutGuess, Steganos, StegHide and UTSA embedding methods, the performance of the system shows a statistically significant increase in classification accuracy of 5%. In addition, this system provides a solution for identifying steganographic fingerprints as well as the ability to include future multi-class classification tools.

AFIT Designator

AFIT-DEE-ENG-08-20

DTIC Accession Number

ADA484934

Download

Share

COinS