"A MultiAlgorithm, High Reliability, Extensible Steganalyzer using Services Oriented Architecture"

Abstract

Network security has received increased attention in the last decades. Encryption has laid itself as the traditional method to transmit information in secrecy. Although strong encryption is a very secure approach for transmitting information, it can be easily identified that transmitted information is encrypted. Once the information is identified as encrypted, an intruder can block the encrypted transmission. In contrast, Steganography is a viable option to hide information in transmission without being identified. It provides a blanket that hides encrypted information. Thus, it becomes essential to develop mechanisms that reveal if the communicated information has any embedded data. Steganalysis is the art of detecting invisible communication and is a very challenging field due to different types of media and embedding techniques involved. Existing research in Steganalysis has focused on developing individual stego detection algorithms for a particular media type or for a particular embedding technique. In this dissertation we are proposing to develop a unified Steganalyzer that can not only work with different media types such as images and audio, but further is capable of providing improved accuracy in stego detection through the use of multiple algorithms. Our proposed system integrates different steganalysis techniques in a reliable Steganalyzer by using a Services Oriented Architecture (SOA). The SOA architecture not only allows for concurrent processing to speed up the system, but also provides higher reliability than those reported in the existing literature because multiple stego detection algorithms are incorporated simultaneously. Furthermore, the extendable nature of the SOA implementation allows for easy addition of new Steganalysis algorithms to the system in terms of services. The universal steganalysis technique proposed in this dissertation involves two processes; feature extraction and feature classification. An improved 2D Mel-Cepstrum implementation is used for wav files feature extraction. Intra-blocks technique is used for jpeg images feature extraction. The feature classification process is implemented using three different classifiers; neural network classifier, Support Vector Machines classifier, and AdaBoost classifier. The unified steganalyzer is tested for jpeg images and wav audio files. The accuracy of classification ranges from 90.0% to 99.9% depending on the object type and the feature extraction method. In particular, an enhancement of 2D Mel-Cepstrum implementation is introduced that achieves an accuracy of 99.9%. This is significantly better result than the average detection accuracy of 89.9% to 96.7% reported by Liu [1]. Finally, an extensible classifier is introduced that allows adding detection of new embedding techniques to the currently supported embedding techniques, so that the framework will maintain its reliability even if new embedding techniques are introduced.