Hashing, pooling and coding: towards optimal embedding in linguistic steganography

Thesis

The goal of steganography, the art of hiding information, is to send hidden messages without revealing their existence to outside observers. So-called cover objects are modified to convey the payload, the hidden message. In the majority of literature, covers are images; this work is concerned with the sub-field of linguistic steganography, where covers are pieces of natural language.

This area has received little attention, and as a result has lagged behind the state-of-the-art approaches present in image steganography. This work identifies the following weaknesses of the current linguistic steganography literature: methods to assign values to cover sentences are inefficient; the amount of data hidden in covers (capacity) is low; security is rarely tested. The first two weaknesses are linked, low capacity caused, in part, by inefficient value assignment techniques. The last weakness is an inexcusable flaw; the evolution of linguistic steganography is impossible without thorough security analysis.

This work applies concepts from image steganography and steganalysis in order to address these flaws, and to provide a robust framework for future linguistic steganography research. A novel linguistic stegosystem is developed, utilising hashing to assign values to cover objects, and manual input to ensure fluency of generated steganography. Security of the system is evaluated twice: first against a human attacker, an approach uniquely appropriate to the linguistic setting; then against a novel automated attack, utilising the paradigm of pooled steganalysis.

The automated attack is powerful against the generated steganography, until adaptive embedding, a technique well known in image steganography, is applied to the linguistic setting for the first time. This utilises coding to minimise distortion (a measure of the detectability of generated steganography). Four linguistic distortion measures are described, the first of their kind. The resulting steganographic objects are shown to be secure both against humans, and the developed attack.

Authors: Wilson, A

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Subjects: Linguistic steganography