On the application of coding to adaptive steganography : robustness, efficiency, and coding loss

• Main Author: Kin-Cleaves, Christopher Stephen
• Format: Dissertation
• Language: English
• Published: University of Oxford 2020
• Subjects: Coding theory; Computer security; Image steganography

Steganography is the process by which we hide payloads (secret messages) into covers (inconspicuous mediums of data) resulting in stego ob- jects (original mediums of data with subtle changes to encode our message), with the aim of these objects appearing to be indistinguishable from the cover. Steganography aims to compromise the adversaries' ability to classify an object as innocent or not. The most successful steganography is adaptive embedding, which uses a distortion function to associate individual cover changes with a cost of change (higher costs implying a more detectable change), and a code which embeds the payload whilst minimising the sum of distortion (a proxy for detectability). This thesis explores approaches to codes, for adaptive embedding on the subjects of robustness, efficiency, and simulating. Since the wide adoption of adaptive embedding, the significant focus of most researchers in the field has been on distortion algorithms. Our work has aimed to explore the potentials around codes instead, building on the highly successful Syndrome Trellis Codes (STC) algorithm. On the area of robustness, we explore the ability to create stego objects able to withstand the noisy (non-adversarial) channel. Since syndrome codes are fragile by design, syndrome-encoded stego objects are a bad choice: amplifying the channel error rate into unacceptably high payload error rates. Our work on improving the efficiency of STCs looks at how H matrix construction can influence the fficiency of the algorithm. Expanding on existing research around minimising outliers, we look at how moving from a block-based matrix to one constructed with rows can improve the embedding efficiency. Finally, we explore the area around simulated steganography. Because of speed constraints, many researchers omit the use of a code, to speed up experiments. The justification is clear: not using a code saves both time and experimental complexity, when you can simulate stego objects on the rate- distortion bound. If we cannot achieve the rate-distortion bound with a code, and more parameters are influencing this than which are taken into account when simulating, is this a sensible option? We propose a method which samples the code with the given parameters and is then able to simulate objects not on the rate-distortion bound, but at the distortion and detectability of a code.