Quantifying steganographic embedding capacity in DCT-based embedding schemes.
Digital image steganography has been made relevant by the rapid increase in media sharing over the Internet and has thus experienced a renaissance. This dissertation starts with a discussion of the role of modern digital image steganography and cell-based digital image stego-systems which are the focus of this work. Of particular interest is the fact that cell-based stego-systems have good security properties but relatively poor embedding capacity. The main research problem is stated as the development of an approach to improve embedding capacity in cell-based systems. The dissertation then tracks the development of digital image stego-systems from spatial and naïve to transform-based and complex, providing the context within which cell-based systems have emerged and re-states the research problem more specifically as the development of an approach to determine more efficient data embedding and error coding schemes in cell-based stego-systems to improve embedding capacity while maintaining security. The dissertation goes on to describe the traditional application of data handling procedures particularly relating to the likely eventuality of JPEG compression of the image containing the hidden information (i.e. stego-image) and proposes a new approach. The approach involves defining a different channel model, empirically determining channel characteristics and using them in conjunction with error coding systems and security selection criteria to find data handling parameters that optimise embedding capacity in each channel. Using these techniques and some reasoning regarding likely cover image size and content, image-global error coding is also determined in order to keep the image error rate below 1% while maximising embedding capacity. The performance of these new data handling schemes is tested within cell-based systems. Security of these systems is shown to be maintained with an up to 7 times improvement in embedding capacity. Additionally, up to 10% of embedding capacity can be achieved versus simple LSB embedding. The 1% image error rate is also confirmed to be upheld. The dissertation ends with a summary of the major points in each chapter and some suggestions of future work stemming from this research.