Digital Image Data Hiding Using Side Information

Robustness is a key issue for many prospective applications of data hiding technologies. The proposal to take into account side information at the encoder has meant a noteworthy advance in the quest for better data hiding methods. Still, the true impact of the advantages theoretically derived from this groundbreaking idea has not been sufficiently founded in practice. Departing from this situation, this thesis provides a new robustness evaluation of the most relevant data hiding technologies with an emphasis on their real-world implementability. To this end, side-informed methods such as Dither Modulation (DM) and Distortion-Compensated DM (DC-DM, SCS) —which implements the near-optimal theoretical scheme devised by Costa— are compared to the previous spread-spectrum-based methods, that do not use side information. Additive random attacks are used as a benchmarking environment, together with the probability of decoding error. Methods are tested using Monte Carlo simulations and both synthetic and real image data. In order to confirm our results, and also for prediction purposes, we give accurate theoretical analyses of most of the methods considered. Multidimensional extensions are also studied to offer a more complete view of the comparisons. Results show that the strategies using side information are better performing for moderate and low levels of distortion, but much more sensitive to the level and type of attack. Also, we show that the impossibility to optimize a priori these methods adds a somewhat less optimistic view on their potential performance. The conclusions from this comparison afford the proposal and analysis of data hiding methods based on the quantization of a projection. Among them, the Quantized Projection (QP) method turns out to be a generalization of the already existing Spread-Transform Dither Modulation (STDM). Nevertheless, the philosophy behind the derivation of quantized projection methods is quite different from the spread-transform ones, using also a more restrictive distortion treatment. We show how to optimize quantized projection methods and we provide their theoretical analyses, demonstrating that QP performs better than STDM in realistic conditions.