Deep Learning for Predictive Analytics in Reversible Steganography

Deep learning is regarded as a promising solution for reversible steganography. The recent development of end-to-end learning has made it possible to bypass multiple intermediate stages of steganographic operations with a pair of encoder and decoder neural networks. This framework is, however, incapable of guaranteeing perfect reversibility since it is difficult for this kind of monolithic machinery, in the form of a black box, to learn the intricate logics of reversible computing. A more reliable way to develop a learning-based reversible steganographic scheme is through a divide-and-conquer paradigm. Prediction-error modulation is a well-established modular framework that consists of an analytics module and a coding module. The former serves the purpose of analysing pixel correlations and predicting pixel intensities, while the latter specialises in reversible coding mechanisms. Given that reversibility is governed independently by the coding module, we narrow our focus to the incorporation of neural networks into the analytics module. The objective of this study is to evaluate the impacts of different training configurations on predictive neural networks and to provide practical insights. Context-aware pixel intensity prediction has a central role in reversible steganography and can be perceived as a low-level computer vision task. Therefore, instead of reinventing the wheel, we can adopt neural network models originally designed for such computer vision tasks to perform intensity prediction. Furthermore, we rigorously investigate the effect of intensity initialisation upon predictive performance and the influence of distributional shift in dual-layer prediction. Experimental results show that state-of-the-art steganographic performance can be achieved with advanced neural network models.