Multivariate Side-Informed Gaussian Embedding Minimizing Statistical Detectability

Steganography schemes based on a deflection criterion for embedding posses clear advantage against heuristics as they provide direct link between theoretical detectability and empirical performance. However, this depends the accuracy of cover stego model underlying scheme. In work we propose an original steganography scheme realistic sensor noise, taking into account camera model, ISO setting processing pipeline. Exploiting statistical allows us to take correlations DCT coefficients account. Several types dependency models are presented, including very general lattice which accurately dependencies introduced by large class pipelines interest. We show in particular that signal minimizes KL divergence under has covariance proportional noise covariance. The resulting achieves state-of-the-art performances go well beyond current standards side-informed JPEG steganography.