Zero memory information sources approximating to video watermarking attacks

Under the framework of the watermarking applications, to accurately model the malicious attacks becomes nowadays a challenging research topic. Following a previous study pointing out the inner limitations of the popular Gaussian model, this paper computes for the first time accurate approximations for the zero memory information sources describing some main real life attacks in video watermarking: linear and non-linear filtering, rotations, and StirMark (which may emulate the in-theatre camcorder capture). In order to obtain these models, the authors advance an original and generic statistical approach for pdf estimation combining EM Gaussian mixtures and confidence limits. This investigation procedure does not rely on any a priori assumption concerning the video/attack stationarity and handles with mathematical rigour the dependency existing among successive frames in a video sequence.