Statistical counter-attacks in MPEG-4 AVC watermarking

This paper presents a method able to optimize watermarking detection in the MPEG-4 AVC compressed domain. The optimization was achieved by introducing a statistical counter -attack based on the MAP criterion applied on the noise matrices corresponding to each attack. As no statistical models are nowadays available for the attack effects in the MPEG-4 AVC domain, they had to be first estimated. In this respect, 95% confidence limits for noise matrices, computed with relative errors lower than 15% have been obtained for geometric attacks (StirMark, small rotations), and linear or non-linear filtering (Gaussian filtering, sharpening). The viability of these statistical results was demonstrated by the watermarking experiments: it was obtained that the counter-attacks based on the attack model approaches its theoretical upper limit.