نبذة مختصرة : Numerical technologies increase the ease of transmission, storage and modification of multimedia content, but also makes authentication and copyright management difficult. Robust watermarking is a solution that appeared about ten years ago. It embeds a message (author name, numerical signature, . . .) within a host multimedia document. This process must not be perceptible in order not to spoil the regular use of the watermarked documents. Moreover, one must be able to correctly extract the message despite modifications of the host (i.e. attacks). Although the first approaches were empirical, watermarking was quickly considered as a communication problem. This work deals with a novel watermarking scheme based on a wide spread spectrum (a communication technique that fits to very noisy channels) and considering SAWGN attacks (scaling and additive white Gaussian noise). The relationship between watermarking and attacks is modeled by a game between an attacker and a defender. A max-min optimization leads to the optimal attack and the corresponding embedding strategy (counter-attack). Experiments using signals from wavelet transformed images confirm the relevance of our approach : the attack is very efficient and the embedding strategy leads to better results than previous literature's approaches. The watermarking channel is a channel with side information : a part of the noise (the host signal) is perfectly known during the embedding process. Previous works on this kind of channel showed better theoretical performances than on a classical Gaussian channels. In the third part of this work, we introduce a new performance measure in our watermarking game, taking into account the side information. We then develop a structured dictionary for this kind of channel. The experiments show great improvements in terms of performance. Finally, the last chapter deals with some game improvements : we study the influence of geometrical desynchronizations on the performance of our watermarking scheme, and we define an informed ...
No Comments.