The inability of existing countermeasures to consistently cope against localized geometric attacks has precluded the widespread acceptance of image watermarking for commercial applications. The efficiency of these attacks against the so-called spread spectrum methods resides in their ability to affect the synchronization between the watermark reference and the extracted watermark at the detector end. In systems based on quantization schemes, geometric attacks have the effect of moving the watermark vector away from its actual quantization centroid, thus causing the watermark decoder to output wrong message symbols. In this paper, our goal is to gain a better understanding of the challenges imposed by the watermark synchronization problem in the context of localized geometric attacks. For that matter, we propose a model for the watermark synchronization problem based on maximum-likelihood (ML) estimation techniques. In that way, we derive theoretically optimal watermark synchronizer structures for either blind or non-blind schemes and based on the Cramer-Rao inequality we set lower bounds on the variance of these attack parameter estimates as a means to assess the accuracy of such synchronizers.