Many watermarking algorithms have been proposed and studied for their robustness and other properties. But, there has been little effort is analyzing these algorithms for their vulnerability against detection by an adversary. As a general philosophy for robust watermarking the host signal and the watermarked signal are well separated in a statistical distance sense. This very nature can be exploited by an adversary to easily detect the watermark and perhaps remove it. We argue that the ability of a watermark to avoid detection by an adversary is a key factor that needs more attention. In this paper, we propose a framework, based on a distributed detection technique, that can be used by the adversary to study signals for the presence/absence of watermarks. We choose a particular spatial domain image watermarking algorithm and explain how the proposed framework can be applied to detect the watermark with very little knowledge about the watermark insertion procedure. The false alarm probability can be optimally traded-off with the probability of detection using the receiver operating characteristic function.