Power: a metric for evaluating watermarking algorithms

An important parameter in evaluating data hiding methods is hiding capacity (Cohen et al., 1999) i.e. the amount of data that a certain algorithm can "hide" until reaching allowable distortion limits. One fundamental difference between watermarking (Cox et al., 1997; 1999) and generic data hiding resides exactly in the main applicability and descriptions of the two domains. In the digital framework, watermarking algorithms that make use of information hiding techniques have been developed and hiding capacity was naturally used as a metric in evaluating their power to hide information. Whereas the maximal amount of information that a certain algorithm can "hide" (while keeping the data within allowable distortion bounds) is certainly related to the ability to assert ownership in court, it is not directly measuring its power of persuasion, in part also because it does not consider directly the existence and power of watermarking attacks. We show why, due to its particularities, watermarking requires a different metric, more closely related to its ultimate purpose, claiming ownership in a court of law. We define one suitable metric (watermarking power) and show how it relates to derivates of hiding capacity. We prove that there are cases where considering hiding capacity is sub-optimal as a metric in evaluating watermarking methods whereas the metric of watermarking power delivers good results.