Multimedia identification via content fingerprints is used in many applications, such as content filtering on user-generated content websites, and automatic multimedia identification and tagging. A compact “fingerprint” is computed for each multimedia signal that captures robust and unique properties of the perceptual content, which is later used for identifying the multimedia. Several different multimedia fingerprinting schemes have been proposed in the literature and have been evaluated through experiments. To complement these experimental evaluations and provide guidelines for choosing system parameters and designing better schemes, this paper develops models for content fingerprinting and provides an analysis of the identification performance under these models. As a first step, bounds on the identification accuracy and the required fingerprint length for the simplest case when the fingerprint bits are modeled as i.i.d. are summarized. Markov Random Fields are then used to address more realistic settings of fingerprints with correlated components. The optimal likelihood ratio detector is derived and a statistical physics inspired approach for computing the probability of detection and probability of false alarm is described. The analysis shows that the commonly used Hamming distance detection criterion is susceptible to correlations among fingerprint bits, whereas the optimal log-likelihood ratio decision rule yields 5-20% improvement in the accuracy over a range of correlations. Simulation results demonstrate the validity of the theoretical predictions.
Read full abstract