Modeling and analysis of ordinal ranking in content fingerprinting

Content fingerprinting provides a compact representation of multimedia objects for copy detection. This paper analyzes the robustness of the ordinal ranking module frequently used in content fingerprinting by examining the changes in ranks as local variations are introduced in feature values. Closed-form expressions to measure such sensitivity are derived when feature values are jointly Gaussian-distributed. The results show that sensitivity depends on the strength of local variation, the total number of blocks, and the correlations among block-based feature values. Experiments with both synthesized data and image data validate the analysis and provide interesting insights, inspiring an approach to reduce the sensitivity.