Abstract
One of the most important information hiding techniques is fingerprinting, which aims to generate new representations for data that are significantly more compact than the original. Fingerprinting is a promising technique for secure and efficient similarity search for multimedia data on the cloud. In this paper, we propose LID-Fingerprint, a simple binary fingerprinting technique for high-dimensional data. The binary fingerprints are derived from sparse representations of the data objects, which are generated using a feature selection criterion, Support-Weighted Intrinsic Dimensionality (support-weighted ID), within a similarity graph construction method, NNWID-Descent. The sparsification process employed by LID-Fingerprint significantly reduces the information content of the data, thus ensuring data suppression and data masking. Experimental results show that LID-Fingerprint is able to generate compact binary fingerprints while allowing a reasonable level of search accuracy.