The analysis of an NMF-based perceptual image hashing scheme

Abstract

This paper analyzes a recently introduced perceptual image hashing scheme. This scheme uses non-negative matrix factorization (NMF) as the feature of the image on the secondary image obtained by the ring partitioning of the original one. The image hash is rotation-invariant with arbitrary degree but it is robust against acceptable content-preserving image modifications. We re-visit the algorithm and show that the algorithm tolerates some serious security flaws. Ignoring specific areas of the image in the feature extraction stage and invariance against individual ring rotation are two existing drawbacks. Practical experiments show the mentioned weaknesses and some solutions will be proposed.