168 Detection of Trivial Image Tampering Via Unmatched Feature Points Clustering
-
Published:2011
Download citation file:
Perceptual image hashing is an emerging solution for multimedia content authentication. Due to most perceptual image hashing techniques focus on the robustness but ignore their sensitivity to trivial tampering area, such techniques might not work well when malicious attack is perceptually insignificant. Through experiments we found that some state-of-the-art image hash algorithms could not distinguish small malicious distortion and some authentic distortion. So we propose a novel method about detection of trivial image tampering via unmatched feature points clustering in this paper. First, we extract SURF (Speeded Up Robust Features) feature points but do not use its original descriptor, we use BRIEF (Binary Robust Independent Elementary Features) descriptor which is more simple and fast. Then we make a clustering to the unmatched feature points after matching, and analyze the density of the clusters. Finally, by comparing the density and a pre-defined threshold, we can correctly determine whether an image is attacked or not. The experimental results show that, compared to V. Monga's algorithm which the TPR(true positive ratio) reaches 89%, but the FPR(false positive ratio) reaches 50%, the TPR of the proposed algorithm reaches 88%, but the FPR is only 15%.