Measuring the Statistical Correlation Inconsistencies in Mobile Images for Tamper Detection

In this paper, we propose a novel framework to statistically measure the correlation inconsistency in mobile images for tamper detection. By first sampling a number of blocks at different image locations, we extract a set of derivative weights as features from each block using partial derivative correlation models. Through regularizing the within-image covariance eigenspectrum and performing eigenfeature transformation, we derive a compact set of eigen weights, which are sensitive to image signal mixing from different source models. A metric is then proposed to quantify the inconsistency among the sampled blocks at different image locations. Through comparison, our eigen weights features show better performance than the eigenfeatures from several other types of forensics features in detecting the presence of tampering. Experimentally, our method shows good tamper detection performance especially when a small percentage of sampled blocks are from a different camera model or brand with different demosaicing processing.