Morphed Face Detection With Wavelet-Based Co-Occurrence Matrices

Identity theft facilitated by the use of morphed images is an emerging forensic concern. This scenario occurs when a digitally morphed image exhibits similarities to two accomplices. Given that the process of morphing alters the correlations between adjacent pixels and color channels, we propose the use of multi-dimensional co-occurrence matrices based on wavelet coefficients as features for detecting digitally morphed images. Experimental results show that the information about the frequencies at which wavelet coefficients and their prediction errors occur together are effective features in: 1) accurately detecting digitally morphed images, 2) uncovering distinctive traces left by various morphing methods in resulting images, and 3) demonstrating the feasibility of developing universal morphing attack detectors.