Comparative analysis of feature extraction and fusion for blind authentication of digital images using chroma channels

Blind authentication is one of the challenging techniques which has attracted considerable attention with the increasing security hacks on digital images. This paper evaluates different models based on feature extraction to detect digitally-altered images. Various feature-extraction methods have been investigated and compared including LBP, MSLBP, CSLBP, SLBP, WLD, LPQ, MSLPQ and DCT. Moreover, we propose a number of ways to combine a variety of features utilizing single and multi-scale representations of images. For classification, SVM is employed with different kernels to identify forged and authentic images. To evaluate the effectiveness of the investigated models, several experiments have been conducted using k-fold cross-validation and computed performance measures for two benchmark image tampering datasets (CASIA v1.0 and CASIA v2.0). Additionally, we have conducted statistical analysis for the top-ten models and the results confirmed that the best models for CASIA v1.0 and CASIA v2.0 are MSLBP-DCT and MSLPQ, respectively. Further improvements have been achieved by integrating features from the three color channels (Y, Cb and Cr) with and without feature reduction using PCA and LLP. In this case, the results show that MSLPQ-DCT achieved a better accuracy of 98.56% on CASIA v1.0 with 1020 features and MSLPQ achieved a slightly better accuracy of 97.4% on CASIA v2.0 with 1536 features.