Holistic Image Manipulation Detection using Pixel Cooccurrence Matrices

Digital image forensics aims to detect images that have been digitally manipulated. Realistic forgeries involve a combination of splicing, resampling, region removal, smoothing and other manipulation methods. While most detection methods in literature focus on detecting particular type manipulation, it is challenging identify doctored host manipulations. In this paper, we propose novel approach holistically tampered using pixel co-occurrence matrices deep learning. We extract horizontal vertical three color channels the domain train model convolutional neural network (CNN) framework. Our method agnostic classifies an as or untampered. validate our dataset more than 86,000 images. Experimental results show promising achieves 0.99 area under curve (AUC) evaluation metric training validation subsets. Further, also generalizes well around 0.81 AUC unseen test comprising 19,740 released part Media Forensics Challenge (MFC) 2020. score was highest among all teams participated challenge, at time announcement challenge results.