Localization of Facial Images Manipulation in Digital Forensics via Convolutional Neural Networks

Throughout digital media forensics, the identification of manipulated images and videos is a key topic. Many methods of detection use a binary classification to assess the likelihood of manipulation of a message. Another significant subject is the position, mainly due to three standard attacks, of the exploited regions (i.e., segmentation): elimination, copy-move and splicing. In order to simultaneously detect manipulated images and videos and locate the region for every question, a convolutional neural network is built which uses the multi-work learning approach. The information gained during the execution of one task would be exchanged and both tasks strengthened. To improve network generation, a semi-supervised learning approach is used. A decoder and a Y-shaped decoder are part of the network. For binary classification, activation of the encoded features is used. The output from one decoder branch is used to segment the areas manipulated and from the other branch to reconstruct the input to boost overall efficiency. Experiments using FaceForensics and FaceForensics++ have shown the network's ability to deal with the flaws in the preceding attacks and to counter facial reenactment attacks. In addition, the network can deal with unknown attacks by fine-tuning just a limited amount of data.