Digital image forensics is rapidly growing science which plays crucial role in different applications including crime investigation, evidence recovery and establishment of truth in presence of chaos. In the presence of plenty of image editing techniques, digital image forensics is indispensable towards detection of image tampering or forgery. The existing methods used for image forensics suffer from inability to deal with limited training data and mediocre performance in tampering detection. There is need for a more comprehensive framework that not only addresses these problems but also strives to localize tampered region in addition to tampering detection. Towards this end, we proposed a framework known as Deep Image Forensics Framework (DIFF) which has mechanisms to leverage quality of training data and uses an enhanced U-Net model for classification and localization accuracy. We proposed an algorithm named Automatic Image Tampering Detection and Localization (AITDL) to realize the framework. This algorithm exploits outcome of Error Level Analysis (ELA) of training images (X) and ground truth labels to train an improved U-Net model. Prior to this data augmentation is carried out to make U-Net robust to different input conditions besides improving quality of training. Our empirical study using IFSTC dataset has revealed that AITDL outperforms many existing methods such as CFA1, NOI1 and ELA with 96.25% accuracy.

Convolutional Neural Network, Deep Learning, Machine Learning, Automatic Image Tampering Detection, U-Net Model