Image forgery detection by transforming local descriptors into deep-derived features

Image forgery is the intentional alteration of digital images, either manually using image editors or through deep fake techniques, for the purpose of disseminating fake information. We propose a forgery detection approach that efficiently detects copy-move and splicing attacks of varying scales in digital images. Our goal is to identify the homogeneous region(s) inconsistent with the rest of the image. This region property has been typically employed in object detection and classification, while we exploit this property to detect forgery in images. Thus, we generate the deep-derived features from the existing hand-crafted features in forgery detection as input to the VGG16, a deep learning method, trained for object classification. We use a binary class SVM trained on the obtained deep-derived features to determine whether an image is real or fake. We perform extensive experiments on three publicly available image manipulation datasets, DVMM, Casia and Korus to validate the effectiveness of the proposed methodology. The results show a better accuracy compared to the state-of-the-art methods.