Digital image forgery detection using deep learning approach

Abstract

This paper presents an algorithm for detecting one of the most commonly used types of digital image forgeries - splicing. The algorithm is based on the use of the VGG-16 convolutional neural network. The proposed network architecture takes image patches as input and obtains classification results for a patch: original or forgery. On the training stage we select patches from original image regions and on the borders of embedded splicing. The obtained results demonstrate high classification accuracy (97.8% accuracy for fine-tuned model and 96.4% accuracy for the zero-stage trained) for a set of images containing artificial distortions in comparison with existing solutions. Experimental research was conducted using CASIA dataset.