Comparing Convolutional Neural Networks to traditional methods and the human eye for copy-move forgery detection

Abstract: A common image forgery technique is called copy-move forgery, in which a part of an image has been copied and placed elsewhere to cover some other part of the image. This study compares implementations of both traditional and deep learning methods for copy-move forgery detection, as well as human ability to do the same. Two Convolutional Neural Networks, one utilizing transfer learning and the other a custom architecture with a so-called constrained layer, are implemented, as well as two traditional methods based on SIFT and BRISK keypoint comparison respectively. The results show that both of the implemented neural networks slightly outperform the traditional methods in terms of accuracy, with a considerably lower computational cost. Additionally, a survey is conducted (n=19) which shows that the neural networks are able to outperform humans to an extent when given the task of classifying images as either forged or authentic. This implies that there is a real practical use for these models, in particular the model utilizing the constrained layer, outperforming the transfer learning model in terms of both accuracy and runtime.