Exposing computer generated images by using deep convolutional neural networks

• Published in: Signal processing. Image communication Vol. 66; pp. 113 - 126
• Main Authors: de Rezende, Edmar R.S., Ruppert, Guilherme C.S., Theóphilo, Antônio, Tokuda, Eric K., Carvalho, Tiago
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01-08-2018; Elsevier BV
• Subjects: Artificial neural networks; Artists; CG detection; Computer graphics; Deep learning; Digital computers; Digital forensics; Digital imaging; Fake news; Feature extraction; Image compression; Image detection; Image processing; Image processing systems; JPEG encoders-decoders; Neural networks; State of the art; Transfer learning; Deep learning; CG detection; Digital forensics; Transfer learning; Fake news
• ISSN: 0923-5965; 1879-2677
• DOI: 10.1016/j.image.2018.04.006

The recent computer graphics developments have upraised the quality of the generated digital content, astonishing the most skeptical viewer. Games and movies have taken advantage of this fact but, at the same time, these advances have brought serious negative impacts like the ones yielded by fake images produced with malicious intents. Digital artists can compose artificial images capable of deceiving the great majority of people, turning this into a very dangerous weapon in a timespan currently know as “Fake News/Post-Truth” Era. In this work, we propose a new approach for dealing with the problem of detecting computer generated images, through the application of deep convolutional networks and transfer learning techniques. We start from Residual Networks and develop different models adapted to the binary problem of identifying if an image was, or not, computer generated. Differently from the current state-of-the-art approaches, we do not rely on hand-crafted features, but provide to the model the raw pixel information, achieving the same 0.97 performance of state-of-the-art methods with three main advantages: (i) executes considerably faster than state-of-the-art methods with equivalent accuracy; (ii) eliminates the laborious and manual step of specialized features extraction and selection, and (iii) is very robust against image processing operations as noise addition, blur and JPEG compression.
[Display omitted]
•Proposal of a new approach based on DNN and transfer learning that achieves an accuracy of 0.97.•Use of an extended dataset (more difficult for the task).•Faster method proved by reducing execution time in more than ten times.•Evaluation of different kinds of classifiers in association with a DNN.•Qualitative analysis of bottleneck features produced by ResNet-50 in CG image detection problem.