Synthetic image detection using a modern CNN and noise patterns

Abstract

Today, people enjoy unprecedented access to state-of-the-art artificial intelligence models. This generates great opportunities for innovation and development but also several problems. Artists working with images are now worried about deep learning models trained to replicate their styles, and public figures are increasingly concerned about AIs being able to mimic their appearance and voice. In general, people are now concerned about whether they can trust anything they see, hear, or read on the internet. One very present problem is to trust if an image is true or computer generated. In the paper, we evaluate whether new methods of image synthesis can still be recognized by deep learning models. This work evaluates the new developments in CNNs and pattern recognition applied to synthetic image detection. The results present improvements in the proposed model compared to previous works on the efficiency gains of new architectures of convolutional neural networks and on the generalization potential of models trained through pattern identification methods.