Detecting Synthetic Images of Faces using Deep Learning
Abstract: Significant progress has been made within human face synthesis due to recent advances in generative adversarial networks. These networks can be used to generate credible high-quality images of faces not belonging to real people, which is something that could be exploited by malicious actors. In this thesis, several state-of-the-art deep learning detection models were evaluated with respect to their robustness and generalization capability, which are two factors that must be taken into consideration for models that are intended to be deployed in the wild. The results show that some classifiers exhibited near-perfect performance when tested on real and synthetic images post-processed heavily using various augmentation techniques. These types of image perturbations further improved robustness when also incorporated in the training data. However, no model generalized well to out-of-distribution images from unseen datasets, although one model showed impressive results after being fine-tuned on a small number of samples from the target distributions. Nevertheless, the limited generalization capability remains a shortcoming that must be overcome before the detection models can become viable in the wild.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)