Abstract:
Digital alterations have become a common trend on social media websites as well as in the entertainment industry, with magazine printing altered images of celebrities to make them look more attractive. Advancements in Generative Adversarial Network (GAN) leads to the generation of realistic images. With GAs, altering images on the basis of attributes and retouching has become an easy task. This paper presents a patch based algorithm to detect retouching and GAN based digital alterations. It exploits textural differences in generated and retouched images to distinguish between different models of GAN, retouching and authentic images. The algorithm uses an SVM to classify images from patch based predictions of the CNN network. The algorithm achieves an accuracy of 99.76% on three class classification
of the two digital alterations and authentic images. Images generated using SRGAN, StarGAN, DCGAN, and Context-Encoder along with retouched images from ND-IIITD dataset have been used for performing the experiments. The paper shows that the algorithm is robust in detecting generated images from unknown GAN models as well.