Generative adversarial networks (GANs) are often used to synthesize realistic looking images, which can be a source of dis/misinformation. To detect GAN-fake images effectively, a robust and lightweight detector is proposed and named RGGID (Robust and Green GAN-fake Image Detector) in this work. RGGID is developed under the assumption that GANs fail to generate high-frequency components of real images in high fidelity. Based on this assumption, we design a set of filters using a specific local neighborhood pattern of a pixel, called a PixelHop, and determine the associated discriminant channels. We obtain multiple PixelHops by varying the local patterns, use the validation data to identify discriminant channels, and ensemble their channel responses to yield state-of-the-art detection performance. RGGID offers a green solution since its model size is significantly smaller than that of deep neural networks. Furthermore, we apply common manipulations to real/fake source images, including JPEG compression, resizing and Gaussian additive noise, and demonstrate the robustness of RGGID to these manipulations.
Companion
APSIPA Transactions on Signal and Information Processing Special Issue - Multi-Disciplinary Dis/Misinformation Analysis and Countermeasures: Articles Overview
See the other articles that are part of this special issue.