Deep learning model to detect deceptive generative adversarial network generated images using multimedia forensic

Abstract

Computer-generated imagery has been made more lifelike and misleading by new Generative Adversarial Network (GAN) models, such as StyleGAN, posing severe risks to people's safety, social order, and privacy. Deceptive content creation, including Deepfakes, image tampering, and information hiding, can be facilitated through the misuse of GANs. To tackle these challenges, a detection model is proposed in this research, employing a spatial-frequency joint dual-stream convolutional neural network. Learnable frequency-domain filtering kernels and frequency-domain networks are leveraged to thoroughly learn and extract frequency-domain features, considering the discernible artifacts left by GAN images in the frequency spectrum due to the upsampling process during production. Lastly, the two sets of traits are combined to identify GAN-created faces. The proposed model outperforms state-of-the-art methods, as evidenced by experimental findings on various datasets, both in terms of detection accuracy on high-quality created datasets and generalization across datasets.