GAN-generated Fake Face Image Detection using Opponent Color Local Binary Pattern and Deep Learning Technique

Advancements in AI techniques like Generative Adversarial Network (GAN) facilitate the creation of realistic-looking fake face images and these images are used to create fake profiles on various social media platforms. In this work, we develop deep learning-based binary classification models to distinguish GAN-generated fake face images from camera-captured real face images. The classification models are developed by fine-tuning three lightweight state-of-the-art pre-trained Convolutional Neural Networks (CNNs) - GoogLeNet, ResNet-18, and MobileNet-v2 -using the transfer learning approach. In this method, instead of RGB images, joint color texture feature maps of the images obtained using Opponent Color-Local Binary Pattern (OC-LBP) are used as input to the CNN. For the experimental analysis, we use datasets that contain fake face images generated by Progressive Growing GAN (PGGAN) and Style-based GAN (StyleGAN2), and camera-captured real face images from CelebFaces Attributes- High Quality (CelebA-HQ) and Flickr Faces High Quality (FFHQ) datasets. The proposed method shows remarkable performance in terms of test accuracy, generalization capability, and robustness against JPEG compression. Also, the method exhibits excellent performance when compared with state-of-the-art methods.