Detecting GAN generated Fake Images using Co-occurrence Matrices

Abstract

The advent of Generative Adversarial Networks (GANs) has brought about completely novel ways of transforming and manipulating pixels in digital images. GAN based techniques such as Image-to-Image translations, DeepFakes, and other automated methods have become increasingly popular in creating fake images. In this paper, we propose a novel approach to detect GAN generated fake images using a combination of co-occurrence matrices and deep learning. We extract co-occurrence matrices on three color channels in the pixel domain and train a model using a deep convolutional neural network (CNN) framework. Experimental results on two diverse and challenging GAN datasets comprising more than 56,000 images based on unpaired image-to-image translations (cycleGAN [1]) and facial attributes/expressions (StarGAN [2]) show that our approach is promising and achieves more than 99% classification accuracy in both datasets. Further, our approach also generalizes well and achieves good results when trained on one dataset and tested on the other.

Detecting and Simulating Artifacts in GAN Fake Images

“…We use horse images from CycleGAN and AutoGAN for training respectively. Since both the JPEG compression and image resize destroy the up-sampling artifact, the model trained with images without post-processing does not generalize to the post-processed images, also as reported in [15]. Training a new model with post-processed images improves the performance on post-processed images.…”

“…Marra et al [4] propose to use raw pixel and conventional forensics features extracted from real and fake images to train a classifier. Nataraj et al [15] propose to use the co-occurrence matrix as features and show better performance than that of classifiers trained over raw pixels on CycleGAN data. McCloskey and Albright [16] observe that GAN generated images have some artifacts in color cues due to the normalization layers.…”

A Comprehensive Tutorial and Survey of Applications of Deep Learning for Cyber Security

“…They have proposed three preprocessing methods and a novel evaluation technique in order to convert the flowbase data into continuous values which are fed into GAN and to evaluate the quality of the generated traffic data. In [419], a novel approach for the detection of fake images generated by GAN is proposed. The proposed approach computes the co-occurrence matrices on the RGB channels of the images and uses those matrices to train the CNN model to detect fake GAN images.…”

Copy-Move Forgery Detection and Localization Using a Generative Adversarial Network and Convolutional Neural-Network

“…The adversarial training paradigm features two main components: the generator (i.e., the image-to-image network) and the discriminator (i.e., the support network). Within the paradigm, the generator's training revolves around learning to deceive the discriminator, while the discriminator is trained to detect real images from forged ones [55][56][57]. In a previous study [23], Zhu et al devised a method for automatically pairing images, thus, overcoming the shortage in genuine image pairs.…”