DeepFake Image Detection

AI-generated fake images, also known as DeepFakes, are designed to spread abusive content and misinformation amongst millions of people, exacerbated by their inherently controversial nature and the reach of modern media. In this paper, we focus on detecting DeepFake images and propose a binary classifier based on a 2-phase learning architecture. The first phase consists of a ResNet-based CNN trained as a Siamese Neural Network, designed to find distinguishing features for fake images. The second phase is a 2-layer CNN that takes the feature encodings from the first phase and outputs a REAL/FAKE classification. We build on top of prior work exploring this architecture and demonstrate 91% validation accuracy on a large, diverse dataset of sophisticated GAN-generated DeepFake images.