Quantifying DeepFake Detection Accuracy for a Variety of Natural Settings

Deep fakes are videos generated from a starting video of a person where that person\u27s face has been swapped for someone else\u27s. In this report, we describe our work to develop general, deep learning-based models to classify Deep Fake content. Our first experiments involved simple Convolution Neural Network (CNN)-based models where we varied how individual frames from the source video were passed to the CNN. These simple models tended to give low accuracy scores for discriminating fake versus non-fake videos of less than 60%. We then developed three more sophisticated models: one based on choosing test frames, one based on video Optical Flow, and one that uses Generative Adversarial Networks (GANs) to determine structural differences in images. This last technique we call MRI-GAN and is new to the literature. We tested our models using the Deep Fake Detection Challenge dataset and found our plain frames-based model achieves 90% test accuracy, our MRI model achieves 79% test accuracy, and Optical Flow-based model achieves 69% test accuracy