Deepfake Video Detection with Convolutional and Recurrent Networks
Abstract
Although deepfakes have many positive applications, they have serious negative effects as well. In this project, we utilize deep learning methods, the technology employed in creating deepfakes, to combat its negative effects. We use the Facebook Deepfake Detection Challenge dataset [1], containing over 100,000 videos, both real and altered, of people. The focus of this project is the development and training of various deep models to effectively detect deepfake videos, as well as the extensive data processing required to reduce the size of the massive Facebook Deepfake Detection Challenge video dataset. By creating an ensemble of deep neural networks, including a 3D CNN and CNN-LSTM models, we were able to detect deepfake videos with 87% accuracy on a balanced test dataset.