Detecting Deepfakes and Forged Videos Using Deep Learning

Abstract: Over just a few years, methods to manipulate videos have become so sophistica- ted that even someone without much expertise or computational resources can forge videos inseparable from pristine ones to the human eye. These methods can for instance insert a person in a video or manipulate their lip movements to make them say anything of the manipulator’s liking. Though there exist harm- less and constructive uses of these technologies, it is not hard to imagine the harm they could cause if put in the wrong hands. This report presents a model to detect forged manipulated videos, more specifically those where faces have been manipulated. Four kinds of manipu- lation videos were taken into consideration: FaceSwap, DeepFakes, Face2Face and Neural Textures. The model proposed consists of a feature extraction CNN followed by an LSTM network. The FaceForensics++ dataset was used, as well as the associated benchmark. The model, though not competing with the state- of-the-art detectors, was able to classify videos with an accuracy higher than or close to that of several models in the benchmark.