Deepfake Video Detection Using Convolutional Neural Network

Abstract

With the advent of new technological enhancements in artificial intelligence, new sophisticated AI techniques are used to create fake videos. Such videos can pose a great threat to the society in various social and political ways and can be used for malicious purposes. These fake videos are called deepfakes. Deepfakes refer to manipulated videos, or other digital representations produced by sophisticated artificial intelligence, that yield fabricated images and sounds that appear to be real. A deep-learning system can produce a persuasive counterfeit by studying photographs and videos of a target person from multiple angles, and then mimicking its behaviour and speech patterns. Detecting these videos is a massive problem because of the increasing developments in more realistic deepfake creation technologies emerging every now and then. The paper aims to solve this problem by proposing a model that analyses the frames of the videos using deep learning approach to detect inconsistencies in facial features, compression rate and discrepancies introduced in the videos while creating them. The model uses a convolutional neural network along with transfer learning to train the model that can catch these instilled errors in the deepfakes. The neural network is trained on these discrepancies induced during deepfake creation around the face. It uses a dataset called “Celeb-DF: A New Dataset for DeepFake Forensics” to train the model. The paper further discusses methods that can be used, in detail, to improve learning by this model.