Deepfake detection in videos detecting face wrapping artifacts with convolutional neural network

Abstract

Alteration of video les by changing the face of a person on frame is Deepfake. In such manipulated contents a person's face is used on a video performing or saying something that they never actually said or did. Deepfake allows using a person's face without their consent in a video they never actually shot. The manipulation is done with the help of an arti cial intelligent method called deep learning. The AI learns the facial features of an individual from their pictures and applies them to another face in the video. As popular people have more of their images on the internet their Deepfakes can be easily created and shared without their knowledge or consent. Previously fake videos were made with simple copy pasting and photo editing, which were easily detectable by simply examining them with our eyes. Now, with arti cial intelligence it is a whole new game; Deepfake videos have become very di cult to detect and judge and a software mechanism has become a necessity to determine the authenticity of possibly manipulated videos. Thus, our team tries to build an arti cial intelligent network that is capable of Deepfake detection and determine the authenticity of a video le. For our solution, we will be attempting to detect face-wrapping artifacts from a subject's face in a particular video frame using Convolutional Neural Networks (CNN). To detect the faces of a subject in frame we will be using Haar-cascade classi ers. For training the network we will use a custom model made using Xception algorithm which trains only by using the faces extracted from video frames. Here, we will use the available dataset for Deepfake detection on Kaggle.com. The video les from the dataset will be compressed to a lower quality to train and validate our models as most Deepfakes being shared online have lower qualities. There are many ways to nd Deepfakes but most of them are not e cient enough in their detection rate. We plan to increase the rate by successfully analyzing a video in poor and good condition to determine whether it has been manipulated or not.