DeepFake Detection Through Key Video Frame Extraction using GAN

Abstract

The emergence of deepfake videos in recent years has made image falsification a real threat. A deepfake video uses deep learning technology to substitute a person's face, emotion, or speech with the face, emotion, or speech of another person. Finding such deceptive deepfake videos on social media is the first step in preventing them. A robust neural network-based technique to identify false videos is presented in this paper. An important video frame extraction approach is used to speed up the process of finding deep fake videos. A model made up of a convolutional neural network (CNN) and a classifier network consisting of GAN technology is provided. Resnet, Resnext50 and LSTM were passed over in favor of the Confusion Matrix when deciding which structure to pair with the classifier while detecting the fake video. The model is a method for detecting visual artefacts. The subsequent classifier network uses the feature vectors from the CNN module as this is the input to categorize the video whether it is fake or real one. The dataset is considered from DeepFake Detection Challenge to get the best model. The key goal is to get high accuracy without using a lot of data to train the model. In comparison to earlier efforts, the key video frame extraction method dramatically decreases computations by achieving 97.2% accuracy using the Deepfake Detection Challenge dataset.