Explainable Deepfake video detection using Convolutional Neural Network and CapsuleNet

Abstract

The term ‘Deepfake’ comes from the deep learning technology. Deepfake technology can easily and smoothly stitch anyone into any digital media where they never took part in reality. The key components of deepfake are machine learning and Artificial Intelligence (AI). At the beginning deepfake was introduced for research, industrial uses and entertainment purposes. The capabilities of deepfakes have existed for decades but the creations were not as realistic as they are today. As time passes, deepfakes are also improving and creating such things which are hard to identify as ‘real’ or as ‘fake’ with bare eyes. Furthermore, the new technologies now allow anyone to make deepfakes even if the creator is unskilled. The ease of accessibility and the increase of availability of deepfake creations have raised the issue of security.The most highly used algorithm to make these deepfake videos is GAN (generative Adversarial network), which is basically a machine learning algorithm which creates a fake image and discriminates itself to reproduce the best possible Fake frame or video. Our Primary goal is to use CNN (Convolutional Neural Network) and CapsuleNet with LSTM to distinguish which frame of the video was generated by the deepfake algorithm and which was the original one. We also want to find out why our model predicted the output of detection and analyze the patterns using Explainable AI. We want to apply this approach to develop a transparent relation between AI and human agents and also to set an applicable example of explainable Ai in real-life-scenarios.