Deep Fake detection using deep learning

Abstract

Deep learning (DL) is, relatively speaking, a new and emerging field of computer science. Its growth is part of the sudden boom that artificial intelligence (AI) has experienced over the past decade or so. In layman terms, deep learning is the use of massive artificial neural networks (ANNs) to carry out various tasks commonly associated with AI. The world of deep learning comes with its own problems, though. One such problem is that of misinformation and disinformation. Misinformation refers to false, inaccurate or misleading information that is communicated regardless of an intention to deceive. On the other hand, disinformation refers to similar information that is communicated specifically with the intention of deceiving the masses. Deep learning is slowly gaining traction as an effective tool to generate such mis/disinformation. In today's time, it is possible to use Deep Learning algorithms to generate false images, videos and audio involving people, where they may be engaged in an activity that never happened or saying words that they never spoke. Such generated media is called a DeepFake. Deepfakes fall under the disinformation category since there is, in most cases, an intention to deceive. Over the years, due to consistent research, these algorithms have become so good at their jobs that a human cannot be expected to identify a deep fake by looking at the media, unless told explicitly that what they are looking at is a deep fake. In this work, the dataset has been collected from pcloud, which contains both real and deep fake images collected from videos publicly available on the web. The dataset has been split into a training and validation set with a total of 5740 deepfakes and 8270 real images in the training set & 1435 deepfakes and 2067 real images in the validation set. In this work, an exponential decay algorithm has been used which provides an accuracy of 92.52% on validation set using LeRu activation function.