A machine learning approach to detect DeepFake videos

Abstract

DeepFake detection is important as the internet is a big part of our lives. DeepFake photos and videos can easily mislead us into thinking something that probably did not happen. It can also reduce trust in the media. As these manipulations become more convincing, celebrities are usually the victim of these kinds of misleading photos and videos. To detect fake videos, we will focus on existing methods and build our model to be more accurate as images of small imperceptible perturbations are su cient to fool the most powerful neural network. In our Machine Learning approach, we rst take the sample videos for training. Then, using open CV2, we have generated images from those videos. After that, we have passed these images to PCA for extracting principal component features. Then we applied VGG-16 and nally we have compared the train-test accuracy using di erent classi ers like SVC, RFC, GNB, CNN etc. After analyzing through our model we will be able to infer whether the input video is real or fake.