An Approach to Identify DeepFakes Using Deep Learning

Abstract

Deepfakes are the manipulated images and videos that are created by performing a face swap using various Artificial Intelligence tools. This creates an illusion that someone either said something that they did not say or are someone they are not. This incorrect way of using the deepfake technology leads to severe consequences. Some scenarios like creating political tension, fake terrorism events, damaging image and dignity of people are some of the negative impacts created by these deep fakes. It is difficult for a naked human eye to detect the results of these deepfake technology. Thus, in this work, a deep learning-based method that can efficiently differentiate deepfake videos is developed. The proposed method uses Res-Next pretrained Convolutional neural network for extracting features from the frames that are divided from the input video. The extracted features are then utilized to train a Long Short Term Memory (LSTM) network to differentiate whether the input video taken from the user is subject to any manipulation or not. To make the model perform better on real time data, it is trained with a balanced dataset. The dataset used to train the model is Deepfake detection challenge dataset. The results of the proposed model can predict the output with a better accuracy. This method can thus encounter the threats and danger produced from the deepfake technology to the society.