Detection of deepfakes using visual artifacts and neural network classifier

Abstract

The field of Artificial Intelligence is so advanced that it made the creation and modification of synthetic images and videos very easy. Tampering of videos attained a new level of refinement due to the deep learning techniques and the availability of high computing power. This contributes to the ‘deepfake’ era. Deepfake is a term coined for the fake videos created using deep learning techniques. With this method, one can create fake videos of people that they never did by replacing their face in some other real videos. There is a great danger of misusing this technique to disseminate false information or fake news. Thus the detection of deepfakes is critical to protect the people’s pride and trust in the digital content. Most of the works in detecting deepfakes are using deep learning methods. In this paper, we are proposing an approach to identify deepfake videos with very less computational power. The proposed method exploits visual artifacts present in the face regions in the generated deepfakes. We use a three-layer neural network to classify the videos as deepfake or real. As a second step of confirmation, the variance of laplacian is calculated for different patches in the face, and based on their comparison, detection of deepfakes is assured. Our approach is tested in two datasets, UADF dataset, and the latest DeepFakeDetection dataset released by the Google AI team. The proposed method achieves better results in terms of computational requirements and accuracy, and are explained in detail in the analysis section.