An improved deepfake detection using deep learning

Abstract

With the emergence of Generative Models for creating Deepfake images, and videos of high quality, which are extremely realistic and hard to recognize, several social, and political issues have come to light. DeepFakes of celebrities and political personalities are used to exploit and spread misinformation that leads to several social and political unrest. Hence, the necessity to develop a methodology to detect such images and videos created with DeepFakes has arisen in recent times. Several state- of-the-art methodologies are in use for the purpose such as CNN, RNN, reverse engineering of GMs, neural networks, ensemble-based learning approaches, etc. As many machine learning-based approaches are already adopted, our aim is to improve the quality of detection of DeepFakes using models that utilize deep learning, in our study. The state-of-the-art methodologies have shown promising results when ap- plied to popular datasets vastly used for training and research purposes. However, most methods are not robust enough to perform well in all kinds of general-purpose DeepFakes. Hence, in this paper, we have developed a new comprehensive and ef- ficient framework that improves the DeepFake detection performance not only on general purpose but also on training purpose data.