Deepfake Video Detection Methods using Deep Neural Networks

Nowadays, humans are dealing with incipient trouble known as deepfake videos, advanced with the usage of deep learning. Due to freely accessible deep fake technology equipment and inexpensive computational power, internet is flooded with fake media like fake images, videos, audios etc. Fake images and videos are causing threats to privacy, reputation and the very identity of common people. Researchers are taking efforts to develop tools using various Convolutional Neural Networks (CNNs) to automatically detect this fake media however the existing tools are not able to cope with the evolution of deep fakes. In this paper, 26 unique deep convolutional models are utilised for the task of deepfake video detection. The models can natively classify objects like table, face, humans, cars etc. However, the paper highlights the use of these models in detection of deepfake/manipulated images and videos by changing the top layer of the model with sigmoid layer hence, detecting artifacts in an image produced by Generative Adversarial Networks (GANs)[11]. Once the models are trained, the paper demonstrates the usefulness of model ensemble to improve the accuracy of proposed system and thereby making the system more reliable.