Abstract:
Deep learning algorithms have become as powerful as computation power has increased that
creating an indistinguishable human synthesized video, also known as deep fakes, has become
very simple. Deepfakes involves images and videos, often obscene, in which a face can be swapped
with someone else’s which can be used to create political distress, criminal activities, revenge
pom, and blackmailing people as individuals as well as communities and organizations. In this
project, we have described a deep learning method that can effectively distinguish between a fake
or Al-based generated video from a real video. The method used in this project comprises of a
Res-Next Convolution neural network to extract the frame-level features in the first phase and for
the second phase these features are further used as an input to train the Long Short-Term Memory
(LSTM) based Recurrent Neural Network (RNN) to classify the videos as fake or real. To make
the model perform better on real time data, we have created a new dataset, by combining various
datasets like FaceForensic-H-, Face Forensics in the Wild (FFIW), Deepfake detection challenge,
and Celeb-DF. The evaluated results show the effectiveness of the proposed method comparing
with existing models like VGG-16 and CapsureForensics on the newly created dataset
