Synthetic Content Detection in Deepfake Video using Deep Learning

In recent years, significant improvements in the field of deep learning have promoted the development of highly realistic AI-based video manipulation and forgery technologies commonly known as “DeepFake” using Autoencoders and Generative Adversarial Networks (GANs). Although this technology has its own set of beneficial applications, it decreases the integrity of digital visual media and can have serious sociopolitical implications. With new improving video manipulation tools, the detection of deepfake is a major challenge. According to the recent surveys that are reported, the deepfake systems are able to give a decision about whether a given video contains fake content or not. The objective of our project is to identify the presence of deepfake content in the digital video and report the same. Our system uses ResNext Convolution Neural Network for extracting the features of the video at frame level and Long Short-Term Memory (LSTM) for training a model to classify if a video is deepfake or pristine. We have trained the network using Deepfake Detection Challenge dataset. The result along with the confidence score of the model is shown to the user using a simple graphical user interface.