Deepfake Detection using Deep Learning based approaches

Abstract

In smart communities, social-media has allowed user easy access to multimedia contents. With the recent advancements in generative adversarial networks (GAN), it has become possible to create fake images/audio/and video streams of a person, or use some persons audio and visual details to fit other environments. They are known as deepfakes. Especially in recent times, deepfakes are becoming more and more realistic as well as becoming easier to create at the same time. Due to the tools with capabilities of creating such highly realistic deepfakes, fake news and syntethic contents will be rampant in the upcoming future. Mostly deepfakes have been used for spreading fake news and pornographic purposes. And hence, developing an efficient deepfake detection system is the need of the hour. But at the same time, developing an efficient deepfake detection system with high generalizability is extremely difficult. In this report, we have proposed a CNN based Neural Network architecture which can detect deepfake contents with good accuracy and have better generalizability. We have tried to use images from multiple sources for the purpose of training the model with the intention to improve generalizability capabilities of the model. When tested over test set consisting of images from more than 7 different data sources, the model yielded accuracy of 97.2% in detecting deepfake images.