Enhancing the Fight against Social Media Misinformation: An Ensemble Deep Learning Framework for Detecting Deepfakes

Abstract

Deepfakes are synthetic media that replace someone's action (source person) with another (target person). Images deepfakes, commonly known as "visual deepfakes," depict a complicated and contentious high-tech avant-garde phenomenon in the sphere of digital trickery and artificial intelligence. These are highly deceitful and computer-based distortions of static images, commonly photographs, where the appearance of a single individual is painstakingly superimposed onto another in a highly sophisticated manner that seems to be real. Image Deepfakes are easy to generate due to easy access to opensource deepfake generation software applications such as FakeApp. Once it is generated, social media becomes its marketplace where it is easily distributed to engage and deceive millions of users. Most research in this area focuses on using a single deep-learning algorithm on a small dataset in the development of the deepfakes detection model. Therefore this research work is focused on building a robust and efficient deepfakes image detection model using a publicly available dataset from Kaggle comprising one hundred and forty thousand (140,000) images. The model was developed using Convolutional Neural Networks (CNN), Support Vector Machines (SVM), Feed Forward Neural Network (FFNN), and Gated Recurrent Unit (GRU). To make the model more robust and efficient, the ensemble technique was employed to ensemble the individual models and an accuracy of 94.91% was achieved.