Fake Face Detection in Deepfake Videos Using Deep Learning Algorithms

Abstract: The easy access to generative adversarial networks (GANs) has resulted in the creation of very realistic deepfake videos. This poses a serious threat to the accuracy of information and public trust. Detecting these altered videos is crucial because traditional methods cannot identify the subtle changes made by deep learning models. This work presents a new hybrid model for detecting deepfake videos. Our approach employs a ResNext convolutional neural network (CNN) to extract important spatial features from individual video frames, particularly focusing on small mismatched areas on faces. These features are then analyzed by a Long Short-Term Memory (LSTM) recurrent neural network (RNN) to track how these features change over time and identify issues between frames that are common in GAN-generated fakes. The model is trained and tested on a large dataset of real and fake videos. We demonstrate how effective our spatiotemporal analysis is, and we also introduce a web-based platform for practical use. Future work will include adding audio and visual analysis to check all types of media.

Keywords: Deepfake, Generative Adversarial Networks, Video Forensics, Convolutional Neural Networks, Long Short-Term Memory, Spatiotemporal Analysis.