Short And Low Resolution Deepfake Video Detection Using CNN
Ashifur Rahman, Nipo Siddique, Mohasina Jannat Moon, Tahera Tasnim, Mazharul Islam, Md. Shahiduzzaman, Samsuddin Ahmed
Abstract
Recently, convincing deepfake videos are growing very fast that can delude even the trained experts. These deepfake videos have huge impacts all over the world covering the political, social, and personal lives. The state-of-the-art machine learning studies are demonstrating noticeable success to detect fake videos in high resolution and long-time video data while the same performance is not observed in low resolution and short-time clips. In this work, we have trained a convolutional neural network (CNN) that demonstrates mentionable accuracy in detecting fake videos in low-resolution and short-time video data. We have exploited Kaggle Deepfake Detection Challenge (DFDC) and the Face Forensics++ datasets in our experiment. Our model shows 94.93% accuracy in detecting fake videos for the DFDC dataset while the same is 93.2% for FaceForensics++ Dataset. We evaluated our models by different performance metrics and compared the performance with state-of-the-art methods. Our model demonstrates comparable performance.