Deep Authenticator: A method for compressed DeepFake detection

Abstract

The rapid advancement of generative AI and social media has enabled the widespread dissemination of compressed DeepFake videos due to storage and bandwidth constraints. However, current DeepFake detection methods struggle with compressed videos, making their identification a critical challenge. We propose Deep Authenticator, an efficient two-stream frame-level approach that first extracts I-Frames from input videos. One stream processes RGB (Red, Green, Blue) colour model information to detect semantic inconsistencies, while the other extracts noise features using Spatial Rich Model (SRM) filters. Both streams utilize deep convolutional neural networks, and our findings suggest that using more than three SRM filters enhances detection performance. Finally, stream fusion was applied. We evaluated our approach on large, diverse DeepFake datasets, including FaceForensics++, Celeb-DF-V1, and Celeb-DF-V2. Experimental results demonstrate that our method achieved state-of-the-art performance. Deep Authenticator achieved 1% to 6% higher accuracy on different datasets. Deep Authenticator achieved state-of-the-art performance while using 3 to 20 times fewer video frames compared to other studies.