Enhancing cross-domain deepfake detection through an Xception-based multi-branch model: challenges, insights and the VeriFake dataset

Abstract

Deepfake technology has advanced rapidly, producing highly realistic synthetic videos that can evade state-of-the-art detectors. However, most models trained on curated benchmarks fail to generalize to unseen datasets. This work examines the causes of such failures and explores solutions, introducing VeriFake, a custom dataset of 500 real and 500 fake videos generated using Roop and FaceFusion to reflect modern high-quality manipulations. We trained CNN and Transformer-based models on various combinations of VeriFake, Celeb-DF, and FaceForensics++ (FF++), with and without domain adaptation techniques such as Gradient Reversal Layers (GRL) and heuristic features. While VeriFake-trained models excelled in-domain, they generalized poorly to DFDC and Celeb-DF due to reliance on dataset-specific artifacts. To address this, we developed a multi-branch Xception-based framework with disentangled feature learning, a reconstruction decoder, and a GRL-powered domainadversarial module and trained on FF++, achieving 0.7509 AUC on DFDC and 0.8183 AUC on Celeb-DF, surpassing results from existing works, though with a trade-off on VeriFake (0.7054 AUC). These findings highlight the importance of diverse, well-designed benchmarks and domain-invariant features for robust real-world deepfake detection.