Share
Export Citation
DS-Xception: Dual-Stream Noise Residual Architecture with Few-Shot Domain Anchor for Generalizable Diffusion-Based Deepfake Video Detection
Slat C.C.
2026 IEEE International Conference on Artificial Intelligence and Mechatronics Systems Aims 2026
Abstract
The rapid shift from Generative Adversarial Networks (GANs) to diffusion models has rendered conventional deepfake detectors ineffective, as GAN-optimized architectures suffer catastrophic cross-domain performance drops. We reveal a critical motion-field divergence where GAN content exhibits erratic temporal fields while diffusion pipelines produce excessively smooth trajectories, creating a motion smoothness trap that yields an average cross-domain predicted fake probability of only 7.43%. To address this, we propose DS-Xception, a dual-stream architecture fusing a spatial stream with a Photo Response Non-Uniformity (PRNU) noise residual stream, augmented by a Few-Shot Domain Anchor strategy injecting only 50 diffusion samples (0.5% of training data) to calibrate cross-paradigm decision boundaries. DS-Xception achieves a cross-domain AUC of 0.9257 and recall of 86.06%, up from a baseline of 0.6234, while preserving an intra-domain AUC of 0.9989.