DualFocusNet: Dual-Stream Spatio-Temporal Transformer for Generalizable Deepfake Detection

The rapid advancement of deep generative models has enabled the creation of highly realistic deepfake videos, posing significant challenges for reliable detection. Existing approaches typically focus either on local spatial artifacts or global temporal inconsistencies, often lacking a unified mechanism to effectively integrate these complementary cues, which limits their generalization across diverse manipulation domains. To address this limitation, we propose DualFocusNet, a dual-branch spatio-temporal transformer designed for robust and generalizable deepfake video detection. The architecture incorporates two specialized attention modules: Short-Range Dependency Attention (SRDA), which captures fine-grained motion and texture inconsistencies in localized facial regions, and Long-Range Dependency Attention (LRDA), which models global structural coherence and long-term temporal dynamics. These complementary representations are adaptively fused through a trainable cross-attention mechanism with dynamic gating, enabling input-dependent integration of local and global cues. Extensive cross-dataset evaluations show that DualFocusNet consistently outperforms state-of-the-art methods under challenging generalization settings. When trained on DeepFake Detection Challenge Preview (DFDC-p), it achieves AUC scores of 82.5% on FaceForensics¸¸ (FF¸¸), 86.2% (Celeb-DF), 87.3% (DeeperForensics), and 82.0% (WildDeepfake). When trained on Celeb-DF, it achieves 79.8% (FF¸¸), 78.5% (DFDC-p), and 80.9% (DeeperForensics), demonstrating robust performance across unseen datasets. It consistently surpasses strong baselines such as GramNet, Swin Transformer, and CNEDetector, highlighting its superior cross-domain generalization. These results confirm that the proposed dual-scale design captures intrinsic forgery inconsistencies rather than dataset-specific artifacts, making it well-suited for real-world deepfake detection.