LGSFNet: A Local–Global Semantic Fusion Network for robust cross-domain deepfake detection

Deepfakes, enabled by recent advances in generative models, pose significant ethical, societal, and security risks. Although many detection methods achieve strong intradataset performance, they often degrade on low-quality or cross-domain data due to compression artifacts and unseen manipulations. To address this, we introduce LGSFNet, a robust deepfake detection framework that fuses local and global forgery semantics in a dual-path architecture. The design integrates a Spatial Resolution Adapter (SRA) to extract local low-level features and a novel Local Semantic Fusion Adapter (LSFA) to inject these cues into the DINOv3 transformer backbone for multi-stage feature fusion with parameter-efficient training. Experiments on FaceForensics++ demonstrate state-of-the-art results across all four manipulation types, achieving up to 99.98% AUC. Cross-corpora evaluations on Celeb-DF, DFD, and DFDC further highlight strong generalization, with improvements of up to +11.2% AUC over prior methods. A t-SNE visualization confirms discriminative representation of forgery features, while ablation studies validate that three LSFA modules achieve the best trade-off between performance and complexity. Overall, LGSFNet provides a robust, efficient, and generalizable solution for detecting low-quality and unseen deepfakes, moving toward reliable real-world deployment. The source code can be accessed using the link: https: //github.com/zulkaifsajjad/LGSFNet