Answers are grounded in this paper and cite their sources.
Existing detection methods struggle to generalize to unknown domains, often due to the distinct nature of cross-domain forgeries, especially with inconsistent audial and visual information, and to high-resolution synthesis, which is becoming more prevalent. Current methods largely still rely on spatial or temporal features that are spatially fixed or temporally static, overlooking semantic drift and artifacts due to resolution. To address these gaps, this study introduces a Cross-modal Multi-resolution Temporal Discrepancy Network (CMTD-Net), a novel architecture designed to exploit misalignment in temporal consistency and multiscale visual patterns. CMTD-Net integrates a Dual-Stream Temporal Discrepancy Module (DTDM) with a Hierarchical Multi-Resolution Patch Embedding (HMRPE) transformer, where DTDM captures latent temporal drifts using audio–video synchronization irregularities, and HMRPE encodes fine-grained spatial inconsistencies across resolutions. Unlike traditional 3D CNNs, CMTD-Net leverages patch-wise feature aggregation and temporal consistency scoring to achieve a higher generalization. The model was trained and validated on a newly compiled hybrid dataset combining high-fidelity cross-domain forgeries, namely DFDC, FaceForensics ++ (FF + + ), and FakeAVCeleb, which were curated to include complex facial dynamics, speech-driven synthesis, and generative editing artifacts. Experimental tests reveal that CMTD-Net, which works far better than baseline models on forgery types, has not been encountered before, and works well across resolution scales and domain changes. This makes it possible to improve the reliability of deepfake detection systems.
No takes yet. Share an insight, caveat, or question.
Panigrahi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: