Beyond frames: 3D-CoAtNet for generalizable deepfake video detection

Deepfakes pose a growing risk to digital integrity and public trust, driving the need for robust video-level forgery-detection methods. Many existing approaches analyse individual frames independently and overlook temporal dependencies, thereby weakening the generalisation to unseen manipulation techniques. This paper introduces 3D-CoAtNet, a spatiotemporal architecture for deepfake video detection that processes multiple frames simultaneously, thereby reducing reliance on single-frame artefacts. The model inflates CoAtNet’s 2D convolutional, residual, pooling, and self-attention layers into their 3D counterparts to learn spatial and temporal representations from multiple frames. We evaluated two input modalities: RGB 15-frame clips sampled from each video, and 15-frame optical-flow sequences that capture motion cues. Extensive experiments on FaceForensics++ (FF++), DFDC, and Celeb-DF under intra- and cross-dataset settings show that 3D-CoAtNet is competitive in intra-dataset evaluations (best in the DeepFakes dataset) and transfers well to Celeb-DF. Moreover, although frame-based CoAtNet16A achieves strong within-dataset accuracy, 3D-CoAtNet improves cross-dataset generalisation. These findings highlight the importance of the proposed 3D-CoAtNet model for deepfake forensics.