3D-CADet: Leveraging 3D Channel Attention for Improving Deepfake Video Detection

With the continuous advancement of deepfake technology, the threats it poses to social stability and information security are increasingly severe, making research on deepfake detection technology particularly critical. Existing spatiotemporal models (3D CNN) often lack the ability to adaptively emphasize key features, making it challenging to comprehensively capture subtle forgery traces within videos, thereby limiting their detection performance. To address these challenges, we propose a 3D Channel Attention Deepfake Detection Network (3D-CADet), which combines 3D CNN with 3D channel attention module (3DCABlock) to enhance the model's sensitivity to critical features. The 3DCABlock module employs 1D convolution with adaptive kernel size to enable local cross-channel interactions and dynamic weight assignment, allowing the model to effectively capture subtle spatiotemporal inconsistencies in deepfake videos. Experimental results validate the effectiveness of 3D-CADet, demonstrating superior detection performance across multiple challenging datasets.