Uncovering visual attention-based multi-level tampering traces for face forgery detection

With the rise of realistic face forgery techniques, the threat of identity fraud is more significant than ever. Several research works have focused on detecting such forgeries, but they extract forgery clues as a preprocessing step to the feature extraction phase of deep neural networks. A novel DenseTrace-Net architecture is designed in this manuscript to extract more comprehensive and refined face tampering traces locally and globally. Specifically, DenseTrace-Net extracts attentional multi-level tampering traces from facial images. A novel 'Local Attentional Tamper Trace Extractor' (LATTE) module extracts face tampering traces locally at the block level. A novel 'Global Attentional Tamper Trace Extractor' (GATTE) module aggregates multi-scale tampering traces globally. The LATTE and GATTE modules use visual depth attention to enhance their feature representation capability. Additionally, the proposed DenseTrace-Net is computationally lightweight with just 1.378 million parameters. DenseTrace-Net is evaluated on three benchmark datasets, the FF + + , CelebDF and DFDC datasets, achieving AUC scores of 0.9784, 0.9843 and 0.9916, respectively. These excellent scores allow the DenseTrace-Net to outperform the existing state-of-the-art face forgery detection methods comfortably.