This paper proposes a Dual-Path Guidance Network (DPGNet) that utilizes large-scale unlabeled data to overcome the labeling data dependency of existing deepfake detection methods. Considering the reality that it is becoming difficult to distinguish deepfake images from real images, we focus on bridging the domain gap between different generative models and effectively utilizing unlabeled data. DPGNet solves this problem through a text-based cross-domain alignment module and a curriculum-based pseudo-label generation module, and prevents the forgetting problem through cross-domain knowledge distillation. Experimental results using 11 representative datasets show that DPGNet outperforms the existing state-of-the-art (SoTA) method by 6.3%.
