To address the labeled data dependency of existing deepfake detection methods, this paper proposes a Dual-Path Guidance Network (DPGNet) that leverages large-scale unlabeled data. DPGNet consists of two core modules: one that bridges the domain gap between face images generated by different generative models and the other that leverages unlabeled image samples. The first module, text-guided cross-domain alignment, integrates visual and textual embeddings into a domain-invariant feature space using learnable prompts. The second module, curriculum-driven pseudo label generation, dynamically leverages information-rich unlabeled samples. Furthermore, it prevents forgetting through cross-domain knowledge distillation. Experimental results on 11 datasets demonstrate that DPGNet outperforms the state-of-the-art (SoTA) method by 6.3%. This demonstrates the effectiveness of DPGNet in solving the increasingly realistic annotation problem of deepfake images by leveraging unlabeled data.
