Leveraging Unlabeled Data from Unknown Sources via Dual-Path Guidance for Deepfake Face Detection

Zhiqiang Yang, Renshuai Tao, Chunjie Zhang, Guoqing Yang, Xiaolong Zheng, Yaoyao Fiona ZhaoarXiv 2025 · DOI · arXiv

Abstract

Existing deepfake detection methods heavily rely on static labeled datasets. However, with the proliferation of generative models, real-world scenarios are flooded with massive amounts of unlabeled fake face data from unknown sources. This presents a critical dilemma: detectors relying solely on existing data face generalization failure, while manual labeling for this new stream is infeasible due to the high realism of fakes. A more fundamental challenge is that, unlike typical unsupervised learning tasks where categories are clearly defined, real and fake faces share the same semantics, which leads to a decline in the performance of traditional unsupervised strategies. This paper proposes a dual-path guided network (DPGNet) to address two key challenges: (1) bridging the domain differences between faces generated by different generative models; and (2) utilizing unlabeled image samples. The method comprises two core modules: text-guided cross-domain alignment, which uses learnable cues to unify visual and textual embeddings into a domain-invariant feature space; and curriculum-driven pseudo-label generation, which dynamically utilizes unlabeled samples. Extensive experiments on multiple mainstream datasets show that DPGNet significantly outperforms existing techniques, highlighting its effectiveness in addressing the challenges posed by deepfakes using unlabeled data.

Introduction

Most deepfake detectors are "brittle," performing exceptionally well on specific training datasets like FaceForensics++ but suffering from significant performance drops when encountering fakes generated by newer, unseen AI models. This generalization failure is a major hurdle for real-world deployment. Traditional solutions involving manual labeling are unscalable given the volume of modern fakes. Furthermore, standard Unsupervised Domain Adaptation (UDA) struggles because real and fake faces share high-level semantics, differing only in subtle, non-semantic artifacts. DPGNet addresses this by shifting from a passive defense posture to actively leveraging massive streams of unlabeled, unknown-source data.

Methodology: Dual-Path Guided Network (DPGNet)

DPGNet processes labeled source data and unlabeled target data simultaneously through two core modules designed to bridge the domain gap and effectively utilize unlabeled samples.

The DPGNet architecture, featuring Text-Guided Cross-Domain Alignment (left) and Curriculum-Driven Pseudo-Label Generation (right).

Text-Guided Cross-Domain Alignment

In the shifting landscape of visual artifacts, language provides a stable "north star." DPGNet uses learnable text prompts (e.g., "real face photo" and "deep fake face photo") to align visual embeddings with textual anchors in a shared space. This alignment helps the model learn domain-invariant features—characteristics of fakes that persist regardless of the generative model used—by ignoring domain-specific visual noise.

Curriculum-Driven Pseudo-Label Generation

To utilize unlabeled data without being led astray by errors, DPGNet employs a "dual-verification" process. It compares a sample's CLIP-based prediction against its distance from a library of known features (Target Feature Bank). A curriculum learning strategy is then applied: starting with a high confidence threshold (e.g., 0.85) to learn from "easy" samples and gradually lowering it (to 0.70) to incorporate more challenging, high-value fakes as training progresses. This "easy-to-hard" strategy prevents error propagation and builds a solid foundation for adaptation.

Experimental Results

The study utilized a comprehensive suite of nine datasets, including standard benchmarks (FF++, DFDC, CelebDF) and newly constructed unlabeled sets (UCDDP, UDF40) to emulate real-world data streams.

Radar chart comparing DPGNet (Ours) against various CLIP-based baselines across multiple deepfake benchmarks.

DPGNet consistently outperformed state-of-the-art (SOTA) methods across multiple protocols. On average, DPGNet achieved a 93.8% AUC, a 4.2% improvement over previous leaders like ForensicsAdapter. On the notoriously difficult DFDC dataset, DPGNet saw a nearly 5% jump in accuracy compared to existing techniques. Ablation studies showed that even with relatively small amounts of unlabeled data (6k–12k samples), the model could bridge the domain gap effectively, demonstrating its scalability for real-world deployment.

Conclusion

DPGNet demonstrates that the flood of unlabeled online data can be transformed from a challenge into a powerful resource. By combining text-guided domain alignment with a dynamic curriculum-driven pseudo-labeling strategy, the framework effectively bridges the gap between known training data and the ever-evolving landscape of AI-generated forgeries. This approach offers a robust and scalable path forward for deepfake detection in the wild.

Summary updated

Abstract

Existing deepfake detection methods heavily rely on static labeled datasets. However, with the proliferation of generative models, real-world scenarios are flooded with massive amounts of unlabeled fake face data from unknown sources. This presents a critical dilemma: detectors relying solely on existing data face generalization failure, while manual labeling for this new stream is infeasible due to the high realism of fakes. A more fundamental challenge is that, unlike typical unsupervised learning tasks where categories are clearly defined, real and fake faces share the same semantics, which leads to a decline in the performance of traditional unsupervised strategies. Therefore, there is an urgent need for a new paradigm designed specifically for this scenario to effectively utilize these unlabeled data. Accordingly, this paper proposes a dual-path guided network (DPGNet) to address two key challenges: (1) bridging the domain differences between faces generated by different generative models; and (2) utilizing unlabeled image samples. The method comprises two core modules: text-guided cross-domain alignment, which uses learnable cues to unify visual and textual embeddings into a domain-invariant feature space; and curriculum-driven pseudo-label generation, which dynamically utilizes unlabeled samples. Extensive experiments on multiple mainstream datasets show that DPGNet significantly outperforms existing techniques,, highlighting its effectiveness in addressing the challenges posed by the deepfakes using unlabeled data.
BibTeX
@misc{yang2025leveraging,
  author = {Zhiqiang Yang and Renshuai Tao and Chunjie Zhang and Guoqing Yang and Xiaolong Zheng and Yaoyao Fiona Zhao},
  title = {{Leveraging Unlabeled Data from Unknown Sources via Dual-Path Guidance for Deepfake Face Detection}},
  year = {2025},
  eprint = {2508.09022},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.09022},
  howpublished = {arXiv},
}
Figure